Cue-based dialogue act classification

Cue-Based Dialogue Act Classification Nick Webb Supervisors: Yorick Wilks & Mark Hepple In this thesis, we will address three research questions relating to the discovery and use of cue phrases for Dialogue Act classification. Cue phrases are single words, or combinations of words in phrases, that can serve as reliable indicators of some discourse function. In our case, we are looking to automatically discover cue phrases in corpora that are useful for the detection of Dialogue Acts (da). Dialogue Acts are labels attached to utterances in dialogue that serve to concisely characterise a speaker’s intention in producing a particular utterance, a notion that most major theories of dialogue take as central. Our first research question is whether or not we can extract cue phrases automatically from a corpus. We apply a method of cue extraction to the switchboard corpus of annotated human-human dialogues, and experiment with thresholds to identify cue phrases. To determine if these automatically extracted cue phrases are reliable indicators of das, we created a novel, cue-based da classification model. In this model, our cue phrases are exploited directly, by determining if they appear in unseen dialogue utterances. This forms our second research question, and we extensively explore