Reconstructing Event Regions for Event Extraction via Graph Attention Networks

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 811–820 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 811 Reconstructing Event Regions for Event Extraction via Graph Attention Networks Pei Chen1∗, Hang Yang2,3, Kang Liu2,3, Ruihong Huang1, Yubo Chen2,3, Taifeng Wang4, and Jun Zhao2,3 1 Texas A&M University, College Station, TX 2 Institute of Automation, Chinese Academy of Sciences, Beijing, China 3 University of Chinese Academy of Sciences, Beijing, China 4 Ant Group, Hangzhou, China {chenpei, huangrh}@tamu.edu, taifeng.wang@antgroup.com {hang.yang, kliu, yubo.chen, jzhao}@nlpr.ia.ac.cn Abstract Event information is usually scattered across multiple sentences within a document. The lo- cal sentence-level event extractors often yield many noisy event role filler extractions in the absence of a broader view of the document- level context. Filtering spurious extractions and aggregating event information in a docu- ment remains a challenging problem. Follow- ing the observation that a document has several relevant event regions densely populated with event role fillers, we build graphs with candi- date role filler extractions enriched by senten- tial embeddings as nodes, and use graph at- tention networks to identify event regions in a document and aggregate event information. We characterize edges between candidate ex- tractions in a graph into rich vector represen- tations to facilitate event region identification. The experimental results on two datasets of two languages show that our approach yields new state-of-the-art performance for the chal- lenging event extraction task. 1 Introduction Event Extraction (EE), a challenging task in Natu- ral Language Processing, aims to extract key types of information (aka event roles, e.g., perpetra- tors and victims of an attack event) that can rep- resent an event in texts and plays a critical role in downstream applications such as Question An- swer (Yang et al., 2003) and Summarizing (Filatova and Hatzivassiloglou, 2004). Existing research on EE mostly focused on sentence-level, such as the evaluation in Automatic Content Extraction (ACE) 20051. However, an event is usually described in ∗Most of the work was done when the first author was a research engineer in the Institute of Automation, CAS. 1http://projects.ldc.upenn.edu/ace/ Event Template Event Roles Role Fillers PerpInd TERRORISTS, HOODED INDIVIDUALS PerpOrg SHINING PATH Victim DOLORES HINOSTROZA, HINOSTROZA Original Document S1: That alleged TERRORISTS today killed DOLORES HINOSTROZA, the mayor of Mulqui district. S2: HINOSTROZA, who was at home, was shot five times. S3: Hinostroza's children told police that four HOODED INDIVIDUALS broke into the HOUSE and shot their mother after having insulted her. S4: And their FATHER was on a business trip then. S5: DOLORES HINOSTROZA deceased when the ambulance came. S6: She is the second woman mayor killed this week by alleged commando groups of the Maoist SHINING PATH. Region1 Region2 Figure 1: An example of document-level event extrac- tion. We need to extract noun phrases from the docu- ment as role fillers for the event roles in the predefined event template. The uppercased noun phrases in the document are role …

Content truncated to 500 words