Learning to Interpret Utterances Using Dialogue History

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 184–192, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Learning to Interpret Utterances Using Dialogue History David DeVault Institute for Creative Technologies University of Southern California Marina del Rey, CA 90292 devault@ict.usc.edu Matthew Stone Department of Computer Science Rutgers University Piscataway, NJ 08845-8019 Matthew.Stone@rutgers.edu Abstract We describe a methodology for learning a disambiguation model for deep pragmatic interpretations in the context of situated task-oriented dialogue. The system accu- mulates training examples for ambiguity resolution by tracking the fates of alter- native interpretations across dialogue, in- cluding subsequent clarificatory episodes initiated by the system itself. We illus- trate with a case study building maxi- mum entropy models over abductive in- terpretations in a referential communica- tion task. The resulting model correctly re- solves 81% of ambiguities left unresolved by an initial handcrafted baseline. A key innovation is that our method draws exclu- sively on a system’s own skills and experi- ence and requires no human annotation. 1 Introduction In dialogue, the basic problem of interpretation is to identify the contribution a speaker is making to the conversation. There is much to recognize: the domain objects and properties the speaker is refer- ring to; the kind of action that the speaker is per- forming; the presuppositions and implicatures that relate that action to the ongoing task. Neverthe- less, since the seminal work of Hobbs et al. (1993), it has been possible to conceptualize pragmatic in- terpretation as a unified reasoning process that se- lects a representation of the speaker’s contribution that is most preferred according to a background model of how speakers tend to behave. In principle, the problem of pragmatic interpre- tation is qualitatively no different from the many problems that have been tackled successfully by data-driven models in NLP. However, while re- searchers have shown that it is sometimes possi- ble to annotate corpora that capture features of in- terpretation, to provide empirical support for the- ories, as in (Eugenio et al., 2000), or to build classifiers that assist in dialogue reasoning, as in (Jordan and Walker, 2005), it is rarely feasible to fully annotate the interpretations themselves. The distinctions that must be encoded are subtle, theoretically-loaded and task-specific—and they are not always signaled unambiguously by the speaker. See (Poesio and Vieira, 1998; Poesio and Artstein, 2005), for example, for an overview of problems of vagueness, underspecification and ambiguity in reference annotation. As an alternative to annotation, we argue here that dialogue systems can and should prepare their own training data by inference from under- specified models, which provide sets of candi- date meanings, and from skilled engagement with their interlocutors, who know which meanings are right. Our specific approach is based on contribu- tion tracking (DeVault, 2008), a framework which casts linguistic inference in situated, task-oriented dialogue in probabilistic terms. In contribution tracking, ambiguous utterances may result in alter- native possible contexts. As subsequent utterances are interpreted in those contexts, ambiguities …

Content truncated to 500 words