Unsupervised Conversation Disentanglement through Co-Training

• DOI: 10.18653/v1/2021.emnlp-main.181
• Publication Date: 2021-12-17T03:56:42Z
• Authors: Hui Liu; Zhan Shi; Xiaodan Zhu
• Keywords: Training set

Conversation disentanglement aims to separate intermingled messages into detached sessions, which is a fundamental task in understanding multi-party conversations. Existing work on conversation relies heavily upon human-annotated datasets, expensive obtain practice. In this work, we explore training model without referencing any human annotations. Our method built the deep co-training algorithm, consists of two neural networks: message-pair classifier and session classifier. The former responsible retrieving local relations between while latter categorizes message by capturing context-aware information. Both networks are initialized respectively with pseudo data from unannotated corpus. During process, use as reinforcement learning component learn assigning policy maximizing rewards given For classifier, enrich its pairs high confidence disentangled sessions predicted Experimental results large Movie Dialogue Dataset demonstrate that our proposed approach achieves competitive performance compared previous supervised methods. Further experiments show conversations can promote downstream response selection.