A method for the approximation of incremental understanding of explicit utterance meaning using predictive models in finite domains

Proceedings of NAACL-HLT 2013, pages 1092–1099, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics A method for the approximation of incremental understanding of explicit utterance meaning using predictive models in finite domains David DeVault and David Traum Institute for Creative Technologies, University of Southern California, 12015 Waterfront Dr., Playa Vista, CA 90094 USA {devault,traum}@ict.usc.edu Abstract This paper explores the relationship between explicit and predictive models of incremental speech under- standing in a dialogue system that supports a finite set of user utterance meanings. We present a method that enables the approximation of explicit under- standing using information implicit in a predictive understanding model for the same domain. We show promising performance for this method in a corpus evaluation, and discuss its practical application and annotation costs in relation to some alternative ap- proaches. 1 Introduction In recent years, there has been a growing interest among researchers in methods for incremental natural language understanding (NLU) for spoken dialogue systems; see e.g. (Skantze and Schlangen, 2009; Sagae et al., 2009; Schlangen et al., 2009; Heintze et al., 2010; DeVault et al., 2011a; Selfridge et al., 2012). This work has gen- erally been motivated by a desire to make dialogue sys- tems more efficient and more natural, by enabling them to provide lower latency responses (Skantze and Schlangen, 2009), human-like feedback such as backchannels that in- dicate how well the system is understanding user speech (DeVault et al., 2011b; Traum et al., 2012), and more in- teractive response capabilities such as collaborative com- pletions of user utterances (DeVault et al., 2011a), more adaptive handling of interruptions (Buschmeier et al., 2012), and others. This paper builds on techniques developed in previous work that has adopted a predictive approach to incremen- tal NLU (DeVault et al., 2011a). On this approach, at specific moments while a user’s speech is in progress, an attempt is made to predict what the full meaning of the complete user utterance will be. Predictive models can be contrasted with explicit approaches to incremen- tal NLU. We use the term explicit understanding to refer to approaches that attempt to determine the meaning that has been expressed explicitly in the user’s partial utter- ance so far (without predicting further aspects of mean- ing to come). Explicit understanding of partial utterances can be implemented using statistical classification or se- quential tagging models (Heintze et al., 2010). Both predictive and explicit incremental NLU capabil- ities can be valuable in a dialogue system. Prediction can support specific response capabilities, such as sys- tem completion of user utterances (DeVault et al., 2011a) and reduced response latency.1 However, explicit models support additional and complementary capabilities. For instance, depending on the application domain (Heintze et al., 2010) and on the individual utterance (DeVault et al., 2011b), it may be difficult for a system to predict a user’s impending meaning with confidence. Neverthe- less, it may often be possible for systems to determine the meaning of what a user has said so far, and to take action based …

Content truncated to 500 words