Joint Document-Level Event Extraction via Token-Token Bidirectional Event Completed Graph

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10481–10492 July 9-14, 2023 ©2023 Association for Computational Linguistics Joint Document-Level Event Extraction via Token-Token Bidirectional Event Completed Graph Qizhi Wan1,2, Changxuan Wan1,2∗, Keli Xiao3∗, Dexi Liu1,2, Chenliang Li4 Bolong Zheng5, Xiping Liu1,2, Rong Hu1,2 1Jiangxi Key Lab of Data and Knowledge Engineering 2School of Information Management, Jiangxi University of Finance and Economics 3College of Business, Stony Brook University 4School of Cyber Science and Engineering, Wuhan University 5School of Comp. Scie. & Tech., Huazhong University of Science and Technology wanqizhi@163.com, wanchangxuan@263.net, keli.xiao@stonybrook.edu Abstract We solve the challenging document-level event extraction problem by proposing a joint exac- tion methodology that can avoid inefficiency and error propagation issues in classic pipeline methods. Essentially, we address the three cru- cial limitations in existing studies. First, the autoregressive strategy of path expansion heav- ily relies on the orders of argument roles. Sec- ond, the number of events in documents must be specified in advance. Last, unexpected er- rors usually exist when decoding events based on the entity-entity adjacency matrix. This pa- per designs a Token-Token Bidirectional Event Completed Graph (TT-BECG) in which the re- lation eType-Role1-Role2 serves as the edge type, precisely revealing which tokens play ar- gument roles in an event of a specific event type. Exploiting the token-token adjacency matrix of the TT-BECG, we develop an edge-enhanced joint document-level event extraction model. Guided by the target token-token adjacency matrix, the predicted token-token adjacency matrix can be obtained during model training. Then, the event records in a document are de- coded based on the predicted matrix, includ- ing the graph structure and edge-type decoding. Extensive experiments are conducted on two public datasets, and the results confirm the ef- fectiveness of our method and its superiority over the state-of-the-art baselines. 1 Introduction Document-level event extraction aims to recognize events in a document with pre-defined types and corresponding event arguments, which includes en- tity extraction, entity-event correlation mapping, event type recognition, and argument role judg- ment. Sentence-level event extraction approaches (Sha et al., 2018; Yang et al., 2019; Lu et al., 2021; ∗Corresponding Author. Wan et al., 2021, 2023a) are difficult to deal with the problem of arguments across sentences. Also, a document usually contains multiple events without clear boundaries, and the corresponding descrip- tions may interact together, increasing the chal- lenges of extracting event information accurately. Figure 1 demonstrates a document example where we can summarize the following challenges in event extraction. (1) Arguments across sentences. The event e1 of EU (EquityUnderweight) type is triggered by the “reduced” with most of the argu- ments in S5, yet the argument acting as “LaterHold- ingShares” is scattered in S8, and the information describing other events are in S6 and S7; that is, the descriptions of different events are mixed to- gether. (2) Multiple events. We must determine the accurate number of events. Also, considering that tokens reflecting the meaning of “reduce” appear more than once, it will be a …

Content truncated to 500 words