DESED: Dialogue-based Explanation for Sentence-level Event Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 2483–2493 October 12–17, 2022. 2483 DESED: Dialogue-based Explanation for Sentence-level Event Detection Yinyi Wei1∗†, Shuaipeng Liu2∗‡, Jianwei Lv2, Xiangyu Xi2, Hailei Yan2, Wei Ye3‡, Tong Mo1, Fan Yang2, Guanglu Wan2 1 Peking University 2 Meituan Group, Beijing, China 3 National Engineering Research Center for Software Engineering, Peking University wyyy@pku.edu.cn, liushuaipeng@meituan.com, wye@pku.edu.cn Abstract Many recent sentence-level event detection ef- forts focus on enriching sentence semantics, e.g., via multi-task or prompt-based learning. Despite the promising performance, these meth- ods commonly depend on label-extensive man- ual annotations or require domain expertise to design sophisticated templates and rules. This paper proposes a new paradigm, named dialogue-based explanation, to enhance sen- tence semantics for event detection. By say- ing dialogue-based explanation of an event, we mean explaining it through a consistent information-intensive dialogue, with the origi- nal event description as the start utterance. We propose three simple dialogue generation meth- ods, whose outputs are then fed into a hybrid attention mechanism to characterize the com- plementary event semantics. Extensive experi- mental results on two event detection datasets verify the effectiveness of our method and sug- gest promising research opportunities in the dialogue-based explanation paradigm. 1 Introduction Event detection (ED) is a crucial task in informa- tion extraction, which aims to identify event trig- gers (words or phrases that indicate events) and classify triggers into predefined event types. For example, we can identify the trigger weddings and classify it into Marry event type from the text “Giu- liani regularly officiated at weddings while in of- fice”. Sentence-level event detection plays a dom- inant role in event detection and is significant for various downstream NLP tasks. However, it is usually challenging to accurately detect events in a single sentence due to the lim- ited information. Therefore, most prior methods on sentence-level event detection make improve- ments by enhancing sentence semantics, being *Equal Contribution. †Work was done when Yinyi Wei interned at Meituan. ‡Corresponding Author. I got fired today. Could i get arrested for protesting? User_1 User_2 If you do it from the street, no. But shouldn't you be trying to find a new job first man? Well that'll be for tomorrow, today im pissed and hurt User_1 User_2 Don't bother protesting. Update your resume, look for a new job. There's no win here. User 肉都变质了，东西也不新鲜，投诉商家 => The meat is spoiled and the food is not fresh, I want to complain to the merchant 亲亲，可以说一下什么商品不新鲜吗 => My sweetie honey, can you tell us what commodities are not fresh? Agent User 肉不新鲜，有一股变质味，土豆也是坏的 => The meat is not fresh, smelled stale and the potatoes are also bad 亲亲，非常抱歉给您带来了不好的用餐感受，您这边现在的诉求是什么呢？ => Sorry for the bad experience my honey. What's your demand now? Agent Figure 1: Two examples of dialogues from Reddit and FOSAED, respectively. divided into two categories. The first category mainly involves leveraging other information ex- traction tasks (e.g., named entity recognition and relation extraction) via multi-task learning (Wad- den et al., 2019; Lin et al., 2020; Van Nguyen et al., …

Content truncated to 500 words