Regional Variation of Domain-Specific Lexical Items: Toward a Pan- Chinese Lexical Resource Oi Yee Kwong and Benjamin K. Tsou Language Information Sciences Research Centre City University of Hong Kong Tat Chee Avenue, Kowloon, Hong Kong {rlolivia,rlbtsou}@cityu.edu.hk

Abstract This paper reports on an initial and nec- essary step toward the construction of a Pan-Chinese lexical resource. We inves- tigated the regional variation of lexical items in two specific domains, finance and sports; and explored how much of such variation is covered in existing Chi- nese synonym dictionaries, in particular the Tongyici Cilin. The domain-specific lexical items were obtained from subsec- tions of a synchronous Chinese corpus, LIVAC. Results showed that 20-40% of the words from various subcorpora are unique to the individual communities, and as much as 70% of such unique items are not yet covered in the Tongyici Cilin. The results suggested great potential for building a Pan-Chinese lexical resource for Chinese language processing. Our next step would be to explore automatic means for extracting related lexical items from the corpus, and to incorporate them into existing semantic classifications. 1 Introduction Many cities have underground railway systems. Somehow one takes the tube in London but the subway in New York. In a more recent edition of the Roget's Thesaurus (Kirkpatrick, 1987), subway , tube , underground railway and metro are found in the same semicolon-separated group under head 624 Way . Similarly if one looks up WordNet (http://wordnet.princeton.edu; Miller et al., 1990), the synset to which subway belongs also contains the words metro , tube , under- ground , and subway system ; and it is further in- dicated that "in Paris the subway system is called the `metro' and in London it is called the `tube' or the `underground''. Such regional lexical variation is also found in Chinese. For instance, the subway system in Hong Kong, known as the Mass Transit Railway or MTR, is called +'J0 in Chinese. The subway systems in Beijing and Shanghai, as well as the one in Singapore, are also known as +'J0 , but that in Taipei is known as L—PI . Their counterpart in Japan is written as +'J__k 4;zP in Kanji. Such regional variation, as part of lexical knowledge, is important and useful for many natural language applications, including natural language understanding, information re- trieval, and machine translation. Unfortunately, existing Chinese lexical resources often lack such comprehensiveness. To fill this gap, Tsou and Kwong (2006) pro- posed a comprehensive Pan-Chinese lexical re- source, based on a large and unique synchronous Chinese corpus as an authentic basis for lexical acquisition and analysis across various Chinese speech communities. For a significant world language like Chinese, a useful lexical resource should have maximum versatility and portability . It is not sufficient to target at one particular community speaking the language and thus cover only language usage observed from that particu- lar community. Instead, such a lexical resource should document the core and universal sub- stances of the language on the one hand, and also the more subtle variations found in different communities on the other. As is evident from the above example on the variation of subway , re- gional variation should be captured for the lexi- cal resource to be useful in a wide range of ap- plications. In this study, we investigate and compare the regional variation of lexical items from two spe-

9 Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing , pages 9–16, Sydney, July 2006. c � 2006 Association for Computational Linguistics

