Controllable Abstractive Dialogue Summarization with Sketch Supervision

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 5108–5122 August 1–6, 2021. ©2021 Association for Computational Linguistics 5108 Controllable Abstractive Dialogue Summarization with Sketch Supervision Chien-Sheng Wu∗1, Linqing Liu∗2, Wenhao Liu1, Pontus Stenetorp2, Caiming Xiong1 1 Salesforce Research 2 University College London {linqing.liu,p.stenetorp}@cs.ucl.ac.uk, {wu.jason, wenhao.liu, cxiong}@salesforce.com Abstract In this paper, we aim to improve abstractive di- alogue summarization quality and, at the same time, enable granularity control. Our model has two primary components and stages: 1) a two-stage generation strategy that generates a preliminary summary sketch serving as the basis for the final summary. This summary sketch provides a weakly supervised signal in the form of pseudo-labeled interrogative pro- noun categories and key phrases extracted us- ing a constituency parser. 2) A simple strategy to control the granularity of the final summary, in that our model can automatically determine or control the number of generated summary sentences for a given dialogue by predicting and highlighting different text spans from the source text. Our model achieves state-of-the- art performance on the largest dialogue sum- marization corpus SAMSum, with as high as 50.79 in ROUGE-L score. In addition, we con- duct a case study and show competitive human evaluation results and controllability to human- annotated summaries. 1 Introduction Text summarization aims to produce an abridged version of the input text by distilling its most crit- ical information. In particular, abstractive – as opposed to extractive – summarization requires generative models with a high level of semantic understanding, as the output words do not nec- essarily appear in the source text. While it is more challenging, it gives more flexibility to a summary compared to extractive summarization models (Zhang et al., 2018). Significant research ef- forts have been focused on summarization of single- speaker documents such as text documents (Liao et al., 2018), News (Hermann et al., 2015; Nal- lapati et al., 2016; See et al., 2017) or scientific ∗Equal contribution. Work mainly done when Linqing Liu was an intern at Salesforce Research. publications (Qazvinian and Radev, 2008; Nikolov et al., 2018). However, dialogue summarization has not received much attention despite the prevalence of dialogues (text messages, email, social media, etc.) and the vast application potential of dialogue summarization systems. Since dialogue language is inherently different from written text, it poses a unique set of chal- lenges (Zechner, 2001): 1) Distributed information across multiple speakers. The most important infor- mation is usually scattered across several conversa- tion turns from different speakers, while in articles it mostly presents in titles or the first few sentences. 2) Boundary detection. In each turn pauses do not always match linguistic sensible segments; it is difficult to identify various critical information across turns due to surrounding non-content noise and disfluency. 3) Modeling interactions between speakers. The speaker interaction plays an impor- tant role as it would imply the current dialog state and the status of the next speaker. If we directly apply neural abstract summarization models which mostly encode the whole input only as a source …

Content truncated to 500 words