Leveraging Domain Information for Data Driven Task-Oriented Dialogue Systems

Task-oriented dialogue systems are designed to interact and assist users to achieve a certain goal. A typical approach to model such systems is a component-based approach that involves multiple components, each performing a sub-task of the dialogue system. These components are generally modelled based on a given domain with domain-specific features making them expensive to port to new domains and requiring large in-domain data. In this thesis we investigate the role of domain information in data-driven task-oriented systems. We note that domain information is widely unused in current neural models, and present several approaches that incorporate and take advantage of domain knowledge in different components of a dialogue system. More specifically, this thesis: i) presents a neural model for the entity recognition task focused on low-resource domains integrating large gazetteers as domain information; ii) presents an approach for single domain dialogue state trackers that incorporates domain information; iii) presents a multi-domain dialogue state tracker, which is able to read directly from domain schema and is able to make predictions even for unseen domains; iv) provides a survey on the recent neural approaches to dialogue state tracker by grouping neural models based on their approach; and v) proposes to incorporate domain knowledge for implementing a proactive strategy of a dialogue policy component. Both results and analysis of the experiments show the effectiveness of incorporating domain information as part of network modelling. Particularly, we address current limitations of neural models for task-oriented systems, such as requiring large training data, high latency response time, and flexibility with respect to domains.