Response Generation Based on Hierarchical Semantic Structure with POMDP Re-ranking for Conversational Dialogue Systems

Conversational spoken dialogue systems can assist individuals to communicate with machine to obtain relevant information to their problems efficiently and effectively. By referring to relevant response, individuals can understand how to interact with an intelligent system according to recommendations of dialogue systems. This work presents a response generation based on hierarchical semantic structure with POMDP Re-ranking for conversational dialogue systems to achieve this aim. The hierarchical semantic structure incorporates the historical information according to dialogue discourse to keep more than one possible values for each slot. According to the status of concept graph, the candidate sentences are generated. The near optimal response selected by POMDP Re-ranking strategy to achieve human-like communication. The MOS and recall/precision rates are considered as the criterion for evaluations. Finally, the proposed method is adopted for dialogue system in travel domain, and indicates its superiority in information retrieval over traditional approaches.