Heterogeneous Transfer Learning for Image Clustering via the Social Web |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Qiang Yang |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Hong Kong University of Science and Technology, Clearway Bay, Kowloon, Hong Kong |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ qyang@cs.ust.hk |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Yuqiang Chen Gui-Rong Xue Wenyuan Dai Yong Yu |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Shanghai Jiao Tong University, 800 Dongchuan Road, Shanghai 200240, China |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ {yuqiangchen,grxue,dwyak,yyu}@apex.sjtu.edu.cn |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In this paper, we present a new learning |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ scenario, heterogeneous transfer learn- |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ing, which improves learning performance |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ when the data can be in different feature |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spaces and where no correspondence be- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tween data instances in these spaces is pro- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vided. In the past, we have classified Chi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nese text documents using English train- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing data under the heterogeneous trans- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fer learning framework. In this paper, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we present image clustering as an exam- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ple to illustrate how unsupervised learning |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ can be improved by transferring knowl- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ edge from auxiliary heterogeneous data |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ obtained from the social Web. Image |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clustering is useful for image sense dis- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ambiguation in query-based image search, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ but its quality is often low due to image- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data sparsity problem. We extend PLSA |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to help transfer the knowledge from social |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Web data, which have mixed feature repre- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentations. Experiments on image-object |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clustering and scene clustering tasks show |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that our approach in heterogeneous trans- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fer learning based on the auxiliary data is |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ indeed effective and promising. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Traditional machine learning relies on the avail- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ability of a large amount of data to train a model, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which is then applied to test data in the same |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ feature space. However, labeled data are often |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scarce and expensive to obtain. Various machine |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning strategies have been proposed to address |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this problem, including semi-supervised learning |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Zhu, 2007), domain adaptation (Wu and Diet- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ terich, 2004; Blitzer et al., 2006; Blitzer et al., |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2007; Arnold et al., 2007; Chan and Ng, 2007; |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Daume, 2007; Jiang and Zhai, 2007; Reichart |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Rappoport, 2007; Andreevskaia and Bergler, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2008), multi-task learning (Caruana, 1997; Re- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ichart et al., 2008; Arnold et al., 2008), self-taught |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning (Raina et al., 2007), etc. A commonality |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ among these methods is that they all require the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training data and test data to be in the same fea- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture space. In addition, most of them are designed |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for supervised learning. However, in practice, we |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ often face the problem where the labeled data are |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scarce in their own feature space, whereas there |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ may be a large amount of labeled heterogeneous |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data in another feature space. In such situations, it |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ would be desirable to transfer the knowledge from |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ heterogeneous data to domains where we have rel- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ atively little training data available. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To learn from heterogeneous data, researchers |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ have previously proposed multi-view learning |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Blum and Mitchell, 1998; Nigam and Ghani, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2000) in which each instance has multiple views in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different feature spaces. Different from previous |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ works, we focus on the problem of heterogeneous |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ transfer learning, which is designed for situation |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ when the training data are in one feature space |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (such as text), and the test data are in another (such |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as images), and there may be no correspondence |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ between instances in these spaces. The type of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ heterogeneous data can be very different, as in the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ case of text and image. To consider how hetero- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ geneous transfer learning relates to other types of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning, Figure 1 presents an intuitive illustration |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of four learning strategies, including traditional |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ machine learning, transfer learning across differ- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent distributions, multi-view learning and hetero- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ geneous transfer learning. As we can see, an |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ important distinguishing feature of heterogeneous |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ transfer learning, as compared to other types of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning, is that more constraints on the problem |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are relaxed, such that data instances do not need to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correspond anymore. This allows, for example, a |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ collection of Chinese text documents to be classi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fied using another collection of English text as the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 1–9, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ training data (c.f. (Ling et al., 2008) and Section |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2.1). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this paper, we will give an illustrative exam- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ple of heterogeneous transfer learning to demon- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ strate how the task of image clustering can ben- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ efit from learning from the heterogeneous social |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Web data. A major motivation of our work is |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Web-based image search, where users submit tex- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tual queries and browse through the returned result |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages. One problem is that the user queries are of- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ten ambiguous. An ambiguous keyword such as |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ “Apple” might retrieve images of Apple comput- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ers and mobile phones, or images of fruits. Im- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ age clustering is an effective method for improv- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the accessibility of image search result. Loeff |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al. (2006) addressed the image clustering prob- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lem with a focus on image sense discrimination. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In their approach, images associated with textual |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features are used for clustering, so that the text |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and images are clustered at the same time. Specif- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ically, spectral clustering is applied to the distance |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ matrix built from a multimodal feature set associ- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ated with the images to get a better feature repre- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentation. This new representation contains both |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ image and text information, with which the per- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formance of image clustering is shown to be im- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proved. A problem with this approach is that when |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ images contained in the Web search results are |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ very scarce and when the textual data associated |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the images are very few, clustering on the im- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ages and their associated text may not be very ef- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fective. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Different from these previous works, in this pa- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ per, we address the image clustering problem as |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a heterogeneous transfer learning problem. We |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ aim to leverage heterogeneous auxiliary data, so- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ cial annotations, etc. to enhance image cluster- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing performance. We observe that the World Wide |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Web has many annotated images in Web sites such |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as Flickr (http : / /www. flickr . com), which |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ can be used as auxiliary information source for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ our clustering task. In this work, our objective |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is to cluster a small collection of images that we |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are interested in, where these images are not suf- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ficient for traditional clustering algorithms to per- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ form well due to data sparsity and the low level of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ image features. We investigate how to utilize the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ readily available socially annotated image data on |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the Web to improve image clustering. Although |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these auxiliary data may be irrelevant to the im- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ages to be clustered and cannot be directly used |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to solve the data sparsity problem, we show that |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ they can still be used to estimate a good latentfea- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture representation, which can be used to improve |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ image clustering. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 Related Works |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2.1 Heterogeneous Transfer Learning |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Between Languages |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this section, we summarize our previous work |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ on cross-language classification as an example of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ heterogeneous transfer learning. This example |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is related to our image clustering problem be- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cause they both rely on data from different feature |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spaces. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As the World Wide Web in China grows rapidly, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ it has become an increasingly important prob- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lem to be able to accurately classify Chinese Web |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages. However, because the labeled Chinese Web |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages are still not sufficient, we often find it diffi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cult to achieve high accuracy by applying tradi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tional machine learning algorithms to the Chinese |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Web pages directly. Would it be possible to make |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the best use of the relatively abundant labeled En- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ glish Web pages for classifying the Chinese Web |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages? |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To answer this question, in (Ling et al., 2008), |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ we developed a novel approach for classifying the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Web pages in Chinese using the training docu- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments in English. In this subsection, we give a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ brief summary of this work. The problem to be |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ solved is: we are given a collection of labeled |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ English documents and a large number of unla- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ beled Chinese documents. The English and Chi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nese texts are not aligned. Our objective is to clas- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sify the Chinese documents into the same label |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ space as the English data. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Our key observation is that even though the data |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ use different text features, they may still share |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ many of the same semantic information. What we |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ need to do is to uncover this latent semantic in- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formation by finding out what is common among |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ them. We did this in (Ling et al., 2008) by us- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the information bottleneck theory (Tishby et |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 1999). In our work, we first translated the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chinese document into English automatically us- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing some available translation software, such as |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Google translate. Then, we encoded the training |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ text as well as the translated target text together, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in terms of the information theory. We allowed all |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the information to be put through a ‘bottleneck’ |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and be represented by a limited number of code- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 1: An intuitive illustration of different kinds learning strategies using classification/clustering of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ image apple and banana as the example. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words (i.e. labels in the classification problem). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Finally, information bottleneck was used to main- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tain most of the common information between the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two data sources, and discard the remaining irrel- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ evant information. In this way, we can approxi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mate the ideal situation where similar training and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ translated test pages shared in the common part are |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ encoded into the same codewords, and are thus as- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ signed the correct labels. In (Ling et al., 2008), we |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ experimentally showed that heterogeneous trans- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fer learning can indeed improve the performance |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of cross-language text classification as compared |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to directly training learning models (e.g., Naive |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Bayes or SVM) and testing on the translated texts. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2.2 Other Works in Transfer Learning |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In the past, several other works made use of trans- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ fer learning for cross-feature-space learning. Wu |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Oard (2008) proposed to handle the cross- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ language learning problem by translating the data |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ into a same language and applying kNN on the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ latent topic space for classification. Most learning |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ algorithms for dealing with cross-language hetero- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ geneous data require a translator to convert the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data to the same feature space. For those data that |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are in different feature spaces where no transla- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tor is available, Davis and Domingos (2008) pro- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ posed a Markov-logic-based transfer learning al- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gorithm, which is called deep transfer, for trans- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferring knowledge between biological domains |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Web domains. Dai et al. (2008a) proposed |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a novel learning paradigm, known as translated |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ learning, to deal with the problem of learning het- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erogeneous data that belong to quite different fea- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture spaces by using a risk minimization frame- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2.3 Relation to PLSA |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Our work makes use of PLSA. Probabilistic la- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tent semantic analysis (PLSA) is a widely used |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ probabilistic model (Hofmann, 1999), and could |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be considered as a probabilistic implementation of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ latent semantic analysis (LSA) (Deerwester et al., |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1990). An extension to PLSA was proposed in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Cohn and Hofmann, 2000), which incorporated |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the hyperlink connectivity in the PLSA model by |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using a joint probabilistic model for connectivity |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and content. Moreover, PLSA has shown a lot |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of applications ranging from text clustering (Hof- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mann, 2001) to image analysis (Sivic et al., 2005). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2.4 Relation to Clustering |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Compared to many previous works on image clus- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tering, we note that traditional image cluster- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing is generally based on techniques such as K- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ means (MacQueen, 1967) and hierarchical clus- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tering (Kaufman and Rousseeuw, 1990). How- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, when the data are sparse, traditional clus- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tering algorithms may have difficulties in obtain- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing high-quality image clusters. Recently, several |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ researchers have investigated how to leverage the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ auxiliary information to improve target clustering |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ P(zIv) P(fIz) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ performance, such as supervised clustering (Fin- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ley and Joachims, 2005), semi-supervised cluster- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing (Basu et al., 2004), self-taught clustering (Dai |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al., 2008b), etc. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Image Clustering with Annotated |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Auxiliary Data |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this section, we present our annotation-based |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ probabilistic latent semantic analysis algorithm |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ (aPLSA), which extends the traditional PLSA |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ model by incorporating annotated auxiliary im- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ age data. Intuitively, our algorithm aPLSA per- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ forms PLSA analysis on the target images, which |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are converted to an image instance-to-feature co- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ occurrence matrix. At the same time, PLSA is |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ also applied to the annotated image data from so- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cial Web, which is converted into a text-to-image- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ feature co-occurrence matrix. In order to unify |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ those two separate PLSA models, these two steps |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are done simultaneously with common latent vari- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ables used as a bridge linking them. Through |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these common latent variables, which are now |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ constrained by both target image data and auxil- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ iary annotation data, a better clustering result is |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ expected for the target data. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.1 Probabilistic Latent Semantic Analysis |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Let F = { fi}!Fi be an image feature space, and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ V = {vi}iv11 be the image data set. Each image |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ vi E V is represented by a bag-of-features {f I f E |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ vi A f E F}. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Based on the image data set V, we can esti- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ mate an image instance-to-feature co-occurrence |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ matrix AIVI x I FI E RIVIx IFI, where each element |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ AiT (1 G i G IV I and 1 G j G IF I) in the matrix |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ A is the frequency of the feature fT appearing in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ the instance vi. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Let W = {wi } I WI i��be a text feature space. The |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ annotated image data allow us to obtain the co- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ occurrence information between images v and text |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features w E W. An example of annotated im- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ age data is the Flickr (http : / /www. f l ickr . |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ com), which is a social Web site containing a large |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ number of annotated images. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ By extracting image features from the annotated |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ images v, we can estimate a text-to-image fea- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture co-occurrence matrix B I WIxIFI E RIW IxIFI, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ where each element BiT (1 G i G I W I and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 1 G j G IFI) in the matrix B is the frequency |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ of the text feature wi and the image feature fT oc- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ curring together in the annotated image data set. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 2: Graphical model representation of PLSA |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ model. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Let Z = {ziffl be the latent variable set in our |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ aPLSA model. In clustering, each latent variable |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ zi E Z corresponds to a certain cluster. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Our objective is to estimate a clustering func- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tion g : V H Z with the help of the two co- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ occurrence matrices A and B as defined above. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To formally introduce the aPLSA model, we |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ start from the probabilistic latent semantic anal- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ysis (PLSA) (Hofmann, 1999) model. PLSA is |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ a probabilistic implementation of latent seman- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic analysis (LSA) (Deerwester et al., 1990). In |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our image clustering task, PLSA decomposes the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ instance-feature co-occurrence matrix A under the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ assumption of conditional independence of image |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ instances V and image features F, given the latent |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ variables Z. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P(f Iv) = 1: P(f I z)P(zI v). (1) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ZEZ |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The graphical model representation of PLSA is |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ shown in Figure 2. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Based on the PLSA model, the log-likelihood can |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ be defined as: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1: L = |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ i 1:AiT |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ T ET, AiT, log P(fT I vi) (2) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ where AIVI x IFI E RIVI xIFI is the image instance- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ feature co-occurrence matrix. The term : A�j |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ j, Azj/ |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ in Equation (2) is a normalization term ensuring |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ each image is giving the same weight in the log- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ likelihood. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Using EM algorithm (Dempster et al., 1977), |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ which locally maximizes the log-likelihood of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the PLSA model (Equation (2)), the probabilities |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P(f I z) and P(zI v) can be estimated. Then, the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ clustering function is derived as |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ g(v) = argmax P(zIv). (3) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ZEZ |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Due to space limitation, we omit the details for the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ PLSA model, which can be found in (Hofmann, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1999). |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.2 aPLSA: Annotation-based PLSA |XML| xmlLoc_7 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In this section, we consider how to incorporate |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ a large number of socially annotated images in a |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ V Z F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 4 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 3: Graphical model representation of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ aPLSA model. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unified PLSA model for the purpose of utilizing |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the correlation between text features and image |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features. In the auxiliary data, each image has cer- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tain textual tags that are attached by users. The |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correlation between text features and image fea- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures can be formulated as follows. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P(f Iw) = X P(f I z)P(zI w). (4) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ZEZ |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ It is clear that Equations (1) and (4) share a same |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ term P(f I z). So we design a new PLSA model by |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ joining the probabilistic model in Equation (1) and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the probabilistic model in Equation (4) into a uni- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fied model, as shown in Figure 3. In Figure 3, the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ latent variables Z depend not only on the corre- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lation between image instances V and image fea- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures F, but also the correlation between text fea- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures W and image features F. Therefore, the aux- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ iliary socially-annotated image data can be used |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to help the target image clustering performance by |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ estimating good set of latent variables Z. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Based on the graphical model representation in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 3, we derive the log-likelihood objective |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ function, in a similar way as in (Cohn and Hof- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mann, 2000), as follows |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ text-to-image occurrence matrix B. In this case, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the aPLSA model degenerates to the traditional |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PLSA model. Therefore, aPLSA is an extension |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to the PLSA model. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Now, the objective is to maximize the log- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ likelihood L of the aPLSA model in Equation (5). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Then we apply the EM algorithm (Dempster et |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 1977) to estimate the conditional probabilities |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P(f I z), P(zI w) and P(zI v) with respect to each |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ dependence in Figure 3 as follows. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ • E-Step: calculate the posterior probability of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ each latent variable z given the observation |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of image features f, image instances v and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ text features w based on the old estimate of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P(f I z), P(zI w) and P(zI v): |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ PIvi,fj) = P(fjIzk)P(zkIvi) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (zk |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ P(zkI wl, fj) = P(fjIzk)P(zkIwl) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Pk, P(fjIzk,)P(zk,Iwl) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (7) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ • M-Step: re-estimates conditional probabili- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ ties P(zkIvi) and P(zkI wl): |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ PAAP(zkIvi,fj) (8) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ PBlBlj, P(zk I wl , fj) (9) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and conditional probability P(fj I zk ), which |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is a mixture portion of posterior probability |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of latent variables |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ W |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ V |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Z F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ P(fIz) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ P |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ k, P(fj Izk,)P(zk, I vi) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (6) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ XP(zkIvi) = |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ j |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ XP(zkIwl) = |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ j |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ P(fjIzk) a AX |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ i |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L=X |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ j |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ P |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ j, Blj, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ X |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ +(1 — A) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ l |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ +(1—A)X |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ l |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Aij P( zkI vi, fj) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Pj,Aij, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Blj P(zkIwl,fj) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (10) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ XAij |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ A i Pj, Aij, log P(fj I vi) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Blj log P(fj I wl ) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Pj, Blj, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (5) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ where AIVIxIFI E RIVIxIFI is the image instance- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ feature co-occurrence matrix, and BIW I xIFI E |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ RIWIxIFI is the text-to-image feature-level co- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ occurrence matrix. Similar to Equation (2), |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PAij and PB`j in Equation (5) are the nor- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ij, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ malization terms to prevent imbalanced cases. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Furthermore, A acts as a trade-off parameter be- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tween the co-occurrence matrices A and B. In |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the extreme case when A = 1, the log-likelihood |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ objective function ignores all the biases from the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Finally, the clustering function for a certain im- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ age v is |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ g(v) = argmax P(zIv). (11) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ZEZ |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ From the above equations, we can derive |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ our annotation-based probabilistic latent semantic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ analysis (aPLSA) algorithm. As shown in Algo- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rithm 1, aPLSA iteratively performs the E-Step |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the M-Step in order to seek local optimal |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ points based on the objective function L in Equa- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion (5). |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Algorithm 1 Annotation-based PLSA Algorithm |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (aPLSA) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Input: The V-F co-occurrence matrix A and W- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ F co-occurrence matrix B. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Output: A clustering (partition) function g : V H |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ i, which maps an image instance v E V to a latent |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ variable z E i. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 1: Initial i so that IiI equals the number clus- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ ters desired. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2: Initialize P(zIv), P(zIw), P(f Iz) randomly. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ 3: while the change of L in Eq. (5) between two |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ sequential iterations is greater than a prede- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fined threshold do |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4: E-Step: Update P(zIv, f) and P(zIw, f) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ based on Eq. (6) and (7) respectively. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ 5: M-Step: Update P(zIv), P(zIw) and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ P(f Iz) based on Eq. (8), (9) and (10) re- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ spectively. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6: end while |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ 7: for all v in V do |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ 8: g(v) +— argmaxP(zIv). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ 9: end for |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ 10: Return g. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ 4 Experiments |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In this section, we empirically evaluate the aPLSA |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ algorithm together with some state-of-art base- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ line methods on two widely used image corpora, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to demonstrate the effectiveness of our algorithm |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ aPLSA. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.1 Data Sets |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In order to evaluate the effectiveness of our algo- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ rithm aPLSA, we conducted experiments on sev- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ eral data sets generated from two image corpora, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Caltech-256 (Griffin et al., 2007) and the fifteen- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scene (Lazebnik et al., 2006). The Caltech-256 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data set has 256 image objective categories, rang- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing from animals to buildings, from plants to au- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tomobiles, etc. The fifteen-scene data set con- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tains 15 scenes such as store and forest. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ From these two corpora, we randomly generated |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ eleven image clustering tasks, including seven 2- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ way clustering tasks, two 4-way clustering task, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ one 5-way clustering task and one 8-way cluster- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing task. The detailed descriptions for these clus- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tering tasks are given in Table 1. In these tasks, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ b i 7 and o c t 1 were generated from fifteen-scene |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data set, and the rest were from Caltech-256 data |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ set. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ DATA SET INVOLVED CLASSES DATA SIZE |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ bi1 skateboard, airplanes 102,800 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ bi2 billiards, mars 278, 155 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi3 cd, greyhound 102, 94 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi4 electric-guitar, snake 122,112 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi5 calculator, dolphin 100,106 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi6 mushroom, teddy-bear 202, 99 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi7 MIThighway, livingroom 260,289 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quad1 calculator, diamond-ring, dolphin, 100, 118, 106,116 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ microscope |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quad2 bonsai, comet, frog, saddle 122, 120, 115, 110 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quint1 frog, kayak, bear, jesus-christ,watch115,102,101,87, 201 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ oct1 MIThighway, MITmountain, 260, 374, 210, 360, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ kitchen, MITcoast, PARoffice, MIT- tallbuilding, livingroom, bedroom 215, 356, 289, 216 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tune1 coin, horse 123,270 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tune2 socks, spider 111,106 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tune3 galaxy, snowmobile 80,112 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tune4 dice, fern 98,110 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tune5 backpack, lightning, mandolin, swan 151, 136, 93,114 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 1: The descriptions of all the image clus- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tering tasks used in our experiment. Among |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these data sets, b i 7 and o c t 1 were generated |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from fifteen-scene data set, and the rest were from |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Caltech-256 data set. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To empirically investigate the parameter A and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the convergence of our algorithm aPLSA, we gen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erated five more date sets as the development sets. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The detailed description of these five development |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sets, namely tune1 to tune5 is listed in Table 1 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as well. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The auxiliary data were crawled from the Flickr |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (http://www.flickr.com/) web site dur- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing August 2007. Flickr is an internet community |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ where people share photos online and express their |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ opinions as social tags (annotations) attached to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each image. From Flicker, we collected 19,959 |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ images and 91,719 related annotations, among |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which 2,600 words are distinct. Based on the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method described in Section 3, we estimated the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ co-occurrence matrix B between text features and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ image features. This co-occurrence matrix B was |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used by all the clustering tasks in our experiments. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For data preprocessing, we adopted the bag-of- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ features representation of images (Li and Perona, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2005) in our experiments. Interesting points were |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ found in the images and described via the SIFT |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ descriptors (Lowe, 2004). Then, the interesting |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ points were clustered to generate a codebook to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ form an image feature space. The size of code- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ book was set to 2, 000 in our experiments. Based |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on the codebook, which serves as the image fea- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture space, each image can be represented as a cor- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ responding feature vector to be used in the next |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ step. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To set our evaluation criterion, we used the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 6 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Data Set KMeancombined PLSA STC aPLSA |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ separate separate combined |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi1 0.645±0.064 0.548±0.031 0.544±0.074 0.537±0.033 0.586±0.139 0.482±0.062 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi2 0.687±0.003 0.662±0.014 0.464±0.074 0.692±0.001 0.577±0.016 0.455±0.096 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi3 1.294±0.060 1.300±0.015 1.085±0.073 1.126±0.036 1.103±0.108 1.029±0.074 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi4 1.227±0.080 1.164±0.053 0.976±0.051 1.038±0.068 1.024±0.089 0.919±0.065 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi5 1.450±0.058 1.417±0.045 1.426±0.025 1.405±0.040 1.411±0.043 1.377±0.040 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi6 1.969±0.078 1.852±0.051 1.514±0.039 1.709±0.028 1.589±0.121 1.503±0.030 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bi7 0.686±0.006 0.683±0.004 0.643±0.058 0.632±0.037 0.651±0.012 0.624±0.066 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quad1 0.591±0.094 0.675±0.017 0.488±0.071 0.662±0.013 0.580±0.115 0.432±0.085 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quad2 0.648±0.036 0.646±0.045 0.614±0.062 0.626±0.026 0.591±0.087 0.515±0.098 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quint1 0.557±0.021 0.508±0.104 0.547±0.060 0.539±0.051 0.538±0.100 0.502±0.067 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ oct1 0.659±0.031 0.680±0.012 0.340±0.147 0.691±0.002 0.411±0.089 0.306±0.101 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ average 0.947±0.029 0.922±0.017 0.786±0.009 0.878±0.006 0.824±0.036 0.741±0.018 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2: Experimental result in term of entropy for all data sets and evaluation methods. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ entropy to measure the quality of our clustering |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ results. In information theory, entropy (Shan- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ non, 1948) is a measure of the uncertainty as- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sociated with a random variable. In our prob- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lem, entropy serves as a measure of randomness |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of clustering result. The entropy of g on a sin- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gle latent variable z is defined to be H(g, z) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ - PcEC P(c�z)lo�2 P(cIz), where C is the class |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ label set of V and P(clz) = ��v�s(v)=zAt(v)=c}� |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 11v1s(v)=z}1 , |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ in which t(v) is the true class label of image v. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Lower entropy H(g, i) indicates less randomness |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and thus better clustering result. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2 Empirical Analysis |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We now empirically analyze the effectiveness of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ our aPLSA algorithm. Because, to our best of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ knowledge, few existing methods addressed the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ problem of image clustering with the help of so- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cial annotation image data, we can only compare |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our aPLSA with several state-of-the-art cluster- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing algorithms that are not directly designed for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our problem. The first baseline is the well-known |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ KMeans algorithm (MacQueen, 1967). Since our |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ algorithm is designed based on PLSA (Hofmann, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1999), we also included PLSA for clustering as a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ baseline method in our experiments. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For each of the above two baselines, we have |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ two strategies: (1) separated: the baseline |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method was applied on the target image data only; |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2) combined: the baseline method was applied |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to cluster the combined data consisting of both |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target image data and the annotated image data. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Clustering results on target image data were used |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for evaluation. Note that, in the combined data, all |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the annotations were thrown away since baseline |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ methods evaluated in this paper do not leverage |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotation information. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In addition, we compared our algorithm aPLSA |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ to a state-of-the-art transfer clustering strategy, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ known as self-taught clustering (STC) (Dai et al., |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2008b). STC makes use of auxiliary data to esti- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mate a better feature representation to benefit the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target clustering. In these experiments, the anno- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tated image data were used as auxiliary data in |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ STC, which does not use the annotation text. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In our experiments, the performance is in the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ form of the average entropy and variance of five |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ repeats by randomly selecting 50 images from |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each of the categories. We selected only 50 im- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ages per category, since this paper is focused on |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clustering sparse data. Table 2 shows the perfor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mance with respect to all comparison methods on |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each of the image clustering tasks measured by |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the entropy criterion. From the tables, we can see |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that our algorithm aPLSA outperforms the base- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ line methods in all the data sets. We believe that is |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ because aPLSA can effectively utilize the knowl- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ edge from the socially annotated image data. On |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ average, aPLSA gives rise to 21.8% of entropy re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ duction and as compared to KMeans, 5.7% of en- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tropy reduction as compared to PLSA, and 10.1% |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of entropy reduction as compared to S TC. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2.1 Varying Data Size |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We now show how the data size affects aPLSA, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ with two baseline methods KMeans and PLSA as |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reference. The experiments were conducted on |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different amounts of target image data, varying |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from 10 to 80. The corresponding experimental |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ results in average entropy over all the 11 clustering |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tasks are shown in Figure 4(a). From this figure, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we observe that aPLSA always yields a significant |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reduction in entropy as compared with two base- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ line methods KMeans and PLSA, regardless of the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ size of target image data that we used. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0.95 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.9 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.85 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.8 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.75 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.7 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 20 30 40 50 60 70 80 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ KMeans |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ PLSA |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ aPLSA |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.75 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0.65 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0.6 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.55 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.45 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0.7 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.4 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0 0.2 0.4 0.6 0.8 1 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ average over 5 development sets |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.75 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.65 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.55 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.7 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.6 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0.5 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 0 50 100 150 200 250 300 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ average over 5 development sets |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Data size per category Number of Iteration |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (a) (b) (c) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 4: (a) The entropy curve as a function of different amounts of data per category. (b) The entropy |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ curve as a function of different number of iterations. (c) The entropy curve as a function of different |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trade-off parameter A. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2.2 Parameter Sensitivity |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In aPLSA, there is a trade-off parameter A that af- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ fects how the algorithm relies on auxiliary data. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ When A = 0, the aPLSA relies only on annotated |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ image data B. When A = 1, aPLSA relies only |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on target image data A, in which case aPLSA de- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ generates to PLSA. Smaller A indicates heavier re- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ liance on the annotated image data. We have done |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ some experiments on the development sets to in- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vestigate how different A affect the performance |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of aPLSA. We set the number of images per cate- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gory to 50, and tested the performance of aPLSA. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The result in average entropy over all development |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sets is shown in Figure 4(b). In the experiments |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ described in this paper, we set A to 0.2, which is |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the best point in Figure 4(b). |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2.3 Convergence |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In our experiments, we tested the convergence |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ property of our algorithm aPLSA as well. Fig- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ure 4(c) shows the average entropy curve given |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by aPLSA over all development sets. From this |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ figure, we see that the entropy decreases very fast |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ during the first 100 iterations and becomes stable |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ after 150 iterations. We believe that 200 iterations |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is sufficient for aPLSA to converge. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 Conclusions |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In this paper, we proposed a new learning scenario |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ called heterogeneous transfer learning and illus- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trated its application to image clustering. Image |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clustering, a vital component in organizing search |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ results for query-based image search, was shown |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to be improved by transferring knowledge from |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unrelated images with annotations in a social Web. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ This is done by first learning the high-quality la- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tent variables in the auxiliary data, and then trans- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferring this knowledge to help improve the cluster- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing of the target image data. We conducted experi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments on two image data sets, using the Flickr data |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ as the annotated auxiliary image data, and showed |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that our aPLSA algorithm can greatly outperform |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ several state-of-the-art clustering algorithms. