Dialogue Segmentation based on Dynamic Context Coherence
Hengfeng Pu, Liqing Wang · 2023
Dialogue segmentation plays a crucial role in various dialogue modeling tasks. Semantic coherence-based methods are a major trend, but previous methods have tended to take fixed-size units in assessing semantic coherence, which may suffer from semantic loss or noise introduction, thus obtaining inaccurate coherence scores. In this paper, we propose a simple but effective approach for dynamically assessing semantic coherence, which can improve the accuracy of coherence scoring by dynamically expanding the assessment unit from a single utterance to a semantically related context. Experimental results on three benchmark datasets in English and Chinese demonstrate that our proposal can achieve competitive performance over strong baselines.