Relationship between dialogue acts and hot spots in meetings

Abstract

We examine the relationship between hot spots (annotated in terms of involvement) and dialogue acts (DAs, annotated in an independent effort) in roughly 32 hours of speech data from naturally-occurring meetings. Results reveal that four independently-motivated involvement categories (non-involved, disagreeing, amused, and other) show statistically significant associations with particular DAs. Further examination shows that involvement is associated with contextual features (such as the speaker or type of meeting), as well as with lexical features (such as utterance length and perplexity). Finally, we found (surprisingly) that perplexities are similar for involved and non-involved utterances. This suggests that it may not be the amount of propositional content, but rather participants' attitudes toward that content, that differentiates hot spots from other regions in a meeting. Overall, these specific correlations, and their relationships to other features, such as perplexity, could provide useful information for the automatic archiving and browsing of natural meetings.