Many researches in the dialogue research community have worked on the constructing goal-oriented dialogue system so far. However though, these previous researches focus on the situation where the dialogue system tries to achieve either the user goal or the system (or its owners) goal through the dialogue. In this thesis, I discuss the realization of the cooperative dialogue system, the dialogue system which try to achieve both the user goal and the system goal through the dialogue for the decision making. As typical dialogues cooperative systems perform, I focus on 1) the negotiation dialogue and 2) the cooperative persuasive dialogue. In order to make systems have ability to converse with multiple-interlocutors in these dialogue, I propose these dialogue model and apply statistical framework to learn its dialogue manager:
[Optimization of the dialogue manager in multi-party negotiation dialogue]:
In this part, I propose the model of multi-party negotiation dialogue, where the both users and the system has right to decide, and optimize its statistical dialogue manager. I examine several combination of reinforcement learning algorithms and reward function and evaluate their performances.
[Modeling cooperative persuasive dialogue and optimization of its dialogue manager]:
In this part, I propose the framework to develop dialogue manager for the situation where the only user has right to decide, and the system needs to perform cooperative dialogue (cooperative persuasive dialogue). At first, I collect and analyze the human cooperative persuasive dialogue corpus. Then, considering the analysis result, I apply reinforcement learning to system dialogue manager, and perform its evaluation.
In both of these part, constructed dialogue manager achieved comparable or better performance than the human or the hand-crafted policy.