Image Source: acquire.io
Understanding the current state of a dialogue is extremely critical in building powerful conversational recommendation systems. However, the biggest problem encountered while developing such systems is lack of labelled data as with every new domain, new slots-value pairs come and thus a labelling cost. Moreover, manual labelling of states in a new domain is costly and time-consuming. Therefore, we need to resort to low resource techniques. However, going for a zero shot approach comes with another problem of high computation cost. Hence, we are working on an approach which utilises weak supervision, optimising the labelling and computation cost.
We attempt to tackle this problem using data augmentation techniques leveraging language generation models to generate more data and using pre-trained models to study syntactic and semantic details of the data for generating possible slots. The proposed method will be tested on low-resource settings of dialogue state tracking tasks in various domains.
- Ojasv Kamal (Undergraduate Intern Student)
- Li Chuang (Ph.D. Student)
- Min-Yen Kan (Professor and Advisor)