Uncertainty in Dialogue Systems
Another research goal is to improve how systems respond to uncertainty about a user’s meaning, by enabling them to use probabilistic inference and more flexible follow-up and clarification capabilities to resolve uncertainty over time.
Many dialogue systems track the status of an ongoing conversation using a single, non-probabilistic information state. This can make it difficult for systems to tolerate substantial uncertainty about a user’s meaning. In practice, when ambiguities arise, most systems either assume the correctness of the highest-ranked interpretation (which can lead to misunderstandings), or else require immediate clarification from the user (which can be tiresome and unnatural). In my Ph.D. research, I developed an improved methodology for modeling the evolving mental state of an uncertain dialogue agent.
Selected Publications
-
Learning to Interpret Utterances Using Dialogue History David DeVault and Matthew Stone, The 12th Conference of the European Chapter of the Association for Computational Linguistics (EACL-09), Athens, Greece
-
Contribution Tracking: Participating in Task-Oriented Dialogue under Uncertainty David DeVault, Ph.D. Dissertation, Department of Computer Science, Rutgers, The State University of New Jersey, New Brunswick, NJ, October, 2008.
Committee: Dr. Matthew Stone, Dr. Chung-chieh Shan, Dr. Michael Littman, Dr. David Traum
-
Managing ambiguities across utterances in dialogue David DeVault and Matthew Stone. The 2007 Workshop on the Semantics and Pragmatics of Dialogue (DECALOG 2007), University of Trento, Italy, May, 2007.