Improving The Probabilistic Framework for Representing Dialogue Systems with User Response Model

A probabilistic framework for goal-driven spoken dialogue systems (SDSs) has been proposed by us in a previous work. In the framework, a target distribution, instead of the frame structure, is used to represent the dialogue state at each turn. The target based state tracking algorithm enables the system to handle uncertainties in the dialogue. By summarizing the target-based state, information from the back-end database can be exploited to develop efficient dialogue strategies. To extend our probabilistic framework and adapt our approach to real application scenarios, a user response model is investigated and integrated into the probabilistic framework to enhance the dialogue policy in this paper. Experiments in both ideal setting and real user test setting are conducted to test the enhanced dialogue policy. The results show that despite an unavoidable mismatch between the user response model based on prior knowledge and real users' behaviors in the experiment, the enhanced dialogue policy works robustly and efficiently. The results further demonstrate that the probabilistic framework is quite flexible and amenable to the integration of additional factors and models of real-world dialogue problems.