An Entropy Minimization Approach to Dialogue Segmentation

Abstract

This paper introduces an approach to segmentation of short dialogue fragments, based on the entropy of linguistic cues. Starting from the assumption that a dialogue fragment consists of two non-overlapping segments, the segment boundary is determined to minimize the maximum interactional entropy amongst the segments. The approach is evaluated on a corpus of 4500 artificially generated two-segment dialogues, each of which containing from 8 to 12 dialogue acts. In 29.20 percent of the dialogues, the detected segment boundary coincides with the actual segment boundary, and in 69.27 percent of the dialogues, the detected segment boundary either coincides with the actual boundary or is immediately preceded or succeeded by it.