User Goal Change Model for Spoken Dialog State Tracking
Yi Ma · 2013
In this paper, a Maximum Entropy Markov Model (MEMM) for dialog state tracking is proposed to efficiently handle user goal evolvement in two steps. The system first predicts the occurrence of a user goal change based on linguistic features and dialog context for each dialog turn, and then the proposed model could utilize this user goal change information to infer the most probable dialog state sequence which underlies the evolvement of user goal during the dialog. It is believed that with the suggested various domain independent feature functions, the proposed model could better exploit not only the intra-dependencies within long ASR N-best lists but also the inter-dependencies of the observations across dialog turns, which leads to more efficient and accurate dialog state inference. best list when the top ASR hypothesis is incorrect. Furthermore, reasoning over different ASR N-best lists is also difficult since it is hard to decide when to detect commonality (when user repeats) and when to look for differences (when user changes her or his mind) among multiple ASR N-best lists. Another challenge is how to handle more complex user actions such as negotiating alternative choices or seeking out other potential solutions when interacting with the system. This proposal presents a probabilistic framework for modeling the evolvement of user goal during the dialog (focusing on the shaded component Dialog State Tracking in Figure 1 that shows a typical diagram for a spoken dialog system), which aims to endow the system with the ability to model natural negotiation strategies, in the hope of leading to more accurate and efficient dialog state tracking performance. 1