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In natural language processing, there are many |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ future opportunities to apply heterogeneous trans- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fer learning. In (Ling et al., 2008) we have shown |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ how to classify the Chinese text using English text |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as the training data. We may also consider cluster- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing, topic modeling, question answering, etc., to |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be done using data in different feature spaces. We |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ can consider data in different modalities, such as |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ video, image and audio, as the training data. Fi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nally, we will explore the theoretical foundations |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and limitations of heterogeneous transfer learning |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as well. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Acknowledgement Qiang Yang thanks Hong |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Kong CERG grant 621307 for supporting the re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ search. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ References |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Alina Andreevskaia and Sabine Bergler. 2008. When spe- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ cialists and generalists work together: Overcoming do- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ main dependence in sentiment tagging. In ACL-08: HLT, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 290–298, Columbus, Ohio, June. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Andrew Arnold, Ramesh Nallapati, and William W. Cohen. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2007. A comparative study of methods for transductive |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ transfer learning. In ICDM 2007 Workshop on Mining |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Management of Biological Data, pages 77-82. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Andrew Arnold, Ramesh Nallapati, and William W. Cohen. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2008. Exploiting feature hierarchy for transfer learning in |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ named entity recognition. In ACL-08: HLT. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Sugato Basu, Mikhail Bilenko, and Raymond J. Mooney. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2004. A probabilistic framework for semi-supervised |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clustering. In ACM SIGKDD 2004, pages 59–68. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ John Blitzer, Ryan Mcdonald, and Fernando Pereira. 2006. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Domain adaptation with structural correspondence learn- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing. In EMNLP 2006, pages 120–128, Sydney, Australia. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 8 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ John Blitzer, Mark Dredze, and Fernando Pereira. 2007. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Biographies, bollywood, boom-boxes and blenders: Do- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ main adaptation for sentiment classification. In ACL 2007, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 440–447, Prague, Czech Republic. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Avrim Blum and Tom Mitchell. 1998. Combining labeled |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and unlabeled data with co-training. In COLT 1998, pages |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 92–100, New York, NY, USA. ACM. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Rich Caruana. 1997. Multitask learning. Machine Learning, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 28(1):41–75. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Yee Seng Chan and Hwee Tou Ng. 2007. Domain adaptation |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ with active learning for word sense disambiguation. In |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ACL 2007, Prague, Czech Republic. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ David A. Cohn and Thomas Hofmann. 2000. The missing |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ link - a probabilistic model of document content and hy- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pertext connectivity. In NIPS 2000, pages 430–436. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Wenyuan Dai, Yuqiang Chen, Gui-Rong Xue, Qiang Yang, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and Yong Yu. 2008a. Translated learning: Transfer learn- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing across different feature spaces. In NIPS 2008, pages |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 353–360. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Wenyuan Dai, Qiang Yang, Gui-Rong Xue, and Yong Yu. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2008b. Self-taught clustering. In ICML 2008, pages 200– |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 207. Omnipress. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hal Daume,III. 2007. Frustratingly easy domain adaptation. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In ACL 2007, pages 256–263, Prague, Czech Republic. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Jesse Davis and Pedro Domingos. 2008. Deep transfer via |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ second-order markov logic. In AAAI 2008 Workshop on |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Transfer Learning, Chicago, USA. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Scott Deerwester, Susan T. Dumais, George W. Furnas, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Thomas K. L, and Richard Harshman. 1990. Indexing by |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ latent semantic analysis. Journal of the American Society |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for Information Science, pages 391–407. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ A. P. Dempster, N. M. Laird, and D. B. Rubin. 1977. Max- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ imum likelihood from incomplete data via the em algo- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rithm. J. of the Royal Statistical Society, 39:1–38. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Thomas Finley and Thorsten Joachims. 2005. Supervised |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ clustering with support vector machines. In ICML 2005, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 217–224, New York, NY, USA. ACM. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ G. Griffin, A. Holub, and P. Perona. 2007. Caltech-256 ob- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ject category dataset. Technical Report 7694, California |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Institute of Technology. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Thomas Hofmann. 1999 Probabilistic latent semantic anal- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ysis. In Proc. of Uncertainty in Artificial Intelligence, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ UAI99. Pages 289–296 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Thomas Hofmann. 2001. Unsupervised learning by proba- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ bilistic latent semantic analysis. Machine Learning. vol- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ume 42, number 1-2, pages 177–196. Kluwer Academic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Publishers. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Jing Jiang and Chengxiang Zhai. 2007. Instance weighting |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ for domain adaptation in NLP. In ACL 2007, pages 264– |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 271, Prague, Czech Republic, June. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Leonard Kaufman and Peter J. Rousseeuw. 1990. Finding |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ groups in data: an introduction to cluster analysis. John |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Wiley and Sons, New York. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. 2006. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Beyond bags of features: Spatial pyramid matching for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ recognizing natural scene categories. In CVPR 2006, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 2169–2178, Washington, DC, USA. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Fei-Fei Li and Pietro Perona. 2005. A bayesian hierarchi- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ cal model for learning natural scene categories. In CVPR |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2005, pages 524–531, Washington, DC, USA. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Xiao Ling, Gui-Rong Xue, Wenyuan Dai, Yun Jiang, Qiang |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Yang, and Yong Yu. 2008. Can chinese web pages be |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ classified with english data source? In WWW 2008, pages |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 969–978, New York, NY, USA. ACM. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Nicolas Loeff, Cecilia Ovesdotter Alm, and David A. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Forsyth. 2006. Discriminating image senses by clustering |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with multimodal features. In COLING/ACL 2006 Main |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conference poster sessions, pages 547–554. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ David G. Lowe. 2004. Distinctive image features from scale- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ invariant keypoints. International Journal of Computer |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Vision (IJCV) 2004, volume 60, number 2, pages 91–110. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ J. B. MacQueen. 1967. Some methods for classification and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ analysis of multivariate observations. In Proceedings of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Fifth Berkeley Symposium on Mathematical Statistics and |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Probability, pages 1:281–297, Berkeley, CA, USA. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Kamal Nigam and Rayid Ghani. 2000. Analyzing the effec- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tiveness and applicability of co-training. In Proceedings |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the Ninth International Conference on Information and |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Knowledge Management, pages 86–93, New York, USA. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Rajat Raina, Alexis Battle, Honglak Lee, Benjamin Packer, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and Andrew Y. Ng. 2007. Self-taught learning: transfer |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning from unlabeled data. In ICML 2007, pages 759– |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 766, New York, NY, USA. ACM. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Roi Reichart and Ari Rappoport. 2007. Self-training for |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ enhancement and domain adaptation of statistical parsers |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trained on small datasets. In ACL 2007. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Roi Reichart, Katrin Tomanek, Udo Hahn, and Ari Rap- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ poport. 2008. Multi-task active learning for linguistic |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotations. In ACL-08: HLT, pages 861–869. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ C. E. Shannon. 1948. A mathematical theory of communi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ cation. Bell system technicaljournal, 27. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ J. Sivic, B. C. Russell, A. A. Efros, A. Zisserman, and W. T. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Freeman. 2005. Discovering object categories in image |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ collections. In ICCV 2005. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Naftali Tishby, Fernando C. Pereira, and William Bialek. The |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ information bottleneck method. 1999. In Proc. of the 37- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ th Annual Allerton Conference on Communication, Con- |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ trol and Computing, pages 368–377. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Pengcheng Wu and Thomas G. Dietterich. 2004. Improving |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ svm accuracy by training on auxiliary data sources. In |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ICML 2004, pages 110–117, New York, NY, USA. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Yejun Wu and Douglas W. Oard. 2008. Bilingual topic as- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ pect classification with a few training examples. In ACM |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SIGIR 2008, pages 203–210, New York, NY, USA. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Xiaojin Zhu. 2007. Semi-supervised learning literature sur- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ vey. Technical Report 1530, Computer Sciences, Univer- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sity of Wisconsin-Madison. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 9 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Investigations on Word Senses and Word Usages |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Katrin Erk Diana McCarthy Nicholas Gaylord |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ University of Texas at Austin University of Sussex University of Texas at Austin |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ katrin.erk@mail.utexas.edu dianam@sussex.ac.uk nlgaylord@mail.utexas.edu |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The vast majority of work on word senses |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ has relied on predefined sense invento- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ries and an annotation schema where each |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word instance is tagged with the best fit- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ting sense. This paper examines the case |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for a graded notion of word meaning in |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two experiments, one which uses WordNet |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ senses in a graded fashion, contrasted with |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the “winner takes all” annotation, and one |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which asks annotators to judge the similar- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ity of two usages. We find that the graded |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ responses correlate with annotations from |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ previous datasets, but sense assignments |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are used in a way that weakens the case for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clear cut sense boundaries. The responses |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from both experiments correlate with the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ overlap of paraphrases from the English |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lexical substitution task which bodes well |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for the use of substitutes as a proxy for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word sense. This paper also provides two |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ novel datasets which can be used for eval- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ uating computational systems. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The vast majority of work on word sense tag- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ging has assumed that predefined word senses |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from a dictionary are an adequate proxy for the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ task, although of course there are issues with |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this enterprise both in terms of cognitive valid- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ity (Hanks, 2000; Kilgarriff, 1997; Kilgarriff, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2006) and adequacy for computational linguis- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tics applications (Kilgarriff, 2006). Furthermore, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ given a predefined list of senses, annotation efforts |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and computational approaches to word sense dis- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ambiguation (WSD) have usually assumed that one |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ best fitting sense should be selected for each us- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ age. While there is usually some allowance made |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for multiple senses, this is typically not adopted by |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ annotators or computational systems. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Research on the psychology of concepts (Mur- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ phy, 2002; Hampton, 2007) shows that categories |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the human mind are not simply sets with clear- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cut boundaries: Some items are perceived as |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ more typical than others (Rosch, 1975; Rosch and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Mervis, 1975), and there are borderline cases on |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which people disagree more often, and on whose |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ categorization they are more likely to change their |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ minds (Hampton, 1979; McCloskey and Glucks- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ berg, 1978). Word meanings are certainly related |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to mental concepts (Murphy, 2002). This raises |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the question of whether there is any such thing as |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the one appropriate sense for a given occurrence. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this paper we will explore using graded re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ sponses for sense tagging within a novel annota- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion paradigm. Modeling the annotation frame- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work after psycholinguistic experiments, we do |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ not train annotators to conform to sense distinc- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions; rather we assess individual differences by |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ asking annotators to produce graded ratings in- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stead of making a binary choice. We perform two |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotation studies. In the first one, referred to |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as WSsim (Word Sense Similarity), annotators |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ give graded ratings on the applicability of Word- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Net senses. In the second one, Usim (Usage Sim- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ilarity), annotators rate the similarity of pairs of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ occurrences (usages) of a common target word. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Both studies explore whether users make use of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a graded scale or persist in making binary deci- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sions even when there is the option for a graded |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ response. The first study additionally tests to what |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ extent the judgments on WordNet senses fall into |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clear-cut clusters, while the second study allows |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ us to explore meaning similarity independently of |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ any lexicon resource. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 10–18, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 2 Related Work |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Manual word sense assignment is difficult for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ human annotators (Krishnamurthy and Nicholls, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2000). Reported inter-annotator agreement (ITA) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for fine-grained word sense assignment tasks has |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ranged between 69% (Kilgarriff and Rosenzweig, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2000) for a lexical sample using the HECTOR dic- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tionary and 78.6.% using WordNet (Landes et al., |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1998) in all-words annotation. The use of more |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ coarse-grained senses alleviates the problem: In |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ OntoNotes (Hovy et al., 2006), an ITA of 90% is |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used as the criterion for the construction of coarse- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grained sense distinctions. However, intriguingly, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for some high-frequency lemmas such as leave |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this ITA threshold is not reached even after mul- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tiple re-partitionings of the semantic space (Chen |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Palmer, 2009). Similarly, the performance |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of WSD systems clearly indicates that WSD is not |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ easy unless one adopts a coarse-grained approach, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and then systems tagging all words at best perform |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a few percentage points above the most frequent |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense heuristic (Navigli et al., 2007). Good perfor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mance on coarse-grained sense distinctions may |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be more useful in applications than poor perfor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mance on fine-grained distinctions (Ide and Wilks, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2006) but we do not know this yet and there is |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ some evidence to the contrary (Stokoe, 2005). |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Rather than focus on the granularity of clus- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ters, the approach we will take in this paper |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is to examine the phenomenon of word mean- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing both with and without recourse to predefined |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ senses by focusing on the similarity of uses of a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word. Human subjects show excellent agreement |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on judging word similarity out of context (Ruben- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stein and Goodenough, 1965; Miller and Charles, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1991), and human judgments have previously been |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used successfully to study synonymy and near- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ synonymy (Miller and Charles, 1991; Bybee and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Eddington, 2006). We focus on polysemy rather |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ than synonymy. Our aim will be to use WSsim |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to determine to what extent annotations form co- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hesive clusters. In principle, it should be possi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ble to use existing sense-annotated data to explore |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this question: almost all sense annotation efforts |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ have allowed annotators to assign multiple senses |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to a single occurrence, and the distribution of these |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense labels should indicate whether annotators |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ viewed the senses as disjoint or not. However, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the percentage of markables that received multi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ple sense labels in existing corpora is small, and it |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ varies massively between corpora: In the SemCor |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corpus (Landes et al., 1998), only 0.3% of all |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ markables received multiple sense labels. In the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SENSEVAL-3 English lexical task corpus (Mihal- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cea et al., 2004) (hereafter referred to as SE-3), the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ratio is much higher at 8% of all markables1. This |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ could mean annotators feel that there is usually a |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ single applicable sense, or it could point to a bias |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ towards single-sense assignment in the annotation |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guidelines and/or the annotation tool. The WSsim |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ experiment that we report in this paper is designed |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to eliminate such bias as far as possible and we |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conduct it on data taken from SemCor and SE-3 so |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that we can compare the annotations. Although we |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use WordNet for the annotation, our study is not a |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ study of WordNet per se. We choose WordNet be- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cause it is sufficiently fine-grained to examine sub- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tle differences in usage, and because traditionally |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotated datasets exist to which we can compare |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our results. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Predefined dictionaries and lexical resources are |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ not the only possibilities for annotating lexical |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ items with meaning. In cross-lingual settings, the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ actual translations of a word can be taken as the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense labels (Resnik and Yarowsky, 2000). Re- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cently, McCarthy and Navigli (2007) proposed |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the English Lexical Substitution task (hereafter |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ referred to as LEXSUB) under the auspices of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SemEval-2007. It uses paraphrases for words in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ context as a way of annotating meaning. The task |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ was proposed following a background of discus- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sions in the WSD community as to the adequacy |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of predefined word senses. The LEXSUB dataset |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comprises open class words (nouns, verbs, adjec- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tives and adverbs) with token instances of each |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word appearing in the context of one sentence |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ taken from the English Internet Corpus (Sharoff, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2006). The methodology can only work where |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ there are paraphrases, so the dataset only contains |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words with more than one meaning where at least |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two different meanings have near synonyms. For |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ meanings without obvious substitutes the annota- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tors were allowed to use multiword paraphrases or |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words with slightly more general meanings. This |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dataset has been used to evaluate automatic sys- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tems which can find substitutes appropriate for the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ context. To the best of our knowledge there has |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ been no study of how the data collected relates to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word sense annotations or judgments of semantic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ similarity. In this paper we examine these relation- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 This is even though both annotation efforts use balanced |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ corpora, the Brown corpus in the case of SemCor, the British |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ National Corpus for SE-3. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 11 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ships by re-using data from LEXSUB in both new |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ annotation experiments and testing the results for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correlation. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Annotation |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We conducted two experiments through an on- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ line annotation interface. Three annotators partic- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ipated in each experiment; all were native British |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ English speakers. The first experiment, WSsim, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ collected annotator judgments about the applica- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bility of dictionary senses using a 5-point rating |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scale. The second, Usim, also utilized a 5-point |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scale but collected judgments on the similarity in |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ meaning between two uses of a word. 2 The scale |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ was 1 – completely different, 2 – mostly different, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 – similar, 4 – very similar and 5 – identical. In |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Usim, this scale rated the similarity of the two uses |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the common target word; in WSsim it rated the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ similarity between the use of the target word and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the sense description. In both experiments, the an- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation interface allowed annotators to revisit and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ change previously supplied judgments, and a com- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment box was provided alongside each item. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim. This experiment contained a total of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 430 sentences spanning 11 lemmas (nouns, verbs |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and adjectives). For 8 of these lemmas, 50 sen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tences were included, 25 of them randomly sam- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pled from SemCor 3 and 25 randomly sampled |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from SE-3 .4 The remaining 3 lemmas in the ex- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ periment each had 10 sentences taken from the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ LEXSUB data. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ WSsim is a word sense annotation task using |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ WordNet senses.5 Unlike previous word sense an- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation projects, we asked annotators to provide |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments on the applicability of every WordNet |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense of the target lemma with the instruction: 6 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2Throughout this paper, a target word is assumed to be a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ word in a given PoS. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3The SemCor dataset was produced alongside WordNet, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ so it can be expected to support the WordNet sense distinc- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions. The same cannot be said for SE-3. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4Sentence fragments and sentences with 5 or fewer words |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ were excluded from the sampling. Annotators were given |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the sentences, but not the original annotation from these re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sources. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5WordNet 1.7.1 was used in the annotation of both SE-3 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and SemCor; we used the more current WordNet 3.0 after |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verifying that the lemmas included in this experiment had the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ same senses listed in both versions. Care was taken addition- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ally to ensure that senses were not presented in an order that |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reflected their frequency of occurrence. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6The guidelines for both experiments are avail- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ able at http://comp.ling.utexas.edu/ |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ people/katrin erk/graded sense and usage |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ annotation |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Your task is to rate, for each of these descriptions, |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ how well they reflect the meaning of the boldfaced |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word in the sentence. |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Applicability judgments were not binary, but were |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ instead collected using the five-point scale given |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ above which allowed annotators to indicate not |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only whether a given sense applied, but to what |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ degree. Each annotator annotated each of the 430 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ items. By having multiple annotators per item and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a graded, non-binary annotation scheme we al- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ low for and measure differences between annota- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tors, rather than training annotators to conform to |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a common sense distinction guideline. By asking |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotators to provide ratings for each individual |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense, we strive to eliminate all bias towards either |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ single-sense or multiple-sense assignment. In tra- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ditional word sense annotation, such bias could be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ introduced directly through annotation guidelines |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or indirectly, through tools that make it easier to |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ assign fewer senses. We focus not on finding the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ best fitting sense but collect judgments on the ap- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ plicability of all senses. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Usim. This experiment used data from LEXSUB. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ For more information on LEXSUB, see McCarthy |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Navigli (2007). 34 lemmas (nouns, verbs, ad- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ jectives and adverbs) were manually selected, in- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cluding the 3 lemmas also used in WSsim. We se- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lected lemmas which exhibited a range of mean- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ings and substitutes in the LEXSUB data, with |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as few multiword substitutes as possible. Each |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lemma is the target in 10 LEXSUB sentences. For |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our experiment, we took every possible pairwise |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comparison of these 10 sentences for a lemma. We |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ refer to each such pair of sentences as an SPAIR. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The resulting dataset comprised 45 SPAIRs per |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lemma, adding up to 1530 comparisons per anno- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tator overall. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this annotation experiment, annotators saw |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ SPAIRs with a common target word and rated the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ similarity in meaning between the two uses of the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target word with the instruction: |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Your task is to rate, for each pair of sentences, how |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ similar in meaning the two boldfaced words are on |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a five -point scale. |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In addition annotators had the ability to respond |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ with “Cannot Decide”, indicating that they were |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unable to make an effective comparison between |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the two contexts, for example because the mean- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing of one usage was unclear. This occurred in |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 9 paired occurrences during the course of anno- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tation, and these items (paired occurrences) were |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 12 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ excluded from further analysis. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The purpose of Usim was to collect judgments |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ about degrees of similarity between a word’s |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ meaning in different contexts. Unlike WSsim, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Usim does not rely upon any dictionary resource |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as a basis for the judgments. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 Analyses |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ This section reports on analyses on the annotated |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ data. In all the analyses we use Spearman’s rank |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correlation coefficient (p), a nonparametric test, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ because the data does not seem to be normally |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ distributed. We used two-tailed tests in all cases, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rather than assume the direction of the relation- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ship. As noted above, we have three annotators |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ per task, and each annotator gave judgments for |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ every sentence (WSsim) or sentence pair (Usim). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Since the annotators may vary as to how they use |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the ordinal scale, we do not use the mean of judg- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments7 but report all individual correlations. All |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ analyses were done using the R package.8 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.1 WSsim analysis |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ In the WSsim experiment, annotators rated the ap- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ plicability of each WordNet 3.0 sense for a given |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target word occurrence. Table 1 shows a sample |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotation for the target argument.n. 9 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Pattern of annotation and annotator agree- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ment. Figure 1 shows how often each of the five |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments on the scale was used, individually and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ summed over all annotators. (The y-axis shows |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ raw counts of each judgment.) We can see from |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this figure that the extreme ratings 1 and 5 are used |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ more often than the intermediate ones, but annota- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tors make use of the full ordinal scale when judg- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the applicability of a sense. Also, the figure |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shows that annotator 1 used the extreme negative |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rating 1 much less than the other two annotators. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 2 shows the percentage of times each judg- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment was used on senses of three lemmas, differ- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent.a, interest.n, and win.v. In WordNet, they have |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5, 7, and 4 senses, respectively. The pattern for |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ win.v resembles the overall distribution of judg- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments, with peaks at the extreme ratings 1 and 5. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The lemma interest.n has a single peak at rating |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1, partly due to the fact that senses 5 (financial |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7We have also performed several of our calculations us- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing the mean judgment, and they also gave highly significant |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ results in all the cases we tested. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 8http://www.r-project.org/ |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 9We use word.PoS to denote a target word (lemma). |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Annotator 1 Annotator 2 Annotator 3 overall |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 1: WSsim experiment: number of times |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ each judgment was used, by annotator and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ summed over all annotators. The y-axis shows raw |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ counts of each judgment. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different.a interest.n win.v |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 2: WSsim experiment: percentage of times |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ each judgment was used for the lemmas differ- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent.a, interest.n and win.v. Judgment counts were |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ summed over all three annotators. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ involvement) and 6 (interest group) were rarely |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ judged to apply. For the lemma different.a, all |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments have been used with approximately the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ same frequency. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We measured the level of agreement between |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ annotators using Spearman’s p between the judg- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments of every pair of annotators. The pairwise |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correlations were p = 0.506, p = 0.466 and p = |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.540, all highly significant with p < 2.2e-16. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Agreement with previous annotation in |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ SemCor and SE-3. 200 of the items in WSsim |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ had been previously annotated in SemCor, and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 200 in SE-3. This lets us compare the annotation |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ results across annotation efforts. Table 2 shows |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the percentage of items where more than one |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense was assigned in the subset of WSsim from |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SemCor (first row), from SE-3 (second row), and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 13 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Sentence 1 2 Senses 5 6 7 Annotator |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 3 4 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ This question provoked arguments in America about the 1 4 4 2 1 1 3 Ann. 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Norton Anthology of Literature by Women, some of the 4 5 4 2 1 1 4 Ann. 2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ contents of which were said to have had little value as literature. 1 4 5 1 1 1 1 Ann. 3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 1: A sample annotation in the WSsim experiment. The senses are: 1:statement, 2:controversy, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 3:debate, 4:literary argument, 5:parameter, 6:variable, 7:line of reasoning |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim judgment |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ >3 >4 5 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 80.2 57.5 28.3 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 78.0 58.3 27.1 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 78.8 57.4 27.7 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Data Orig. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ W Ssim/SemCor 0.0 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim/SE-3 24.0 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ All WSsim |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p<0.05 pos neg p<0.01 pos neg |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Ann. 1 30.8 11.4 23.2 5.9 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ann. 2 22.2 24.1 19.6 19.6 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ann. 3 12.7 12.0 10.0 6.0 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2: Percentage of items with multiple senses |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ assigned. Orig: in the original SemCor/SE-3 data. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim judgment: items with judgments at or |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ above the specified threshold. The percentages for |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim are averaged over the three annotators. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ all of WSsim (third row). The Orig. column |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ indicates how many items had multiple labels in |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the original annotation (SemCor or SE-3)10. Note |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that no item had more than one sense label in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SemCor. The columns under WSsim judgment |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ show the percentage of items (averaged over |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the three annotators) that had judgments at or |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ above the specified threshold, starting from rating |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 – similar. Within WSsim, the percentage of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ multiple assignments in the three rows is fairly |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ constant. WSsim avoids the bias to one sense |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by deliberately asking for judgments on the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ applicability of each sense rather than asking |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotators to find the best one. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To compute the Spearman’s correlation between |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the original sense labels and those given in the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim annotation, we converted SemCor and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SE-3 labels to the format used within WSsim: As- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ signed senses were converted to a judgment of 5, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and unassigned senses to a judgment of 1. For the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim/SemCor dataset, the correlation between |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ original and WSsim annotation was p = 0.234, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p = 0.448, and p = 0.390 for the three anno- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tators, each highly significant with p < 2.2e-16. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For the WSsim/SE-3 dataset, the correlations were |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p = 0.346, p = 0.449 and p = 0.338, each of them |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ again highly significant at p < 2.2e-16. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Degree of sense grouping. Next we test to what |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ extent the sense applicability judgments in the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10Overall, 0.3% of tokens in SemCor have multiple labels, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and 8% of tokens in SE-3, so the multiple label assignment in |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our sample is not an underestimate. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3: Percentage of sense pairs that were sig- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ nificantly positively (pos) or negatively (neg) cor- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ related at p < 0.05 and p < 0.01, shown by anno- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tator. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ j>3 j>4 j=5 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Ann. 1 71.9 49.1 8.1 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ann. 2 55.3 24.7 8.1 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ann. 3 42.8 24.0 4.9 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 4: Percentage of sentences in which at least |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ two uncorrelated (p > 0.05) or negatively corre- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lated senses have been annotated with judgments |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ at the specified threshold. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim task could be explained by more coarse- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ grained, categorial sense assignments. We first |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ test how many pairs of senses for a given lemma |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ show similar patterns in the ratings that they re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ceive. Table 3 shows the percentage of sense pairs |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that were significantly correlated for each anno- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tator.11 Significantly positively correlated senses |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ can possibly be reduced to more coarse-grained |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ senses. Would annotators have been able to des- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ignate a single appropriate sense given these more |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ coarse-grained senses? Call two senses groupable |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ if they are significantly positively correlated; in or- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ der not to overlook correlations that are relatively |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ weak but existent, we use a cutoff of p = 0.05 for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ significant correlation. We tested how often anno- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tators gave ratings of at least similar, i.e. ratings |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ > 3, to senses that were not groupable. Table 4 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shows the percentages of items where at least two |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ non-groupable senses received ratings at or above |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the specified threshold. The table shows that re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gardless of which annotator we look at, over 40% |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of all items had two or more non-groupable senses |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ receive judgments of at least 3 (similar). There |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 11 We exclude senses that received a uniform rating of 1 on |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ all items. This concerned 4 senses for annotator 2 and 6 for |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotator 3. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 14 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 1) We study the methods and concepts that each writer uses to |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ defend the cogency of legal, deliberative, or more generally |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ political prudence against explicit or implicit charges that |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ practical thinking is merely a knack or form of cleverness. |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2) Eleven CIRA members have been convicted of criminal |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ charges and others are awaiting trial. |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 3: An SPAIR for charge.n. Annotator judg- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ments: 2,3,4 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ were even several items where two or more non- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ groupable senses each got a judgment of 5. The |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence in table 1 is a case where several non- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ groupable senses got ratings > 3. This is most |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pronounced for Annotator 2, who along with sense |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 (controversy) assigned senses 1 (statement), 7 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (line of reasoning), and 3 (debate), none of which |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are groupable with sense 2. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2 Usim analysis |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ In this experiment, ratings between 1 and 5 were |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ given for every pairwise combination of sentences |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for each target lemma. An example of an SPAIR |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for charge.n is shown in figure 3. In this case the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verdicts from the annotators were 2, 3 and 4. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Pattern of Annotations and Annotator Agree- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ment Figure 4 gives a bar chart of the judgments |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for each annotator and summed over annotators. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We can see from this figure that the annotators |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use the full ordinal scale when judging the simi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ larity of a word’s usages, rather than sticking to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the extremes. There is variation across words, de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pending on the relatedness of each word’s usages. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 5 shows the judgments for the words bar.n, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work.v and raw.a. We see that bar.n has predom- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ inantly different usages with a peak for category |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1, work.v has more similar judgments (category 5) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ compared to any other category and raw.a has a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ peak in the middle category (3). 12 There are other |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words, like for example fresh.a, where the spread |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is more uniform. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To gauge the level of agreement between anno- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tators, we calculated Spearman’s p between the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments of every pair of annotators as in sec- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion 4.1. The pairwise correlations are all highly |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ significant (p < 2.2e-16) with Spearman’s p = |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.502, 0.641 and 0.501 giving an average corre- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lation of 0.548. We also perform leave-one-out re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sampling following Lapata (2006) which gave us |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a Spearman’s correlation of 0.630. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 12For figure 5 we sum the judgments over annotators. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 4: Usim experiment: number of times each |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ judgment was used, by annotator and summed |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ over all annotators |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bar.n raw.a work.v |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 5: Usim experiment: number of times each |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ judgment was used for bar.n, work.v and raw. a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Comparison with LEXSUB substitutions Next |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ we look at whether the Usim judgments on sen- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tence pairs (SPAIRs) correlate with LEXSUB sub- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stitutes. To do this we use the overlap of substi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tutes provided by the five LEXSUB annotators be- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tween two sentences in an SPAIR. In LEXSUB the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotators had to replace each item (a target word |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ within the context of a sentence) with a substitute |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that fitted the context. Each annotator was permit- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ted to supply up to three substitutes provided that |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ they all fitted the context equally. There were 10 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentences per lemma. For our analyses we take |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ every SPAIR for a given lemma and calculate the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ overlap (inter) of the substitutes provided by the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotators for the two usages under scrutiny. Let |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s1 and s2 be a pair of sentences in an SPAIR and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Annotator 4 Annotator 5 Annotator 6 overall |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 4 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 1 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 3 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 4 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 5 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 15 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ x1 and x2 be the multisets of substitutes for the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ respective sentences. Let freq(w,x) be the fre- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quency of a substitute w in a multiset x of sub- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stitutes for a given sentence. 