cific domains, finance and sports, as an initial and necessary step toward the more important undertaking of building a Pan-Chinese lexical resource. In addition, we make use of an exist- ing Chinese synonym dictionary, the Tongyici Cilin (Mei et al., 1984) as leverage, and explore its coverage of such variation and thus the poten- tial for enriching it. The lexical items under study were obtained from a synchronous Chinese corpus, LIVAC, which will be further introduced in Section 4. Corpus data from four Chinese speech communities were compared with respect to their commonality and uniqueness, and also against Cilin for their coverage. Results showed that 20-40% of the words extracted from the cor- pus are unique to the individual communities, and as much as 70% of such unique items are not yet covered in Cilin. It therefore suggests that the synchronous corpus is a rich source for mining region-specific lexical items, and there is great potential for building a Pan-Chinese lexical resource for Chinese language processing. In Section 2, we will briefly review existing resources and related work. Then in Section 3, we will briefly outline the design and architec- ture of the Pan-Chinese lexical resource pro- posed by Tsou and Kwong (2006). In Section 4, we will further describe the Chinese synonym dictionary and the synchronous Chinese corpus used in this study. The comparison of their lexi- cal items will be discussed in Section 5. Future directions will be presented in Section 6, fol- lowed by a conclusion. 2 Existing Resources and Related Work The construction and development of large lexical resources is relying more and more on corpus-based approaches, not only as a result of the increased availability of large corpora, but also for the authoritativeness and authenticity allowed by the approach. The Collins COBUILD English Dictionary (Sinclair, 1987) is amongst the most well-known lexicographic fruit based on large corpora. For natural language applications, much of the information in conventional dictionaries targeted at human readers must be made explicit. Lexical resources for computer use thus need consider- able manipulation, customisation, and supple- mentation (e.g. Calzolari, 1982). WordNet (Miller et al., 1990), grouping words into synsets and linking them up with relational pointers, is probably the first broad coverage general compu- tational lexical database. In view of the intensive time and effort required in resource building, some researchers have taken an alternative route by extracting information from existing machine- readable dictionaries and corpora semi- automatically (e.g. Vossen et al., 1989; Riloff and Shepherd, 1999; Lin et al, 2003). Compared to the development of thesauri and lexical databases, and research into semantic networks for major languages such as English, similar work for the Chinese language is less mature. This gap was partly due to the lack of authoritative Chinese corpora as a basis for analysis, but has been gradually reduced with the recent availability of large Chinese corpora in- cluding the LIVAC synchronous corpus (Tsou and Lai, 2003) used in this work and further de- scribed below, the Sinica Corpus (Chen et al., 1996), the Chinese Penn Treebank (Xia et al., 2000), and the like. An important issue which is seldom addressed in the construction of Chinese lexical databases is the problem of versatility and portability. For a language such as Chinese which is spoken in many different communities, different linguistic norms have emerged as a result of the individual- istic evolution and development of the language within a particular community and culture. Such variations are seldom adequately reflected in ex- isting lexical resources, which often only draw reference from one particular source. For in- stance, Tongyici Cilin( �� ) (Mei et al., 1984) is a thesaurus containing some 70,000 Chinese lexical items in the tradition of the Ro- get's Thesaurus for English, that is, in a hierar- chy of broad conceptual categories. First pub- lished in the 1980s, it was based exclusively on Chinese as used in post-1949 Mainland China. Thus for the subway example above, the closest word group found is ' F 41 , 1741 (train) only, let alone the subway itself and its regional variations. With the recent availability of large corpora, especially synchronous ones, to construct an au- thoritative and timely lexical resource for Chi- nese is less distant than it was in the past. A large synchronous corpus provides authentic ex- amples of the language as used in a variety of locations. It thus enables us to attempt a com- prehensive and in-depth analysis of the core common language in constructing a lexical re- source; and to incorporate useful information relating to location-sensitive linguistic variations.

