Dialogue Acts: One or More Dimensions?

This report surveys the main theories of dialogue and communication that have been used to devise dialogue act tagsets, distinguishing theories that deal with a specific level of communication from theories that integrate several levels. The report proceeds to analyse four dialogue act tagsets that have been used to annotate large scale dialogue corpora (damsl, swbd-damsl, icsi-mrda and maltus), paying particular attention to the difference between the range of possible combinations of tags and the range of combinations that do occur in the hand-labelled data. The problem of the dimensionality of tagsets is then introduced, and discussed in relation with other factors that influence the performance of automatic dialogue act taggers. After a brief discussion of empirical arguments relevant to the dimensionality problem, derived from human and automatic labelling experiments, the report proposes a synthesis of guidelines for the definition of dialogue act tagsets, based on multi-dimensional theoretical inspiration, and cross-dimensional constraints and the notion of dominant utterance-function, in order to reduce the search for automatic dialogue act taggers. As a compromise between the divergent needs for theoretical grounding and tagging accuracy, the Dominant Function Approximation proposes that automatic dialogue act taggers could focus initially on finding the main dialogue function of each utterance. This approximation is shown to be empirically acceptable and to have significant practical relevance.