Relating the Semantics of Dialogue Acts to Linguistic Properties: A Machine Learning Perspective through Lexical Cues

This paper describes a corpus-based investigation of dialogue acts. In particular, it attempts to answer questions about the empirical distribution of dialogue acts and to what extent dialogue acts can be automatically predicted from their lexical features. The Switchboard Dialogue Act Corpus is adopted and the SWBD-DAMSL tags used for automatic prediction. We show that 60-70% of the dialogue acts can be predicted from lexical features alone depending on different levels of granularity. We also present a mapping from SWBD-DAMSL tags to the tags of the new ISO standard for dialogue act annotation, as part of an ongoing investigation into the relationship between the structure and granularity of the tag set and classification accuracy. The paper concludes with discussions and suggestions for future work.