3 Proposal of a Pan-Chinese Thesaurus The Pan-Chinese lexicon proposed by Tsou and Kwong (2006) is expected to capture not only the core senses of lexical items but also senses and uses specific to individual Chinese speech communities. The lexical database will be organised into a core database and a supplementary one. The core database will contain the core lexical infor- mation for word senses and usages which are common to most Chinese speech communities, whereas the supplementary database will contain the language uses specific to individual commu- nities, including "marginal" and "sublanguage" uses. A network structure will be adopted for the lexical items. The nodes could be sets of near- synonyms or single lexical items (in which case synonymy will be one type of links). The links will not only represent the paradigmatic semantic relations but also syntagmatic ones (such as se- lectional restrictions). We thus begin by investigating in depth the regional variation of lexical items, especially domain-specific words, among several Chinese speech communities. In addition, we explore the potential of enriching existing resources as a start. In the following section, we will discuss the Tongyici Cilin and the synchronous Chinese cor- pus used in this study in greater details. 4 Materials and Method 4.1 The Tongyici Cilin The Tongyici Cilin( 同義詞詞林 ) (Mei et al., 1984) is a Chinese synonym dictionary, or more often known as a Chinese thesaurus in the tradi- tion of the Roget's Thesaurus for English. The Roget's Thesaurus has about 1,000 numbered semantic heads, more generally grouped under higher level semantic classes and subclasses, and more specifically differentiated into paragraphs and semicolon-separated word groups. Similarly, some 70,000 Chinese lexical items are organized into a hierarchy of broad conceptual categories in the Tongyici Cilin. Its classification consists of 12 top-level semantic classes, 94 sub-classes, 1,248 semantic heads and 3,925 paragraphs. 4.2 The LIVAC Synchronous Corpus LIVAC (http://www.livac.org) stands for Lin- guistic Variation in Chinese Speech Communi- ties. It is a synchronous corpus developed by the Language Information Sciences Research Centre of the City University of Hong Kong since 1995 (Tsou and Lai, 2003). The corpus consists of newspaper articles collected regularly and syn- chronously from six Chinese speech communi- ties, namely Hong Kong, Beijing, Taipei, Singa- pore, Shanghai, and Macau. Texts collected cover a variety of domains, including front page news stories, local news, international news, edi- torials, sports news, entertainment news, and fi- nancial news. Up to December 2005, the corpus has already accumulated about 180 million char- acter tokens which, upon automatic word seg- mentation and manual verification, amount to over 900K word types. For the present study, we make use of the sub- corpora collected over the 9-year period 1995- 2004 from Hong Kong (HK), Beijing (BJ), Taipei (TW), and Singapore (SG). In particular, we focus on the financial news and sports news to investigate the commonality and uniqueness of the lexical items used in these specific domains in the various communities. We also evaluate the adequacy of the Tongyici Cilin in terms of its coverage of such domain-specific terms espe- cially from the Pan-Chinese perspective, and thus assess the room for its enrichment with the synchronous corpus. Table 1 shows the sizes of the subcorpora used for this study.

1133 1488 1219 1325 1320 1464 1272 494 240 240 240 240 269 Subcorpus Overall Overall (rounded to nearest 0.01M) Financial News (rounded to nearest 1K) Sports News (rounded to nearest 1K) Word Token Word Type Word Token Word Type Word Token Word Type HK 14.39M 0.22M 970K 38K 1041K 39K BJ 11.70M 0.19M 232K 20K 443K 28K TW 12.32M 0.20M 254K 22K 657K 33K SG 13.22M 0.21M 621K 28K 998K 34K

Table 1 Sizes of individual subcorpora

4.3 Procedures Word-frequency lists were generated from the financial and sports subcorpora from each indi- vidual community. For each resulting list, the steps below were followed to remove irrelevant items and retain only the potentially useful con- tent words: (a) Remove all numbers and non-Chinese words. (b) Remove all proper names, including those annotated as personal names, geographical names, and organisation names. Proper names have been annotated in the corpora during the process of word segmentation. (c) Remove function words. (d) Remove lexical items with frequency 5 or below. The numbers of remaining items in each sub- corpus after the above steps are listed in Tables 2 and 3 for the two domains respectively. The lexical items retained, which are expected to con- tain a substantial amount of content words, are potentially useful for the current study. The lists in each domain (from the various subcorpora) were compared in terms of the items they share and those unique to individual communities. Their unique items were also compared against the Tongyici Cilin to investigate its adequacy and explore how it might be enriched with the synchronous corpus.

