Event Detection as Graph Parsing

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1630–1640 August 1–6, 2021. ©2021 Association for Computational Linguistics 1630 Event Detection as Graph Parsing Jianye Xie1, Haotong Sun1, Junsheng Zhou1∗, Weiguang Qu1, Xinyu Dai2 1School of Computer and Electronic Information, Nanjing Normal University, China 2National Key Laboratory for Novel Software Technology, Nanjing University, China {zhoujs,wgqu}@njnu.edu.cn,daixinyu@nju.edu.cn {xiejy,sunht}@nnu.edu.cn Abstract Event detection is a fundamental task in infor- mation extraction. Most previous approaches typically view event detection as a trigger- based classification problem, focusing on us- ing syntactic dependency structure or exter- nal knowledge to boost the classification per- formance. To overcome the inherent issues with existing trigger classification based mod- els, we propose a novel approach to event de- tection by formulating it as a graph parsing problem, which can explicitly model the mul- tiple event correlations and naturally utilize the rich information conveyed by event type and subtype. Furthermore, to cope with data sparsity, we employ a pretrained sequence-to- sequence (seq2seq) model to transduce an in- put sentence into an accurate event graph with- out the need for trigger words. Extensive experimental results on the public ACE2005 dataset show that, our approach outperforms all previous state-of-the-art models for event detection by a large margin, obtaining an im- provement of 4.2% F1 score. The result is very encouraging since we achieve this with a con- ceptually simple seq2seq model; moreover, by extending the graph structure, this proposed ar- chitecture can be flexibly applied to more in- formation extraction problems for sentences. 1 Introduction Event Detection (ED) is an important task in Information Extraction (IE) that seeks to identify instances of specified types of events in text (Ji and Grishman, 2008; Li et al., 2013). For example, for the input sentence shown in Figure 1, the ED model aims to predict three event types expressed by this sentence, each of which consists of an event type label and its subtype label, according to the ACE2005 Guidelines. * Corresponding author. Figure 1: An example of event detection ED is an actively studied task in IE where deep learning models have been the dominant approach to deliver the state-of-the-art performance (Chen et al., 2015; Nguyen et al., 2016; Sha et al., 2018; Chen et al., 2018). The last few years witness the success of graph convolutional neural networks for ED (Nguyen and Grishman, 2018; Liu et al., 2018; Yan et al., 2019; Lai et al., 2020) where the depen- dency trees are employed to boost the performance. Also, another line of research focused on exploiting external knowledge to improve classification (Lu et al., 2019; Liu et al., 2019a; Tong et al., 2020). Nevertheless, most previous work typically treats ED as the identification and classification of trigger words, focusing on using various syn- tactic dependency structure or external knowledge to boost classification performance. Methodologi- cally speaking, this type of trigger-based ED mod- els suffer from the following inherent drawbacks. Firstly, most previous approaches depend heav- ily on the trigger word. On the one hand, trig- …

Content truncated to 500 words