Multi-View Sequence-to-Sequence Models with Conversational Structure for Abstractive Dialogue Summarization

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4106–4118, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4106 Multi-View Sequence-to-Sequence Models with Conversational Structure for Abstractive Dialogue Summarization Jiaao Chen School of Interactive Computing Georgia Institute of Technology jiaaochen@gatech.edu Diyi Yang School of Interactive Computing Georgia Institute of Technology dyang888@gatech.edu Abstract Text summarization is one of the most chal- lenging and interesting problems in NLP. Although much attention has been paid to summarizing structured text like news re- ports or encyclopedia articles, summarizing conversations—an essential part of human- human/machine interaction where most impor- tant pieces of information are scattered across various utterances of different speakers— remains relatively under-investigated. This work proposes a multi-view sequence-to- sequence model by first extracting conversa- tional structures of unstructured daily chats from different views to represent conversa- tions and then utilizing a multi-view de- coder to incorporate different views to gen- erate dialogue summaries. Experiments on a large-scale dialogue summarization corpus demonstrated that our methods significantly outperformed previous state-of-the-art mod- els via both automatic evaluations and hu- man judgment. We also discussed specific challenges that current approaches faced with this task. We have publicly released our code at https://github.com/GT-SALT/ Multi-View-Seq2Seq. 1 Introduction We live in an information age where communi- cations between human and human/machine are increasing exponentially in the form of textual di- alogues between users and users-agents (Kester, 2004). It is challenging and time-consuming to review all the content before starting any conversa- tions especially when the chatting history becomes very long (Gao et al., 2020). How to process and or- ganize those interaction activities into concise and structured data, i.e. conversation summarization, becomes technically and socially important. Most existing research efforts on text summa- rization have been focused on single-speaker doc- uments like news reports (Nallapati et al., 2016; See et al., 2017), scientific publications (Nikolov et al., 2018) or encyclopedia articles (Liu* et al., 2018), where structured text is usually used to elab- orate a core idea in the third-person point of view, and the information flow is very clear through para- graphs or sections. Different from these structured documents, conversations are often informal, ver- bose and repetitive, sprinkled with false-starts, back channeling, reconfirmations, hesitations, speaker interruptions (Sacks et al., 1978) and the salient information is scattered in the whole chat, mak- ing current summarization models hard to focus on many informative utterances. Take the conversation in Table 1 as an example, turns, informal words, ab- breviations, and emoticons all introduce new forms of challenges to the task of summarization. This calls for the design and development of new meth- ods for dialogue summarization instead of directly applying current document summarization models. There has been some recent research on conver- sation summarization such as directly deploying existing document summarization models (Gliwa et al., 2019) and exploring multi-sentence compres- sion (Shang et al., 2018), however, most of them haven’t utilized specific conversational structures, which refer to the way utterances are organized in order to make the conversation …

Content truncated to 500 words