1210 1132 1138 1133 1137 1148 264 240 240 240 264 Subcorpus All After (a) After (b) After (c) After(d) HK 37,525 27,937 20,422 17,162 5,238 BJ 20,025 17,361 14,460 12,134 2,791 TW 22,142 19,428 16,316 13,496 3,088 SG 28,193 22,829 16,863 13,822 3,836

Table 2 Number of word types remaining after various data cleaning steps for the financial domain 1210 1132 1138 1133 1137 1148 264 240 240 240 269 Subcorpus All After (a) After (b) After (c) After(d) HK 39,190 35,720 25,289 21,502 6,316 BJ 27,971 26,049 19,799 16,598 3,878 TW 32,706 30,231 20,361 17,248 4,601 SG 34,040 31,974 19,995 16,780 5,120

Table 3 Number of word types remaining after various data cleaning steps for the sports domain

5 Results and Discussion 5.1 Lexical Items from LIVAC The four subcorpora of the financial domain differ considerably in their sizes, and slightly less so for the sports domain. Despite this, we ob- served for both domains from Tables 2 and 3 that in general about 40-50% of all word types are numbers, non-Chinese words, proper names, and function words. Of the remaining items, about 20-30% have frequency greater than 5. These several thousand word types from each subcor- pus are expected to be amongst the more interest- ing items and form the "candidate sets" for fur- ther investigation. 5.2 Commonality among Various Regions Comparing the candidate sets from various sub- corpora, which reflect the use of Chinese in vari- ous Chinese speech communities, Tables 4 and 5 show the sizes of the intersection sets among different places for the two domains respectively. The intersection set for all four places contains slightly more than 1,000 lexical items in the fi- nancial domain. A quick skim through these common lexical items suggests that they contain, on the one hand, the many general concepts in the financial domain (e.g. FlJ company, VA market, �,KLi 7 bank, JOV:r invest / investment, � � business, gW develop / development, PRI corporation, J!U f,j stock shares, J!U4 shareholder, V:r� capital, etc.); and on the other hand, many reportage and cognitive verbs often used in news articles (e.g. i, express, � d reckon, << , ,! appear, '- - M reflect, etc.). In the sports domain, more than 1,700 lexical items were found in all of the four subcorpora. Like its financial counterpart, we found many general concepts at the top of the list (e.g. 5N _J player, 5N�- team, rjf match, r competi-

tion, r league, " t coach, W opponent, ice'#i champion, etc.). The numbers of overlaps in Tables 4 suggest that lexical items used in Mainland China (as evident from BJ data) seem to have the least in common with the rest. For instance, compared to the overlap amongst all four regions (i.e. 1,039), the overlap has increased most when BJ was not included in the comparison; and when we com- pare any two regions, the overlap between BJ and TW is smallest. Nevertheless, such unique- ness of BJ data is less apparent in the sports do- main. In particular, the difference between HK/BJ and BJ/TW is even slightly less than that in the financial domain. If we look at the individual regions, HK ap- parently shares most (about 50%) with SG, and vice versa (about 68%), in the financial domain. At the same time, BJ also shares more with HK than with the other two regions, and so does TW. But surprisingly, BJ has over 60% overlap with SG and about 55% with TW in the sports domain. The overlaps of TW with HK and with BJ differ by more than 20% in the finance domain, but only by about 10% in the sports domain. All these patterns might suggest lexical items in the financial domain are more versatile and have more varied focus in different communities, whereas those in the sports domain reflect the more common interests of different places.

1848 994 969 975 969 984 264 240 240 240 240 240 240 240 240 240 240 240 269 Regions Overlap Proportion to individual lists (%) HK BJ TW SG HK / BJ / TW / SG 1039 19.84 37.23 33.65 27.09 HK / BJ / TW 1126 21.50 40.34 36.46 HK / BJ / SG 1327 25.33 47.55 34.59 HK / TW / SG 1581 30.18 51.20 41.21 BJ / TW / SG 1092 39.13 35.36 28.47 HK / BJ 1609 30.72 57.65 HK / TW 1912 36.50 61.92 HK / SG 2607 49.77 67.96 BJ / TW 1250 44.79 40.48 BJ / SG 1505 53.92 39.23 TW / SG 1795 58.13 46.79

