Event Extraction as Multi-turn Question Answering

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 829–838 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 829 Event Extraction as Multi-turn Question Answering Fayuan Li, Weihua Peng, Yuguang Chen, Quan Wang, Lu Pan, Yajuan Lyu, Yong Zhu Baidu Inc., China {lifayuan, pengweihua, chenyuguang, wangquan05, panlu01, lvyajuan, zhuyong }@baidu.com Abstract Event extraction, which aims to identify event triggers of pre-defined event types and their ar- guments of specific roles, is a challenging task in NLP. Most traditional approaches formulate this task as classification problems, with event types or argument roles taken as golden labels. Such approaches fail to model rich interactions among event types and arguments of different roles, and cannot generalize to new types or roles. This work proposes a new paradigm that formulates event extraction as multi-turn ques- tion answering. Our approach, MQAEE, casts the extraction task into a series of reading com- prehension problems, by which it extracts trig- gers and arguments successively from a given sentence. A history answer embedding strat- egy is further adopted to model question an- swering history in the multi-turn process. By this new formulation, MQAEE makes full use of dependency among arguments and event types, and generalizes well to new types with new argument roles. Empirical results on ACE 2005 shows that MQAEE outperforms current state-of-the-art, pushing the final F1 of argu- ment extraction to 53.4% (+2.0%). And it also has a good generalization ability, achiev- ing competitive performance on 13 new event types even if trained only with a few samples of them. 1 Introduction Event extraction is an important yet challenging task in natural language understanding. Given a sentence, an event extraction system ought to iden- tify event triggers with specific event types, as well as their corresponding arguments with specific ar- gument roles. As an example, Figure 1 presents an event mention of the type Movement Transport, triggered by “left”. There are three arguments: “Saddam’s family” playing the role of Artifact, “that city” the role of Origin, and “three days ago” the role of Time-Within. Figure 1: An example of event extraction. Typically, event extraction can be divided into two subtasks: trigger extraction (trigger identifi- cation and classification) and argument extraction (argument identification and classification), as de- fined by the standard Automatic Content Extraction (ACE) 2005 benchmark (Grishman et al., 2005). Current approaches to event extraction can thus be roughly categorized into two groups: (1) pipelined approaches that perform trigger extraction and argu- ment extraction in separate stages (Liao and Grish- man, 2010; Hong et al., 2011; Lu and Roth, 2012; Chen et al., 2015; Yang et al., 2019); (2) joint ap- proaches that perform all subtasks simultaneously in a joint learning fashion (Li et al., 2013; Nguyen et al., 2016; Liu et al., 2018; Sha et al., 2018). Most of these approaches, whether pipelined or joint, formulate event extraction as classification tasks, by classifying event triggers into pre-defined event types (trigger extraction), and further event ar- guments into pre-defined argument roles (argument extraction). …

Content truncated to 500 words