RepSum: Unsupervised Dialogue Summarization based on Replacement Strategy

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 6042–6051 August 1–6, 2021. ©2021 Association for Computational Linguistics 6042 RepSum: Unsupervised Dialogue Summarization based on Replacement Strategy Xiyan Fu1, Yating Zhang2, Tianyi Wang2 Xiaozhong Liu3, Changlong Sun2 and Zhenglu Yang1 1 Nankai University, Tianjin, China 2 Alibaba Group, Hangzhou, China 3 Indiana University, Bloomington, USA fuxiyan@mail.nankai.edu.cn, {ranran.zyt,will.wty}@alibaba-inc.com liu237@indiana.edu, changlong.scl@taobao.com, yangzl@nankai.edu.cn Abstract In the field of dialogue summarization, due to the lack of training data, it is often difficult for supervised summary generation methods to learn vital information from dialogue context. Several works on unsupervised summarization for document by leveraging semantic informa- tion solely or auto-encoder strategy (i.e., sen- tence compression), they however cannot be adapted to the dialogue scene due to the lim- ited words in utterances and huge gap between the dialogue and its summary. In this study, we propose a novel unsupervised strategy to address this challenge, which roots from the hypothetical foundation that a superior sum- mary approximates a replacement of the orig- inal dialogue, and they are roughly equivalent for auxiliary (self-supervised) tasks, e.g., dia- logue generation. The proposed strategy Rep- Sum is applied to generate both extractive and abstractive summary with the guidance of the followed nth utterance generation and classi- fication tasks. Extensive experiments on vari- ous datasets demonstrate the superiority of the proposed model compared with other unsuper- vised methods. 1 Introduction Dialogue summarization distills key information from a dialogue context and synopsizes it into a concise summary. As a novel topic of critical im- portance, it offers powerful potentials for a number of scenarios, e.g, the court debate in civil trial, the customer service calls arisen from agent(s) and cus- tomer, the business meeting engaged with multi- members. It also assists users in quick access and consumes the essential content in the dialogue. Major attempts on dialogue summarization are template-based (Wang and Cardie, 2013; Oya et al., 2014) in the primitive stage by extracting key in- formation and filling it into the learned templates. However, these template-based techniques limit the Figure 1: A summary is generated from the input di- alogue firstly, and then the original dialogue and its corresponding summary are exploited for nth utterance prediction, respectively. J is the ground truth, and J in different colors are the decoded utterances based on the original dialogue and the generated summary respec- tively. The difference between decoded J is employed for optimization of summary generation. The motiva- tion is that a superior summary approximates a replace- ment of the original dialogue, and they are roughly equivalent for auxiliary tasks. scope of their applications and cannot be adapted to a wider range of conversational data since their input structure is predefined and the learned tem- plates are domain-specific. Later, various works ex- plore the assistance from labeled auxiliary informa- tion for summary generation, by leveraging either dialogue act (Goo and Chen, 2018), or key point sequence (Liu et al., …

Content truncated to 500 words