13 INTER(s1,s2) = |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ �wEx1f1x2 min (freq(w,x1), freq(w,x2)) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ max(1x1 1, 1x2 1) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Using this calculation for each SPAIR we can |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ now compute the correlation between the Usim |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments for each annotator and the INTER val- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ues, again using Spearman’s. The figures are |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shown in the leftmost block of table 5. The av- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erage correlation for the 3 annotators was 0.488 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the p-values were all < 2.2e-16. This shows |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a highly significant correlation of the Usim judg- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments and the overlap of substitutes. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We also compare the WSsim judgments against |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the LEXSUB substitutes, again using the INTER |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ measure of substitute overlap. For this analysis, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we only use those WSsim sentences that are origi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nally from LEXSUB. In WSsim, the judgments for |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a sentence comprise judgments for each WordNet |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sense of that sentence. In order to compare against |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ INTER, we need to transform these sentence-wise |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ratings in WSsim to a WSsim-based judgment of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence similarity. To this end, we compute the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Euclidean Distance14 (ED) between two vectors J1 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and J2 of judgments for two sentences s1, s2 for the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ same lemma E. Each of the n indexes of the vector |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ represent one of the n different WordNet senses |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for E. The value at entry i of the vector J1 is the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgment that the annotator in question (we do not |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ average over annotators here) provided for sense i |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of E for sentence s1. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (J1[i]—J2[i])2) (1) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We correlate the Euclidean distances with |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ INTER. We can only test correlation for the subset |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of WSsim that overlaps with the LEXSUB data: the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 30 sentences for investigator.n, function.n and or- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ der.v, which together give 135 unique SPAIRs. We |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ refer to this subset as Wf1U. The results are given |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the third block of table 5. Note that since we are |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ measuring distance between SPAIRs for WSsim |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 13The frequency of a substitute in a multiset depends on |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the number of LEXSUB annotators that picked the substitute |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for this item. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 14We use Euclidean Distance rather than a normalizing |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ measure like Cosine because a sentence where all ratings are |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 should be very different from a sentence where all senses |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ received a rating of 1. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Usim All Usim Wf1U WSsim Wf1U |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ann. p p ann. p |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 0.383 0.330 1 -0.520 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 0.498 0.635 2 -0.503 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 0.584 0.631 3 -0.463 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 5: Annotator correlation with LEXSUB sub- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ stitute overlap (inter) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ whereas INTER is a measure of similarity, the cor- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ relation is negative. The results are highly signif- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ icant with individual p-values from < 1.067e-10 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to < 1.551e-08 and a mean correlation of -0.495. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The results in the first and third block of table 5 are |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ not directly comparable, as the results in the first |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ block are for all Usim data and not the subset of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ LEXSUB with WSsim annotations. We therefore |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ repeated the analysis for Usim on the subset of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data in WSsim and provide the correlation in the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ middle section of table 5. The mean correlation |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for Usim on this subset of the data is 0.532, which |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is a stronger relationship compared to WSsim, al- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ though there is more discrepancy between individ- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ual annotators, with the result for annotator 4 giv- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing a p-value = 9.139e-05 while the other two an- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notators had p-values < 2.2e-16. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The LEXSUB substitute overlaps between dif- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ferent usages correlate well with both Usim and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim judgments, with a slightly stronger rela- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tionship to Usim, perhaps due to the more compli- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cated representation of word meaning in WSsim |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which uses the full set of WordNet senses. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.3 Correlation between WSsim and Usim |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ As we showed in section 4.1, WSsim correlates |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ with previous word sense annotations in SemCor |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and SE-3 while allowing the user a more graded |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ response to sense tagging. As we saw in sec- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion 4.2, Usim and WSsim judgments both have a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ highly significant correlation with similarity of us- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ages as measured using the overlap of substitutes |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from LEXSUB. Here, we look at the correlation |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of WSsim and Usim, considering again the sub- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ set of data that is common to both experiments. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We again transform WSsim sense judgments for |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ individual sentences to distances between SPAIRs |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using Euclidean Distance (ED). The Spearman’s |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p range between —0.307 and —0.671, and all re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults are highly significant with p-values between |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0.0003 and < 2.2e-16. As above, the correla- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion is negative because ED is a distance measure |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ between sentences in an SPAIR, whereas the judg- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ED(J1,J2) = V1( |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ n |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ i=1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 16 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ments for Usim are similarity judgments. We see |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ that there is highly significant correlation for every |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pairing of annotators from the two experiments. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 Discussion |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Validity of annotation scheme. Annotator rat- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ings show highly significant correlation on both |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tasks. This shows that the tasks are well-defined. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In addition, there is a strong correlation between |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSsim and Usim, which indicates that the poten- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tial bias introduced by the use of dictionary senses |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in WSsim is not too prominent. However, we note |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that WSsim only contained a small portion of 3 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lemmas (30 sentences and 135 SPAIRs) in com- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mon with Usim, so more annotation is needed to |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be certain of this relationship. Given the differ- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ences between annotator 1 and the other annota- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tors in Fig. 1, it would be interesting to collect |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments for additional annotators. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Graded judgments of use similarity and sense |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ applicability. The annotators made use of the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ full spectrum of ratings, as shown in Figures 1 and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4. This may be because of a graded perception of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the similarity of uses as well as senses, or because |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ some uses and senses are very similar. Table 4 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shows that for a large number of WSsim items, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ multiple senses that were not significantly posi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tively correlated got high ratings. This seems to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ indicate that the ratings we obtained cannot sim- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ply be explained by more coarse-grained senses. It |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ may hence be reasonable to pursue computational |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ models of word meaning that are graded, maybe |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ even models that do not rely on dictionary senses |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ at all (Erk and Pado, 2008). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Comparison to previous word sense annotation. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Our graded WSsim annotations do correlate with |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ traditional “best fitting sense” annotations from |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SemCor and SE-3; however, if annotators perceive |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ similarity between uses and senses as graded, tra- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ditional word sense annotation runs the risk of in- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ troducing bias into the annotation. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Comparison to lexical substitutions. There is a |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ strong correlation between both Usim and WSsim |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the overlap in paraphrases that annotators gen- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erated for LEXSUB. This is very encouraging, and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ especially interesting because LEXSUB annotators |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ freely generated paraphrases rather than selecting |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ them from a list. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 Conclusions |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We have introduced a novel annotation paradigm |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ for word sense annotation that allows for graded |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ judgments and for some variation between anno- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tators. We have used this annotation paradigm |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in two experiments, WSsim and Usim, that shed |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ some light on the question of whether differences |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ between word usages are perceived as categorial |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or graded. Both datasets will be made publicly |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ available. There was a high correlation between |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotator judgments within and across tasks, as |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ well as with previous word sense annotation and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with paraphrases proposed in the English Lex- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ical Substitution task. Annotators made ample |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use of graded judgments in a way that cannot |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be explained through more coarse-grained senses. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ These results suggest that it may make sense to |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ evaluate WSD systems on a task of graded rather |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ than categorial meaning characterization, either |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ through dictionary senses or similarity between |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ uses. In that case, it would be useful to have more |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ extensive datasets with graded annotation, even |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ though this annotation paradigm is more time con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ suming and thus more expensive than traditional |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word sense annotation. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As a next step, we will automatically cluster the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ judgments we obtained in the WSsim and Usim |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ experiments to further explore the degree to which |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the annotation gives rise to sense grouping. We |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ will also use the ratings in both experiments to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ evaluate automatically induced models of word |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ meaning. The SemEval-2007 word sense induc- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion task (Agirre and Soroa, 2007) already allows |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for evaluation of automatic sense induction sys- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tems, but compares output to gold-standard senses |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from OntoNotes. We hope that the Usim dataset |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ will be particularly useful for evaluating methods |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which relate usages without necessarily producing |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hard clusters. Also, we will extend the current |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dataset using more annotators and exploring ad- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ditional lexicon resources. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Acknowledgments. We acknowledge support |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ from the UK Royal Society for a Dorothy Hodkin |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Fellowship to the second author. We thank Sebas- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tian Pado for many helpful discussions, and An- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ drew Young for help with the interface. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ References |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ E. Agirre and A. Soroa. 2007. SemEval-2007 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ task 2: Evaluating word sense induction and dis- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 17 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ crimination systems. In Proceedings of the 4th |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ International Workshop on Semantic Evaluations |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ (SemEval-2007), pages 7–12, Prague, Czech Repub- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ lic. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ J. Bybee and D. Eddington. 2006. A usage-based ap- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ proach to Spanish verbs of ’becoming’. Language, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 82(2):323–355. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ J. Chen and M. Palmer. 2009. Improving English |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ verb sense disambiguation performance with lin- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guistically motivated features and clear sense dis- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tinction boundaries. Journal of Language Resources |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ and Evaluation, Special Issue on SemEval-2007. in |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ press. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ K. Erk and S. Pado. 2008. A structured vector space |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ model for word meaning in context. In Proceedings |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of EMNLP-08, Waikiki, Hawaii. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ J. A. Hampton. 1979. Polymorphous concepts in se- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ mantic memory. Journal of Verbal Learning and |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Verbal Behavior, 18:441–461. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ J. A. Hampton. 2007. Typicality, graded membership, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and vagueness. Cognitive Science, 31:355–384. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P. Hanks. 2000. Do word meanings exist? Computers |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and the Humanities, 34(1-2):205–215(11). |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ E. H. Hovy, M. Marcus, M. Palmer, S. Pradhan, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L. Ramshaw, and R. Weischedel. 2006. OntoNotes: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The 90% solution. In Proceedings of the Hu- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ man Language Technology Conference of the North |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ American Chapter of the ACL (NAACL-2006), pages |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 57–60, New York. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ N. Ide and Y. Wilks. 2006. Making sense about |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ sense. In E. Agirre and P. Edmonds, editors, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Word Sense Disambiguation, Algorithms and Appli- |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ cations, pages 47–73. Springer. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ A. Kilgarriff and J. Rosenzweig. 2000. Framework |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and results for English Senseval. Computers and the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Humanities, 34(1-2):15–48. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ A. Kilgarriff. 1997. I don’t believe in word senses. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Computers and the Humanities, 31(2):91–113. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ A. Kilgarriff. 2006. Word senses. In E. Agirre |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and P. Edmonds, editors, Word Sense Disambigua- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion, Algorithms and Applications, pages 29–46. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Springer. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ R. Krishnamurthy and D. Nicholls. 2000. Peeling |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ an onion: the lexicographers’ experience of man- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ual sense-tagging. Computers and the Humanities, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ 34(1-2). |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ S. Landes, C. Leacock, and R. Tengi. 1998. Build- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing semantic concordances. In C. Fellbaum, editor, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WordNet: An Electronic Lexical Database. The MIT |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ Press, Cambridge, MA. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ M. Lapata. 2006. Automatic evaluation of information |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ordering. Computational Linguistics, 32(4):471– |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 484. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ D. McCarthy and R. Navigli. 2007. SemEval-2007 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ task 10: English lexical substitution task. In Pro- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ceedings of the 4th International Workshop on Se- |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ mantic Evaluations (SemEval-2007), pages 48–53, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Prague, Czech Republic. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ M. McCloskey and S. Glucksberg. 1978. Natural cat- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ egories: Well defined or fuzzy sets? Memory & |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Cognition, 6:462–472. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ R. Mihalcea, T. Chklovski, and A. Kilgarriff. 2004. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The Senseval-3 English lexical sample task. In |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3rd International Workshop on Semantic Evalua- |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ tions (SensEval-3) atACL-2004, Barcelona, Spain. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ G. Miller and W. Charles. 1991. Contextual correlates |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ of semantic similarity. Language and cognitive pro- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ cesses, 6(1):1–28. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ G. L. Murphy. 2002. The Big Book of Concepts. MIT |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Press. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ R. Navigli, K. C. Litkowski, and O. Hargraves. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2007. SemEval-2007 task 7: Coarse-grained En- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ glish all-words task. In Proceedings of the 4th |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ International Workshop on Semantic Evaluations |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ (SemEval-2007), pages 30–35, Prague, Czech Re- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ public. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P. Resnik and D. Yarowsky. 2000. Distinguishing |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ systems and distinguishing senses: New evaluation |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ methods for word sense disambiguation. Natural |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Language Engineering, 5(3):113–133. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ E. Rosch and C. B. Mervis. 1975. Family resem- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ blance: Studies in the internal structure of cate- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gories. Cognitive Psychology, 7:573–605. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ E. Rosch. 1975. Cognitive representations of seman- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tic categories. Journal of Experimental Psychology: |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ General, 104:192–233. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ H. Rubenstein and J. Goodenough. 1965. Contextual |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ correlates of synonymy. Computational Linguistics, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 8:627–633. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ S. Sharoff. 2006. Open-source corpora: Using the net |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ to fish for linguistic data. International Journal of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Corpus Linguistics, 11(4):435–462. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ C. Stokoe. 2005. Differentiating homonymy and pol- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ysemy in information retrieval. In Proceedings of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ HLT/EMNLP-05, pages 403–410, Vancouver, B.C., |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Canada. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 18 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ A Comparative Study on Generalization of Semantic Roles in FrameNet |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Yuichiroh Matsubayashit Naoaki Okazakit Jun’ichi Tsujiit$* |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ tDepartment of Computer Science, University of Tokyo, Japan |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ $School of Computer Science, University of Manchester, UK |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ *National Centre for Text Mining, UK |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ {y-matsu,okazaki,tsujii}@is.s.u-tokyo.ac.jp |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ A number of studies have presented |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ machine-learning approaches to semantic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role labeling with availability of corpora |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ such as FrameNet and PropBank. These |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corpora define the semantic roles of predi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cates for each frame independently. Thus, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ it is crucial for the machine-learning ap- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proach to generalize semantic roles across |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different frames, and to increase the size |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of training instances. This paper ex- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ plores several criteria for generalizing se- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic roles in FrameNet: role hierar- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ chy, human-understandable descriptors of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles, semantic types of filler phrases, and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mappings from FrameNet roles to the- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ matic roles of VerbNet. We also pro- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pose feature functions that naturally com- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bine and weight these criteria, based on |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the training data. The experimental result |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the role classification shows 19.16% |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and 7.42% improvements in error reduc- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion rate and macro-averaged F 1 score, re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spectively. We also provide in-depth anal- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ yses of the proposed criteria. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Semantic Role Labeling (SRL) is a task of analyz- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing predicate-argument structures in texts. More |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ specifically, SRL identifies predicates and their |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ arguments with appropriate semantic roles. Re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ solving surface divergence of texts (e.g., voice |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of verbs and nominalizations) into unified seman- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic representations, SRL has attracted much at- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tention from researchers into various NLP appli- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cations including question answering (Narayanan |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Harabagiu, 2004; Shen and Lapata, 2007; |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ buy.v PropBank FrameNet |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Frame buy.01 Commerce buy |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Roles ARG0: buyer Buyer Goods Seller Money Recipient |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG1: thing bought ... |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG2: seller |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG3: paid |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG4: benefactive |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ... |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 1: A comparison of frames for buy.v de- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ fined in PropBank and FrameNet |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Moschitti et al., 2007), and information extrac- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tion (Surdeanu et al., 2003). |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In recent years, with the wide availability of cor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ pora such as PropBank (Palmer et al., 2005) and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ FrameNet (Baker et al., 1998), a number of stud- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ies have presented statistical approaches to SRL |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (M`arquez et al., 2008). Figure 1 shows an exam- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ple of the frame definitions for a verb buy in Prop- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Bank and FrameNet. These corpora define a large |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ number of frames and define the semantic roles for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each frame independently. This fact is problem- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ atic in terms of the performance of the machine- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning approach, because these definitions pro- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ duce many roles that have few training instances. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PropBank defines a frame for each sense of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ predicates (e.g., buy.01), and semantic roles are |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ defined in a frame-specific manner (e.g., buyer and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ seller for buy.01). In addition, these roles are asso- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ciated with tags such as ARG0-5 and AM-*, which |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are commonly used in different frames. Most |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SRL studies on PropBank have used these tags |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in order to gather a sufficient amount of training |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data, and to generalize semantic-role classifiers |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ across different frames. However, Yi et al. (2007) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reported that tags ARG2 –ARG5 were inconsis- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tent and not that suitable as training instances. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Some recent studies have addressed alternative ap- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proaches to generalizing semantic roles across dif- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferent frames (Gordon and Swanson, 2007; Zapi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 19 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 19–27, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Commerce_sell::Buyer Commerce_buy::Buyer |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Givi ng:: Reci pi ent |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Recipient |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Transfer::Recipient |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Buyer |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Agent |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Commerce_sell::Seller Commerce_buy::Seller |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Giving::Donor |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Transfer::Donor |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Donor |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Seller |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ role-to-role relation |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ hierarchical class |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ thematic role |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role descriptor |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 2: An example of role groupings using different criteria. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ rain et al., 2008). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ FrameNet designs semantic roles as frame spe- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ cific, but also defines hierarchical relations of se- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic roles among frames. Figure 2 illustrates |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ an excerpt of the role hierarchy in FrameNet; this |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ figure indicates that the Buyer role for the Com- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ merce buy frame (Commerce buy::Buyer here- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ after) and the Commerce sell::Buyer role are in- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ herited from the Transfer:: Recipient role. Al- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ though the role hierarchy was expected to gener- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ alize semantic roles, no positive results for role |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ classification have been reported (Baldewein et al., |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2004). Therefore, the generalization of semantic |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles across different frames has been brought up |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as a critical issue for FrameNet (Gildea and Juraf- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sky, 2002; Shi and Mihalcea, 2005; Giuglea and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Moschitti, 2006) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this paper, we explore several criteria for gen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ eralizing semantic roles in FrameNet. In addi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion to the FrameNet hierarchy, we use various |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pieces of information: human-understandable de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scriptors of roles, semantic types of filler phrases, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and mappings from FrameNet roles to the thematic |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles of VerbNet. We also propose feature func- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions that naturally combines these criteria in a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ machine-learning framework. Using the proposed |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method, the experimental result of the role classi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fication shows 19.16% and 7.42% improvements |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in error reduction rate and macro-averaged F1, re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spectively. We provide in-depth analyses with re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spect to these criteria, and state our conclusions. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 Related Work |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Moschitti et al. (2005) first classified roles by us- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing four coarse-grained classes (Core Roles, Ad- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ juncts, Continuation Arguments and Co-referring |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Arguments), and built a classifier for each coarse- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grained class to tag PropBank ARG tags. Even |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ though the initial classifiers could perform rough |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ estimations of semantic roles, this step was not |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ able to solve the ambiguity problem in PropBank |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG2-5. When training a classifier for a seman- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic role, Baldewein et al. (2004) re-used the train- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing instances of other roles that were similar to the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target role. As similarity measures, they used the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ FrameNet hierarchy, peripheral roles of FrameNet, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and clusters constructed by a EM-based method. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Gordon and Swanson (2007) proposed a general- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ization method for the PropBank roles based on |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactic similarity in frames. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Many previous studies assumed that thematic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ roles bridged semantic roles in different frames. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Gildea and Jurafsky (2002) showed that classifica- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion accuracy was improved by manually replac- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing FrameNet roles into 18 thematic roles. Shi |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Mihalcea (2005) and Giuglea and Moschitti |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2006) employed VerbNet thematic roles as the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target of mappings from the roles defined by the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different semantic corpora. Using the thematic |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles as alternatives of ARG tags, Loper et al. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2007) and Yi et al. (2007) demonstrated that the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ classification accuracy of PropBank roles was im- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proved for ARG2 roles, but that it was diminished |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for ARG1. Yi et al. (2007) also described that |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG2–5 were mapped to a variety of thematic |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles. Zapirain et al. (2008) evaluated PropBank |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG tags and VerbNet thematic roles in a state-of- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the-art SRL system, and concluded that PropBank |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARG tags achieved a more robust generalization of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the roles than did VerbNet thematic roles. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Role Classification |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ SRL is a complex task wherein several problems |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ are intertwined: frame-evoking word identifica- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ tion, frame disambiguation (selecting a correct |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ frame from candidates for the evoking word), role- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ phrase identification (identifying phrases that fill |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles), and role classification (assigning |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correct roles to the phrases). In this paper, we fo- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cus on role classification, in which the role gen- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ eralization is particularly critical to the machine |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ learning approach. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In the role classification task, we are given a |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ sentence, a frame evoking word, a frame, and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 20 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Hierarchical-relation groups |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Role-descriptor groups |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Thematic-role groups |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Semantic-type groups |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 4: Examples for each type of role group. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ INPUT: |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ frame = Commerce_sell |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ candidate roles = { Seller, Buyer, Goods, Reason, Time, ... , Place} |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ sentence = Can't [you] [sell Commerce_sell] [the factory] [to some other |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ company]? |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ OUTPUT: |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ sentence = Can't [you Seller] [sell Commerce_sell] [the factory Goods] |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ [to some other company Buyer] ? |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 3: An example of input and output of role |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ classification. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ phrases that take semantic roles. We are inter- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ested in choosing the correct role from the can- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ didate roles for each phrase in the frame. Figure 3 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shows a concrete example of input and output; the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles for the phrases are chosen from the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ candidate roles: Seller, Buyer, Goods, Reason, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ... , and Place. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role at a node in the hierarchy inherits the char- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ acteristics of the roles of its ancestor nodes. For |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ example, Commerce sell::Seller in Figure 2 in- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ herits the property of Giving:: Donor. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For Inheritance, Using, Perspective on, and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Subframe relations, we assume that descendant |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles in these relations have the same or special- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ized properties of their ancestors. Hence, for each |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role yi, we define the following two role groups, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hchild = {y y = yi V y is a child of yi}, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ yi = {yly = yi V y is a descendant of yij. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hdesc |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ yi |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The hierarchical-relation groups in Figure 4 are |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the illustrations of Hdesc |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ yi . |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ For the relation types Inchoative of and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Causative of, we define role groups in the oppo- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ site direction of the hierarchy, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 Design of Role Groups Hparent = {yjy = yi V y is a parent of yi}, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ yi = {yjy = yi V y is anancestor of yij. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hance |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ yi |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We formalize the generalization of semantic roles |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ as the act of grouping several roles into a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ class. We define a role group as a set of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role labels grouped by a criterion. Figure 4 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shows examples of role groups; a group Giv- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing::Donor (in the hierarchical-relation groups) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ contains the roles Giving::Donor and Com- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ merce pay::Buyer. The remainder of this section |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ describes the grouping criteria in detail. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.1 Hierarchical relations among roles |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ FrameNet defines hierarchical relations among |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ frames (frame-to-frame relations). Each relation |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is assigned one of the seven types of directional |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relationships (Inheritance, Using, Perspective on, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Causative of, Inchoative of, Subframe, and Pre- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ cedes). Some roles in two related frames are also |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ connected with role-to-role relations. We assume |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that this hierarchy is a promising resource for gen- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ eralizing the semantic roles; the idea is that the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ This is because lower roles of Inchoative of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and Causative of relations represent more neu- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tral stances or consequential states; for example, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Killing::Victim is a parent of Death:: Protagonist |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the Causative of relation. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Finally, the Precedes relation describes the se- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ quence of states and events, but does not spec- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ify the direction of semantic inclusion relations. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Therefore, we simply try Hchild Hdesc Hparent |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ yi , yi,yi , |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and Hynce for this relation type. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 4.2 Human-understandable role descriptor |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ FrameNet defines each role as frame-specific; in |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ other words, the same identifier does not appear |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in different frames. However, in FrameNet, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ human experts assign a human-understandable |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ name to each role in a rather systematic man- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ner. Some names are shared by the roles in |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different frames, whose identifiers are dif- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferent. Therefore, we examine the semantic |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 21 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ commonality of these names; we construct an |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ equivalence class of the roles sharing the same |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ name. We call these human-understandable |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ names role descriptors. In Figure 4, the role- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ descriptor group Buyer collects the roles Com- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ merce pay::Buyer, Commerce buy::Buyer, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ and Commerce sell::Buyer. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ This criterion may be effective in collecting |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ similar roles since the descriptors have been anno- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tated by intuition of human experts. As illustrated |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in Figure 2, the role descriptors group the seman- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic roles which are similar to the roles that the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ FrameNet hierarchy connects as sister or parent- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ child relations. However, role-descriptor groups |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cannot express the relations between the roles |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as inclusions since they are equivalence classes. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For example, the roles Commerce sell::Buyer |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Commerce buy::Buyer are included in the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role descriptor group Buyer in Figure 2; how- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, it is difficult to merge Giving:: Recipient |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Commerce sell::Buyer because the Com- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ merce sell:: Buyer has the extra property that one |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ gives something of value in exchange and a hu- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ man assigns different descriptors to them. We ex- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pect that the most effective weighting of these two |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ criteria will be determined from the training data. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.3 Semantic type of phrases |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We consider that the selectional restriction is help- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ful in detecting the semantic roles. FrameNet pro- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vides information concerning the semantic types |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of role phrases (fillers); phrases that play spe- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cific roles in a sentence should fulfill the se- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic constraint from this information. For |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ instance, FrameNet specifies the constraint that |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Self motion::Area should be filled by phrases |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ whose semantic type is Location. Since these |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ types suggest a coarse-grained categorization of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles, we construct role groups that con- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tain roles whose semantic types are identical. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.4 Thematic roles of VerbNet |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ VerbNet thematic roles are 23 frame-independent |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ semantic categories for arguments of verbs, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ such as Agent, Patient, Theme and Source. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ These categories have been used as consis- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tent labels across verbs. We use a partial |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mapping between FrameNet roles and Verb- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Net thematic roles provided by SemLink. 1 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Each group is constructed as a set Tti = |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1http://verbs.colorado.edu/semlink/ |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ {yISemLink maps y into the thematic role ti}. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ SemLink currently maps 1,726 FrameNet roles |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ into VerbNet thematic roles, which are 37.61% of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles appearing at least once in the FrameNet cor- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus. This may diminish the effect of thematic-role |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ groups than its potential. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 Role classification method |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 5.1 Traditional approach |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ We are given a frame-evoking word e, a frame f |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and a role phrase x detected by a human or some |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ automatic process in a sentence s. Let Yf be the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ set of semantic roles that FrameNet defines as be- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing possible role assignments for the frame f, and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ let x = {x1, ... , x,,,} be observed features for x |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from s, e and f . The task of semantic role classifi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cation can be formalized as the problem of choos- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the most suitable role y� from Yf. Suppose we |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ have a model P(yIf, x) which yields the condi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tional probability of the semantic role y for given |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ f and x. Then we can choose y� as follows: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ y� = argmax P(y� f, x). (1) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ yEYf |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ A traditional way to incorporate role groups |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ into this formalization is to overwrite each role |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ y in the training and test data with its role |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ group m(y) according to the memberships of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the group. For example, semantic roles Com- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ merce sell::Seller and Giving::Donor can be re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ placed by their thematic-role group Theme::Agent |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in this approach. We determine the most suitable |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role group c� as follows: |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ c� = argmax P-(c�f, x). (2) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ cE �-(y)�yEYf} |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Here, P-(c�f, x) presents the probability of the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ role group c for f and x. The role y� is determined |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ uniquely iff a single role y E Yf is associated |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with �c. Some previous studies have employed this |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ idea to remedy the data sparseness problem in the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training data (Gildea and Jurafsky, 2002). How- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, we cannot apply this approach when multi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ple roles in Yf are contained in the same class. For |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ example, we can construct a semantic-type group |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ St::State of affairs in which Giving::Reason and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Giving:: Means are included, as illustrated in Fig- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ure 4. If c� = St:: State of affairs, we cannot dis- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ambiguate which original role is correct. In ad- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dition, it may be more effective to use various |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 22 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ groupings of roles together in the model. For in- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ stance, the model could predict the correct role |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Commerce sell::Seller for the phrase “you” in |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 3 more confidently, if it could infer its |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ thematic-role group as Theme::Agent and its par- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent group Giving::Donor correctly. Although the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ensemble of various groupings seems promising, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we need an additional procedure to prioritize the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ groupings for the case where the models for mul- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tiple role groupings disagree; for example, it is un- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ satisfactory if two models assign the groups Giv- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing:: Theme and Theme::Agent to the same phrase. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5.2 Role groups as feature functions |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We thus propose another approach that incorpo- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ rates group information as feature functions. We |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ model the conditional probability P(y� f, x) by us- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the maximum entropy framework, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ exp(Ei Aigi(x, y)) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ �( y� f, x) = Ey,Y, exp(Ei Aigi(x, y)). (3) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Here, G = Jgi� denotes a set of n feature func- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tions, and A = JAi� denotes a weight vector for |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the feature functions. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In general, feature functions for the maximum |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ entropy model are designed as indicator functions |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for possible pairs of xj and y. For example, the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ event where the head word of x is “you” (xi = 1) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and x plays the role Commerce sell::Seller in a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence is expressed by the indicator function, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 (xi =1A |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ y = Commerce sell::Seller) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0 (otherwise) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (4) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We call this kind of feature function an x-role. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In order to incorporate role groups into the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ model, we also include all feature functions for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ possible pairs of xj and role groups. Equation 5 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is an example of a feature function for instances |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ where the head word of x is “you” and y is in the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role group Theme::Agent, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 (xi=1A |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ y E Theme::Agent) . (5) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0 (otherwise) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Thus, this feature function fires for the roles wher- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ever the head word “you” plays Agent (e.g., Com- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ merce sell::Seller, Commerce buy::Buyer and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Giving:: Donor). We call this kind of feature func- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ tion an x-group function. