HMEAE: Hierarchical Modular Event Argument Extraction

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5777–5783, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5777 HMEAE: Hierarchical Modular Event Argument Extraction Xiaozhi Wang1∗, Ziqi Wang1∗, Xu Han1, Zhiyuan Liu1† , Juanzi Li1, Peng Li2, Maosong Sun1, Jie Zhou2, Xiang Ren3 1Department of Computer Science and Technology, Tsinghua University, Beijing, China 2Pattern Recognition Center, WeChat AI, Tencent Inc, China 3Department of Computer Science, University of Southern California, CA, USA {xz-wang16, ziqi-wan16, hanxu17}@mails.tsinghua.edu.cn Abstract Existing event extraction methods classify each argument role independently, ignoring the conceptual correlations between different argument roles. In this paper, we propose a Hierarchical Modular Event Argument Ex- traction (HMEAE) model, to provide effec- tive inductive bias from the concept hierarchy of event argument roles. Specifically, we de- sign a neural module network for each basic unit of the concept hierarchy, and then hierar- chically compose relevant unit modules with logical operations into a role-oriented mod- ular network to classify a specific argument role. As many argument roles share the same high-level unit module, their correlation can be utilized to extract specific event arguments better. Experiments on real-world datasets show that HMEAE can effectively leverage useful knowledge from the concept hierar- chy and significantly outperform the state- of-the-art baselines. The source code can be obtained from https://github.com/ thunlp/HMEAE. 1 Introduction Event argument extraction (EAE) aims to iden- tify the entities serving as event arguments and classify the roles they play in an event. For instance, given that the word “sold” triggers a Transfer-Ownership event in the sentence “Steve Jobs sold Pixar to Disney”, EAE aims to identify that “Steve Jobs” is an event argument and its argument role is “Seller”. Most event extraction (EE) methods treat EE as a two-stage problem, including event detection (ED, to iden- tify the trigger word and determine the event type) and EAE. As ED is well-studied (Nguyen and Gr- ishman, 2018; Zhao et al., 2018) in recent years, EAE becomes the bottleneck of EE. ∗ indicates equal contribution † Corresponding author: Z.Liu(liuzy@tsinghua.edu.cn) Steve Jobs sold Pixar to Disney in 2006. Person Org Time Buyer Seller Time- within Superordinate Concept Argument Role Instance Figure 1: An example of the concept hierarchy. Since EE benefits many NLP applica- tions (Yang et al., 2003; Basile et al., 2014; Cheng and Erk, 2018), intensive efforts have been devoted to detecting events and extracting their event arguments. Traditional feature-based methods (Patwardhan and Riloff, 2009; Liao and Grishman, 2010b,a; Huang and Riloff, 2012; Li et al., 2013) rely on hand-crafted features and patterns. With the ongoing development of neural networks, various neural networks have been used to automatically represent textual semantics with low-dimensional vectors, and further extract event arguments based on those semantic vectors, including convolutional neu- ral networks (Chen et al., 2015) and recurrent neural networks (Nguyen et al., 2016; Sha et al., 2018). Advanced techniques also have been adopted to further improve EE, such as …

Content truncated to 500 words