CLIO: Role-interactive Multi-event Head Attention Network for Document-level Event Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2504–2514 October 12–17, 2022. 2504 CLIO: Role-interactive Multi-event Head Attention Network for Document-level Event Extraction Yubing Ren1,2, Yanan Cao1,2, Fang Fang1,2*, Ping Guo1,2 Zheng Lin1,2, Wei Ma1,2∗and Yi Liu3 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China 3National Computer Network Emergency Response Technical Team/Coordination Center of China {renyubing,caoyanan,fangfang0703,guoping,linzheng,mawei}@iie.ac.cn liuyi@cert.org.cn Abstract Transforming the large amounts of unstructured text on the Internet into structured event knowl- edge is a critical, yet unsolved goal of NLP, es- pecially when addressing document-level text. Existing methods struggle in Document-level Event Extraction (DEE) due to its two intrin- sic challenges: (a) Nested arguments, which means one argument is the substring of another one. (b) Multiple events, which indicates we should identify multiple events and assemble the arguments for them. In this paper, we pro- pose a role-interactive multi-event head atten- tion network (CLIO) to solve these two chal- lenges jointly. The key idea is to map different events to multiple subspaces (i.e., multi-event head). In each event subspace, we draw the se- mantic representation of each role closer to its corresponding arguments, then we determine whether the current event exists. To further optimize event representation, we propose an event representation enhancing strategy to regu- larize pre-trained embedding space to be more isotropic. Our experiments on two widely used DEE datasets show that CLIO achieves consis- tent improvements over previous methods. 1 Introduction Cognitive scientists believe that humans remem- ber and understand reality primarily in terms of events (Shipley and Zacks, 2008). Event studies are justifiably popular in Natural Language Pro- cessing (NLP), such as Event Coreference Reso- lution, Event Causality Identification, and Event Extraction. Event extraction is the process of ex- tracting structured event knowledge from unstruc- tured text and can be divided into sentence-level and document-level. Sentence-level Event Extrac- tion has demonstrated promising results in empiri- cal evaluations. However, in real-world scenarios, a large number of event elements are expressed ∗Fang Fang and Wei Ma are the co-corresponding authors Meet witnesses Oklahoma City Eldon Elliott, prosecutors Topic Place Participant Role Argument Argument ExchangeBuySell Beneficiary Recipient Giver PaymentBarter AcquireEntity - McVeigh Eldon Elliott $280.32 Mercury Marquis Role The owner of a body shop and truck … who rented the Ryder truck used in the bombing of Oklahoma City... Eldon Elliott…recalling the two occasions when McVeigh entered his shop--on April 15, 1995, to reserve a truck and pay the $280.32 …when prosecutors presented a series of witnesses and document…The Building in Oklahoma City was shattered by…and purchased a 1977 Mercury Marquts… Prosecutors say he drove the truck to Geary Lake in Kansas… Input document DEE Transportation Transporter PassengerArtifact Vehicle Origin Destination Role Argument McVeigh, him McVeigh Ryder truck truck, restaurant, shop Junction City, Geary Lake Figure 1: An illustration of DEE task. Different colored tabels indicate different event types. DEE needs to detect multiple event types and extract arguments for the roles of …

Content truncated to 500 words