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this way, we obtain x-group functions for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ all grouping methods, e.g., gtheme |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ k, ghierarchy |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ k . |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The role-group features will receive more training |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ instances by collecting instances for fine-grained |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles. Thus, semantic roles with few training in- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stances are expected to receive additional clues |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from other training instances via role-group fea- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures. Another advantage of this approach is that |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the usefulness of the different role groups is de- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ termined by the training processes in terms of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ weights of feature functions. Thus, we do not need |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to assume that we have found the best criterion for |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grouping roles; we can allow a training process to |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ choose the criterion. We will discuss the contribu- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions of different groupings in the experiments. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5.3 Comparison with related work |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Baldewein et al. (2004) suggested an approach |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ that uses role descriptors and hierarchical rela- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions as criteria for generalizing semantic roles |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in FrameNet. They created a classifier for each |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ frame, additionally using training instances for the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role A to train the classifier for the role B, if the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles A and B were judged as similar by a crite- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rion. This approach performs similarly to the over- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ writing approach, and it may obscure the differ- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ences among roles. Therefore, they only re-used |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the descriptors as a similarity measure for the roles |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ whose coreness was peripheral. 2 |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In contrast, we use all kinds of role descriptors |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ to construct groups. Since we use the feature func- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions for both the original roles and their groups, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ appropriate units for classification are determined |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ automatically in the training process. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 Experiment and Discussion |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We used the training set of the Semeval-2007 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Shared task (Baker et al., 2007) in order to ascer- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tain the contributions of role groups. This dataset |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ consists of the corpus of FrameNet release 1.3 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (containing roughly 150,000 annotations), and an |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ additional full-text annotation dataset. We ran- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ domly extracted 10% of the dataset for testing, and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used the remainder (90%) for training. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Performance was measured by micro- and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ macro-averaged F 1(Chang and Zheng, 2008) with |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ respect to a variety of roles. The micro average bi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ases each F1 score by the frequencies of the roles, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2In FrameNet, each role is assigned one of four different |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ types of coreness (core, core-unexpressed, peripheral, extra- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ thematic) It represents the conceptual necessity of the roles |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the frame to which it belongs. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ giole(x, y) = { |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ . |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ gtheme |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2(x, y) = |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ { |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 23 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and the average is equal to the classification accu- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ racy when we calculate it with all of the roles in |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the test set. In contrast, the macro average does |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ not bias the scores, thus the roles having a small |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ number of instances affect the average more than |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the micro average. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6.1 Experimental settings |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We constructed a baseline classifier that uses |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ only the x-role features. The feature de- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sign is similar to that of the previous stud- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ies (M`arquez et al., 2008). The characteristics |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of x are: frame, frame evoking word, head |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word, content word (Surdeanu et al., 2003), |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first/last word, head word of left/right sister, |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ phrase type, position, voice, syntactic path (di- |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rected/undirected/partial), governing category |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Gildea and Jurafsky, 2002), WordNet super- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ sense in the phrase, combination features of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ frame evoking word & headword, combination |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ features of frame evoking word & phrase type, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ and combination features of voice & phrase type. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We also used PoS tags and stem forms as extra |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features of any word-features. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We employed Charniak and Johnson’s rerank- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing parser (Charniak and Johnson, 2005) to an- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ alyze syntactic trees. As an alternative for the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ traditional named-entity features, we used Word- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Net supersenses: 41 coarse-grained semantic cate- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gories of words such as person, plant, state, event, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ time, location. We used Ciaramita and Altun’s Su- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ per Sense Tagger (Ciaramita and Altun, 2006) to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tag the supersenses. The baseline system achieved |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 89.00% with respect to the micro-averaged F1. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The x-group features were instantiated similarly |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ to the x-role features; the x-group features com- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bined the characteristics of x with the role groups |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ presented in this paper. The total number of fea- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures generated for all x-roles and x-groups was |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 74,873,602. The optimal weights A of the fea- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures were obtained by the maximum a poste- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rior (MAP) estimation. We maximized an L2- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ regularized log-likelihood of the training set us- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the Limited-memory BFGS (L-BFGS) method |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Nocedal, 1980). |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6.2 Effect of role groups |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Table 1 shows the micro and macro averages of F 1 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ scores. Each role group type improved the micro |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ average by 0.5 to 1.7 points. The best result was |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ obtained by using all types of groups together. The |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ result indicates that different kinds of group com- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Feature Micro Macro -Err. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Baseline 89.00 68.50 0.00 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role descriptor 90.78 76.58 16.17 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role descriptor (replace) 90.23 76.19 11.23 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hierarchical relation 90.25 72.41 11.40 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic type 90.36 74.51 12.38 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ VN thematic role 89.50 69.21 4.52 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ All 91.10 75.92 19.16 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 1: The accuracy and error reduction rate of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ role classification for each type of role group. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Feature #instances Pre. Rec. Micro |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ baseline < 10 63.89 38.00 47.66 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ < 20 69.01 51.26 58.83 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ <50 75.84 65.85 70.50 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ + all groups < 10 72.57 55.85 63.12 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ <20 76.30 65.41 70.43 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ < 50 80.86 74.59 77.60 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2: The effect of role groups on the roles with |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ few instances. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ plement each other with respect to semantic role |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ generalization. Baldewein et al. (2004) reported |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that hierarchical relations did not perform well for |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ their method and experimental setting; however, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we found that significant improvements could also |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be achieved with hierarchical relations. We also |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tried a traditional label-replacing approach with |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role descriptors (in the third row of Table 1). The |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comparison between the second and third rows in- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dicates that mixing the original fine-grained roles |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the role groups does result in a more accurate |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ classification. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ By using all types of groups together, the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ model reduced 19.16 % of the classification errors |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from the baseline. Moreover, the macro-averaged |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ F1 scores clearly showed improvements resulting |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from using role groups. In order to determine |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the reason for the improvements, we measured |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the precision, recall, and F1-scores with respect |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to roles for which the number of training instances |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ was at most 10, 20, and 50. In Table 2, we show |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that the micro-averaged F1 score for roles hav- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing 10 instances or less was improved (by 15.46 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ points) when all role groups were used. This result |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ suggests the reason for the effect of role groups; by |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bridging similar semantic roles, they supply roles |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ having a small number of instances with the infor- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mation from other roles. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6.3 Analyses of role descriptors |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In Table 1, the largest improvement was obtained |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ by the use of role descriptors. We analyze the ef- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fect of role descriptors in detail in Tables 3 and 4. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3 shows the micro-averaged F 1 scores of all |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 24 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Coreness #roles #instances/#role #groups #instances/#group #roles/#group |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Core 1902 122.06 655 354.4 2.9 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Peripheral 1924 25.24 250 194.3 7.7 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Extra-thematic 763 13.90 171 62.02 4.5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 4: The analysis of the numbers of roles, instances, and role-descriptor groups, for each type of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ coreness. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Coreness Micro |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Baseline 89.00 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Core 89.51 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Peripheral 90.12 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Extra-thematic 89.09 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ All 90.77 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3: The effect of employing role-descriptor |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ groups of each type of coreness. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles when we use role-descriptor groups |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ constructed from each type of coreness (core3, pe- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ripheral, and extra-thematic) individually. The pe- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ripheral type generated the largest improvements. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 4 shows the number of roles associated |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ with each type of coreness (#roles), the number of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ instances for the original roles (#instances/#role), |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the number of groups for each type of coreness |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (#groups), the number of instances for each group |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (#instances/#group), and the number of roles per |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each group (#roles/#group). In the peripheral |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type, the role descriptors subdivided 1,924 distinct |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles into 250 groups, each of which contained 7.7 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles on average. The peripheral type included |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles such as place, time, reason, dura- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion. These semantic roles appear in many frames, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ because they have general meanings that can be |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shared by different frames. Moreover, the seman- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic roles of peripheral type originally occurred in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only a small number (25.24) of training instances |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on average. Thus, we infer that the peripheral |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type generated the largest improvement because |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles in this type acquired the greatest |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ benefit from the generalization. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6.4 Hierarchical relations and relation types |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We analyzed the contributions of the FrameNet hi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ erarchy for each type of role-to-role relations and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for different depths of grouping. Table 5 shows |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the micro-averaged F1 scores obtained from var- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ious relation types and depths. The Inheritance |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Using relations resulted in a slightly better ac- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ curacy than the other types. We did not observe |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ any real differences among the remaining five re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lation types, possibly because there were few se- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 We include Core-unexpressed in core, because it has a |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ property of core inside one frame. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ No. Relation Type Micro |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ - baseline 89.00 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 + Inheritance (children) 89.52 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 + Inheritance (descendants) 89.70 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 +Using (children) 89.35 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 +Using (descendants) 89.37 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 + Perspective on (children) 89.01 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 + Perspective on (descendants) 89.01 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7 + Subframe (children) 89.04 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 8 + Subframe (descendants) 89.05 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 9 + Causative of (parents) 89.03 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 + Causative of (ancestors) 89.03 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 11 + Inchoative of (parents) 89.02 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 12 + Inchoative of (ancestors) 89.02 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 13 +Precedes (children) 89.01 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 14 +Precedes (descendants) 89.03 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 15 +Precedes (parents) 89.00 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 16 +Precedes (ancestors) 89.00 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 18 +all relations (2,4,6,8,10,12,14) 90.25 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 5: Comparison of the accuracy with differ- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ent types of hierarchical relations. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic roles associated with these types. We ob- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tained better results by using not only groups for |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parent roles, but also groups for all ancestors. The |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ best result was obtained by using all relations in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the hierarchy. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6.5 Analyses of different grouping criteria |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Table 6 reports the precision, recall, and micro- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ averaged F1 scores of semantic roles with respect |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to each coreness type .4 In general, semantic roles |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the core coreness were easily identified by all |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the grouping criteria; even the baseline system |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ obtained an F 1 score of 91.93. For identifying se- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic roles of the peripheral and extra-thematic |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ types of coreness, the simplest solution, the de- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scriptor criterion, outperformed other criteria. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In Table 7, we categorize feature functions |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ whose weights are in the top 1000 in terms of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ greatest absolute value. The behaviors of the role |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ groups can be distinguished by the following two |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ characteristics. Groups of role descriptors and se- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic types have large weight values for the first |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word and supersense features, which capture the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ characteristics of adjunctive phrases. The original |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles and hierarchical-relation groups have strong |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4The figures of role descriptors in Tables 4 and 6 differ. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In Table 4, we measured the performance when we used one |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or all types of coreness for training. In contrast, in Table 6, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we used all types of coreness for training, but computed the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ performance of semantic roles for each coreness separately. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 25 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Feature Type Pre. Rec. Micro |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ baseline c 91.07 92.83 91.93 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p 81.05 76.03 78.46 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ e 78.17 66.51 71.87 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ + descriptor group c 92.50 93.41 92.95 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p 84.32 82.72 83.51 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ e 80.91 69.59 74.82 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ +hierarchical c 92.10 93.28 92.68 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relation p 82.23 79.84 81.01 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ class e 77.94 65.58 71.23 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ + semantic c 92.23 93.31 92.77 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type group p 83.66 81.76 82.70 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ e 80.29 67.26 73.20 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ + VN thematic c 91.57 93.06 92.31 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role group p 80.66 76.95 78.76 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ e 78.12 66.60 71.90 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ + all group c 92.66 93.61 93.13 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p 84.13 82.51 83.31 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ e 80.77 68.56 74.17 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 6: The precision and recall of each type of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ coreness with role groups. Type represents the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type of coreness; c denotes core, p denotes periph- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ eral, and e denotes extra-thematic. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ associations with lexical and structural character- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ istics such as the syntactic path, content word, and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ head word. Table 7 suggests that role-descriptor |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ groups and semantic-type groups are effective for |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ peripheral or adjunctive roles, and hierarchical re- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lation groups are effective for core roles. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7 Conclusion |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We have described different criteria for general- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ izing semantic roles in FrameNet. They were: |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role hierarchy, human-understandable descriptors |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of roles, semantic types of filler phrases, and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mappings from FrameNet roles to thematic roles |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of VerbNet. We also proposed a feature design |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that combines and weights these criteria using the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training data. The experimental result of the role |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ classification task showed a 19.16% of the error |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reduction and a 7.42% improvement in the macro- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ averaged F1 score. In particular, the method we |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ have presented was able to classify roles having |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ few instances. We confirmed that modeling the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role generalization at feature level was better than |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the conventional approach that replaces semantic |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role labels. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Each criterion presented in this paper improved |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the accuracy of classification. The most success- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ful criterion was the use of human-understandable |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role descriptors. Unfortunately, the FrameNet hi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erarchy did not outperform the role descriptors, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ contrary to our expectations. A future direction |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of this study would be to analyze the weakness of |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the FrameNet hierarchy in order to discuss possi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ble improvement of the usage and annotations of |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features of x class type |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ or hr rl st vn |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ frame 0 4 0 1 0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ evoking word 3 4 7 3 0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ew & hw stem 9 34 20 8 0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ew & phrase type 11 7 11 3 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ head word 13 19 8 3 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hw stem 11 17 8 8 1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ content word 7 19 12 3 0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cw stem 11 26 13 5 0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cw PoS 4 5 14 15 2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ directed path 19 27 24 6 7 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ undirected path 21 35 17 2 6 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ partial path 15 18 16 13 5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ last word 15 18 12 3 2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first word 11 23 53 26 10 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supersense 7 7 35 25 4 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ position 4 6 30 9 5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ others 27 29 33 19 6 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ total 188 298 313 152 50 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 7: The analysis of the top 1000 feature func- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tions. Each number denotes the number of feature |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ functions categorized in the corresponding cell. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Notations for the columns are as follows. ‘or’: |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ original role, ‘hr’: hierarchical relation, ‘rd’: role |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ descriptor, ‘st’: semantic type, and ‘vn’: VerbNet |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ thematic role. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the hierarchy. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Since we used the latest release of FrameNet |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ in order to use a greater number of hierarchical |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role-to-role relations, we could not make a direct |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comparison of performance with that of existing |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ systems; however we may say that the 89.00% F 1 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ micro-average of our baseline system is roughly |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comparable to the 88.93% value of Bejan and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hathaway (2007) for SemEval-2007 (Baker et al., |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2007). 5 In addition, the methodology presented in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this paper applies generally to any SRL resources; |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we are planning to determine several grouping cri- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ teria from existing linguistic resources and to ap- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ply the methodology to the PropBank corpus. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Acknowledgments |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The authors thank Sebastian Riedel for his useful |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ comments on our work. This work was partially |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supported by Grant-in-Aid for Specially Promoted |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Research (MEXT, Japan). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ References |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Collin F. Baker, Charles J. Fillmore, and John B. Lowe. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 1998. The berkeley framenet project. In Proceed- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ings of Coling -ACL 1998, pages 86–90. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Collin Baker, Michael Ellsworth, and Katrin Erk. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2007. Semeval-2007 task 19: Frame semantic struc- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5There were two participants that performed whole SRL |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ in SemEval-2007. Bejan and Hathaway (2007) evaluated role |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ classification accuracy separately for the training data. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 26 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ture extraction. In Proceedings of SemEval-2007, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ pages 99–104. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ulrike Baldewein, Katrin Erk, Sebastian Pad´o, and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Detlef Prescher. 2004. Semantic role labeling |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with similarity based generalization using EM-based |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clustering. In Proceedings of Senseval-3, pages 64– |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 68. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Cosmin Adrian Bejan and Chris Hathaway. 2007. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ UTD-SRL: A Pipeline Architecture for Extract- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing Frame Semantic Structures. In Proceedings |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of SemEval-2007, pages 460–463. Association for |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Computational Linguistics. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ X. Chang and Q. Zheng. 2008. Knowledge Ele- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ment Extraction for Knowledge-Based Learning Re- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sources Organization. Lecture Notes in Computer |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Science, 4823:102–113. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Eugene Charniak and Mark Johnson. 2005. Coarse- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ to-fine n-best parsing and MaxEnt discriminative |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reranking. In Proceedings of the 43rd Annual Meet- |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ing on Association for Computational Linguistics, |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 173–180. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Massimiliano Ciaramita and Yasemin Altun. 2006. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Broad-coverage sense disambiguation and informa- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion extraction with a supersense sequence tagger. In |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings ofEMNLP-2006, pages 594–602. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Daniel Gildea and Daniel Jurafsky. 2002. Automatic |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ labeling of semantic roles. Computational Linguis- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tics, 28(3):245–288. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ana-Maria Giuglea and Alessandro Moschitti. 2006. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Semantic role labeling via FrameNet, VerbNet and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PropBank. In Proceedings of the 21st International |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Conference on Computational Linguistics and the |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 44th Annual Meeting of the ACL, pages 929–936. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Andrew Gordon and Reid Swanson. 2007. General- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ izing semantic role annotations across syntactically |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ similar verbs. In Proceedings of ACL-2007, pages |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 192–199. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Edward Loper, Szu-ting Yi, and Martha Palmer. 2007. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Combining lexical resources: Mapping between |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ propbank and verbnet. In Proceedings of the 7th In- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ternational Workshop on Computational Semantics, |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ pages 118–128. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Lluis M`arquez, Xavier Carreras, Kenneth C. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Litkowski, and Suzanne Stevenson. 2008. Se- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic role labeling: an introduction to the special |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ issue. Computational linguistics, 34(2):145–159. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Alessandro Moschitti, Ana-Maria Giuglea, Bonaven- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tura Coppola, and Roberto Basili. 2005. Hierar- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ chical semantic role labeling. In Proceedings of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CoNLL-2005, pages 201–204. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Alessandro Moschitti, Silvia Quarteroni, Roberto |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Basili, and Suresh Manandhar. 2007. Exploiting |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactic and shallow semantic kernels for question |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ answer classification. In Proceedings of ACL-07, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 776–783. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Srini Narayanan and Sanda Harabagiu. 2004. Ques- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tion answering based on semantic structures. In Pro- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ceedings of Coling-2004, pages 693–701. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Jorge Nocedal. 1980. Updating quasi-newton matrices |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ with limited storage. Mathematics of Computation, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 35(151):773–782. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Martha Palmer, Daniel Gildea, and Paul Kingsbury. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2005. The proposition bank: An annotated cor- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus of semantic roles. Computational Linguistics, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 31(1):71–106. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Dan Shen and Mirella Lapata. 2007. Using semantic |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ roles to improve question answering. In Proceed- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ings ofEMNLP-CoNLL 2007, pages 12–21. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Lei Shi and Rada Mihalcea. 2005. Putting Pieces To- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ gether: Combining FrameNet, VerbNet and Word- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Net for Robust Semantic Parsing. In Proceedings of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CICLing-2005, pages 100–111. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Mihai Surdeanu, Sanda Harabagiu, John Williams, and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Paul Aarseth. 2003. Using predicate-argument |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structures for information extraction. In Proceed- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ings ofACL-2003, pages 8–15. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Szu-ting Yi, Edward Loper, and Martha Palmer. 2007. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Can semantic roles generalize across genres? In |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings ofHLT-NAACL 2007, pages 548–555. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Be˜nat Zapirain, Eneko Agirre, and Lluis M`arquez. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2008. Robustness and generalization of role sets: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PropBank vs. VerbNet. In Proceedings of ACL-08: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ HLT, pages 550–558. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 27 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Unsupervised Argument Identification for Semantic Role Labeling |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Omri Abend' Roi Reichart2 Ari Rappoport' |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ 'Institute of Computer Science, 2ICNC |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ Hebrew University of Jerusalem |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ {omria01|roiri|arir}@cs.huji.ac.il |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The task of Semantic Role Labeling |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (SRL) is often divided into two sub-tasks: |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verb argument identification, and argu- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment classification. Current SRL algo- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rithms show lower results on the identifi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cation sub-task. Moreover, most SRL al- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gorithms are supervised, relying on large |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ amounts of manually created data. In |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this paper we present an unsupervised al- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gorithm for identifying verb arguments, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ where the only type of annotation required |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is POS tagging. The algorithm makes use |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of a fully unsupervised syntactic parser, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using its output in order to detect clauses |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and gather candidate argument colloca- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion statistics. We evaluate our algorithm |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on PropBank10, achieving a precision of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 56%, as opposed to 47% of a strong base- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ line. We also obtain an 8% increase in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ precision for a Spanish corpus. This is |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the first paper that tackles unsupervised |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verb argument identification without using |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ manually encoded rules or extensive lexi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cal or syntactic resources. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Semantic Role Labeling (SRL) is a major NLP |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ task, providing a shallow sentence-level semantic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ analysis. SRL aims at identifying the relations be- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tween the predicates (usually, verbs) in the sen- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tence and their associated arguments. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The SRL task is often viewed as consisting of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ two parts: argument identification (ARGID) and ar- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gument classification. The former aims at identi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fying the arguments of a given predicate present |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the sentence, while the latter determines the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type of relation that holds between the identi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ fied arguments and their corresponding predicates. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The division into two sub-tasks is justified by |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the fact that they are best addressed using differ- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent feature sets (Pradhan et al., 2005). Perfor- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mance in the ARGID stage is a serious bottleneck |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for general SRL performance, since only about |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 81% of the arguments are identified, while about |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 95% of the identified arguments are labeled cor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rectly (M`arquez et al., 2008). |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SRL is a complex task, which is reflected by the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ algorithms used to address it. A standard SRL al- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gorithm requires thousands to dozens of thousands |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentences annotated with POS tags, syntactic an- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation and SRL annotation. Current algorithms |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ show impressive results but only for languages and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ domains where plenty of annotated data is avail- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ able, e.g., English newspaper texts (see Section 2). |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Results are markedly lower when testing is on a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ domain wider than the training one, even in En- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ glish (see the WSJ-Brown results in (Pradhan et |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 2008)). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Only a small number of works that do not re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ quire manually labeled SRL training data have |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ been done (Swier and Stevenson, 2004; Swier and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Stevenson, 2005; Grenager and Manning, 2006). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ These papers have replaced this data with the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ VerbNet (Kipper et al., 2000) lexical resource or |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a set of manually written rules and supervised |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsers. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ A potential answer to the SRL training data bot- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tleneck are unsupervised SRL models that require |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ little to no manual effort for their training. Their |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ output can be used either by itself, or as training |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ material for modern supervised SRL algorithms. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this paper we present an algorithm for unsu- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ pervised argument identification. The only type of |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotation required by our algorithm is POS tag- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 28 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 28–36, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ging, which needs relatively little manual effort. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The algorithm consists of two stages. As pre- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ processing, we use a fully unsupervised parser to |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parse each sentence. Initially, the set of possi- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ble arguments for a given verb consists of all the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ constituents in the parse tree that do not contain |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that predicate. The first stage of the algorithm |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ attempts to detect the minimal clause in the sen- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tence that contains the predicate in question. Us- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing this information, it further reduces the possible |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ arguments only to those contained in the minimal |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clause, and further prunes them according to their |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ position in the parse tree. In the second stage we |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use pointwise mutual information to estimate the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ collocation strength between the arguments and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the predicate, and use it to filter out instances of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ weakly collocating predicate argument pairs. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We use two measures to evaluate the perfor- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ mance of our algorithm, precision and F-score. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Precision reflects the algorithm’s applicability for |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ creating training data to be used by supervised |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SRL models, while the standard SRL F-score mea- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sures the model’s performance when used by it- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ self. The first stage of our algorithm is shown to |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ outperform a strong baseline both in terms of F- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ score and of precision. The second stage is shown |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to increase precision while maintaining a reason- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ able recall. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We evaluated our model on sections 2-21 of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Propbank. As is customary in unsupervised pars- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing work (e.g. (Seginer, 2007)), we bounded sen- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tence length by 10 (excluding punctuation). Our |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first stage obtained a precision of 52.8%, which is |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ more than 6% improvement over the baseline. Our |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ second stage improved precision to nearly 56%, a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 9.3% improvement over the baseline. In addition, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we carried out experiments on Spanish (on sen- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tences of length bounded by 15, excluding punctu- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ation), achieving an increase of over 7.5% in pre- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cision over the baseline. Our algorithm increases |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ F–score as well, showing an 1.8% improvement |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ over the baseline in English and a 2.2% improve- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment in Spanish. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Section 2 reviews related work. In Section 3 we |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ detail our algorithm. Sections 4 and 5 describe the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ experimental setup and results. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 Related Work |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The advance of machine learning based ap- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ proaches in this field owes to the usage of large |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scale annotated corpora. English is the most stud- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ied language, using the FrameNet (FN) (Baker et |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ al., 1998) and PropBank (PB) (Palmer et al., 2005) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ resources. PB is a corpus well suited for evalu- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ation, since it annotates every non-auxiliary verb |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in a real corpus (the WSJ sections of the Penn |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Treebank). PB is a standard corpus for SRL eval- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ uation and was used in the CoNLL SRL shared |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tasks of 2004 (Carreras and M`arquez, 2004) and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2005 (Carreras and M`arquez, 2005). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Most work on SRL has been supervised, requir- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing dozens of thousands of SRL annotated train- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing sentences. In addition, most models assume |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that a syntactic representation of the sentence is |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ given, commonly in the form of a parse tree, a de- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pendency structure or a shallow parse. Obtaining |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these is quite costly in terms of required human |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotation. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The first work to tackle SRL as an indepen- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ dent task is (Gildea and Jurafsky, 2002), which |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ presented a supervised model trained and evalu- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ated on FrameNet. The CoNLL shared tasks of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2004 and 2005 were devoted to SRL, and stud- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ied the influence of different syntactic annotations |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and domain changes on SRL results. Computa- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tional Linguistics has recently published a special |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ issue on the task (M`arquez et al., 2008), which |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ presents state-of-the-art results and surveys the lat- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ est achievements and challenges in the field. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Most approaches to the task use a multi-level |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ approach, separating the task to an ARGID and an |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ argument classification sub-tasks. They then use |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the unlabeled argument structure (without the se- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic roles) as training data for the ARGID stage |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the entire data (perhaps with other features) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for the classification stage. Better performance |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is achieved on the classification, where state- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of-the-art supervised approaches achieve about |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 81% F-score on the in-domain identification task, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of which about 95% are later labeled correctly |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (M`arquez et al., 2008). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ There have been several exceptions to the stan- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ dard architecture described in the last paragraph. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ One suggestion poses the problem of SRL as a se- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quential tagging of words, training an SVM clas- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sifier to determine for each word whether it is in- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ side, outside or in the beginning of an argument |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Hacioglu and Ward, 2003). Other works have in- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tegrated argument classification and identification |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ into one step (Collobert and Weston, 2007), while |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ others went further and combined the former two |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ along with parsing into a single model (Musillo |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 29 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and Merlo, 2006). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Work on less supervised methods has been |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ scarce. Swier and Stevenson (2004) and Swier |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Stevenson (2005) presented the first model |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that does not use an SRL annotated corpus. How- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, they utilize the extensive verb lexicon Verb- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Net, which lists the possible argument structures |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ allowable for each verb, and supervised syntac- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic tools. Using VerbNet along with the output of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a rule-based chunker (in 2004) and a supervised |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactic parser (in 2005), they spot instances in |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the corpus that are very similar to the syntactic |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ patterns listed in VerbNet. They then use these as |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ seed for a bootstrapping algorithm, which conse- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ quently identifies the verb arguments in the corpus |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and assigns their semantic roles. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Another less supervised work is that |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ of (Grenager and Manning, 2006), which presents |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a Bayesian network model for the argument |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structure of a sentence. They use EM to learn |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the model’s parameters from unannotated data, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and use this model to tag a test corpus. However, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARGID was not the task of that work, which dealt |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ solely with argument classification. ARGID was |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ performed by manually-created rules, requiring a |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supervised or manual syntactic annotation of the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corpus to be annotated. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The three works above are relevant but incom- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ parable to our work, due to the extensive amount |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of supervision (namely, VerbNet and a rule-based |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or supervised syntactic system) they used, both in |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ detecting the syntactic structure and in detecting |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the arguments. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Work has been carried out in a few other lan- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ guages besides English. Chinese has been studied |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in (Xue, 2008). Experiments on Catalan and Span- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ish were done in SemEval 2007 (M`arquez et al., |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2007) with two participating systems. Attempts |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to compile corpora for German (Burdchardt et al., |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2006) and Arabic (Diab et al., 2008) are also un- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ derway. The small number of languages for which |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ extensive SRL annotated data exists reflects the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ considerable human effort required for such en- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ deavors. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Some SRL works have tried to use unannotated |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ data to improve the performance of a base su- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pervised model. Methods used include bootstrap- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ping approaches (Gildea and Jurafsky, 2002; Kate |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Mooney, 2007), where large unannotated cor- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pora were tagged with SRL annotation, later to |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be used to retrain the SRL model. Another ap- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proach used similarity measures either between |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ verbs (Gordon and Swanson, 2007) or between |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nouns (Gildea and Jurafsky, 2002) to overcome |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lexical sparsity. These measures were estimated |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using statistics gathered from corpora augmenting |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the model’s training data, and were then utilized |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to generalize across similar verbs or similar argu- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Attempts to substitute full constituency pars- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing by other sources of syntactic information have |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ been carried out in the SRL community. Sugges- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions include posing SRL as a sequence labeling |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ problem (M`arquez et al., 2005) or as an edge tag- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ging problem in a dependency representation (Ha- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cioglu, 2004). Punyakanok et al. (2008) provide |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a detailed comparison between the impact of us- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing shallow vs. full constituency syntactic infor- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mation in an English SRL system. Their results |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clearly demonstrate the advantage of using full an- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The identification of arguments has also been |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ carried out in the context of automatic subcatego- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rization frame acquisition. Notable examples in- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clude (Manning, 1993; Briscoe and Carroll, 1997; |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Korhonen, 2002) who all used statistical hypothe- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sis testing to filter a parser’s output for arguments, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the goal of compiling verb subcategorization |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lexicons. However, these works differ from ours |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as they attempt to characterize the behavior of a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verb type, by collecting statistics from various in- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stances of that verb, and not to determine which |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are the arguments of specific verb instances. