Deep Generation Techniques in Task-Oriented Dialogue Systems

Building a task-oriented dialogue system like a restaurant booking is meaningful since it can largely reduce the working load of human and serve multiple users at the same time. A task-oriented dialogue system is often composed of a few modules, such as natural language understanding, dialogue state tracking, knowledgebase (KB) query, dialogue policy engine and response generation. Language understanding aims to convert the input to some predefined semantic frame. State tracking models explicitly the input semantic frame and the dialogue history for producing KB queries. Dialogue policy model decides on the system action which is then realized by a natural language generation component. The natural language generation component, particularly style-variation text generation, aims to map the meaning representations (MRs) and style (such as personality), we call them together as themes, to one or more corresponding natural language (NL) texts. A novel Focal-Variation Network (FVN) that learns latent distributions that closely follow the given themes are proposed for diverse text generation. Besides the language generation module, the other modules can also adopt the deep generation technique to achieve better performance: (1) multi-act generation in the policy engine module, (2) a flexible-structured end-to-end dialogue system based on a two-stage-decoder network. A future work that extends multiple-act to the natural language will also be discussed.