Automatic Dialogue Segmentation Using Discourse Chunking

This study explains a method for arranging dialogues into discourse chunks. Discourse chunking is a simple way to segment dialogues according to how dialogue participants raise topics and negotiate them. It has been used successfully to improve performance in dialogue act tagging, a classification task where utterances are classified according to the intentions of the speaker. Earlier work showed that discourse chunking improved performance on the dialogue act tagging task when the chunk information was correct and hand-coded. The goal for the current study is two-fold: first, to investigate how accurately dialogues can be marked with discourse chunks automatically, and second, to determine the effect of the discourse chunk information on dialogue act tagging. We present evidence which shows that discourse chunking improves the performance of the dialogue act tagger, even when the chunk information is imperfect. The dialogue act tagger for this study uses case-based reasoning, a machine learning technique which classifies utterances by comparing their similarity to examples from a knowledge base.