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The algorithm presented in this paper performs |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ unsupervised clause detection as an intermedi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ate step towards argument identification. Super- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vised clause detection was also tackled as a sepa- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rate task, notably in the CoNLL 2001 shared task |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Tjong Kim Sang and D`ejean, 2001). Clause in- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formation has been applied to accelerating a syn- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tactic parser (Glaysher and Moldovan, 2006). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Algorithm |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In this section we describe our algorithm. It con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ sists of two stages, each of which reduces the set |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of argument candidates, which a-priori contains all |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ consecutive sequences of words that do not con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tain the predicate in question. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.1 Algorithm overview |XML| xmlLoc_7 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ As pre-processing, we use an unsupervised parser |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ that generates an unlabeled parse tree for each sen- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 30 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tence (Seginer, 2007). This parser is unique in that |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ it is able to induce a bracketing (unlabeled pars- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing) from raw text (without even using POS tags) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ achieving state-of-the-art results. Since our algo- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rithm uses millions to tens of millions sentences, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we must use very fast tools. The parser’s high |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ speed (thousands of words per second) enables us |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to process these large amounts of data. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The only type of supervised annotation we |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ use is POS tagging. We use the taggers MX- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ POST (Ratnaparkhi, 1996) for English and Tree- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Tagger (Schmid, 1994) for Spanish, to obtain POS |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tags for our model. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The first stage of our algorithm uses linguisti- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ cally motivated considerations to reduce the set of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ possible arguments. It does so by confining the set |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of argument candidates only to those constituents |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which obey the following two restrictions. First, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ they should be contained in the minimal clause |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ containing the predicate. Second, they should be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ k-th degree cousins of the predicate in the parse |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tree. We propose a novel algorithm for clause de- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tection and use its output to determine which of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the constituents obey these two restrictions. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The second stage of the algorithm uses point- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ wise mutual information to rule out constituents |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that appear to be weakly collocating with the pred- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ icate in question. Since a predicate greatly re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stricts the type of arguments with which it may |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ appear (this is often referred to as “selectional re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ strictions”), we expect it to have certain character- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ istic arguments with which it is likely to collocate. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.2 Clause detection stage |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The main idea behind this stage is the observation |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ that most of the arguments of a predicate are con- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tained within the minimal clause that contains the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ predicate. We tested this on our development data |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ – section 24 of the WSJ PTB, where we saw that |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 86% of the arguments that are also constituents |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (in the gold standard parse) were indeed contained |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in that minimal clause (as defined by the tree la- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bel types in the gold standard parse that denote |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a clause, e.g., S, SBAR). Since we are not pro- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vided with clause annotation (or any label), we at- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tempted to detect them in an unsupervised manner. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Our algorithm attempts to find sub-trees within the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parse tree, whose structure resembles the structure |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of a full sentence. This approximates the notion of |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a clause. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ VBP L |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ VBP L |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 1: An example of an unlabeled POS tagged |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ parse tree. The middle tree is the ST of ‘reach’ |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the root as the encoded ancestor. The bot- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tom one is the ST with its parent as the encoded |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ancestor. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Statistics gathering. In order to detect which |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ of the verb’s ancestors is the minimal clause, we |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ score each of the ancestors and select the one that |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ maximizes the score. We represent each ancestor |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using its Spinal Tree (ST). The ST of a given |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verb’s ancestor is obtained by replacing all the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ constituents that do not contain the verb by a leaf |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ having a label. This effectively encodes all the k- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ th degree cousins of the verb (for every k). The |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ leaf labels are either the word’s POS in case the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ constituent is a leaf, or the generic label “L” de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ noting a non-leaf. See Figure 1 for an example. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this stage we collect statistics of the occur- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ rences of STs in a large corpus. For every ST in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the corpus, we count the number of times it oc- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ curs in a form we consider to be a clause (positive |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ examples), and the number of times it appears in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ other forms (negative examples). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Positive examples are divided into two main |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ types. First, when the ST encodes the root an- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cestor (as in the middle tree of Figure 1); second, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ when the ancestor complies to a clause lexico- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactic pattern. In many languages there is a |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ small set of lexico-syntactic patterns that mark a |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clause, e.g. the English ‘that’, the German ‘dass’ |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the Spanish ‘que’. The patterns which were |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used in our experiments are shown in Figure 2. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For each verb instance, we traverse over its an- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ IN |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ DT |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ NNS |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ materials |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ VBP |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ L |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ in DT NN |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ NNS |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ IN |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ students |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ CD |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ about 90 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ each set |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ reach |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ L L |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ L L |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 31 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ English |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ TO + VB. The constituent starts with “to” followed by a verb in infinitive form. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WP. The constituent is preceded by a Wh-pronoun. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ That. The constituent is preceded by a “that” marked by an “IN” POS tag indicating that it is a subordinating conjunction. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Spanish |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CQUE. The constituent is preceded by a word with the POS “CQUE” which denotes the word “que” as a con-junction. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ INT. The constituent is preceded by a word with the POS “INT” which denotes an interrogative pronoun. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CSUB. The constituent is preceded by a word with one of the POSs “CSUBF”, “CSUBI” or “CSUBX”, which denote a subordinating conjunction. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 2: The set of lexico-syntactic patterns that |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ mark clauses which were used by our model. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cestors from top to bottom. For each of them we |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ update the following counters: sentence(5T) for |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the root ancestor’s 5T, patternz (5T) for the ones |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ complying to the i-th lexico-syntactic pattern and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ negative(5T) for the other ancestors1. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Clause detection. At test time, when detecting |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the minimal clause of a verb instance, we use |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the statistics collected in the previous stage. De- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ note the ancestors of the verb with A1 ... Am. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For each of them, we calculate clause(5TA, ) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and total (5TA, ). clause(5TA,) is the sum |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of sentence(5TA,) and patternz (5TA,) if this |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ancestor complies to the i-th pattern (if there |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is no such pattern, clause(5TA,) is equal to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence (5TA, )). total (5TA,) is the sum of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clause(5TA,) and negative(5TA, ). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The selected ancestor is given by: |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ clause(STA, ) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (1) Amax = argmaxA, total(STA,) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ An 5T whose total(5T) is less than a small |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ threshold2 is not considered a candidate to be the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ minimal clause, since its statistics may be un- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reliable. In case of a tie, we choose the low- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ est constituent that obtained the maximal score. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1If while traversing the tree, we encounter an ancestor |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ whose first word is preceded by a coordinating conjunction |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (marked by the POS tag “CC”), we refrain from performing |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ any additional counter updates. Structures containing coor- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dinating conjunctions tend not to obey our lexico-syntactic |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rules. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2We used 4 per million sentences, derived from develop- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ment data. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ If there is only one verb in the sentence3 or if |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ clause(5TA,) = 0 for every 1 G j G m, we |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ choose the top level constituent by default to be |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the minimal clause containing the verb. Other- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ wise, the minimal clause is defined to be the yield |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the selected ancestor. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Argument identification. For each predicate in |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the corpus, its argument candidates are now de- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fined to be the constituents contained in the min- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ imal clause containing the predicate. However, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these constituents may be (and are) nested within |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each other, violating a major restriction on SRL |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ arguments. Hence we now prune our set, by keep- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing only the siblings of all of the verb’s ancestors, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as is common in supervised SRL (Xue and Palmer, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2004). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.3 Using collocations |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We use the following observation to filter out some |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ superfluous argument candidates: since the argu- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments of a predicate many times bear a semantic |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ connection with that predicate, they consequently |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tend to collocate with it. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We collect collocation statistics from a large |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ corpus, which we annotate with parse trees and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ POS tags. We mark arguments using the argu- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment detection algorithm described in the previous |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two sections, and extract all (predicate, argument) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pairs appearing in the corpus. Recall that for each |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence, the arguments are a subset of the con- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituents in the parse tree. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We use two representations of an argument: one |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ is the POS tag sequence of the terminals contained |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the argument, the other is its head word4. The |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ predicate is represented as the conjunction of its |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lemma with its POS tag. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Denote the number of times a predicate x |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ appeared with an argument y by nxy. Denote |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the total number of (predicate, argument) pairs |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by N. Using these notations, we define the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ following quantities: nx = Eynxy, ny = Exnxy, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ p(x) = n�N , p(y) = n�N and p(x, y) = nx N . The |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pointwise mutual information of x and y is then |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ given by: |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3In this case, every argument in the sentence must be re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ lated to that verb. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4Since we do not have syntactic labels, we use an approx- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ imate notion. For English we use the Bikel parser default |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ head word rules (Bikel, 2004). For Spanish, we use the left- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ most word. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 32 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (2) PMI(x, y) = log p( x) P(y) =log n�y |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (n� �ny)/N |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ PMI effectively measures the ratio between |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the number of times x and y appeared together and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the number of times they were expected to appear, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ had they been independent. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ At test time, when an (x, y) pair is observed, we |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ check if PMI (x, y), computed on the large cor- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus, is lower than a threshold a for either of x’s |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ representations. If this holds, for at least one rep- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ resentation, we prune all instances of that (x, y) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pair. The parameter a may be selected differently |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for each of the argument representations. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In order to avoid using unreliable statistics, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ we apply this for a given pair only if n .ny N> |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ r, for some parameter r. That is, we consider |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PMI (x, y) to be reliable, only if the denomina- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tor in equation (2) is sufficiently large. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 Experimental Setup |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Corpora. We used the PropBank corpus for de- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ velopment and for evaluation on English. Section |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 24 was used for the development of our model, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and sections 2 to 21 were used as our test data. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The free parameters of the collocation extraction |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ phase were tuned on the development data. Fol- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lowing the unsupervised parsing literature, multi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ple brackets and brackets covering a single word |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are omitted. We exclude punctuation according |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to the scheme of (Klein, 2005). As is customary |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in unsupervised parsing (e.g. (Seginer, 2007)), we |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bounded the lengths of the sentences in the cor- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus to be at most 10 (excluding punctuation). This |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ results in 207 sentences in the development data, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ containing a total of 132 different verbs and 173 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verb instances (of the non-auxiliary verbs in the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SRL task, see ‘evaluation’ below) having 403 ar- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guments. The test data has 6007 sentences con- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ taining 1008 different verbs and 5130 verb in- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stances (as above) having 12436 arguments. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Our algorithm requires large amounts of data |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ to gather argument structure and collocation pat- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ terns. For the statistics gathering phase of the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clause detection algorithm, we used 4.5M sen- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tences of the NANC (Graff, 1995) corpus, bound- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing their length in the same manner. In order |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to extract collocations, we used 2M sentences |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from the British National Corpus (Burnard, 2000) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and about 29M sentences from the Dmoz cor- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus (Gabrilovich and Markovitch, 2005). Dmoz |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is a web corpus obtained by crawling and clean- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the URLs in the Open Directory Project |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (dmoz.org). All of the above corpora were parsed |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using Seginer’s parser and POS-tagged by MX- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ POST (Ratnaparkhi, 1996). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For our experiments on Spanish, we used 3.3M |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ sentences of length at most 15 (excluding punctua- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion) extracted from the Spanish Wikipedia. Here |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we chose to bound the length by 15 due to the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ smaller size of the available test corpus. The |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ same data was used both for the first and the sec- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ond stages. Our development and test data were |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ taken from the training data released for the Se- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mEval 2007 task on semantic annotation of Span- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ish (M`arquez et al., 2007). This data consisted |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of 1048 sentences of length up to 15, from which |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 200 were randomly selected as our development |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data and 848 as our test data. The development |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data included 313 verb instances while the test |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data included 1279. All corpora were parsed us- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the Seginer parser and tagged by the “Tree- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Tagger” (Schmid, 1994). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Baselines. Since this is the first paper, to our |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ knowledge, which addresses the problem of unsu- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pervised argument identification, we do not have |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ any previous results to compare to. We instead |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ compare to a baseline which marks all k-th degree |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cousins of the predicate (for every k) as arguments |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (this is the second pruning we use in the clause |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ detection stage). We name this baseline the ALL |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ COUSINS baseline. We note that a random base- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ line would score very poorly since any sequence of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ terminals which does not contain the predicate is |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a possible candidate. Therefore, beating this ran- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dom baseline is trivial. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Evaluation. Evaluation is carried out using |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ standard SRL evaluation software5. The algorithm |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is provided with a list of predicates, whose argu- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments it needs to annotate. For the task addressed |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in this paper, non-consecutive parts of arguments |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are treated as full arguments. A match is consid- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ered each time an argument in the gold standard |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data matches a marked argument in our model’s |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ output. An unmatched argument is an argument |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which appears in the gold standard data, and fails |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to appear in our model’s output, and an exces- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sive argument is an argument which appears in |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our model’s output but does not appear in the gold |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ standard. Precision and recall are defined accord- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ingly. We report an F-score as well (the harmonic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mean of precision and recall). We do not attempt |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5http://www.lsi.upc.edu/—srlconll/soft.html#software. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 33 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ to identify multi-word verbs, and therefore do not |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ report the model’s performance in identifying verb |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ boundaries. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Since our model detects clauses as an interme- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ diate product, we provide a separate evaluation |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of this task for the English corpus. We show re- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults on our development data. We use the stan- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dard parsing F-score evaluation measure. As a |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gold standard in this evaluation, we mark for each |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the verbs in our development data the minimal |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clause containing it. A minimal clause is the low- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ est ancestor of the verb in the parse tree that has |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a syntactic label of a clause according to the gold |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ standard parse of the PTB. A verb is any terminal |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ marked by one of the POS tags of type verb ac- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cording to the gold standard POS tags of the PTB. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 Results |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Our results are shown in Table 1. The left section |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ presents results on English and the right section |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ presents results on Spanish. The top line lists re- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults of the clause detection stage alone. The next |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two lines list results of the full algorithm (clause |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ detection + collocations) in two different settings |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the collocation stage. The bottom line presents |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the performance of the ALL COUSINS baseline. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In the “Collocation Maximum Precision” set- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ting the parameters of the collocation stage (a and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ r) were generally tuned such that maximal preci- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sion is achieved while preserving a minimal recall |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ level (40% for English, 20% for Spanish on the de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ velopment data). In the “Collocation Maximum F- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ score” the collocation parameters were generally |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tuned such that the maximum possible F-score for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the collocation algorithm is achieved. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The best or close to best F-score is achieved |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ when using the clause detection algorithm alone |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (59.14% for English, 23.34% for Spanish). Note |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that for both English and Spanish F-score im- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ provements are achieved via a precision improve- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment that is more significant than the recall degra- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dation. F-score maximization would be the aim of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a system that uses the output of our unsupervised |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ARGID by itself. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The “Collocation Maximum Precision” |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ achieves the best precision level (55.97% for |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ English, 21.8% for Spanish) but at the expense |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the largest recall loss. Still, it maintains a |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reasonable level of recall. The “Collocation |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Maximum F-score” is an example of a model that |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ provides a precision improvement (over both the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ baseline and the clause detection stage) with a |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ relatively small recall degradation. In the Spanish |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ experiments its F-score (23.87%) is even a bit |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ higher than that of the clause detection stage |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (23.34%). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The full two–stage algorithm (clause detection |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ + collocations) should thus be used when we in- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tend to use the model’s output as training data for |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supervised SRL engines or supervised ARGID al- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gorithms. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In our algorithm, the initial set of potential ar- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ guments consists of constituents in the Seginer |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parser’s parse tree. Consequently the fraction |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of arguments that are also constituents (81.87% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for English and 51.83% for Spanish) poses an |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ upper bound on our algorithm’s recall. Note |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that the recall of the ALL COUSINS baseline is |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 74.27% (45.75%) for English (Spanish). This |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ score emphasizes the baseline’s strength, and jus- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tifies the restriction that the arguments should be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ k-th cousins of the predicate. The difference be- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tween these bounds for the two languages provides |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a partial explanation for the corresponding gap in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the algorithm’s performance. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 3 shows the precision of the collocation |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ model (on development data) as a function of the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ amount of data it was given. We can see that |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the algorithm reaches saturation at about 5M sen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tences. It achieves this precision while maintain- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing a reasonable recall (an average recall of 43.1% |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ after saturation). The parameters of the colloca- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion model were separately tuned for each corpus |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ size, and the graph displays the maximum which |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ was obtained for each of the corpus sizes. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To better understand our model’s performance, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ we performed experiments on the English cor- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus to test how well its first stage detects clauses. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Clause detection is used by our algorithm as a step |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ towards argument identification, but it can be of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ potential benefit for other purposes as well (see |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Section 2). The results are 23.88% recall and 40% |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ precision. As in the ARGID task, a random se- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lection of arguments would have yielded an ex- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tremely poor result. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 Conclusion |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In this work we presented the first algorithm for ar- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ gument identification that uses neither supervised |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactic annotation nor SRL tagged data. We |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ have experimented on two languages: English and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Spanish. The straightforward adaptability of un- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 34 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ English (Test Data) Spanish (Test Data) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Precision Recall F1 Precision Recall F1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Clause Detection 52.84 67.14 59.14 18.00 33.19 23.34 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Collocation Maximum F–score 54.11 63.53 58.44 20.22 29.13 23.87 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Collocation Maximum Precision 55.97 40.02 46.67 21.80 18.47 20.00 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ALL COUSINS baseline 46.71 74.27 57.35 14.16 45.75 21.62 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 1: Precision, Recall and F 1 score for the different stages of our algorithm. Results are given for English (PTB, sentences |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ length bounded by 10, left part of the table) and Spanish (SemEval 2007 Spanish SRL task, right part of the table). The results |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the collocation (second) stage are given in two configurations, Collocation Maximum F-score and Collocation Maximum |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Precision (see text). The upper bounds on Recall, obtained by taking all arguments output by our unsupervised parser, are |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 81.87% for English and 51.83% for Spanish. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Number of Sentences (Millions) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 3: The performance of the second stage on English |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (squares) vs. corpus size. The precision of the baseline (trian- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gles) and of the first stage (circles) is displayed for reference. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The graph indicates the maximum precision obtained for each |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corpus size. The graph reaches saturation at about 5M sen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tences. The average recall of the sampled points from there |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on is 43.1%. Experiments were performed on the English |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ development data. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supervised models to different languages is one |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ of their most appealing characteristics. The re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cent availability of unsupervised syntactic parsers |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ has offered an opportunity to conduct research on |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SRL, without reliance on supervised syntactic an- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation. This work is the first to address the ap- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ plication of unsupervised parses to an SRL related |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ task. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Our model displayed an increase in precision of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 9% in English and 8% in Spanish over a strong |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ baseline. Precision is of particular interest in this |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ context, as instances tagged by high quality an- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation could be later used as training data for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supervised SRL algorithms. In terms of F–score, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our model showed an increase of 1.8% in English |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and of 2.2% in Spanish over the baseline. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Although the quality of unsupervised parses is |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ currently low (compared to that of supervised ap- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proaches), using great amounts of data in identi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fying recurring structures may reduce noise and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in addition address sparsity. The techniques pre- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sented in this paper are based on this observation, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using around 35M sentences in total for English |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and 3.3M sentences for Spanish. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ As this is the first work which addressed un- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ supervised ARGID, many questions remain to be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ explored. Interesting issues to address include as- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sessing the utility of the proposed methods when |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supervised parses are given, comparing our model |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to systems with no access to unsupervised parses |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and conducting evaluation using more relaxed |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ measures. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Unsupervised methods for syntactic tasks have |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ matured substantially in the last few years. No- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ table examples are (Clark, 2003) for unsupervised |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ POS tagging and (Smith and Eisner, 2006) for un- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ supervised dependency parsing. Adapting our al- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gorithm to use the output of these models, either to |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reduce the little supervision our algorithm requires |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (POS tagging) or to provide complementary syn- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tactic information, is an interesting challenge for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ future work. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ References |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Collin F. Baker, Charles J. Fillmore and John B. Lowe, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 1998. The Berkeley FrameNet Project. ACL- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ COLING ’98. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Daniel M. Bikel, 2004. Intricacies of Collins’ Parsing |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Model. Computational Linguistics, 30(4):479–511. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Ted Briscoe, John Carroll, 1997. Automatic Extraction |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ of Subcategorization from Corpora. Applied NLP |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 1997. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Aljoscha Burchardt, Katrin Erk, Anette Frank, Andrea |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Kowalski, Sebastian Pad and Manfred Pinkal, 2006 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The SALSA Corpus: a German Corpus Resource for |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Lexical Semantics. LREC ’06. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Lou Burnard, 2000. User Reference Guide for the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ British National Corpus. Technical report, Oxford |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ University. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Xavier Carreras and Lluis M`arquez, 2004. Intro- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ duction to the CoNLL–2004 Shared Task: Semantic |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Role Labeling. CoNLL ’04. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 0 2 4 6 8 10 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 48 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 46 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 44 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 42 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 52 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 50 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Second Stage |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ First Stage |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Baseline |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 35 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Xavier Carreras and Lluis M`arquez, 2005. Intro- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ duction to the CoNLL –2005 Shared Task: Semantic |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Role Labeling. CoNLL ’05. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Alexander Clark, 2003. Combining Distributional and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Morphological Information for Part of Speech In- |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ duction. EACL ’03. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Ronan Collobert and Jason Weston, 2007. Fast Se- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ mantic Extraction Using a Novel Neural Network |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Architecture. ACL ’07. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Mona Diab, Aous Mansouri, Martha Palmer, Olga |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Babko-Malaya, Wajdi Zaghouani, Ann Bies and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Mohammed Maamouri, 2008. A pilot Arabic Prop- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Bank. LREC ’08. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Evgeniy Gabrilovich and Shaul Markovitch, 2005. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Feature Generation for Text Categorization using |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ World Knowledge. IJCAI ’05. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Daniel Gildea and Daniel Jurafsky, 2002. Automatic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Labeling of Semantic Roles. Computational Lin- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guistics, 28(3):245–288. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Elliot Glaysher and Dan Moldovan, 2006. Speed- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing Up Full Syntactic Parsing by Leveraging Partial |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Parsing Decisions. COLING/ACL ’06 poster ses- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ sion. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Andrew Gordon and Reid Swanson, 2007. Generaliz- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing Semantic Role Annotations across Syntactically |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Similar Verbs. ACL ’07. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ David Graff, 1995. North American News Text Cor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ pus. Linguistic Data Consortium. LDC95T21. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Trond Grenager and Christopher D. Manning, 2006. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Unsupervised Discovery of a Statistical Verb Lexi- |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ con. EMNLP ’06. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Kadri Hacioglu, 2004. Semantic Role Labeling using |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Dependency Trees. COLING’04. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Kadri Hacioglu and Wayne Ward, 2003. Target Word |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Detection and Semantic Role Chunking using Sup- |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ port Vector Machines. HLT-NAACL ’03. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Rohit J. Kate and Raymond J. Mooney, 2007. Semi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Supervised Learning for Semantic Parsing using |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Support Vector Machines. HLT–NAACL ’07. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Karin Kipper, Hoa Trang Dang and Martha Palmer, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2000. Class-Based Construction of a Verb Lexicon. |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ AAAI ’00. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Dan Klein, 2005. The Unsupervised Learning ofNatu- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ral Language Structure. Ph.D. thesis, Stanford Uni- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ versity. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Anna Korhonen, 2002. Subcategorization Acquisition. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Ph.D. thesis, University of Cambridge. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Christopher D. Manning, 1993. Automatic Acquisition |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ of a Large Subcategorization Dictionary. ACL ’93. |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Lluis M`arquez, Xavier Carreras, Kenneth C. Lit- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tkowski and Suzanne Stevenson, 2008. Semantic |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Role Labeling: An introdution to the Special Issue. |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Computational Linguistics, 34(2):145–159 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Lluis M`arquez, Jesus Gim`enez Pere Comas and Neus |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Catal`a, 2005. Semantic Role Labeling as Sequential |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Tagging. CoNLL’05. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Lluis M`arquez, Lluis Villarejo, M. A. Marti and Mar- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ iona Taul`e, 2007. SemEval–2007 Task 09: Multi- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ level Semantic Annotation of Catalan and Spanish. |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ The 4th international workshop on Semantic Evalu- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ations (SemEval ’07). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Gabriele Musillo and Paula Merlo, 2006. Accurate |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Parsing of the proposition bank. HLT-NAACL ’06. |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Martha Palmer, Daniel Gildea and Paul Kingsbury, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2005. The Proposition Bank: A Corpus Annotated |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ with Semantic Roles. Computational Linguistics, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 31(1):71–106. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Sameer Pradhan, Kadri Hacioglu, Valerie Krugler, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Wayne Ward, James H. Martin and Daniel Jurafsky, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2005. Support Vector Learning for Semantic Argu- |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ment Classification. Machine Learning, 60(1):11– |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 39. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Sameer Pradhan, Wayne Ward, James H. Martin, 2008. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Towards Robust Semantic Role Labeling. Computa- |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ tional Linguistics, 34(2):289–310. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Adwait Ratnaparkhi, 1996. Maximum Entropy Part- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Of-Speech Tagger. EMNLP ’96. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Helmut Schmid, 1994. Probabilistic Part-of-Speech |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Tagging Using Decision Trees International Confer- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ence on New Methods in Language Processing. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Yoav Seginer, 2007. Fast Unsupervised Incremental |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Parsing. ACL ’07. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Noah A. Smith and Jason Eisner, 2006. Annealing |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Structural Bias in Multilingual Weighted Grammar |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Induction. ACL ’06. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Robert S. Swier and Suzanne Stevenson, 2004. Unsu- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ pervised Semantic Role Labeling. EMNLP ’04. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Robert S. Swier and Suzanne Stevenson, 2005. Ex- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ploiting a Verb Lexicon in Automatic Semantic Role |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Labelling. EMNLP ’05. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Erik F. Tjong Kim Sang and Herv´e D´ejean, 2001. In- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ troduction to the CoNLL-2001 Shared Task: Clause |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Identification. CoNLL ’01. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Nianwen Xue and Martha Palmer, 2004. Calibrating |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Features for Semantic Role Labeling. EMNLP ’04. |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Nianwen Xue, 2008. Labeling Chinese Predicates |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ with Semantic Roles. Computational Linguistics, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 34(2):225–255. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 36 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Brutus: A Semantic Role Labeling System Incorporating CCG, CFG, and |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Dependency Features |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ Stephen A. Boxwell, Dennis Mehay, and Chris Brew |XML| xmlLoc_0 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Department of Linguistics |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ The Ohio State University |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ {boxwe11,mehay,cbrew}@1ing.ohio-state.edu |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We describe a semantic role labeling system |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ that makes primary use of CCG-based fea- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures. Most previously developed systems |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are CFG-based and make extensive use of a |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treepath feature, which suffers from data spar- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sity due to its use of explicit tree configura- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions. CCG affords ways to augment treepath- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ based features to overcome these data sparsity |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ issues. By adding features over CCG word- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word dependencies and lexicalized verbal sub- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ categorization frames (“supertags”), we can |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ obtain an F-score that is substantially better |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ than a previous CCG-based SRL system and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ competitive with the current state of the art. A |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ manual error analysis reveals that parser errors |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ account for many of the errors of our system. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ This analysis also suggests that simultaneous |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ incremental parsing and semantic role labeling |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ may lead to performance gains in both tasks. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Semantic Role Labeling (SRL) is the process of assign- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing semantic roles to strings of words in a sentence ac- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cording to their relationship to the semantic predicates |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ expressed in the sentence. The task is difficult because |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the relationship between syntactic relations like “sub- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ject” and “object” do not always correspond to seman- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic relations like “agent” and “patient”. An effective |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic role labeling system must recognize the dif- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferences between different configurations: |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (a) [The man]Arg0 opened [the door]A�g1 [for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ him]Arg3 [today]ArgM-TMP. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (b) [The door]A�g1 opened. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ (c) [The door]A�g1 was opened by [a man]A�g0. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We use Propbank (Palmer et al., 2005), a corpus of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ newswire text annotated with verb predicate semantic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role information that is widely used in the SRL litera- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture (M`arquez et al., 2008). Rather than describe se- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantic roles in terms of “agent” or “patient”, Propbank |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ defines semantic roles on a verb-by-verb basis. For ex- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ample, the verb open encodes the OPENER as Arg0, the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ OPENEE as Arg1, and the beneficiary of the OPENING |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ action as Arg3. Propbank also defines a set of adjunct |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ roles, denoted by the letter M instead of a number. For |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ example, ArgM-TMP denotes a temporal role, like “to- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ day”. By using verb-specific roles, Propbank avoids |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ specific claims about parallels between the roles of dif- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferent verbs. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We follow the approach in (Punyakanok et al., 2008) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ in framing the SRL problem as a two-stage pipeline: |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ identification followed by labeling. During identifica- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion, every word in the sentence is labeled either as |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bearing some (as yet undetermined) semantic role or |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ not . This is done for each verb. Next, during label- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing, the precise verb-specific roles for each word are |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ determined. In contrast to the approach in (Punyakanok |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al., 2008), which tags constituents directly, we tag |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ headwords and then associate them with a constituent, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as in a previous CCG-based approach (Gildea and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hockenmaier, 2003). Another difference is our choice |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of parsers. Brutus uses the CCG parser of (Clark and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Curran, 2007, henceforth the C&C parser), Charniak’s |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parser (Charniak, 2001) for additional CFG-based fea- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tures, and MALT parser (Nivre et al., 2007) for de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pendency features, while (Punyakanok et al., 2008) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use results from an ensemble of parses from Char- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ niak’s Parser and a Collins parser (Collins, 2003; Bikel, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2004). Finally, the system described in (Punyakanok et |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 2008) uses a joint inference model to resolve dis- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ crepancies between multiple automatic parses. We do |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ not employ a similar strategy due to the differing no- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions of constituency represented in our parsers (CCG |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ having a much more fluid notion of constituency and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the MALT parser using a different approach entirely). