Unsupervised Conversation Disentanglement through Co-Training

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2345–2356 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2345 Unsupervised Conversation Disentanglement through Co-Training Hui Liu Zhan Shi Xiaodan Zhu Ingenuity Labs Research Institute & ECE, Queen’s University, Canada {hui.liu, z.shi, xiaodan.zhu}@queensu.ca Abstract Conversation disentanglement aims to sepa- rate intermingled messages into detached ses- sions, which is a fundamental task in un- derstanding multi-party conversations. Exist- ing work on conversation disentanglement re- lies heavily upon human-annotated datasets, which are expensive to obtain in practice. In this work, we explore to train a conversa- tion disentanglement model without referenc- ing any human annotations. Our method is built upon a deep co-training algorithm, which consists of two neural networks: a message- pair classifier and a session classifier. The for- mer is responsible for retrieving local relations between two messages while the latter cate- gorizes a message to a session by capturing context-aware information. Both networks are initialized respectively with pseudo data built from an unannotated corpus. During the deep co-training process, we use the session classi- fier as a reinforcement learning component to learn a session assigning policy by maximiz- ing the local rewards given by the message- pair classifier. For the message-pair classi- fier, we enrich its training data by retrieving message pairs with high confidence from the disentangled sessions predicted by the session classifier. Experimental results on the large Movie Dialogue Dataset demonstrate that our proposed approach achieves competitive per- formance compared to the previous supervised methods. Further experiments show that the predicted disentangled conversations can pro- mote the performance on the downstream task of multi-party response selection. 1 Introduction With the continuing growth of Internet and social media, online group chat channels, e.g., Slack1 and Whatsapp2, among many others, have become in- creasingly popular and played a significant social 1https://slack.com/ 2https://www.whatsapp.com/ Anyone finished the assignment? Not yet. Working on it When is the due date? Hmm… tonight? I’d like to get a new keyboard. Any suggestions? The one I have is pretty good. S1 S1 S1 S1 S2 S2 Session Message Figure 1: An example of conversation disentanglement with two sessions occurring at the same time. and economic role. Along with the convenience of instant communication brought by these applica- tions, the inherent property that multiple topics are often discussed in one channel hinders an efficient access to the conversational content. In the exam- ple shown in Figure 1, people or intelligent systems have to selectively read the messages related to the topics they are interested in from hundreds of mes- sages in the chat channel. With the goal of automatically grouping mes- sages with the same topic into one session, conver- sation disentanglement has proved to be a prereq- uisite for understanding multi-party conversations and solving the corresponding downstream tasks such as response selection (Elsner and Charniak, 2008; Lowe et al., 2017; Jia et al., 2020; Wang et al., 2020). Previous research on conversation disentan- glement can be roughly divided into two categories: (1) …

Content truncated to 500 words