Unsupervised Conversation Disentanglement with GCN Clustering

Conversation disentanglement is a fundamental task for understanding multi-participant and multi-round conversations. The primary objective of conversation disentanglement is to separate entangled utterances into different sessions, in which all utterances are discussing the same topic. Supervised methods for conversation disentanglement have been widely used, yet obtaining the necessary human-annotated datasets can be costly and time-consuming. Motivated by this, an unsupervised conversation disentanglement method based on GCN clustering is proposed in this paper, which is divided into two steps, coarse-grained segmentation and deep clustering based on GCN. Our method explores structural and semantic information between utterances, then fuses them to enhance the performance of conversation disentanglement. The experiments that are evaluated on two datasets from real chat groups show that our method has a better performance than strong baselines.