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For the identification and labeling steps, we train |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ a maximum entropy classifier (Berger et al., 1996) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ over sections 02-21 of a version of the CCGbank cor- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus (Hockenmaier and Steedman, 2007) that has been |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ augmented by projecting the Propbank semantic anno- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tations (Boxwell and White, 2008). We evaluate our |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ SRL system’s argument predictions at the word string |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ level, making our results directly comparable for each |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ argument labeling.1 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In the following, we briefly introduce the CCG |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ grammatical formalism and motivate its use in SRL |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Sections 2–3). Our main contribution is to demon- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ strate that CCG — arguably a more expressive and lin- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1This is guaranteed by our string-to-string mapping from |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the original Propbank to the CCGbank. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 37 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 37–45, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ guistically appealing syntactic framework than vanilla |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ CFGs — is a viable basis for the SRL task. This is sup- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ported by our experimental results, the setup and details |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of which we give in Sections 4–10. In particular, using |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CCG enables us to map semantic roles directly onto |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verbal categories, an innovation of our approach that |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ leads to performance gains (Section 7). We conclude |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with an error analysis (Section 11), which motivates |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our discussion of future research for computational se- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mantics with CCG (Section 12). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 Combinatory Categorial Grammar |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Combinatory Categorial Grammar (Steedman, 2000) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is a grammatical framework that describes syntactic |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structure in terms of the combinatory potential of the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lexical (word-level) items. Rather than using standard |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ part-of-speech tags and grammatical rules, CCG en- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ codes much of the combinatory potential of each word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by assigning a syntactically informative category. For |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ example, the verb loves has the category (s\np)/np, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ which could be read “the kind of word that would be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a sentence if it could combine with a noun phrase on |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the right and a noun phrase on the left”. Further, CCG |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ has the advantage of a transparent interface between the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ way the words combine and their dependencies with |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ other words. Word-word dependencies in the CCG- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank are encoded using predicate-argument (PARG) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relations. PARG relations are defined by the functor |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word, the argument word, the category of the functor |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word and which argument slot of the functor category |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is being filled. For example, in the sentence John loves |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Mary (figure 1), there are two slots on the verbal cat- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ egory to be filled by NP arguments. The first argu- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment (the subject) fills slot 1. This can be encoded |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as , indicating the head of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the functor, the head of the argument, the functor cat- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ egory and the argument slot. The second argument |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (the direct object) fills slot 2. This can be encoded as |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ . One of the potential ad- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vantages to using CCGbank-style PARG relations is |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that they uniformly encode both local and long-range |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependencies — e.g., the noun phrase the Mary that |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ John loves expresses the same set of two dependencies. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We will show this to be a valuable tool for semantic |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role prediction. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Potential Advantages to using CCG |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ There are many potential advantages to using the CCG |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ formalism in SRL. One is the uniformity with which |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CCG can express equivalence classes of local and long- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ range (including unbounded) dependencies. CFG- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ based approaches often rely on examining potentially |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ long sequences of categories (or treepaths) between the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verb and the target word. Because there are a number of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ different treepaths that correspond to a single relation |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (figure 2), this approach can suffer from data sparsity. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CCG, however, can encode all treepath-distinct expres- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sions of a single grammatical relation into a single |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ predicate-argument relationship (figure 3). This fea- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ture has been shown (Gildea and Hockenmaier, 2003) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to be an effective substitute for treepath-based features. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ But while predicate-argument-based features are very |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ effective, they are still vulnerable both to parser er- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rors and to cases where the semantics of a sentence |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ do not correspond directly to syntactic dependencies. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To counteract this, we use both kinds of features with |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the expectation that the treepath feature will provide |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ low-level detail to compensate for missed, incorrect or |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactically impossible dependencies. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Another advantage of a CCG-based approach (and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ lexicalist approaches in general) is the ability to en- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ code verb-specific argument mappings. An argument |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mapping is a link between the CCG category and the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic roles that are likely to go with each of its ar- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guments. The projection of argument mappings onto |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CCG verbal categories is explored in (Boxwell and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ White, 2008). We describe this feature in more detail |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in section 7. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 Identification and Labeling Models |XML| xmlLoc_2 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ As in previous approaches to SRL, Brutus uses a two- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ stage pipeline of maximum entropy classifiers. In ad- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dition, we train an argument mapping classifier (de- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scribed in more detail below) whose predictions are |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used as features for the labeling model. The same |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features are extracted for both treebank and automatic |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parses. Automatic parses were generated using the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ C&C CCG parser (Clark and Curran, 2007) with its |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ derivation output format converted to resemble that of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the CCGbank. This involved following the derivational |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bracketings of the C&C parser’s output and recon- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structing the backpointers to the lexical heads using an |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in-house implementation of the basic CCG combina- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tory operations. All classifiers were trained to 500 iter- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ations of L-BFGS training — a quasi-Newton method |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from the numerical optimization literature (Liu and No- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cedal, 1989) — using Zhang Le’s maxent toolkit. 2 To |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ prevent overfitting we used Gaussian priors with global |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ variances of 1 and 5 for the identifier and labeler, re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spectively.3 The Gaussian priors were determined em- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pirically by testing on the development set. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Both the identifier and the labeler use the following |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ features: |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (1) Words. Words drawn from a 3 word window |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ around the target word ,4 with each word asso- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ciated with a binary indicator feature. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2) Part of Speech. Part of Speech tags drawn |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ from a 3 word window around the target word, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2Available for download at http://homepages. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ inf.ed.ac.uk/s0450736/maxent_toolkit. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ html. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 3Gaussian priors achieve a smoothing effect (to prevent |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ overfitting) by penalizing very large feature weights. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4The size of the window was determined experimentally |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ on the development set – we use the same window sizes |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ throughout. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 38 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Robin fixed the car |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np (s\np)/np np/n n |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ John loves Mary |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ > |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ > |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np (s[dcl]\np)/np np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ > |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s[dcl]\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s[dcl] |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 1: This sentence has two depen- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ dencies: and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fixed |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 2: The semantic relation (Arg1) between ‘car’ |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and ‘fixed’ in both phrases is the same, but the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treepaths — traced with arrows above — are differ- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent: (V>VPVP>S>RC>N the car that Robin fixed |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np/n n (np\np)/(s/np) np (s\np)/np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ >T |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s/(s\np) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ > >s |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s/np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ > |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 3: CCG word-word dependencies are passed |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ up through subordinate clauses, encoding the rela- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion between car and fixed the same in both cases: |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (s\np)/np.2.—> (Gildea and Hockenmaier, 2003) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with each associated with a binary indicator |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ feature. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (3) CCG Categories. CCG categories drawn from |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ a 3 word window around the target word, with |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each associated with a binary indicator feature. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (4) Predicate. The lemma of the predicate we are |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ tagging. E.g. fix is the lemma offixed. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (5) Result Category Detail. The grammatical fea- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ ture on the category of the predicate (indicat- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing declarative, passive, progressive, etc). This |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ can be read off the verb category: declarative |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for eats: (s[dcl]\np)/np or progressive for run- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ning:s[ng]\np. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (6) Before/After. A binary indicator variable indi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ cating whether the target word is before or after |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the verb. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (7) Treepath. The sequence of CCG categories |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ representing the path through the derivation |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from the predicate to the target word. For |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the relationship between fixed and car in the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first sentence of figure 3, the treepath is |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (s[dcl]\np)/np>s[dcl]\np and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ < indicating movement up and down the tree, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ respectively. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (8) Short Treepath. Similar to the above treepath |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ feature, except the path stops at the highest |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ node under the least common subsumer that |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is headed by the target word (this is the con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituent that the role would be marked on if we |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ identified this terminal as a role-bearing word). |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Again, for the relationship between fixed and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ car in the first sentence of figure 3, the short |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treepath is (s[dcl]\np)/np>s[dcl]\np (9) NP Modified. A binary indicator feature indi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ cating whether the target word is modified by |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ an NP modifier.5 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5This is easily read off of the CCG PARG relationships. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ S |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ �� � ��� |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ VP |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � �� |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ V NP |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ � � |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ fixed Det N |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ NP |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Robin |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ car |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ N |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ RC |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ car |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Rel |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ^S |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ that |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ NP |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ VP |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Robin |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ V |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ NP |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ �� � ��� |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Det |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ N |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ �� � ��� |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np\np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 39 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (10) Subcategorization. A sequence of the cate- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ gories that the verb combines with in the CCG |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ derivation tree. For the first sentence in fig- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ure 3, the correct subcategorization would be |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np,np. Notice that this is not necessarily a re- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ statement of the verbal category – in the second |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence of figure 3, the correct subcategoriza- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion is s/(s\np),(np\np)/(s[dcl]/np),np. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (11) PARG feature. We follow a previous CCG- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ based approach (Gildea and Hockenmaier, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2003) in using a feature to describe the PARG |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relationship between the two words, if one ex- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ists. If there is a dependency in the PARG |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structure between the two words, then this fea- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture is defined as the conjunction of (1) the cat- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ egory of the functor, (2) the argument slot that |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is being filled in the functor category, and (3) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ an indication as to whether the functor (—>) or |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the argument (+—) is the lexical head. For ex- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ample, to indicate the relationship between car |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and fixed in both sentences of figure 3, the fea- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture is (s\np)/np.2.—>. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The labeler uses all of the previous features, plus the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ following: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (12) Headship. A binary indicator feature as to |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ whether the functor or the argument is the lex- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ical head of the dependency between the two |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words, if one exists. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (13) Predicate and Before/After. The conjunction |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ of two earlier features: the predicate lemma |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ and the Before/After feature. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (14) Rel Clause. Whether the path from predicate |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ to target word passes through a relative clause |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (e.g., marked by the word ‘that’ or any other |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word with a relativizer category). |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (15) PP features. When the target word is a prepo- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ sition, we define binary indicator features for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the word, POS, and CCG category of the head |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the topmost NP in the prepositional phrase |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ headed by a preposition (a.k.a. the ‘lexical |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ head’ of the PP). So, if on heads the phrase ‘on |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the third Friday’, then we extract features re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lating to Friday for the preposition on. This is |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ null when the target word is not a preposition. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (16) Argument Mappings. If there is a PARG rela- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tion between the predicate and the target word, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the argument mapping is the most likely pre- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dicted role to go with that argument. These |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mappings are predicted using a separate classi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fier that is trained primarily on lexical informa- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion of the verb, its immediate string-level con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ text, and its observed arguments in the train- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing data. This feature is null when there is |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ no PARG relation between the predicate and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the target word. The Argument Mapping fea- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture can be viewed as a simple prediction about |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ some of the non-modifier semantic roles that a |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ verb is likely to express. We use this informa- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion as a feature and not a hard constraint to |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ allow other features to overrule the recommen- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dation made by the argument mapping classi- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fier. The features used in the argument map- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ping classifier are described in detail in section |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 CFG based Features |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In addition to CCG-based features, features can be |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ drawn from a traditional CFG-style approach when |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ they are available. Our motivation for this is twofold. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ First, others (Punyakanok et al., 2008, e.g.), have found |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that different parsers have different error patterns, and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ so using multiple parsers can yield complementary |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sources of correct information. Second, we noticed |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that, although the CCG-based system performed well |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on head word labeling, performance dropped when |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ projecting these labels to the constituent level (see sec- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions 8 and 9 for more). This may have to do with the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fact that CCG is not centered around a constituency- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ based analysis, as well as with inconsistencies between |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CCG and Penn Treebank-style bracketings (the latter |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ being what was annotated in the original Propbank). |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Penn Treebank-derived features are used in the iden- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tifier, labeler, and argument mapping classifiers. For |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ automatic parses, we use Charniak’s parser (Charniak, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2001). For gold-standard parses, we remove func- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tional tag and trace information from the Penn Tree- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank parses before we extract features over them, so as |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to simulate the conditions of an automatic parse. The |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Penn Treebank features are as follows: |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (17) CFG Treepath. A sequence of traditional |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ CFG-style categories representing the path |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from the verb to the target word. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (18) CFG Short Treepath. Analogous to the CCG- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ based short treepath feature. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (19) CFG Subcategorization. Analogous to the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ CCG-based subcategorization feature. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (20) CFG Least Common Subsumer. The cate- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ gory of the root of the smallest tree that domi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_continue +L+ nates both the verb and the target word. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 Dependency Parser Features |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Finally, several features can be extracted from a de- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ pendency representation of the same sentence. Au- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tomatic dependency relations were produced by the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ MALT parser. We incorporate MALT into our col- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lection of parses because it provides detailed informa- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion on the exact syntactic relations between word pairs |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (subject, object, adverb, etc) that is not found in other |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ automatic parsers. The features used from the depen- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dency parses are listed below: |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 40 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (21) DEP-Exists A binary indicator feature show- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ ing whether or not there is a dependency be- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tween the target word and the predicate. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (22) DEP-Type If there is a dependency between |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ the target word and the predicate, what type of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependency it is (SUBJ, OBJ, etc). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7 Argument Mapping Model |XML| xmlLoc_1 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ An innovation in our approach is to use a separate clas- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ sifier to predict an argument mapping feature. An ar- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gument mapping is a mapping from the syntactic argu- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ments of a verbal category to the semantic arguments |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that should correspond to them (Boxwell and White, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2008). In order to generate examples of the argument |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mapping for training purposes, it is necessary to em- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ploy the PARG relations for a given sentence to identify |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the headwords of each of the verbal arguments. That is, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we use the PARG relations to identify the headwords of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each of the constituents that are arguments of the verb. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Next, the appropriate semantic role that corresponds to |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that headword (given by Propbank) is identified. This |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is done by climbing the CCG derivation tree towards |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the root until we find a semantic role corresponding to |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the verb in question — i.e., by finding the point where |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the constituent headed by the verbal category combines |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the constituent headed by the argument in ques- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion. These semantic roles are then marked on the cor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ responding syntactic argument of the verb. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As an example, consider the sentence The boy loves |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ a girl. (figure 4). By examining the arguments that the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verbal category combines with in the treebank, we can |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ identify the corresponding semantic role for each argu- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment that is marked on the verbal category. We then use |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these tags to train the Argument Mapping model, which |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ will predict likely argument mappings for verbal cate- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gories based on their local surroundings and the head- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words of their arguments, similar to the supertagging |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ approaches used to label the informative syntactic cat- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ egories of the verbs (Bangalore and Joshi, 1999; Clark, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2002), except tagging “one level above” the syntax. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The Argument Mapping Predictor uses the following |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ features: |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (23) Predicate. The lemma of the predicate, as be- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ fore. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (24) Words. Words drawn from a 5 word window |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ around the target word, with each word associ- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ated with a binary indicator feature, as before. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (25) Parts of Speech. Part of Speech tags drawn |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ from a 5 word window around the target word, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with each tag associated with a binary indicator |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ feature, as before. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (26) CCG Categories. CCG categories drawn from |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ a 5 word window around the target word, with |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ each category associated with a binary indica- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tor feature, as before. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the boy loves a girl |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np/n n (s[dcl]\npArg0)/npArg1 np/n n |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np — Arga np — Argl |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s[dcl] |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 4: By looking at the constituents that the verb |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ combines with, we can identify the semantic roles cor- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ responding to the arguments marked on the verbal cat- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ egory. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (27) Argument Data. The word, POS, and CCG |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ category, and treepath of the headwords of each |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of the verbal arguments (i.e., PARG depen- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dents), each encoded as a separate binary in- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dicator feature. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (28) Number of arguments. The number of argu- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ ments marked on the verb. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (29) Words of Arguments. The head words of each |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ of the verb’s arguments. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (30) Subcategorization. The CCG categories that |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ combine with this verb. This includes syntactic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ adjuncts as well as arguments. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (31) CFG-Sisters. The POS categories of the sis- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ ters of this predicate in the CFG representation. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (32) DEP-dependencies. The individual depen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ dency types of each of the dependencies re- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lating to the verb (SBJ, OBJ, ADV, etc) taken |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from the dependency parse. We also incorpo- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rate a single feature representing the entire set |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of dependency types associated with this verb |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ into a single feature, representing the set of de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pendencies as a whole. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Given these features with gold standard parses, our |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ argument mapping model can predict entire argument |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mappings with an accuracy rate of 87.96% on the test |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ set, and 87.70% on the development set. We found the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ features generated by this model to be very useful for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic role prediction, as they enable us to make de- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cisions about entire sets of semantic roles associated |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with individual lemmas, rather than choosing them in- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependently of each other. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 8 Enabling Cross-System Comparison |XML| xmlLoc_5 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The Brutus system is designed to label headwords of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ semantic roles, rather than entire constituents. How- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, because most SRL systems are designed to label |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ constituents rather than headwords, it is necessary to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ project the roles up the derivation to the correct con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituent in order to make a meaningful comparison of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the system’s performance. This introduces the poten- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tial for further error, so we report results on the ac- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ curacy of headwords as well as the correct string of |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words. We deterministically move the role to the high- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ est constituent in the derivation that is headed by the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s[dcl]\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 41 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ P R F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ P. et al (treebank) 86.22% 87.40% 86.81% |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Brutus (treebank) 88.29% 86.39% 87.33% |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P. et al (automatic) 77.09% 75.51% 76.29% |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Brutus (automatic) 76.73% 70.45% 73.45% |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a man with glasses spoke |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np/n n (np\np)/np np s\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ �np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np - speak.Arg0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ s |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 5: The role is moved towards the root until the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ original node is no longer the head of the marked con- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituent. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P R F |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ G&H (treebank) 67.5% 60.0% 63.5% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Brutus (treebank) 88.18% 85.00% 86.56% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ G&H (automatic) 55.7% 49.5% 52.4% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Brutus (automatic) 76.06% 70.15% 72.99% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 1: Accuracy of semantic role prediction using |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ only CCG based features. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ originally tagged terminal. In most cases, this corre- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ sponds to the node immediately dominated by the low- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ est common subsuming node of the the target word and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the verb (figure 5). In some cases, the highest con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituent that is headed by the target word is not imme- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ diately dominated by the lowest common subsuming |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ node (figure 6). |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 9 Results |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Using a version of Brutus incorporating only the CCG- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ based features described above, we achieve better re- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults than a previous CCG based system (Gildea and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Hockenmaier, 2003, henceforth G&H). This could be |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ due to a number of factors, including the fact that our |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ system employs a different CCG parser, uses a more |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ complete mapping of the Propbank onto the CCGbank, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ uses a different machine learning approach,6 and has a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ richer feature set. The results for constituent tagging |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ accuracy are shown in table 1. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As expected, by incorporating Penn Treebank-based |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ features and dependency features, we obtain better re- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults than with the CCG-only system. The results for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gold standard parses are comparable to the winning |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ system of the CoNLL 2005 shared task on semantic |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ role labeling (Punyakanok et al., 2008). Other systems |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Toutanova et al., 2008; Surdeanu et al., 2007; Johans- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ son and Nugues, 2008) have also achieved comparable |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ results – we compare our system to (Punyakanok et |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 2008) due to the similarities in our approaches. The |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ performance of the full system is shown in table 2. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3 shows the ability of the system to predict |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the correct headwords of semantic roles. This is a nec- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ essary condition for correctness of the full constituent, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ but not a sufficient one. In parser evaluation, Carroll, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Minnen, and Briscoe (Carroll et al., 2003) have argued |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6G&H use a generative model with a back-off lattice, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ whereas we use a maximum entropy classifier. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2: Accuracy of semantic role prediction using |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ CCG, CFG, and MALT based features. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P R F |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Headword (treebank) 88.94% 86.98% 87.95% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Boundary (treebank) 88.29% 86.39% 87.33% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Headword (automatic) 82.36% 75.97% 79.04% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Boundary (automatic) 76.33% 70.59% 73.35% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3: Accuracy of the system for labeling semantic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ roles on both constituent boundaries and headwords. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Headwords are easier to predict than boundaries, re- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ flecting CCG’s focus on word-word relations rather |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ than constituency. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for dependencies as a more appropriate means of eval- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ uation, reflecting the focus on headwords from con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituent boundaries. We argue that, especially in the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ heavily lexicalized CCG framework, headword evalu- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ation is more appropriate, reflecting the emphasis on |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ headword combinatorics in the CCG formalism. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 The Contribution of the New Features |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Two features which are less frequently used in SRL |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ research play a major role in the Brutus system: The |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PARG feature (Gildea and Hockenmaier, 2003) and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the argument mapping feature. Removing them has |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a strong effect on accuracy when labeling treebank |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parses, as shown in our feature ablation results in ta- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ble 4. We do not report results including the Argu- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment Mapping feature but not the PARG feature, be- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cause some predicate-argument relation information is |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ assumed in generating the Argument Mapping feature. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P R F |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ +PARG +AM 88.77% 86.15% 87.44% |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ +PARG -AM 88.42% 85.78% 87.08% |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ -PARG -AM 87.92% 84.65% 86.26% |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 4: The effects of removing key features from the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ system on gold standard parses. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The same is true for automatic parses, as shown in ta- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ble 5. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 11 Error Analysis |XML| xmlLoc_6 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Many of the errors made by the Brutus system can be |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ traced directly to erroneous parses, either in the auto- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ matic or treebank parse. In some cases, PP attachment |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 42 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ with even brief exposures causing symptoms |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (((vp\vp)/vp[ng])/np n/n n/n n (s[ng]\np)/np np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ n |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s[ng]\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ n |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np — cause.Arg0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (vp\vp)/vp[ng] � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vp\vp |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 6: In this case, with is the head of with even brief exposures, so the role is correctly marked on even brief |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ exposures (based on wsj 0003.2). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ P R F |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ +PARG +AM 74.14% 62.09% 67.58% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ +PARG -AM 70.02% 64.68% 67.25% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ -PARG -AM 73.90% 61.15% 66.93% |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a form of asbestos used to make filters |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np (np\np)/np np np\np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ � |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ � |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np\np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np — Arg1 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ � |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ np |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 5: The effects of removing key features from the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ system on automatic parses. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ambiguities cause a role to be marked too high in the |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ derivation. In the sentence the company stopped using |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ asbestos in 1956 (figure 7), the correct Arg 1 of stopped |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is using asbestos. However, because in 1956 is erro- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ neously modifying the verb using rather than the verb |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stopped in the treebank parse, the system trusts the syn- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tactic analysis and places Arg1 of stopped on using as- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bestos in 1956. This particular problem is caused by an |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ annotation error in the original Penn Treebank that was |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ carried through in the conversion to CCGbank. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Another common error deals with genitive construc- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tions. Consider the phrase a form of asbestos used |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to make filters. By CCG combinatorics, the relative |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ clause could either attach to asbestos or to a form of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ asbestos. The gold standard CCG parse attaches the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relative clause to a form of asbestos (figure 8). Prop- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank agrees with this analysis, assigning Arg1 of use |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to the constituent a form of asbestos. The automatic |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parser, however, attaches the relative clause low – to |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ asbestos (figure 9). When the system is given the au- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tomatically generated parse, it incorrectly assigns the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semantic role to asbestos. In cases where the parser at- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ taches the relative clause correctly, the system is much |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ more likely to assign the role correctly. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Problems with relative clause attachment to genitives |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ are not limited to automatic parses – errors in gold- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ standard treebank parses cause similar problems when |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Treebank parses disagree with Propbank annotator in- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tuitions. In the phrase a group of workers exposed to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ asbestos (figure 10), the gold standard CCG parse at- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ taches the relative clause to workers. Propbank, how- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, annotates a group of workers as Arg 1 of exposed, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rather than following the parse and assigning the role |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only to workers. The system again follows the parse |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and incorrectly assigns the role to workers instead of a |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ group of workers. Interestingly, the C&C parser opts |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for high attachment in this instance, resulting in the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 8: CCGbank gold-standard parse of a relative |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ clause attachment. The system correctly identifies a |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ form of asbestos as Arg1 of used. (wsj 0003.1) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a form of asbestos used to make filters |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np (np\np)/np np — Arg1 np\np |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np\np |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ � |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 9: Automatic parse of the noun phrase in fig- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ure 8. Incorrect relative clause attachment causes the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ misidentification of asbestos as a semantic role bearing |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unit. (wsj 0003.1) |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ correct prediction of a group of workers as Arg1 of ex- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ posed in the automatic parse. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 12 Future Work |XML| xmlLoc_4 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ As described in the error analysis section, a large num- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ber of errors in the system are attributable to errors in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the CCG derivation, either in the gold standard or in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ automatically generated parses. Potential future work |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ may focus on developing an improved CCG parser us- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing the revised (syntactic) adjunct-argument distinc- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions (guided by the Propbank annotation) described in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Boxwell and White, 2008). This resource, together |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the reasonable accuracy (,: 90%) with which ar- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ gument mappings can be predicted, suggests the possi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ bility of an integrated, simultaneous syntactic-semantic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsing process, similar to that of (Musillo and Merlo, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2006; Merlo and Musillo, 2008). We expect this would |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ improve the reliability and accuracy of both the syntac- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic and semantic analysis components. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 13 Acknowledgments |XML| xmlLoc_7 xmlBold_yes xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ This research was funded by NSF grant IIS-0347799. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We are deeply indebted to Julia Hockenmaier for the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 43 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the company stopped using asbestos in 1956 |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np ((s[dcl]\np)/(s[ng]\np)) (s[ng]\np)/np np (s\np)\(s\np) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s[ng]\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s[ng]\np - stop.Arg1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ s[dcl]\np |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s[dcl] |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ � |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 7: An example of how incorrect PP attachment can cause an incorrect labeling. Stop.Arg1 should cover us- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing asbestos rather than using asbestos in 1956. This sentence is based on wsj 0003.3, with the structure simplified |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for clarity. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a group of workers exposed to asbestos |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ np (np\np)/np np - exposed.Arg1 np\np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ � |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ � |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np\np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ np |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 10: Propbank annotates a group of workers as Arg1 of exposed, while CCGbank attaches the relative clause |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ low. The system incorrectly labels workers as a role bearing unit. (Gold standard – wsj 0003.1) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use of her PARG generation tool. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ References |XML| xmlLoc_3 xmlBold_yes xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Srinivas Bangalore and Aravind Joshi. 1999. Su- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ pertagging: An approach to almost parsing. Com- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ putational Linguistics, 25(2):237–265. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Adam L. Berger, S. Della Pietra, and V. Della Pietra. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 1996. A maximum entropy approach to natural |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ language processing. Computational Linguistics, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 22(1):39–71. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ D.M. Bikel. 2004. Intricacies of Collins’ parsing |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ model. Computational Linguistics, 30(4):479–511. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Stephen A. Boxwell and Michael White. 2008. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Projecting propbank roles onto the ccgbank. In |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings of the Sixth International Language |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Resources and Evaluation Conference (LREC-08), |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Marrakech, Morocco. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ J. Carroll, G. Minnen, and T. Briscoe. 2003. Parser |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ evaluation. Treebanks: Building and Using Parsed |XML| xmlLoc_5 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Corpora, pages 299–316. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ E. Charniak. 2001. Immediate-head parsing for lan- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ guage models. In Proc. ACL-01, volume 39, pages |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 116–123. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Stephen Clark and James R. Curran. 2007. Wide- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ coverage Efficient Statistical Parsing with CCG and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Log-linear Models. Computational Linguistics, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 33(4):493–552. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Stephen Clark. 2002. Supertagging for combinatory |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ categorial grammar. In Proceedings of the 6th In- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ternational Workshop on Tree Adjoining Grammars |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ and Related Frameworks (TAG+6), pages 19–24, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Venice, Italy. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ M. Collins. 2003. Head-driven statistical models for |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ natural language parsing. Computational Linguis- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tics, 29(4):589–637. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Daniel Gildea and Julia Hockenmaier. 2003. Identi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ fying semantic roles using Combinatory Categorial |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Grammar. In Proc. EMNLP-03. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Julia Hockenmaier and Mark Steedman. 2007. CCG- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ bank: A Corpus of CCG Derivations and Depen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dency Structures Extracted from the Penn Treebank. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Computational Linguistics, 33(3):355–396. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ R. Johansson and P. Nugues. 2008. Dependency- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ based syntactic–semantic analysis with PropBank |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and NomBank. Proceedings of CoNLL –2008. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ D C Liu and Jorge Nocedal. 1989. On the limited |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ memory method for large scale optimization. Math- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ematical Programming B, 45(3). |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Lluis M`arquez, Xavier Carreras, Kenneth C. Litowski, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and Suzanne Stevenson. 2008. Semantic Role La- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ beling: An Introduction to the Special Issue. Com- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ putational Linguistics, 34(2):145–159. