Summarizing Dialogues with Negative Cues
4695Authors
nan
Abstract
AbstractAbstractive dialogue summarization aims to convert a long dialogue content into its short form where the salient information is preserved while the redundant pieces are ignored. Different from the well-structured text, such as news and scientific articles, dialogues often consist of utterances coming from two or more interlocutors, where the conversations are often informal, verbose, and repetitive, sprinkled with false-starts, backchanneling, reconfirmations, hesitations, speaker interruptions and the salient information is often scattered across the whole chat. The above properties of conversations make it difficult to directly concentrate on scattered outstanding utterances and thus present new challenges of summarizing dialogues. In this work, rather than directly forcing a summarization system to merely pay more attention to the salient pieces, we propose to explicitly have the model perceive the redundant parts of an input dialogue history during the training phase. To be specific, we design two strategies to construct examples without salient pieces as negative cues. Then, the sequence-to-sequence likelihood loss is cooperated with the unlikelihood objective to drive the model to focus less on the unimportant information and also pay more attention to the salient pieces. Extensive experiments on the benchmark dataset demonstrate that our simple method significantly outperforms the baselines with regard to both semantic matching and factual consistent based metrics. The human evaluation also proves the performance gains.
Keywords
Full Content
Proceedings of the 29th International Conference on Computational Linguistics, pages 6050–6056 October 12–17, 2022. 6050 Summarizing Dialogues with Negative Cues Junpeng Liu†∗ Beijing University of Posts and Telecommunications, Beijing, China jeepliu@bupt.edu.cn Yanyan Zou† JD.com, Beijing, China zouyanyan6@jd.com Yuxuan Xi∗ Peking University, Beijing, China xiyuxuan@pku.edu.cn Shengjie Li and Mian Ma and Zhuoye Ding JD.com, Beijing, China {lishengjie1,mamian,dingzhuoye}@jd.com Abstract Abstractive dialogue summarization aims to convert a long dialogue content into its short form where the salient information is preserved, while the redundant pieces are ignored. Dif- ferent from the well-structured text, such as news and scientific articles, dialogues often consist of utterances coming from two or more interlocutors, where the conversations are of- ten informal, verbose, and repetitive, sprinkled with false-starts, backchanneling, reconfirma- tions, hesitations, speaker interruptions and the salient information is often scattered across the whole chat. The above properties of conversa- tions make it difficult to directly concentrate on scattered outstanding utterances and thus present new challenges of summarizing dia- logues. In this work, we propose to explicitly have the model perceive the redundant parts of an input dialogue history, so that the model is able to pay more attention to the salient pieces. To be specific, we design two strategies to con- struct examples without salient pieces as nega- tive cues. Then, the sequence-to-sequence like- lihood loss is cooperated with the unlikelihood objective to drive the model focus less on the unimportant information as well as pay more attention to the salient pieces. Extensive exper- iments on the benchmark dataset demonstrate that our simple method outperforms baselines with regard to both semantic matching and fac- tual consistent based metrics. The human eval- uation also proves the performance gains led by our approach. 1 Introduction Online conversations have become an indispens- able manner of communication in our daily work and life, where people tend to exchange their ideas, share information, consult via textual messages. Especially in the era of information explosion, it is ∗Work done during internship at JD.com. †The first two authors made equal contributions. Corre- sponding to Yanyan Zou. Molly: Guys, do you think it’s a very bad idea to go to Sweden for a week in January? Margaret: We bought some cheep tickets half a year ago and now we’re hesitating. Peter: Haha, no but it will be just dark and cold. Margaret: Rainy? Kal: Possibly. But if you stay in Stockholm, there are always nice things to do. Museums, bars etc Kal: Not so much nature though which is truly stunning around Stockholm. Margaret: Yes, but it’s January, one would have to go to Argentina to enjoy nature. Kal: Exactly. Peter: Visit the Vasa Museum, it’s really fun. Molly: We will:) Thanks :) Peter: Enjoy! Summary: Molly and Margaret are going to Sweden in Jan- uary. Kal and Peter advise them to stay in Stock- holm and visit Vasa Museum. Figure 1: An example of dialogue with its summary. Green: nouns, italic: salient utterances. much more challenging and time-consuming to go through all the conversation content …
Content truncated to 500 words
Labeled Text
AbstractAbstractive dialogue summarization aims to convert a long dialogue content into its short form where the salient information is preserved while the redundant pieces are ignored. Different from the well-structured text, such as news and scientific articles, dialogues often consist of utterances coming from two or more interlocutors, where the conversations are often informal, verbose, and repetitive, sprinkled with false-starts, backchanneling, reconfirmations, hesitations, speaker interruptions and the salient information is often scattered across the whole chat. The above properties of conversations make it difficult to directly concentrate on scattered outstanding utterances and thus present new challenges of summarizing dialogues. In this work, rather than directly forcing a summarization system to merely pay more attention to the salient pieces, we propose to explicitly have the model perceive the redundant parts of an input dialogue history during the training phase. To be specific, we design two strategies to construct examples without salient pieces as negative cues. Then, the sequence-to-sequence likelihood loss is cooperated with the unlikelihood objective to drive the model to focus less on the unimportant information and also pay more attention to the salient pieces. Extensive experiments on the benchmark dataset demonstrate that our simple method significantly outperforms the baselines with regard to both semantic matching and factual consistent based metrics. The human evaluation also proves the performance gains.
Publication Details
- Paper ID:
- 4695
- Year:
- 2022
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan