Modeling response relevance using dialog act and multi-modal utterance-design features

In this study, we focus on response relevance in conversation—specifically, the extent to which a speaker's utterance contributes to the relevance of the recipient's response. Using a corpus of casual conversations, we construct statistical models to predict whether an utterance will elicit a response from recipients, based on dialog act and utterance-design features. The utterance-design features include linguistic characteristics such as word count and part-of-speech tags, acoustic properties including the fundamental frequency and speech rate of accentual phrases, and the speaker's gaze behavior during the utterance as a non-linguistic feature. The results indicate a strong effect of dialog acts: utterances belonging to the high-response group, including various types of questions, are more likely to receive a response than those in the low-response group, such as statements and self-directed speech. Furthermore, the findings reveal the influence of utterance-design features, showing trends in response probability associated with specificlinguistic attributes and correlations with certain acoustic features. Finally, we present an illustrative example that goes beyond the primary findings and discuss possible extensions to the response relevance model.