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Paola Merlo and Gabrile Musillo. 2008. Semantic |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ parsing for high-precision semantic role labelling. In |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings of CONLL-08, Manchester, UK. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Gabriele Musillo and Paola Merlo. 2006. Robust pars- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing of the proposition bank. In Proceedings of the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ EACL 2006 Workshop ROMAND, Trento. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ J. Nivre, J. Hall, J. Nilsson, A. Chanev, G. Eryigit, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ S. K¨ubler, S. Marinov, and E. Marsi. 2007. Malt- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Parser: A language-independent system for data- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ driven dependency parsing. Natural Language En- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gineering, 13(02):95–135. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Martha Palmer, Daniel Gildea, and Paul Kingsbury. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2005. The Proposition Bank: An Annotated Cor- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus of Semantic Roles. Computational Linguistics, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 31(1):71–106. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 44 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Vasin Punyakanok, Dan Roth, and Wen tau Yih. 2008. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The Importance of Syntactic Parsing and Inference |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in Semantic Role Labeling. Computational Linguis- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tics, 34(2):257–287. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Mark Steedman. 2000. The Syntactic Process. MIT |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Press. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ M. Surdeanu, L. M`arquez, X. Carreras, and P. Comas. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2007. Combination strategies for semantic role la- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ beling. Journal of Artificial Intelligence Research, |XML| xmlLoc_1 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 29:105–151. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ K. Toutanova, A. Haghighi, and C.D. Manning. 2008. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ A global joint model for semantic role labeling. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Computational Linguistics, 34(2):161–191. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 45 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Exploiting Heterogeneous Treebanks for Parsing |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Zheng-Yu Niu, Haifeng Wang, Hua Wu |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Toshiba (China) Research and Development Center |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ 5/F., Tower W2, Oriental Plaza, Beijing, 100738, China |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ {niuzhengyu,wanghaifeng,wuhua}@rdc.toshiba.com.cn |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We address the issue of using heteroge- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ neous treebanks for parsing by breaking |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ it down into two sub-problems, convert- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing grammar formalisms of the treebanks |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to the same one, and parsing on these |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ homogeneous treebanks. First we pro- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pose to employ an iteratively trained tar- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ get grammar parser to perform grammar |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formalism conversion, eliminating prede- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fined heuristic rules as required in previ- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ous methods. Then we provide two strate- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gies to refine conversion results, and adopt |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a corpus weighting technique for parsing |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on homogeneous treebanks. Results on the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Penn Treebank show that our conversion |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method achieves 42% error reduction over |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the previous best result. Evaluation on |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the Penn Chinese Treebank indicates that a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ converted dependency treebank helps con- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituency parsing and the use of unlabeled |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data by self-training further increases pars- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing f-score to 85.2%, resulting in 6% error |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reduction over the previous best result. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The last few decades have seen the emergence of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ multiple treebanks annotated with different gram- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mar formalisms, motivated by the diversity of lan- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guages and linguistic theories, which is crucial to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the success of statistical parsing (Abeille et al., |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2000; Brants et al., 1999; Bohmova et al., 2003; |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Han et al., 2002; Kurohashi and Nagao, 1998; |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Marcus et al., 1993; Moreno et al., 2003; Xue et |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 2005). Availability of multiple treebanks cre- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ates a scenario where we have a treebank anno- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tated with one grammar formalism, and another |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treebank annotated with another grammar formal- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ism that we are interested in. We call the first |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a source treebank, and the second a target tree- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ bank. We thus encounter a problem of how to |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use these heterogeneous treebanks for target gram- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mar parsing. Here heterogeneous treebanks refer |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to two or more treebanks with different grammar |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formalisms, e.g., one treebank annotated with de- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pendency structure (DS) and the other annotated |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with phrase structure (PS). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ It is important to acquire additional labeled data |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ for the target grammar parsing through exploita- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion of existing source treebanks since there is of- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ten a shortage of labeled data. However, to our |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ knowledge, there is no previous study on this is- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sue. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Recently there have been some works on us- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing multiple treebanks for domain adaptation of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsers, where these treebanks have the same |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grammar formalism (McClosky et al., 2006b; |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Roark and Bacchiani, 2003). Other related works |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ focus on converting one grammar formalism of a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treebank to another and then conducting studies on |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the converted treebank (Collins et al., 1999; Forst, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2003; Wang et al., 1994; Watkinson and Manand- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ har, 2001). These works were done either on mul- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tiple treebanks with the same grammar formalism |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or on only one converted treebank. We see that |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ their scenarios are different from ours as we work |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with multiple heterogeneous treebanks. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For the use of heterogeneous treebanks1, we |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ propose a two-step solution: (1) converting the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grammar formalism of the source treebank to the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target one, (2) refining converted trees and using |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ them as additional training data to build a target |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grammar parser. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For grammar formalism conversion, we choose |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ the DS to PS direction for the convenience of the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comparison with existing works (Xia and Palmer, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2001; Xia et al., 2008). Specifically, we assume |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that the source grammar formalism is dependency |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1Here we assume the existence of two treebanks. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 46 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 46–54, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ grammar, and the target grammar formalism is |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ phrase structure grammar. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Previous methods for DS to PS conversion |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (Collins et al., 1999; Covington, 1994; Xia and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Palmer, 2001; Xia et al., 2008) often rely on pre- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ defined heuristic rules to eliminate converison am- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ biguity, e.g., minimal projection for dependents, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lowest attachment position for dependents, and the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ selection of conversion rules that add fewer num- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ber of nodes to the converted tree. In addition, the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ validity of these heuristic rules often depends on |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ their target grammars. To eliminate the heuristic |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rules as required in previous methods, we propose |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to use an existing target grammar parser (trained |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on the target treebank) to generate N-best parses |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for each sentence in the source treebank as conver- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sion candidates, and then select the parse consis- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tent with the structure of the source tree as the con- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verted tree. Furthermore, we attempt to use con- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verted trees as additional training data to retrain |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the parser for better conversion candidates. The |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ procedure of tree conversion and parser retraining |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ will be run iteratively until a stopping condition is |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ satisfied. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Since some converted trees might be imper- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ fect from the perspective of the target grammar, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we provide two strategies to refine conversion re- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults: (1) pruning low-quality trees from the con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verted treebank, (2) interpolating the scores from |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the source grammar and the target grammar to se- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lect better converted trees. Finally we adopt a cor- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus weighting technique to get an optimal combi- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nation of the converted treebank and the existing |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target treebank for parser training. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We have evaluated our conversion algorithm on |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ a dependency structure treebank (produced from |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the Penn Treebank) for comparison with previous |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work (Xia et al., 2008). We also have investi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gated our two-step solution on two existing tree- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ banks, the Penn Chinese Treebank (CTB) (Xue et |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 2005) and the Chinese Dependency Treebank |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (CDT)2 (Liu et al., 2006). Evaluation on WSJ data |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ demonstrates that it is feasible to use a parser for |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grammar formalism conversion and the conversion |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ benefits from converted trees used for parser re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training. Our conversion method achieves 93.8% |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ f-score on dependency trees produced from WSJ |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ section 22, resulting in 42% error reduction over |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the previous best result for DS to PS conversion. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Results on CTB show that score interpolation is |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2Available at http://ir.hit.edu.cn/. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ more effective than instance pruning for the use |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ of converted treebanks for parsing and converted |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CDT helps parsing on CTB. When coupled with |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ self-training technique, a reranking parser with |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CTB and converted CDT as labeled data achieves |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 85.2% f-score on CTB test set, an absolute 1.0% |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ improvement (6% error reduction) over the previ- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ous best result for Chinese parsing. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The rest of this paper is organized as follows. In |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Section 2, we first describe a parser based method |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for DS to PS conversion, and then we discuss pos- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sible strategies to refine conversion results, and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ finally we adopt the corpus weighting technique |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for parsing on homogeneous treebanks. Section |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 provides experimental results of grammar for- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ malism conversion on a dependency treebank pro- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ duced from the Penn Treebank. In Section 4, we |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ evaluate our two-step solution on two existing het- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erogeneous Chinese treebanks. Section 5 reviews |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ related work and Section 6 concludes this work. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 Our Two-Step Solution |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2.1 Grammar Formalism Conversion |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Previous DS to PS conversion methods built a |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ converted tree by iteratively attaching nodes and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ edges to the tree with the help of conversion |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rules and heuristic rules, based on current head- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependent pair from a source dependency tree and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the structure of the built tree (Collins et al., 1999; |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Covington, 1994; Xia and Palmer, 2001; Xia et |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ al., 2008). Some observations can be made on |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ these methods: (1) for each head-dependent pair, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only one locally optimal conversion was kept dur- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing tree-building process, at the risk of pruning |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ globally optimal conversions, (2) heuristic rules |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are required to deal with the problem that one |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ head-dependent pair might have multiple conver- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sion candidates, and these heuristic rules are usu- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ally hand-crafted to reflect the structural prefer- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ence in their target grammars. To overcome these |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ limitations, we propose to employ a parser to gen- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erate N-best parses as conversion candidates and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ then use the structural information of source trees |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to select the best parse as a converted tree. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We formulate our conversion method as fol- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ lows. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Let CDS be a source treebank annotated with |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ DS and CPS be a target treebank annotated with |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ PS. Our goal is to convert the grammar formalism |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of CDS to that of CPS. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We first train a constituency parser on CPS |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 47 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Input: CPS, CDS, Q, and a constituency parser Output: Converted trees Cps |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 1. Initialize: |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ — Set Cps,0 as null, DevScore=0, q=0; |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ — Split CPS into training set CPS,train and development set CPS,dev; |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ — Train the parser on CPS,train and denote it by Pq–l; |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2. Repeat: |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_new bi_xmlPara_new +L+ — Use Pq_i to generate N-best PS parses for each sentence in CDS, and convert PS to DS for each parse; |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ — For each sentence in CDS Do |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ o �t=argmaxt Score (xi,t), and select the �t-th parse as a converted tree for this sentence; |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ — Let CD S,q S represent these converted trees, and let Ctrain=CPS,train U CDSS,q ; |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ — Train the parser on Ctrain, and denote the updated parser by Pq; |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ — Let DevScoreq be the f-score of Pq on CPS,dev; |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ — If DevScoreq > DevScore Then DevScore=DevScoreq, and Cps=Cps,q; |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ — Else break; |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ — q++; |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Until q > Q |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Table 1: Our algorithm for DS to PS conversion. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (90% trees in CPS as training set CPS,train, and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ other trees as development set CPS,dev) and then |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ let the parser generate N-best parses for each sen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tence in CDS. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Let n be the number of sentences (or trees) in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ CDS and ni be the number of N-best parses gen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erated by the parser for the i-th (1 < i < n) sen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tence in CDS. Let xi,t be the t-th (1 < t < ni) |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parse for the i-th sentence. Let yi be the tree of the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ i-th (1 < i < n) sentence in CDS. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To evaluate the quality of xi,t as a conversion |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ candidate for yi, we convert xi,t to a dependency |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tree (denoted as xDS) and then use unlabeled de- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pendency f-score to measure the similarity be- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tween xDS and yi. Let Score(xi,t) denote the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unlabeled dependency f-score of xDS against yi. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Then we determine the converted tree for yi by |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ maximizing Score(xi,t) over the N-best parses. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The conversion from PS to DS works as fol- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ lows: |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Step 1. Use a head percolation table to find the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ head of each constituent in xi,t. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Step 2. Make the head of each non-head child |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ depend on the head of the head child for each con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ stituent. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Unlabeled dependency f-score is a harmonic |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ mean of unlabeled dependency precision and unla- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ beled dependency recall. Precision measures how |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ many head-dependent word pairs found in xDS |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ i,t |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ are correct and recall is the percentage of head- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ dependent word pairs defined in the gold-standard |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tree that are found in xDS. Here we do not take |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ dependency tags into consideration for evaluation |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ since they cannot be obtained without more so- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ phisticated rules. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ To improve the quality of N-best parses, we at- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tempt to use the converted trees as additional train- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing data to retrain the parser. The procedure of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tree conversion and parser retraining can be run it- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ eratively until a termination condition is satisfied. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Here we use the parser’s f-score on CPS,dev as a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ termination criterion. If the update of training data |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hurts the performance on CPS,dev, then we stop |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the iteration. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 1 shows this DS to PS conversion algo- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ rithm. Q is an upper limit of the number of loops, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Q>0. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2.2 Target Grammar Parsing |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Through grammar formalism conversion, we have |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ successfully turned the problem of using hetero- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ geneous treebanks for parsing into the problem of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsing on homogeneous treebanks. Before using |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ converted source treebank for parsing, we present |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two strategies to refine conversion results. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Instance Pruning For some sentences in |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ CDS, the parser might fail to generate high qual- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ity N-best parses, resulting in inferior converted |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trees. To clean the converted treebank, we can re- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ move the converted trees with low unlabeled de- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pendency f-scores (defined in Section 2.1) before |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using the converted treebank for parser training |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 48 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Figure 1: A parse tree in CTB for a sentence of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ " t1t A -X � A |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ V� N 4E H A |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ R hJ # A " with |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ "People from all over the world are cast- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing their eyes on Hong Kong" as its English |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ translation. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ because these trees are "misleading" training in- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ stances. The number of removed trees will be de- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ termined by cross validation on development set. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Score Interpolation Unlabeled dependency |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ f-scores used in Section 2.1 measure the quality of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ converted trees from the perspective of the source |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ grammar only. In extreme cases, the top best |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parses in the N-best list are good conversion can- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ didates but we might select a parse ranked quite |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ low in the N-best list since there might be con- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ flicts of syntactic structure definition between the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ source grammar and the target grammar. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Figure 1 shows an example for illustration of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ a conflict between the grammar of CDT and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that of CTB. According to Chinese head percola- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion tables used in the PS to DS conversion tool |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ "Penn2Malt" 3 and Charniak’s parser4, the head |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of VP-2 is the word " 4E " (a preposition, with |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ "BA" as its POS tag in CTB), and the head of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ IP-OBJ is R hJ " . Therefore the word " R |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hJ" depends on the word "4E" . But according |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to the annotation scheme in CDT (Liu et al., 2006), |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the word "4E" is a dependent of the word "R |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hJ " . The conflicts between the two grammars |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ may lead to the problem that the selected parses |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ based on the information of the source grammar |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ might not be preferred from the perspective of the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3Available at http://w3.msi.vxu.se/—nivre/. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 4Available at http://www.cs.brown.edu/—ec/. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target grammar. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Therefore we modified the selection metric in |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Section 2.1 by interpolating two scores, the prob- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ability of a conversion candidate from the parser |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and its unlabeled dependency f-score, shown as |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ follows: |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Score(xi,t) = AxProb(xi,t)+(1—A)xScore(xi,t). (1) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The intuition behind this equation is that converted |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ trees should be preferred from the perspective of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ both the source grammar and the target grammar. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Here 0 < A < 1. Prob(xi,t) is a probability pro- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ duced by the parser for xi,t (0 < Prob(xi,t) < 1). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The value of A will be tuned by cross validation on |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ development set. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ After grammar formalism conversion, the prob- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ lem now we face has been limited to how to build |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsing models on multiple homogeneous tree- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank. A possible solution is to simply concate- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nate the two treebanks as training data. However |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this method may lead to a problem that if the size |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of CPS is significantly less than that of converted |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CDS, converted CDS may weaken the effect CPS |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ might have. One possible solution is to reduce the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ weight of examples from converted CDS in parser |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training. Corpus weighting is exactly such an ap- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ proach, with the weight tuned on development set, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that will be used for parsing on homogeneous tree- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ banks in this paper. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Experiments of Grammar Formalism |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Conversion |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.1 Evaluation on WSJ section 22 |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Xia et al. (2008) used WSJ section 19 from the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Penn Treebank to extract DS to PS conversion |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rules and then produced dependency trees from |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSJ section 22 for evaluation of their DS to PS |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion algorithm. They showed that their |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion algorithm outperformed existing meth- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ods on the WSJ data. For comparison with their |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work, we conducted experiments in the same set- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ting as theirs: using WSJ section 19 (1844 sen- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tences) as CPS, producing dependency trees from |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ WSJ section 22 (1700 sentences) as CDS5, and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ using labeled bracketing f-scores from the tool |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ "EVALB" on WSJ section 22 for performance |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ evaluation. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5We used the tool "Penn2Malt" to produce dependency |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ structures from the Penn Treebank, which was also used for |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ PS to DS conversion in our conversion algorithm. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 49 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ DevScore All the sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ LR LP F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Models (%) (%) (%) (%) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The best result of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Xia et al. (2008) - 90.7 88.1 89.4 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Q-0-method 86.8 92.2 92.8 92.5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Q-10-method 88.0 93.4 94.1 93.8 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2: Comparison with the work of Xia et al. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (2008) on WSJ section 22. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ All the sentences |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ DevScore LR LP F |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Models (%) (%) (%) (%) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Q-0-method 91.0 91.6 92.5 92.1 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Q-10-method 91.6 93.1 94.1 93.6 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3: Results of our algorithm on WSJ section |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2-18 and 20-22. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We employed Charniak’s maximum entropy in- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ spired parser (Charniak, 2000) to generate N-best |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (N=200) parses. Xia et al. (2008) used POS |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tag information, dependency structures and depen- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dency tags in test set for conversion. Similarly, we |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used POS tag information in the test set to restrict |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ search space of the parser for generation of better |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ N-best parses. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We evaluated two variants of our DS to PS con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ version algorithm: |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Q-0-method: We set the value of Q as 0 for a |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ baseline method. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Q-10-method: We set the value of Q as 10 to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ see whether it is helpful for conversion to retrain |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the parser on converted trees. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2 shows the results of our conversion al- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ gorithm on WSJ section 22. In the experiment |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of Q-10-method, DevScore reached the highest |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ value of 88.0% when q was 1. Then we used |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Cps,1 as the conversion result. Finally Q-10- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method achieved an f-score of 93.8% on WSJ sec- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion 22, an absolute 4.4% improvement (42% er- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ror reduction) over the best result of Xia et al. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2008). Moreover, Q-10-method outperformed Q- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 0-method on the same test set. These results indi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cate that it is feasible to use a parser for DS to PS |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion and the conversion benefits from the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use of converted trees for parser retraining. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.2 Evaluation on WSJ section 2-18 and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 20-22 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In this experiment we evaluated our conversion al- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ gorithm on a larger test set, WSJ section 2-18 and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 20-22 (totally 39688 sentences). Here we also |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ used WSJ section 19 as CPS. Other settings for |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Training data All the sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ LR LP F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (%) (%) (%) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 x CTB + CDTPS 84.7 85.1 84.9 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 x CTB + CDTPS 85.1 85.6 85.3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 x CTB + CDTPS 85.0 85.5 85.3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 x CTB +CDTPS 85.3 85.8 85.6 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 20 x CTB +CDTPS 85.1 85.3 85.2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 50 x CTB +CDTPS 84.9 85.3 85.1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 4: Results of the generative parser on the de- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ velopment set, when trained with various weight- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing of CTB training set and CDTPS. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this experiment are as same as that in Section 3. 1, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ except that here we used a larger test set. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3 provides the f-scores of our method with |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Q equal to 0 or 10 on WSJ section 2-18 and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 20-22. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ With Q-10-method, DevScore reached the high- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ est value of 91.6% when q was 1. Finally Q- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10-method achieved an f-score of 93.6% on WSJ |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ section 2-18 and 20-22, better than that of Q-0- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method and comparable with that of Q-10-method |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in Section 3.1. It confirms our previous finding |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that the conversion benefits from the use of con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verted trees for parser retraining. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 Experiments of Parsing |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We investigated our two-step solution on two ex- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ isting treebanks, CDT and CTB, and we used CDT |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as the source treebank and CTB as the target tree- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CDT consists of 60k Chinese sentences, anno- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tated with POS tag information and dependency |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structure information (including 28 POS tags, and |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 24 dependency tags) (Liu et al., 2006). We did not |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ use POS tag information as inputs to the parser in |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ our conversion method due to the difficulty of con- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ version from CDT POS tags to CTB POS tags. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We used a standard split of CTB for perfor- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ mance evaluation, articles 1-270 and 400-1151 as |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training set, articles 301-325 as development set, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and articles 271-300 as test set. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We used Charniak’s maximum entropy inspired |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ parser and their reranker (Charniak and Johnson, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2005) for target grammar parsing, called a gener- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ative parser (GP) and a reranking parser (RP) re- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ spectively. We reported ParseVal measures from |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the EVALB tool. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 50 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ All the sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ LR LP F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Models Training data (%) (%) (%) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ GP CTB 79.9 82.2 81.0 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ RP CTB 82.0 84.6 83.3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ GP 10 x CTB + CDTPS 80.4 82.7 81.5 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ RP 10 x CTB + CDTPS 82.8 84.7 83.8 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 5: Results of the generative parser (GP) and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the reranking parser (RP) on the test set, when |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trained on only CTB training set or an optimal |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ combination of CTB training set and CDTPS. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.1 Results of a Baseline Method to Use CDT |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We used our conversion algorithm6 to convert the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ grammar formalism of CDT to that of CTB. Let |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CDTPS denote the converted CDT by our method. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The average unlabeled dependency f-score of trees |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in CDTPS was 74.4%, and their average index in |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 200-best list was 48. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We tried the corpus weighting method when |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ combining CDTPS with CTB training set (abbre- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ viated as CTB for simplicity) as training data, by |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ gradually increasing the weight (including 1, 2, 5, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10, 20, 50) of CTB to optimize parsing perfor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mance on the development set. Table 4 presents |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the results of the generative parser with various |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ weights of CTB on the development set. Consid- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ering the performance on the development set, we |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ decided to give CTB a relative weight of 10. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Finally we evaluated two parsing models, the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ generative parser and the reranking parser, on the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ test set, with results shown in Table 5. When |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trained on CTB only, the generative parser and the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reranking parser achieved f-scores of 81.0% and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 83.3%. The use of CDTPS as additional training |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data increased f-scores of the two models to 81.5% |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and 83.8%. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2 Results of Two Strategies for a Better Use |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ of CDT |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2.1 Instance Pruning |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We used unlabeled dependency f-score of each |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ converted tree as the criterion to rank trees in |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CDTPS and then kept only the top M trees |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with high f-scores as training data for pars- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing, resulting in a corpus CDTPMS. M var- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ied from 100% xICDTPSI to 10% xICDTPSI |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with 10%xICDTPSI as the interval. ICDTPSI |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6The setting for our conversion algorithm in this experi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ment was as same as that in Section 3.1. In addition, we used |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CTB training set as CPS,tr�i�, and CTB development set as |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CPS,dev. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ All the sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ LR LP F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Models Training data (%) (%) (%) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ GP CTB + CDTa S 81.4 82.8 82.1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ RP CTB + CDTa S 83.0 85.4 84.2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 6: Results of the generative parser and the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ reranking parser on the test set, when trained on |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ an optimal combination of CTB training set and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ converted CDT. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is the number of trees in CDTPS. Then |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ we tuned the value of M by optimizing the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ parser’s performance on the development set with |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 x CTB+CDTPMS as training data. Finally the op- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ timal value of M was 100%x I CDT I. It indicates |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that even removing very few converted trees hurts |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the parsing performance. A possible reason is that |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ most of non-perfect parses can provide useful syn- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tactic structure information for building parsing |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ models. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2.2 Score Interpolation |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We used �Score(xj,t)7 to replace Score(xj,t) in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ our conversion algorithm and then ran the updated |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ algorithm on CDT. Let CDTP S denote the con- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verted CDT by this updated conversion algorithm. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The values of A (varying from 0.0 to 1.0 with 0.1 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as the interval) and the CTB weight (including 1, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2, 5, 10, 20, 50) were simultaneously tuned on the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ development set8. Finally we decided that the op- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ timal value of A was 0.4 and the optimal weight of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CTB was 1, which brought the best performance |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on the development set (an f-score of 86.1%). In |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ comparison with the results in Section 4.1, the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ average index of converted trees in 200-best list |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ increased to 2, and their average unlabeled depen- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dency f-score dropped to 65.4%. It indicates that |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ structures of converted trees become more consis- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tent with the target grammar, as indicated by the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ increase of average index of converted trees, fur- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ther away from the source grammar. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 6 provides f-scores of the generative |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ parser and the reranker on the test set, when |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ trained on CTB and CDTP S. We see that the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ performance of the reranking parser increased to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 7Before calculating �Score(xi,t), we normal- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ized the values of Prob(xi,t) for each N-best list |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ by (1) Prob(xi,t)=Prob(xi,t)-Min(Prob(xi,*)), |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ (2)Prob(xi,t)=Prob(xi,t)/Max(Prob(xi,*)), resulting |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ in that their maximum value was 1 and their minimum value |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ was 0. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 8Due to space constraint, we do not show f-scores of the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ parser with different values of A and the CTB weight. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 51 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ All the sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ LR LP F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Models Training data (%) (%) (%) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Self-trained GP 10 � T+10 � D+P 83.0 84.5 83.7 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Updated RP CTB+CDT. S 84.3 86.1 85.2 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 7: Results of the self-trained gen- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ erative parser and updated reranking parser |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ on the test set. 10 x T+10 x D+P stands for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 10 x CTB+10 x CDTP s+PDC. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 84.2% f-score, better than the result of the rerank- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing parser with CTB and CDTPS as training data |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (shown in Table 5). It indicates that the use of |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ probability information from the parser for tree |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion helps target grammar parsing. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.3 Using Unlabeled Data for Parsing |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Recent studies on parsing indicate that the use of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ unlabeled data by self-training can help parsing |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on the WSJ data, even when labeled data is rel- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ atively large (McClosky et al., 2006a; Reichart |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Rappoport, 2007). It motivates us to em- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ploy self-training technique for Chinese parsing. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ We used the POS tagged People Daily corpus9 |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Jan. 1998—Jun. 1998, and Jan. 2000—Dec. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2000) (PDC) as unlabeled data for parsing. First |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we removed the sentences with less than 3 words |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or more than 40 words from PDC to ease pars- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing, resulting in 820k sentences. Then we ran the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reranking parser in Section 4.2.2 on PDC and used |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the parses on PDC as additional training data for |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the generative parser. Here we tried the corpus |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ weighting technique for an optimal combination |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of CTB, CDTP s and parsed PDC, and chose the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relative weight of both CTB and CDTP s as 10 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by cross validation on the development set. Fi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nally we retrained the generative parser on CTB, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ CDTP s and parsed PDC. Furthermore, we used |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this self-trained generative parser as a base parser |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to retrain the reranker on CTB and CDTP s. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 7 shows the performance of self-trained |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ generative parser and updated reranker on the test |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ set, with CTB and CDTP s as labeled data. We see |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that the use of unlabeled data by self-training fur- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ther increased the reranking parser’s performance |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from 84.2% to 85.2%. Our results on Chinese data |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ confirm previous findings on English data shown |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in (McClosky et al., 2006a; Reichart and Rap- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ poport, 2007). |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 9Available at http://icl.pku.edu.cn/. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 4.4 Comparison with Previous Studies for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Chinese Parsing |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 8 and 9 present the results of previous stud- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ies on CTB. All the works in Table 8 used CTB |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ articles 1-270 as labeled data. In Table 9, Petrov |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Klein (2007) trained their model on CTB ar- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ticles 1-270 and 400-1151, and Burkett and Klein |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2008) used the same CTB articles and parse trees |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of their English translation (from the English Chi- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nese Translation Treebank) as training data. Com- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ paring our result in Table 6 with that of Petrov |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Klein (2007), we see that CDTP s helps pars- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing on CTB, which brought 0.9% f-score improve- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ment. Moreover, the use of unlabeled data further |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ boosted the parsing performance to 85.2%, an ab- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ solute 1.0% improvement over the previous best |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ result presented in Burkett and Klein (2008). |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 Related Work |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Recently there have been some studies address- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing how to use treebanks with same grammar for- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ malism for domain adaptation of parsers. Roark |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and Bachiani (2003) presented count merging and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ model interpolation techniques for domain adap- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tation of parsers. They showed that their sys- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tem with count merging achieved a higher perfor- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mance when in-domain data was weighted more |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ heavily than out-of-domain data. McClosky et al. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (2006b) used self-training and corpus weighting to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ adapt their parser trained on WSJ corpus to Brown |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corpus. Their results indicated that both unla- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ beled in-domain data and labeled out-of-domain |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data can help domain adaptation. In comparison |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with these works, we conduct our study in a dif- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferent setting where we work with multiple het- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ erogeneous treebanks. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Grammar formalism conversion makes it possi- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ble to reuse existing source treebanks for the study |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of target grammar parsing. Wang et al. (1994) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ employed a parser to help conversion of a tree- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank from a simple phrase structure to a more in- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formative phrase structure and then used this con- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ verted treebank to train their parser. Collins et al. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (1999) performed statistical constituency parsing |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of Czech on a treebank that was converted from |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the Prague Dependency Treebank under the guid- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ance of conversion rules and heuristic rules, e.g., |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ one level of projection for any category, minimal |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ projection for any dependents, and fixed position |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of attachment. Xia and Palmer (2001) adopted bet- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ter heuristic rules to build converted trees, which |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 52 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Models < 40 words All the sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ LR LP F LR LP F |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (%) (%) (%) (%) (%) (%) |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Bikel & Chiang (2000) 76.8 77.8 77.3 - - - |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chiang & Bikel (2002) 78.8 81.1 79.9 - - - |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Levy & Manning (2003) 79.2 78.4 78.8 - - - |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Bikel’s thesis (2004) 78.0 81.2 79.6 - - - |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Xiong et. al. (2005) 78.7 80.1 79.4 - - - |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chen et. al. (2005) 81.0 81.7 81.2 76.3 79.2 77.7 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Wang et. al. (2006) 79.2 81.1 80.1 76.2 78.0 77.1 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 8: Results of previous studies on CTB with CTB articles 1-270 as labeled data. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ < 40 words All the sentences |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ LR LP F LR LP F |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Models (%) (%) (%) (%) (%) (%) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Petrov & Klein (2007) 85.7 86.9 86.3 81.9 84.8 83.3 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Burkett & Klein (2008) - - - - - 84.2 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 9: Results of previous studies on CTB with more labeled data. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ reflected the structural preference in their target |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ grammar. For acquisition of better conversion |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rules, Xia et al. (2008) proposed to automati- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cally extract conversion rules from a target tree- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank. Moreover, they presented two strategies to |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ solve the problem that there might be multiple |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion rules matching the same input depen- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dency tree pattern: (1) choosing the most frequent |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rules, (2) preferring rules that add fewer number |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of nodes and attach the subtree lower. