This paper presents a generic dialogue state tracker that maintains beliefs over user goals based on a few simple domain-independent rules applying to observed system actions and partially observable user acts, without the support of any knowledge obtained from external resources, but only using basic probability operations. The core insight is to maximise the amount of information directly gainable from an error-prone dialogue problem itself, so as to better lower-bound one's expectations on the performance of more advanced statistical techniques for the task. The proposed method is evaluated in the Dialog State Tracking Challenge, where it achieves comparable performance in hypothesis accuracy to machine learning based systems. Consequently, with respect to different scenarios for the belief tracking problem, the potential superiority and weakness of machine learning approaches in general are investigated.
|Title of host publication||Proceedings of SIGDIAL 2013|
|Subtitle of host publication||14th Annual Meeting of the Special Interest Group on Discourse and Dialogue|
|Publisher||Association for Computational Linguistics|
|Publication status||Published - Aug 2013|