Date of Award


Publication Type

Master Thesis

Degree Name



Computer Science


Yuan, Xiaobu




This thesis applies the theory of history information space for a thorough study of dialogue management in major approaches, ranging from the classical approach based upon finite state machine to the most recent approach using partially observable Markov decision process (PODMP). While most of the approaches use various techniques to estimate system state, the POMDP-based approach avoids state estimation and uses belief state for decision making. In addition, it provides a mechanism to model uncertainty and allows for errorrecovery. PODMP-based dialogue management demonstrates undeniable advantages in the handling of input uncertainty over all the other approaches.

However, applying Markovian over the belief-state space in the current POMDP models causes significant loss of valuable information in dialogue history, leading to untruthful recognition of user's intention. To improve the performance of POMDP-based dialogue management this thesis introduces belief history into the planning process, and uses not only the current but also the previous belief state for the determination of actions. In the new approach, all changes of belief state require a validation with domain constraints, and an invalid change results in a modification to the actions provided by the POMDP solver. Experiments show that this new approach is able to handle uncertainty caused by user's lack of domain knowledge and practical constraints, thus becoming more accurate in intention recognition.