Table 4 Commonality amongst various regions for the financial domain 1848 994 969 975 969 984 264 240 240 240 240 240 240 240 240 240 240 240 269 Regions Overlap Proportion to individual lists (%) HK BJ TW SG HK / BJ / TW / SG 1668 26.41 43.01 36.25 32.58 HK / BJ / TW 1782 28.21 45.95 38.73 HK / BJ / SG 2047 32.41 52.78 39.98 HK / TW / SG 2249 35.61 48.88 43.93 BJ / TW / SG 1864 48.07 40.51 36.41 HK / BJ 2318 36.70 59.77 HK / TW 2693 42.64 58.53 HK / SG 3305 52.33 64.55 BJ / TW 2124 54.77 46.16 BJ / SG 2554 65.86 49.88 TW / SG 2709 58.88 52.91

Table 5 Commonality amongst various regions for the sports domain

5.3 Uniqueness of Various Regions Next we compared the lists with respect to what they have unique to themselves. Table 6 shows the numbers of unique items found in each list, together with examples from the most frequent 20 unique items in each case. Again, taking the size difference among the candidate sets into account, about 40% of the lexical items found in HK data are unique to the region, which re-echoes the versatility and wide

coverage of interests of HK data. This is espe- cially evident when compared to only about 20% of the candidate sets for SG are unique to Singa- pore. 1 Looking at the unique lexical items found in individual regions, it is not difficult to see the region-specific lexicalisation of certain concepts. For instance, in terms of housing, f z (housing under the Home Ownership Scheme) is a specific kind of housing in Hong Kong, � z is a specific term in Singapore (as seen in SG data), whereas housing is generally expressed as (� Y in Mainland China (as seen in BJ data). Similarly, �;; (HK) and '_F ON " (BJ) both refer to training, but may relate to different practice in the two communities. Such regional variation lends strong support to the importance of a Pan- Chinese lexical resource. The lists of unique items also suggest the vari- ous focus and orientation in different Chinese speech communities. For example, while Hong Kong pays much attention to the real estate mar- ket and stock market, Mainland China may be focusing more on the basic needs like water, farming, poverty alleviation, etc., and Singapore is relatively more concerned with local affairs like port management. The passion for baseball, among other more popular sports like soccer, is most obvious from the unique lexical items found in TW data. 5.4 Comparison with Tongyici Cilin As mentioned earlier, the Tongyici Cilin contains some 70,000 lexical items under 12 broad se- mantic classes, 94 subclasses, and 1,428 heads. It was first published in the 1980s and was based on lexical usages mostly of post-1949 Mainland China. In this section, we discuss the results ob- tained from comparing the unique lexical items found from individual subcorpora with Cilin, which are shown in Table 7. On the one hand, Cilin's collection of words may be considerably dated and obviously will not include new concepts and neologisms arising in the last two decades. On the other hand, the data in LIVAC come from newspaper materials in the 1990s. So overall speaking, for each of the unique word lists, much less than 50% are cov- ered in Cilin. 1 Upon further analysis, on average about 60% of these "unique" items were actually found in one or more of the other regions, but with frequency 5 or below. Since the difference in frequency is quite large for most items, we can reasonably treat them as unique to a particular community. Nevertheless, there is still an apparent gap be- tween Cilin's coverage of the unique items from various places. About 40% of the unique items found in BJ for both domains are covered; but for other places, the coverage is more often less than 30% in either or both domains. Again, this could be considered a result of Cilin's bias to- ward lexical usages in Mainland China. In addition, while almost 40% of the unique items in BJ data are found in Cilin, many of these unique items covered are amongst the most frequent items. On the contrary, even though about 560 unique items in HK data are also found in Cilin, only 3 out of the 20 most frequent items are amongst them. In addition, the appar- ent coverage does not necessarily suggest the correct match of word senses. For instance, fz is found under head Bnl together with other items like � Y , (�=�' , etc., all of which only re- fer to the general concept of housing, instead of the housing specifically under the Home Owner- ship Scheme as known in Hong Kong. Also, coverage of words like V_ , �1�' and -1—f - in the sports domain does not match their actual usages which refer to team names. A more inter- esting example might be 'F `'' , which is used in the basketball context in TW data, and in no way refers to the literal "hot pot" sense. Results from the above comparisons thus sup- port that (1) different Chinese speech communi- ties have their distinct usage of Chinese lexical items, in terms of both form and sense; (2) such variation is found in different domains, such as the financial and sports domain; (3) existing lexical resources, the Tongyici Cilin in particular as in our current study, should be enriched and enhanced by capturing lexical usages from a va- riety of Chinese speech communities, to repre- sent the lexical items from a Pan-Chinese per- spective; and (4) lexical items obtained from the synchronous Chinese corpus can supplement the existing content of the Tongyici Cilin, with more contemporarily lexicalised concepts, as well as variant expressions of similar and related con- cepts from various Chinese speech communities. Hence it remains for us to further investigate how the related lexical items obtained from the synchronous corpus should be grouped and in- corporated into the semantic classification of ex- isting lexical resources; and to further explore how they might be extracted in a large scale by automatic means. These will definitely be amongst the most important future directions as discussed in the next section.

