*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Learning Dialogue States of Task-Oriented Dialogues at Scale
Committee:
Dr. Juang, Advisor
Dr. Anderson, Chair
Dr. Yang
Abstract: The objective of the proposed research is to extend and develop new ways to improve the reliability and generalization capacities of dialogue state tracking (DST) models. As an essential component in task-oriented dialogue systems, dialogue state tracking aims to track human-machine interactions and identify user goals for dialogue management, and is an important task for understanding how dialogue systems process natural-language dialogues. This proposal reveals challenges and inadequacies in task-oriented dialogue systems concerning services in real-world scenarios and the bottleneck of enhancing existing dialogue systems, and proposes methods to overcome weaknesses in previous dialogue modeling. In this proposal, we focus on neural-based task-oriented dialogue state tracking, which utilizes neural networks to manage the interaction between the system and a user to accomplish pre-defined tasks and services. Compared to traditional finite state-based models, neural DST models have proven to be more effective in learning rich and contextualized representations and provide a unified architecture for tackling task-oriented dialogues in multi-domains. This proposal addresses two parts of the research. In the first part, we describe the dialogue state tracking framework with extra supervisions to save the cost when scaling up to new domains and services, and improve its capacities for synthesizing and generalizing domain characteristics. In the second part, we further investigate updates of dialogue states in task-oriented dialogues with the help of graph-based dialogue state tracking. We do not only model state relations to improve the interpretability of the dialogue state updating process, but also enhance the scalability of adapting to unseen dialogue states. The idea of constructing graph-based state representations and modeling transformations of user goals will be beneficial for future dialogue system designs.
