Abstract
Dialogue systems have been developed for several years. As dialogue systems become ubiquitous, dialogue strategies of virtual agents are receiving more and more attention. However, to know how to select a proper dialogue in a specific social context is not a trivial task since the world is complex. In this thesis, we propose reinforcement learning to learn the strategy of “interrogation dialogue” in virtual drama. Our first contribution is describing a new reinforcement learning framework that can learn dialogue strategies from the interrogation dialogue. The second contribution is bringing the social context and emotion states of agents into the dialogue strategies. In order to demonstrate and simulate the performance, we based on a scenario from a detective novel to build the background knowledge of the world. In particular, we model the emotion variations of a suspect using a generation function of human emotion based on psychological literature so that the detective can learn the dialogue strategies based on the suspect emotion context. And the result of the learned dialogue policy is very sensitive in detecting lying of a suspect, and the superintendent gets more correct answer.