CSDS: A Fine-Grained Chinese Dataset for Customer Service Dialogue Summarization

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 4436–4451 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 4436 CSDS: A Fine-Grained Chinese Dataset for Customer Service Dialogue Summarization Haitao Lin1,2, Liqun Ma1,2, Junnan Zhu1,2, Lu Xiang1,2, Yu Zhou1,3∗, Jiajun Zhang1,2, Chengqing Zong1,2 1 National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 3 Fanyu AI Laboratory, Zhongke Fanyu Technology Co., Ltd, Beijing, China {haitao.lin, junnan.zhu, lu.xiang, yzhou, jjzhang, cqzong}@nlpr.ia.ac.cn, maliqun2020@ia.ac.cn Abstract Dialogue summarization has drawn much at- tention recently. Especially in the customer service domain, agents could use dialogue summaries to help boost their works by quickly knowing customer’s issues and ser- vice progress. These applications require sum- maries to contain the perspective of a sin- gle speaker and have a clear topic flow struc- ture, while neither are available in existing datasets. Therefore, in this paper, we intro- duce a novel Chinese dataset for Customer Ser- vice Dialogue Summarization (CSDS). CSDS improves the abstractive summaries in two as- pects: (1) In addition to the overall summary for the whole dialogue, role-oriented sum- maries are also provided to acquire different speakers’ viewpoints. (2) All the summaries sum up each topic separately, thus containing the topic-level structure of the dialogue. We define tasks in CSDS as generating the over- all summary and different role-oriented sum- maries for a given dialogue. Next, we compare various summarization methods on CSDS, and experiment results show that existing methods are prone to generate redundant and incoher- ent summaries. Besides, the performance be- comes much worse when analyzing the perfor- mance on role-oriented summaries and topic structures. We hope that this study could benchmark Chinese dialogue summarization and benefit further studies. 1 Introduction Text summarization aims to compress a long in- put text and generate a condensed summary (Zong et al., 2021). It can help people capture the gist of a long document quickly. Traditional summariza- tion tasks mainly focus on news reports (Nallapati et al., 2016; Narayan et al., 2018; Zhu et al., 2018). However, as the communication tools become con- venient, enormous information is presented in a conversational format, such as meeting records, ∗Corresponding author. daily chatting, and customer service logs. These dialogues usually cost more time to read since they are longer and have more complicated structures. Thus, summarizing information from dialogue be- comes essential in practical use. Compared with news and documents, dialogues have two main unique features. First, dialogues have multiple speakers, and each of them has dif- ferent viewpoints. In some cases, we only focus on the main viewpoint of one participant. A role- oriented summary will help achieve this goal. Sec- ond, a dialogue often has multiple topics, and each topic concerns a different issue. Since the summary expresses the primary information of the dialogue briefly and clearly, it should contain the topic flow of the dialogue by summing up each topic sepa- rately and be organized in a …

Content truncated to 500 words