Edge-Enhanced Graph Convolution Networks for Event Detection with Syntactic Relation

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2329–2339 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2329 Edge-Enhanced Graph Convolution Networks for Event Detection with Syntactic Relation Shiyao Cui1,2 Bowen Yu1,2 Tingwen Liu1,2∗ Zhenyu Zhang1,2 Xuebin Wang1,2 and Jinqiao Shi3,1 1Institute of Information Engineering, Chinese Academy of Sciences. Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences. Beijing, China 3Beijing University of Posts and Telecommunications. Beijing, China {cuishiyao, yubowen, zhangzhenyu1996}@iie.ac.cn {liutingwen, wangxuebin}@iie.ac.cn shijinqiao@bupt.edu.cn Abstract Event detection (ED), a key subtask of infor- mation extraction, aims to recognize instances of specific event types in text. Previous s- tudies on the task have verified the effective- ness of integrating syntactic dependency in- to graph convolutional networks. However, these methods usually ignore dependency la- bel information, which conveys rich and use- ful linguistic knowledge for ED. In this pa- per, we propose a novel architecture named Edge-Enhanced Graph Convolution Network- s (EE-GCN), which simultaneously exploit- s syntactic structure and typed dependency label information to perform ED. Specifical- ly, an edge-aware node update module is de- signed to generate expressive word represen- tations by aggregating syntactically-connected words through specific dependency types. Fur- thermore, to fully explore clues hidden in de- pendency edges, a node-aware edge update module is introduced, which refines the rela- tion representations with contextual informa- tion. These two modules are complementary to each other and work in a mutual promo- tion way. We conduct experiments on the widely used ACE2005 dataset and the result- s show significant improvement over competi- tive baseline methods1. 1 Introduction Event Detection (ED) is an important information extraction task that seeks to recognize events of specific types from given text. Specifically, each event in a sentence is marked by a word or phrase called “event trigger”. The task of ED is to detect event triggers and classify them into specific types of interest. Taking Figure 1 as an example, ED is supposed to recognize the event trigger “visited’’ and classify it to the event type Meet. ∗Corresponding Author 1Source code of this paper could be obtained from http- s://github.com/cuishiyao96/eegcned nmod Putin last visited Bush at his Texas ranch in November 2001 nsubj nmod:poss case nummod case dobj advmod compound nmod ROOT Figure 1: An example of syntactic dependency parsing, which contains an event of Meet triggered by “visited”. Dependency trees convey rich structural infor- mation that is proven useful for ED (Nguyen and Grishman, 2018; Liu et al., 2018b; Yan et al., 2019). Recent works on ED focus on building Graph Con- volutional Networks (GCNs) over the dependen- cy tree of a sentence to exploit syntactic depen- dencies (Nguyen and Grishman, 2018; Liu et al., 2018b; Yan et al., 2019). Compared to sequence- based models, GCN-based models are able to cap- ture non-local syntactic relations that are obscure from the surface form alone (Guo et al., 2019), and usually achieve better performance. Nevertheless, existing GCN-based ED method- s do not consider dependency labels, which may serve as significant …

Content truncated to 500 words