6 Future Work In the current study, we have investigated the regional variation of lexical items from the fi- nancial and sports domain, and the coverage of the Tongyici Cilin for such variation. The results suggested great potential for building a Pan- Chinese lexical resource for Chinese language processing. Our next step would thus be to fur- ther investigate more automatic means for ex- tracting the near-synonymous or closely related items from the various subcorpora. To this end, we would explore algorithms like those used in Lin et al. (2003). Of similar importance is the mechanism for grouping the related lexical items and incorporating them into the semantic classi- fications of existing lexical resources. In this regard we will proceed with further in-depth analysis of the classificatory structures of indi- vidual resources and fit in our Pan-Chinese ar- chitecture. Apart from the Tongyici Cilin, there are other existing Chinese lexical resources such as HowNet (Dong and Dong, 2000), SUMO and Chinese WordNet (Huang et al., 2004), as well as other synonym dictionaries from which we might draw reference to build up our Pan- Chinese lexical resource. 7 Conclusion In this paper, we have investigated the regional variation of lexical items in two specific domains from a synchronous Chinese corpus, and ex- plored their coverage in a Chinese synonym dic- tionary. Results are encouraging in the sense that 20-40% of the candidate words from various subcorpora are unique to the individual commu- nities, and as much as 70% of such unique items are not yet covered in the Tongyici Cilin. It therefore suggests great importance and potential for a Pan-Chinese lexical resource which we aim to construct. The synchronous corpus is a valu- able resource for mining the region-specific expressions while existing synonym dictionaries might provide a ready-made semantic classifica- tory structure. Our next step would be to explore automatic means for extracting related lexical items from the corpus, and to incorporate them into existing semantic classifications. Acknowledgements This work is supported by Competitive Ear- marked Research Grant (CERG) of the Research Grants Council of Hong Kong under grant No. CityU1317/03H. The authors would like to thank the anonymous reviewers for comments. References Calzolari, N. (1982) Towards the organization of lexical definitions on a database structure. In E. Hajicova (Ed.), COLING '82 Abstracts , Charles University, Prague, pp.61-64. Caraballo, S.A. (1999) Automatic construction of a hypernym-labeled noun hierarchy. In Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics (ACL'99) , College Park, Maryland, pp. 120-126. Chen, K-J., Huang, C-R., Chang, L-P. and Hsu, H-L. (1996) Sinica Corpus: Design Methodology for Balanced Corpora. In Proceedings of the 11th Pa- cific Asia Conference on Language, Information, and Computation (PACLIC 11) , Seoul, Korea, pp.167-176. Dong, Z. and Dong, Q. (2000) HowNet . http://www.keenage.com. Huang, C-R., Chang, R-Y. and Lee, S-B. (2004) Sinica BOW (Bilingual Ontological Wordnet): In- tegration of Bilingual WordNet and SUMO . In Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC2004) , Lisbon, Portugal. Kirkpatrick, B. (1987) Roget's Thesaurus of English Words and Phrases . Penguin Books. Lin, D., Zhao, S., Qin, L. and Zhou, M. (2003) Iden- tifying Synonyms among Distributionally Similar Words. In Proceedings of the 18th Joint Interna- tional Conference on Artificial Intelligence (IJCAI- 03) , Acapulco, pp. 1492-1493 . Mei et al. fV�� - � FP, - r1��- � d� (1984) (( fFi1�a fa f�) ( Tongyici Cilin ). ��ff ff M (Commerical Press) / �N�f��r . Miller, G.A., Beckwith, R., Fellbaum, C., Gross, D. and Miller, K.J. (1990) Introduction to WordNet: An online lexical database. International Journal ofLexicography, 3(4) :235-244. Riloff, E. and Shepherd, J. (1999) A corpus-based bootstrapping algorithm for semi-automated se- mantic lexicon construction. Natural Language Engineering, 5(2) :147-156. Sinclair, J. (1987) Collins COBUILD English Lan- guage Dictionary . London, UK: HarperCollins. Tsou, B.K. and Kwong, O.Y. (2006) Toward a Pan- Chinese Thesaurus. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC 2006) , Genoa, Italy. Tsou, B.K. and Lai, T.B.Y. ��>M_ - '' f� (2003) Il�� l _j' M— �� . In B. Xu, M. Sun

