Collaborative agents for task-oriented dialogue systems

Dialogue systems (a.k.a. conversational agents) aim to help people interact with machines through natural language. They are playing an increasingly important role in our daily life. There are two categories of approaches: modularized pipeline agents and end-to-end single-module agents. A challenge of the former is error accumulation because multiple modules are sequentially dependent. And concerning the latter, it is impractical to use a single general agent to handle all complex cases. In this thesis, we introduce a new framework, namely collaborative task-oriented dialogue systems. Within this framework, we propose a series of approaches where a group of collaborative specialized agents outperforms a single general agent, in terms of four dimensions: (i) model collaboration, (ii) user collaboration, (iii) language collaboration, and (iv) uncertainty estimation.