Mind the Gap! Injecting Commonsense Knowledge for Abstractive Dialogue Summarization

Proceedings of the 29th International Conference on Computational Linguistics, pages 6285–6300 October 12–17, 2022. 6285 Mind the Gap! Injecting Commonsense Knowledge for Abstractive Dialogue Summarization Seungone Kim1∗ Se June Joo1∗ Hyungjoo Chae1,2∗ Chaehyeong Kim1∗ Seung-won Hwang3 Jinyoung Yeo1,2† Yonsei University1 Tutoring at Market Designers2 Seoul National University3 {louisdebroglie, sr7418, mapoout, cheris8, jinyeo}@yonsei.ac.kr seungwonh@snu.ac.kr Abstract In this paper, we propose to leverage the unique characteristics of dialogues sharing common- sense knowledge across participants, to re- solve the difficulties in summarizing them. We present SICK, a framework that uses common- sense inferences as additional context. Com- pared to previous work that solely relies on the input dialogue, SICK uses an external knowl- edge model to generate a rich set of common- sense inferences and selects the most probable one with a similarity-based selection method. Built upon SICK, SICK++ utilizes common- sense as supervision, where the task of gener- ating commonsense inferences is added upon summarizing the dialogue in a multi-task learn- ing setting. Experimental results show that with injected commonsense knowledge, our frame- work generates more informative and consis- tent summaries than existing methods. 1 Introduction Abstractive dialogue summarization is a task of generating a shorter summary while preserving the context of a conversation (Li et al., 2017; Gliwa et al., 2019). Unlike conventional document- to-document summarization (e.g., news articles and scientific publications) (Nallapati et al., 2016; Gehrmann et al., 2018), such dialogue-to-document summarization suffers from the discrepancy be- tween input and output forms, which makes learn- ing their mapping patterns more challenging. There are two key challenges that make sum- marizing dialogues harder than documents. First, detecting unspoken intention is crucial for under- standing an utterance (Mendelsohn, 1994; Ram et al., 2018). As shown in Figure 1, without under- standing the intent “to make fun of someone”, it is hard to write a correct summary. Second, there ex- ists information that can only be understood when ∗Equal contribution †Corresponding author Alyssa: What do you think about it? Alyssa: The best part is that she acts like she nailed it. But at least it's funny in a good way. Derek: It is Derek: I can fart bright stripes and bright stars better then she sings. => xWant, to make fun of someone Golden summary Derek and Alyssa Fergie's performance of the national anthem. make fun of Golden summary She won't be repairing it, because . Melody's 5-year-old laptop is broken. Tomorrow she'll know what's wrong. Instead, she'll buy a new one. her laptop is too old Melody: youre probably due for a new one anyway, no? Peggy: you're right. 5 years is a long time to own one. Peggy: ok. i might just not bother getting it repaired after all. Melody: sounds like a good idea Melody: yes, thats ancient by laptop standards => HinderedBy, the laptop is too old Figure 1: Example of dialogue-summary pairs. Captur- ing the intention and hidden meaning is important to generate a novel summary. its hidden meaning is revealed (Talmy, 1988). For example, it is important …

Content truncated to 500 words