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In comparison with the works of Wang et al. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ (1994) and Collins et al. (1999), we went fur- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ther by combining the converted treebank with the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ existing target treebank for parsing. In compar- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ison with previous conversion methods (Collins |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al., 1999; Covington, 1994; Xia and Palmer, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2001; Xia et al., 2008) in which for each head- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependent pair, only one locally optimal conver- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sion was kept during tree-building process, we |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ employed a parser to generate globally optimal |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ syntactic structures, eliminating heuristic rules for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion. In addition, we used converted trees to |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ retrain the parser for better conversion candidates, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ while Wang et al. (1994) did not exploit the use of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ converted trees for parser retraining. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 6 Conclusion |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ We have proposed a two-step solution to deal with |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the issue of using heterogeneous treebanks for |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsing. First we present a parser based method |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to convert grammar formalisms of the treebanks to |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the same one, without applying predefined heuris- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tic rules, thus turning the original problem into the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ problem of parsing on homogeneous treebanks. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Then we present two strategies, instance pruning |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and score interpolation, to refine conversion re- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sults. Finally we adopt the corpus weighting tech- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nique to combine the converted source treebank |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the existing target treebank for parser train- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The study on the WSJ data shows the benefits of |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ our parser based approach for grammar formalism |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conversion. Moreover, experimental results on the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Penn Chinese Treebank indicate that a converted |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependency treebank helps constituency parsing, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and it is better to exploit probability information |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ produced by the parser through score interpolation |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ than to prune low quality trees for the use of the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ converted treebank. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Future work includes further investigation of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ our conversion method for other pairs of grammar |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ formalisms, e.g., from the grammar formalism of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the Penn Treebank to more deep linguistic formal- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ism like CCG, HPSG, or LFG. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ References |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Anne Abeille, Lionel Clement and Francois Toussenel. 2000. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Building a Treebank for French. In Proceedings of LREC |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2000, pages 87-94. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Daniel Bikel and David Chiang. 2000. Two Statistical Pars- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing Models Applied to the Chinese Treebank. In Proceed- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ings of the Second SIGHAN workshop, pages 1-6. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Daniel Bikel. 2004. On the Parameter Space of Generative |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Lexicalized Statistical Parsing Models. Ph.D. thesis, Uni- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ versity of Pennsylvania. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Alena Bohmova, Jan Hajic, Eva Hajicova and Barbora |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Vidova-Hladka. 2003. The Prague Dependency Tree- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank: A Three-Level Annotation Scenario. Treebanks: |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 53 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Building and Using Annotated Corpora. Kluwer Aca- |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ demic Publishers, pages 103-127. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Thorsten Brants, Wojciech Skut and Hans Uszkoreit. 1999. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Syntactic Annotation of a German Newspaper Corpus. In |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings of the ATALA Treebank Workshop, pages 69- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 76. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ David Burkett and Dan Klein. 2008. Two Languages are |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Better than One (for Syntactic Parsing). In Proceedings of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ EMNLP 2008, pages 877-886. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Eugene Charniak. 2000. A Maximum Entropy Inspired |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Parser. In Proceedings of NAACL 2000, pages 132-139. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Eugene Charniak and Mark Johnson. 2005. Coarse-to-Fine |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ N-Best Parsing and MaxEnt Discriminative Reranking. In |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings ofACL 2005, pages 173-180. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ying Chen, Hongling Sun and Dan Jurafsky. 2005. A Cor- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ rigendum to Sun and Jurafsky (2004) Shallow Semantic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Parsing of Chinese. University of Colorado at Boulder |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ CSLR Tech Report TR-CSLR-2005-01. |XML| xmlLoc_2 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ David Chiang and Daniel M. Bikel. 2002. Recovering La- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tent Information in Treebanks. In Proceedings of COL- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ING 2002, pages 1-7. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Micheal Collins, Lance Ramshaw, Jan Hajic and Christoph |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Tillmann. 1999. A Statistical Parser for Czech. In Pro- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ceedings ofACL 1999, pages 505-512. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Micheal Covington. 1994. GB Theory as Dependency |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Grammar. Research Report AI-1992-03. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Martin Forst. 2003. Treebank Conversion - Establishing |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ a Testsuite for a Broad-Coverage LFG from the TIGER |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Treebank. In Proceedings of LINC at EACL 2003, pages |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 25-32. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Chunghye Han, Narae Han, Eonsuk Ko and Martha Palmer. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2002. Development and Evaluation of a Korean Treebank |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and its Application to NLP. In Proceedings ofLREC 2002, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 1635-1642. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Sadao Kurohashi and Makato Nagao. 1998. Building a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Japanese Parsed Corpus While Improving the Parsing Sys- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tem. In Proceedings of LREC 1998, pages 719-724. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Roger Levy and Christopher Manning. 2003. Is It Harder to |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Parse Chinese, or the Chinese Treebank? In Proceedings |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ofACL 2003, pages 439-446. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Ting Liu, Jinshan Ma and Sheng Li. 2006. Building a Depen- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ dency Treebank for Improving Chinese Parser. Journal of |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chinese Language and Computing, 16(4):207-224. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Mitchell P. Marcus, Beatrice Santorini and Mary Ann |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Marcinkiewicz. 1993. Building a Large Annotated Cor- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus of English: The Penn Treebank. Computational Lin- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guistics, 19(2):313-330. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ David McClosky, Eugene Charniak and Mark Johnson. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2006a. Effective Self-Training for Parsing. In Proceed- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ings of NAACL 2006, pages 152-159. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ David McClosky, Eugene Charniak and Mark Johnson. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2006b. Reranking and Self-Training for Parser Adapta- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion. In Proceedings of COLING/ACL 2006, pages 337- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 344. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Antonio Moreno, Susana Lopez, Fernando Sanchez and |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Ralph Grishman. 2003. Developing a Syntactic Anno- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tation Scheme and Tools for a Spanish Treebank. Tree- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ banks: Building and Using Annotated Corpora. Kluwer |XML| xmlLoc_0 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Academic Publishers, pages 149-163. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Slav Petrov and Dan Klein. 2007. Improved Inference for |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Unlexicalized Parsing. In Proceedings of HLT/NAACL |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ 2007, pages 404-411. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Roi Reichart and Ari Rappoport. 2007. Self-Training for En- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ hancement and Domain Adaptation of Statistical Parsers |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Trained on Small Datasets. In Proceedings of ACL 2007, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pages 616-623. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Brian Roark and Michiel Bacchiani. 2003. Supervised and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Unsupervised PCFG Adaptation to Novel Domains. In |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Proceedings of HLT/NAACL 2003, pages 126-133. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Jong-Nae Wang, Jing-Shin Chang and Keh-Yih Su. 1994. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ An Automatic Treebank Conversion Algorithm for Corpus |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Sharing. In Proceedings ofACL 1994, pages 248-254. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Mengqiu Wang, Kenji Sagae and Teruko Mitamura. 2006. A |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Fast, Accurate Deterministic Parser for Chinese. In Pro- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ceedings of COLING/ACL 2006, pages 425-432. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Stephen Watkinson and Suresh Manandhar. 2001. Translat- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ing Treebank Annotation for Evaluation. In Proceedings |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of ACL Workshop on Evaluation Methodologies for Lan- |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ guage and Dialogue Systems, pages 1-8. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Fei Xia and Martha Palmer. 2001. Converting Dependency |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Structures to Phrase Structures. In Proceedings of HLT |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2001, pages 1-5. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Fei Xia, Rajesh Bhatt, Owen Rambow, Martha Palmer |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and Dipti Misra. Sharma. 2008. Towards a Multi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Representational Treebank. In Proceedings of the 7th In- |XML| xmlLoc_3 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ ternational Workshop on Treebanks and Linguistic Theo- |XML| xmlLoc_4 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ries, pages 159-170. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ Deyi Xiong, Shuanglong Li, Qun Liu, Shouxun Lin and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Yueliang Qian. 2005. Parsing the Penn Chinese Tree- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bank with Semantic Knowledge. In Proceedings of IJC- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ NLP 2005, pages 70-81. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Nianwen Xue, Fei Xia, Fu-Dong Chiou and Martha Palmer. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ 2005. The Penn Chinese TreeBank: Phrase Structure An- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ notation of a Large Corpus. Natural Language Engineer- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing, 11(2):207-238. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 54 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Cross Language Dependency Parsing using a Bilingual Lexicon* |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest0 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Hai Zhao(O— )tt, Yan Song(*,,O)t, Chunyu Kitt, Guodong Zhout |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ tDepartment of Chinese, Translation and Linguistics |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ City University of Hong Kong |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ 83 Tat Chee Avenue, Kowloon, Hong Kong, China |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ $School of Computer Science and Technology |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ Soochow University, Suzhou, China 2'5006 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_header +L+ {haizhao,yansong,ctckit}@cityu.edu.hk, gdzhou@suda.edu.cn |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-1 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_header +L+ Abstract |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ This paper proposes an approach to en- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ hance dependency parsing in a language |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by using a translated treebank from an- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ other language. A simple statistical ma- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ chine translation method, word-by-word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ decoding, where not a parallel corpus but |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a bilingual lexicon is necessary, is adopted |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ for the treebank translation. Using an en- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semble method, the key information ex- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tracted from word pairs with dependency |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relations in the translated text is effectively |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ integrated into the parser for the target lan- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guage. The proposed method is evaluated |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in English and Chinese treebanks. It is |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shown that a translated English treebank |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ helps a Chinese parser obtain a state-of- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the-art result. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1 Introduction |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Although supervised learning methods bring state- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ of-the-art outcome for dependency parser infer- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ring (McDonald et al., 2005; Hall et al., 2007), a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ large enough data set is often required for specific |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsing accuracy according to this type of meth- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ods. However, to annotate syntactic structure, ei- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ther phrase- or dependency-based, is a costly job. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Until now, the largest treebanks' in various lan- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ guages for syntax learning are with around one |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ million words (or some other similar units). Lim- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ited data stand in the way of further performance |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ enhancement. This is the case for each individual |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ language at least. But, this is not the case as we |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ observe all treebanks in different languages as a |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ whole. For example, of ten treebanks for CoNLL- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2007 shared task, none includes more than 500K |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ �The study is partially supported by City University of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Hong Kong through the Strategic Research Grant 7002037 |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ and 7002388. The first author is sponsored by a research fel- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lowship from CTL, City University of Hong Kong. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 'It is a tradition to call an annotated syntactic corpus as |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ treebank in parsing community. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tokens, while the sum of tokens from all treebanks |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is about two million (Nivre et al., 2007). |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As different human languages or treebanks |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ should share something common, this makes it |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ possible to let dependency parsing in multiple lan- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guages be beneficial with each other. In this pa- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ per, we study how to improve dependency parsing |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by using (automatically) translated texts attached |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with transformed dependency information. As a |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ case study, we consider how to enhance a Chinese |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependency parser by using a translated English |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treebank. What our method relies on is not the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ close relation of the chosen language pair but the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ similarity of two treebanks, this is the most differ- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ent from the previous work. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Two main obstacles are supposed to confront in |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ a cross-language dependency parsing task. The |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first is the cost of translation. Machine translation |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ has been shown one of the most expensive lan- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guage processing tasks, as a great deal of time and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ space is required to perform this task. In addition, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ a standard statistical machine translation method |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ based on a parallel corpus will not work effec- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tively if it is not able to find a parallel corpus that |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ right covers source and target treebanks. How- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ever, dependency parsing focuses on the relations |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of word pairs, this allows us to use a dictionary- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ based translation without assuming a parallel cor- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ pus available, and the training stage of translation |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ may be ignored and the decoding will be quite fast |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in this case. The second difficulty is that the out- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ puts of translation are hardly qualified for the pars- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ing purpose. The most challenge in this aspect is |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ morphological preprocessing. We regard that the |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ morphological issue should be handled aiming at |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the specific language, our solution here is to use |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ character-level features for a target language like |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chinese. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The rest of the paper is organized as follows. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The next section presents some related existing |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work. Section 3 describes the procedure on tree- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 55 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 55–63, |XML| xmlLoc_7 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Suntec, Singapore, 2-7 August 2009. c�2009 ACL and AFNLP |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ bank translation and dependency transformation. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Section 4 describes a dependency parser for Chi- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nese as a baseline. Section 5 describes how a |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parser can be strengthened from the translated |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treebank. The experimental results are reported in |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Section 6. Section 7 looks into a few issues con- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cerning the conditions that the proposed approach |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is suitable for. Section 8 concludes the paper. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2 The Related Work |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ As this work is about exploiting extra resources to |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ enhance an existing parser, it is related to domain |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ adaption for parsing that has been draw some in- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ terests in recent years. Typical domain adaptation |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tasks often assume annotated data in new domain |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ absent or insufficient and a large scale unlabeled |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ data available. As unlabeled data are concerned, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ semi-supervised or unsupervised methods will be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ naturally adopted. In previous works, two basic |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ types of methods can be identified to enhance an |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ existing parser from additional resources. The first |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is usually focus on exploiting automatic generated |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ labeled data from the unlabeled data (Steedman |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al., 2003; McClosky et al., 2006; Reichart and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Rappoport, 2007; Sagae and Tsujii, 2007; Chen |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al., 2008), the second is on combining super- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ vised and unsupervised methods, and only unla- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ beled data are considered (Smith and Eisner, 2006; |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Wang and Schuurmans, 2008; Koo et al., 2008). |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Our purpose in this study is to obtain a further |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ performance enhancement by exploiting treebanks |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in other languages. This is similar to the above |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first type of methods, some assistant data should |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be automatically generated for the subsequent pro- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cessing. The differences are what type of data are |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ concerned with and how they are produced. In our |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ method, a machine translation method is applied |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to tackle golden-standard treebank, while all the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ previous works focus on the unlabeled data. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Although cross-language technique has been |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ used in other natural language processing tasks, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ it is basically new for syntactic parsing as few |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ works were concerned with this issue. The rea- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ son is straightforward, syntactic structure is too |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ complicated to be properly translated and the cost |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ of translation cannot be afforded in many cases. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ However, we empirically find this difficulty may |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ be dramatically alleviated as dependencies rather |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ than phrases are used for syntactic structure repre- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentation. Even the translation outputs are not so |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ good as the expected, a dependency parser for the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ target language can effectively make use of them |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ by only considering the most related information |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ extracted from the translated text. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The basic idea to support this work is to make |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ use of the semantic connection between different |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ languages. In this sense, it is related to the work of |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Merlo et al., 2002) and (Burkett and Klein, 2008). |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The former showed that complementary informa- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion about English verbs can be extracted from |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ their translations in a second language (Chinese) |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the use of multilingual features improves clas- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sification performance of the English verbs. The |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ latter iteratively trained a model to maximize the |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ marginal likelihood of tree pairs, with alignments |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ treated as latent variables, and then jointly parsing |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bilingual sentences in a translation pair. The pro- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ posed parser using features from monolingual and |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mutual constraints helped its log-linear model to |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ achieve better performance for both monolingual |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsers and machine translation system. In this |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work, cross-language features will be also adopted |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as the latter work. However, although it is not es- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentially different, we only focus on dependency |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parsing itself, while the parsing scheme in (Bur- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ kett and Klein, 2008) based on a constituent rep- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ resentation. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Among of existing works that we are aware of, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ we regard that the most similar one to ours is (Ze- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ man and Resnik, 2008), who adapted a parser to a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ new language that is much poorer in linguistic re- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sources than the source language. However, there |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are two main differences between their work and |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ours. The first is that they considered a pair of suf- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ficiently related languages, Danish and Swedish, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and made full use of the similar characteristics of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two languages. Here we consider two quite dif- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ferent languages, English and Chinese. As fewer |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ language properties are concerned, our approach |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ holds the more possibility to be extended to other |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ language pairs than theirs. The second is that a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parallel corpus is required for their work and a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ strict statistical machine translation procedure was |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ performed, while our approach holds a merit of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ simplicity as only a bilingual lexicon is required. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3 Treebank Translation and Dependency |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Transformation |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.1 Data |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ As a case study, this work will be conducted be- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ tween the source language, English, and the tar- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ get language, Chinese, namely, we will investigate |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 56 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ how a translated English treebank enhances a Chi- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ nese dependency parser. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ For English data, the Penn Treebank (PTB) 3 |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ is used. The constituency structures is converted |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to dependency trees by using the same rules as |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (Yamada and Matsumoto, 2003) and the standard |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ training/development/test split is used. However, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only training corpus (sections 2-21) is used for |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ this study. For Chinese data, the Chinese Treebank |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (CTB) version 4.0 is used in our experiments. The |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ same rules for conversion and the same data split |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is adopted as (Wang et al., 2007): files 1-270 and |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 400-931 as training, 271-300 as testing and files |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 301-325 as development. We use the gold stan- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dard segmentation and part-of-speech (POS) tags |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in both treebanks. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As a bilingual lexicon is required for our task |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and none of existing lexicons are suitable for trans- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lating PTB, two lexicons, LDC Chinese-English |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Translation Lexicon Version 2.0 (LDC2002L27), |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and an English to Chinese lexicon in StarDict2, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are conflated, with some necessary manual exten- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sions, to cover 99% words appearing in the PTB |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (the most part of the untranslated words are named |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ entities.). This lexicon includes 123K entries. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3.2 Translation |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ A word-by-word statistical machine translation |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ strategy is adopted to translate words attached |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ with the respective dependency information from |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the source language to the target one. In detail, a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word-based decoding is used, which adopts a log- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ linear framework as in (Och and Ney, 2002) with |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only two features, translation model and language |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ model, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ exp[E2i� 1 Aihi(c, e)] |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ E, exp[E2i� 1 Aihi(c, e)] |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Where |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ h1 (c, e) = log(p .y(c�e)) |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is the translation model, which is converted from |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ the bilingual lexicon, and |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ h2 (c, e) = log (pO (c)) |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is the language model, a word trigram model |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ trained from the CTB. In our experiment, we set |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ two weights A1 = A2 = 1. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 2StarDict is an open source dictionary software, available |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ at http://stardict.sourceforge.net/. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The conversion process of the source treebank |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is completed by three steps as the following: |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 1. Bind POS tag and dependency relation of a |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ word with itself; |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2. Translate the PTB text into Chinese word by |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_new +L+ word. Since we use a lexicon rather than a parallel |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corpus to estimate the translation probabilities, we |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ simply assign uniform probabilities to all transla- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion options. Thus the decoding process is actu- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ally only determined by the language model. Sim- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ilar to the “bag translation” experiment in (Brown |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ et al., 1990), the candidate target sentences made |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ up by a sequence of the optional target words are |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ranked by the trigram language model. The output |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ sentence will be generated only if it is with maxi- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mum probability as follows, |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_yes bi_xmlSFBIA_continue bi_xmlPara_continue +L+ c = argmax{pO(c)p.y(c�e)} |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ = argmax pO (c) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ = argmaxn pO (w,) |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ A beam search algorithm is used for this process |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ to find the best path from all the translation op- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions; As the training stage, especially, the most |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ time-consuming alignment sub-stage, is skipped, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the translation only includes a decoding procedure |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that takes about 4.5 hours for about one million |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words of the PTB in a 2.8GHz PC. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3. After the target sentence is generated, the at- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tached POS tags and dependency information of |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each English word will also be transferred to each |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ corresponding Chinese word. As word order is of- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ten changed after translation, the pointer of each |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dependency relationship, represented by a serial |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ number, should be re-calculated. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Although we try to perform an exact word-by- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ word translation, this aim cannot be fully reached |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in fact, as the following case is frequently encoun- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tered, multiple English words have to be translated |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ into one Chinese word. To solve this problem, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ we use a policy that lets the output Chinese word |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ only inherits the attached information of the high- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ est syntactic head in the original multiple English |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4 Dependency Parsing: Baseline |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_largest-2 xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 4.1 Learning Model and Features |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ According to (McDonald and Nivre, 2007), all |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ data-driven models for dependency parsing that |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ have been proposed in recent years can be de- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scribed as either graph-based or transition-based. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ P(cle) = |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 57 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Table 1: Feature Notations |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Although the former will be also used as compari- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ son, the latter is chosen as the main parsing frame- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ work by this study for the sake of efficiency. In de- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tail, a shift-reduce method is adopted as in (Nivre, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2003), where a classifier is used to make a parsing |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ decision step by step. In each step, the classifier |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ checks a word pair, namely, s, the top of a stack |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that consists of the processed words, and, i, the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ first word in the (input) unprocessed sequence, to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ determine if a dependent relation should be estab- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lished between them. Besides two dependency arc |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ building actions, a shift action and a reduce ac- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion are also defined to maintain the stack and the |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unprocessed sequence. In this work, we adopt a |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ left-to-right arc-eager parsing model, that means |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ that the parser scans the input sequence from left |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to right and right dependents are attached to their |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ heads as soon as possible (Hall et al., 2007). |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ While memory-based and margin-based learn- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ ing approaches such as support vector machines |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are popularly applied to shift-reduce parsing, we |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ apply maximum entropy model as the learning |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ model for efficient training and adopting over- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lapped features as our work in (Zhao and Kit, |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2008), especially, those character-level ones for |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chinese parsing. Our implementation of maxi- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mum entropy adopts L-BFGS algorithm for pa- |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rameter optimization as usual. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ With notations defined in Table 1, a feature set |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ as shown in Table 2 is adopted. Here, we explain |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ some terms in Tables 1 and 2. We used a large |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scale feature selection approach as in (Zhao et al., |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2009) to obtain the feature set in Table 2. Some |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ feature notations in this paper are also borrowed |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ from that work. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The feature curroot returns the root of a par- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tial parsing tree that includes a specified node. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The feature charseq returns a character sequence |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ whose members are collected from all identified |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ children for a specified word. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In Table 2, as for concatenating multiple sub- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ strings into a feature string, there are two ways, |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ seq and bag. The former is to concatenate all sub- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ strings without do something special. The latter |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ will remove all duplicated substrings, sort the rest |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and concatenate all at last. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Note that we systemically use a group of |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ character-level features. Surprisingly, as to our |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ best knowledge, this is the first report on using this |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type of features in Chinese dependency parsing. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Although (McDonald et al., 2005) used the pre- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fix of each word form instead of word form itself |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ as features, character-level features here for Chi- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nese is essentially different from that. As Chinese |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ is basically a character-based written language. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Character plays an important role in many means, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ most characters can be formed as single-character |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ words, and Chinese itself is character-order free |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rather than word-order free to some extent. In ad- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ dition, there is often a close connection between |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the meaning of a Chinese word and its first or last |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ character. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 4.2 Parsing using a Beam Search Algorithm |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ In Table 2, the feature preact� returns the previous |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ parsing action type, and the subscript n stands for |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the action order before the current action. These |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ are a group of Markovian features. Without this |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ type of features, a shift-reduce parser may directly |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ scan through an input sequence in linear time. |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Otherwise, following the work of (Duan et al., |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 2007) and (Zhao, 2009), the parsing algorithm is |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to search a parsing action sequence with the max- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ imal probability. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Y5di = argmax p(di �di-1di-2...)� |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ i |XML| xmlLoc_6 xmlBold_no xmlItalic_yes xmlFontSize_smaller xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ where 5di is the object parsing action sequence, |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ p(di � di-1...) is the conditional probability, and di |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Meaning |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The word in the top of stack |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The first word below the top of stack. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The first word before(after) the word |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ in the top of stack. |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The first (second) word in the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ unprocessed sequence, etc. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Dependent direction |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Head |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Leftmost child |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Rightmost child |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Right nearest child |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ word form |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ POS tag of word |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ coarse POS: the first letter of POS tag of word |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ coarse POS: the first two POS tags of word |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the left nearest verb |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The first character of a word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ The first two characters of a word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ The last character of a word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The last two characters of a word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ ’s, i.e., ‘s.dprel’ means dependent label |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ of character in the top of stack |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Feature combination, i.e., ‘s.char+i.char’ |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ means both s.char and i.char work as a |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ feature function. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ Notation |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ s |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s' |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ s-1,s1... |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ i, i+1,... |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ dir |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ h |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ lm |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ rm |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ rn |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ form |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ pos |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ cpos1 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ cpos2 |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ lnverb |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ char1 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ char2 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ char-1 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ char-2 |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ . |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ + |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 58 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Figure 1: A comparison before and after translation |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ is i-th parsing action. We use a beam search algo- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ rithm to find the object parsing action sequence. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 2: Features for Parsing |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ in . f orm, n = 0, 1 i.f orm + i1.form |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ in.char2 + in+1.char2, n = —1, 0 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ i.char_1 + i1.char_1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in.char_2 n = 0, 3 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ i1.char_2 +i2.char_2 +i3.char_2 i.lnverb.char_2 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ i3.pos |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in.pos + in+1.pos, n = 0, 1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ i_2.cpos1 + i_1.cpos1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ i1 .cpos1 + i2.cpos1 + i3.cpos1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s'2.char1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s'.char_2 + s'1.char_2 s'_2.cpos2 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s'_1.cpos2 + s'1.cpos2 s'.cpos2 + s'1.cpos2 s’. children.cpos2.seq s’. children. dprel.seq s’.subtree.depth |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s'.h. f orm + s'.rm.cpos1 s'.lm.char2 + s'.char2 s.h. children.dprel.seq s.lm.dprel |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s.char_2 + i1.char_2 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s.charn + i.charn, n = —1,1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s _ 1.pos + i1 .pos |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s.pos + in.pos, n = —1, 0, 1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s : illinePath. f orm.bag s'.form + i.form |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s'.char2 + in.char2, n = —1, 0, 1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s.curroot.pos + i.pos |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s.curroot.char2 + i.char2 s.children.cpos2.seq + i.children.cpos2.seq s.children.cpos2.seq + i.children.cpos2.seq + s.cpos2 + i.cpos2 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ s'.children.dprel.seq + i.children.dprel.seq |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ preact_ 1 preact_2 preact_2+preact_ 1 |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_no xmlTable_yes xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 5 Exploiting the Translated Treebank |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ As we cannot expect too much for a word-by-word |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ translation, only word pairs with dependency rela- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tion in translated text are extracted as useful and |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ reliable information. Then some features based |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ on a query in these word pairs according to the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ current parsing state (namely, words in the cur- |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ rent stack and input) will be derived to enhance |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the Chinese parser. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ A translation sample can be seen in Figure 1. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ Although most words are satisfactorily translated, |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to generate effective features, what we still have to |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ consider at first is the inconsistence between the |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ translated text and the target text. |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ In Chinese, word lemma is always its word form |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ itself, this is a convenient characteristic in com- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ putational linguistics and makes lemma features |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ unnecessary for Chinese parsing at all. However, |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Chinese has a special primary processing task, i.e., |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word segmentation. Unfortunately, word defini- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ tions for Chinese are not consistent in various lin- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ guistical views, for example, seven segmentation |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ conventions for computational purpose are for- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mally proposed since the first Bakeoff3. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Note that CTB or any other Chinese treebank |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ has its own word segmentation guideline. Chi- |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ nese word should be strictly segmented according |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ to the guideline before POS tags and dependency |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ relations are annotated. However, as we say the |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 3Bakeoff is a Chinese processing share task held by |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ SIGHAN. |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ 59 |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_none xmlPic_yes xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ English treebank is translated into Chinese word |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ by word, Chinese words in the translated text are |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ exactly some entries from the bilingual lexicon, |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ they are actually irregular phrases, short sentences |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ or something else rather than words that follows |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ any existing word segmentation convention. If the |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ bilingual lexicon is not carefully selected or re- |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ fined according to the treebank where the Chinese |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ parser is trained from, then there will be a serious |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ inconsistence on word segmentation conventions |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ between the translated and the target treebanks. |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ As all concerned feature values here are calcu- |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ lated from the searching result in the translated |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ word pair list according to the current parsing |XML| xmlLoc_1 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ state, and a complete and exact match cannot be |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ always expected, our solution to the above seg- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mentation issue is using a partial matching strat- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ egy based on characters that the words include. |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Above all, a translated word pair list, L, is ex- |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ tracted from the translated treebank. Each item in |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ the list consists of three elements, dependant word |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ (dp), head word (hd) and the frequency of this pair |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the translated treebank, f . |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ There are two basic strategies to organize the |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ features derived from the translated word pair list. |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ The first is to find the most matching word pair |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ in the list and extract some properties from it, |XML| xmlLoc_3 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ such as the matched length, part-of-speech tags |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and so on, to generate features. Note that a |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ matching priority serial should be defined afore- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ hand in this case. The second is to check every |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ matching models between the current parsing state |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ and the partially matched word pair. In an early |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ version of our approach, the former was imple- |XML| xmlLoc_4 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ mented. However, It is proven to be quite inef- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ficient in computation. Thus we adopt the sec- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ond strategy at last. Two matching model fea- |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ ture functions, 0(•) and 0(•), are correspondingly |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ defined as follows. The return value of 0(•) or |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ 0(•) is the logarithmic frequency of the matched |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ item. There are four input parameters required |XML| xmlLoc_5 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ by the function 0(•). Two parameters of them |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ are about which part of the stack(input) words is |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ chosen, and other two are about which part of |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ each item in the translated word pair is chosen. |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ These parameters could be set to full or charn as |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ shown in Table 1, where n = ..., —2, —1, 1, 2, .... |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ For example, a possible feature could be |XML| xmlLoc_6 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_continue +L+ �(s. f ull, i.chari, dp. f ull, hd.char1 ), it tries to |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_larger xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ find a match in L by comparing stack word and |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_new bi_xmlPara_new +L+ dp word, and the first character of input word |XML| xmlLoc_7 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ Table 3: Features based on the translated treebank |XML| xmlLoc_0 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ and the first character of hd word. If such |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_new +L+ a match item in L is found, then 0(•) returns |XML| xmlLoc_2 xmlBold_no xmlItalic_no xmlFontSize_common xmlPic_no xmlTable_no xmlBullet_no bi_xmlSFBIA_continue bi_xmlPara_continue +L+ log(f ). There are three input parameters required |XML| xmlLoc_2 x