Conversational system for information navigation based on POMDP with user focus tracking

摘要: HighlightsWe address a spoken dialogue system which conducts information navigation.We formulate the problem of management as module selection with POMDP.The reward function POMDP is defined by quality interaction.The tracks user's focus attention to make appropriate actions.The proposed model outperformed conventional systems without information. We navigation in style small talk. The uses Web news articles an source, and user can receive about day through interaction. goal procedure this kind are not well defined. An empirical approach based on partially observable Markov decision process (POMDP) has recently been widely used for management, but it assumes definite task slots, does hold our application system. In work, we modules optimize tracking state attention. POMDP-based manager receives intention that classified language understanding (SLU) component logistic regression (LR). also detected SLU conditional random fields (CRFs). These states selecting policy function, optimized reinforcement learning. interaction encourage long users. responds queries similarity predicate-argument (P-A) structures automatically from domain corpus. It allows flexible response generation even if cannot find exact matching query. proactively presents following retrieving article measure any utterance. Experimental evaluations real sessions demonstrate rule-based terms action selection. Effect detection framework confirmed.