and G. Jin 徐波、孫茂松、靳光瑾 (Eds.), 《中文 z nguage 息瑗理若干重要問題》 ( Issues in Chinese Processing ). 北京：科學出版社 , pp. 147-165. Vossen, P., Meijs, W. and den Broeder, M. (1989) Meaning and structure in dictionary definitions. In B. Boguraev and T. Briscoe (Eds.), Computational Lexicography for Natural Language Processing . Essex, UK: Longman Group.

854 4191 4195 264 1541 1541 1540 1383 206 Region Unique Items and Examples (Financial) Unique Items and Examples (Sports) HK 2105 (40.19%) 2410 (38.16%) BJ 933 (33.43%) 907 (23.39%) 中棋院 TW891 (28.85%{){t+人1!3{02 (28.30%{)�{ 隊獅隊控 75` 信經理降息 �Y 7 I J`I 牛投資成長率團隊立- 委 Jet V� ' 戰神 SG 890 (23.20%) 1044 (20.39%) rr� 璔

Table 6 Uniqueness of individual subcorpora Table 7 Coverage of the Tongyici Cilin for the unique lexical items in individual subcorpora Xia, F., Palmer, M., Xue, N., Okwrowski, M.E., Kovarik, J., Huang, S., Kroch, T. and Marcus, M. (2000) Developing Guidelines and Ensuring Con- sistency for Chinese Text Annotation. In Proceed- ings of the 2nd International Conference on Lan- guage Resources and Evaluation (LREC-2000), Athens, Greece. 16 Region Financial Sports Found in Cilin Not in Cilin Found in Cilin Not in Cilin HK 560 (26.60%) 1545 884 (36.68%) 1526 戶 P 拆息 (73.40%) 現時操練披甲晨攻堅 (62.32%) BJ 369 (39.55%) 564 355 (39.14%) 552 村抽查住房運行黨 (60.45%) 中學生自治區 (60.86%) 走私品種林業鄉鎮森林彩琒農民解放軍 it'll 幹璔 TW 626 354 (27.19%) 948 (70.26%) 剛雷公�送 (72.81%) �-j{�� 水手瓫導"Y"`和晚間 SG (62.58%) 效勞星期六星期三城門補足 (73.08%) 星期四星期五星期天腳踏車星期二

契約專案不動產改選股利通路關卡週終瓨席次 333(37.42%) 557 281 (26.91%) 763

265 (29.74%)