Research And Implementation Of Dialogue Management System Based On Pomdp

With the development of science, people can directly interact with all kinds of computing devices to have more access to the service which might be achieved through interaction with only human, like intelligent advisory services and getting a ticket reservation under the help of cus-tomer service staff. In the dialogue system, Dialogue Management(DM) is the core that used to control the interaction between the system and us-ers. Finite-state machine, slot filling, information status update and the probabilistic model are the four existing methods of DM technology.The paper studies dialogue management technology based on Par-tially Observable Markov Decision Process(POMDP), including the two following aspects. First, the paper proposes a state transition probability estimation method based on Maximum Entropy, in which treats state transition as a classification task that can solved zero probability situation. The experiment results show that the model using maximum entropy for state transition probability estimation can achieve greater expectation re-wards. Second, the paper puts forward a correction method for observa-tion probability using information entropy, in which makes the belief distribution could describe the surrounding information more precisely by introducing information entropy. The experiment results show that POMDP model with corrected observation probability on account of information entropy could make more natural decision.Finally, this paper implements a human-computer dialogue system based on POMDP management model combining those methods above, which can be applied to teach robot object-attribute knowledge.