Modeling Response Relevance using Dialog Act and Utterance-Design Features: A Corpus-Based Analysis

In this paper, we focus on response relevance in conversation, that is, to what degree the speaker's utterance can make the re-cipient's response relevant. Specifically, using a corpus of ca-sual conversations, we construct statistical models that predict whether or not an utterance is responded by recipients based on the dialog act as well as the linguistic and acoustic features of the utterance. The results indicate a marked effect of dialog acts—high response probabilities for a family of questions—as well as effects of utterance-design features—trends in the response probability associated with some linguistic features and correlated with some acoustic features. We illustrate an example beyond our findings and discuss a possible extension of the response relevance model.