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

Addressee detection (AD) is an important problem for dialog systems in human-humancomputer scenarios (contexts involving multiple people and a system) because systemdirected 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 investigate how well out-of-domain data (either outside the domain, or from single-user scenarios) 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 utterance 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 approximated by combining data from H-C and H-H scenarios only.