Using Out-of-Domain Data for Lexical Addressee Detection in Human-Human-Computer Dialog

Proceedings of NAACL-HLT 2013, pages 221–229, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Using Out-of-Domain Data for Lexical Addressee Detection in Human-Human-Computer Dialog Heeyoung Lee1∗ Andreas Stolcke2 Elizabeth Shriberg2 1Dept. of Electrical Engineering, Stanford University, Stanford, California, USA 2Microsoft Research, Mountain View, California, USA heeyoung@stanford.edu, {anstolck,elshribe}@microsoft.com Abstract Addressee detection (AD) is an important problem for dialog systems in human-human- computer scenarios (contexts involving mul- tiple people and a system) because system- directed speech must be distinguished from human-directed speech. Recent work on AD (Shriberg et al., 2012) showed good results using prosodic and lexical features trained on in-domain data. In-domain data, however, is expensive to collect for each new domain. In this study we focus on lexical models and in- vestigate how well out-of-domain data (either outside the domain, or from single-user sce- narios) can fill in for matched in-domain data. We find that human-addressed speech can be modeled using out-of-domain conversational speech transcripts, and that human-computer utterances can be modeled using single-user data: the resulting AD system outperforms a system trained only on matched in-domain data. Further gains (up to a 4% reduction in equal error rate) are obtained when in-domain and out-of-domain models are interpolated. Finally, we examine which parts of an utter- ance are most useful. We find that the first 1.5 seconds of an utterance contain most of the lexical information for AD, and analyze which lexical items convey this. Overall, we conclude that the H-H-C scenario can be ap- proximated by combining data from H-C and H-H scenarios only. ∗Work done while first author was an intern with Microsoft. 1 Introduction Before a spoken dialog system can recognize and in- terpret a user’s speech, it should ideally determine if speech was even meant to be interpreted by the system. We refer to this task as addressee detec- tion (AD). AD is often overlooked, especially in tra- ditional single-user scenarios, because with the ex- ception of self-talk, side-talk or background speech, the majority of speech is usually system-directed. As dialog systems expand to more natural contexts and multiperson environments, however, AD can be- come a crucial part of the system’s operational re- quirements. This is particularly true for systems in which explicit system addressing (e.g., push-to-talk or required keyword addressing) is undesirable. Past research on addressee detection has focused on human-human (H-H) settings, such as meetings, sometimes with multimodal cues (op den Akker and Traum, 2009). Early systems relied primarily on re- jection of H-H utterances either because they could not be interpreted (Paek et al., 2000), or because they yielded low speech recognition confidence (Dowd- ing et al., 2006). Some systems combine gaze with lexical and syntactic cues to detect H-H speech (Katzenmaier et al., 2004). Others use relatively simple prosodic features based on pitch and energy in addition to those derived from automatic speech recognition (ASR) (Reich et al., 2011). With some exceptions (Bohus and Horvitz, 2011; Shriberg et al., 2012), relatively little work has looked at the human-human-computer (H-H-C) sce- …

Content truncated to 500 words