Distilling Discrimination and Generalization Knowledge for Event Detection via Delta-Representation Learning

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4366–4376 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4366 Distilling Discrimination and Generalization Knowledge for Event Detection via ∆-Representation Learning Yaojie Lu1,3, Hongyu Lin1,3, Xianpei Han1,2∗, Le Sun1,2 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China {yaojie2017,hongyu2016,xianpei,sunle}@iscas.ac.cn Abstract Event detection systems rely on discrimina- tion knowledge to distinguish ambiguous trig- ger words and generalization knowledge to detect unseen/sparse trigger words. Cur- rent neural event detection approaches fo- cus on trigger-centric representations, which work well on distilling discrimination knowl- edge, but poorly on learning generalization knowledge. To address this problem, this pa- per proposes a ∆-learning approach to dis- till discrimination and generalization knowl- edge by effectively decoupling, incrementally learning and adaptively fusing event represen- tation. Experiments show that our method significantly outperforms previous approaches on unseen/sparse trigger words, and achieves state-of-the-art performance on both ACE2005 and KBP2017 datasets. 1 Introduction Event detection (ED) aims to identify triggers of specific event types. For instance, an ED system will identify fired as an Attack event trigger in the sentence “An American tank fired on the Palestine Hotel.” Event detection plays an important role in Automatic Content Extraction (Ahn, 2006), Infor- mation Retrieval (Allan, 2012), and Text Under- standing (Chambers and Jurafsky, 2008). Due to the ambiguity and the diversity of natural language expressions (Li et al., 2013; Nguyen and Grishman, 2015), an effective approach should be able to distill both discrimination and general- ization knowledge for event detection. Discrimi- nation knowledge aims to distinguish ambiguous triggers in different contexts. As shown in Figure 1, to identify fired in S4 as an EndPosition trig- ger rather than an Attack trigger, an ED system needs to distill the discrimination knowledge from S1 and S2 that (fired, Attack) usually co-occurs ∗Corresponding Author The airline firedEndPosition that pilot for fault in work. S1: An American tank firedAttack on the Palestine Hotel. S2: A man was hacked to death by the criminal. Attack S5: Discrimination Generalization S3:A heavily armed soldier shotAttack the enemy to death. Training Evaluation S4:That officer was fired from his job. EndPosition Figure 1: Examples of event instances. Identifing am- biguous word fired requires discrimination knowledge and identifing unseen word hacked requires generaliza- tion knowledge. with {tank, death, enemy, ...} and (fired, EndPo- sition) usually co-occurs with {work, fault, job, ...}. Unlike discrimination knowledge, generaliza- tion knowledge aims to detect unseen or sparsely labeled triggers, thus needs to be transferred be- tween different trigger words. For example, to identify the unseen word hacked in S5 as an At- tack trigger, an ED system needs to distill the gen- eralized Attack pattern “[Trigger] to death” from S3. Currently, most neural network ED methods (Chen et al., 2015; Nguyen and Grishman, 2015, 2016; Duan et al., 2017; Yang and Mitchell, 2017) work well on distilling discrimination knowledge, but …

Content truncated to 500 words