Event Extraction via Dynamic Multi-Pooling Convolutional Neural Networks

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 167–176, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Event Extraction via Dynamic Multi-Pooling Convolutional Neural Networks Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng and Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China {yubo.chen,lhxu,kliu,djzeng,jzhao}@nlpr.ia.ac.cn Abstract Traditional approaches to the task of ACE event extraction primarily rely on elabo- rately designed features and complicated natural language processing (NLP) tools. These traditional approaches lack gener- alization, take a large amount of human effort and are prone to error propaga- tion and data sparsity problems. This paper proposes a novel event-extraction method, which aims to automatically ex- tract lexical-level and sentence-level fea- tures without using complicated NLP tools. We introduce a word-representation model to capture meaningful semantic reg- ularities for words and adopt a framework based on a convolutional neural network (CNN) to capture sentence-level clues. However, CNN can only capture the most important information in a sentence and may miss valuable facts when considering multiple-event sentences. We propose a dynamic multi-pooling convolutional neu- ral network (DMCNN), which uses a dy- namic multi-pooling layer according to event triggers and arguments, to reserve more crucial information. The experimen- tal results show that our approach signif- icantly outperforms other state-of-the-art methods. 1 Introduction Event extraction is an important and challenging task in Information Extraction (IE), which aims to discover event triggers with specific types and their arguments. Current state-of-the-art methods (Li et al., 2014; Li et al., 2013; Hong et al., 2011; Liao and Grishman, 2010; Ji and Grishman, 2008) often use a set of elaborately designed features that are extracted by textual analysis and linguistic knowledge. In general, we can divide the features into two categories: lexical features and contex- tual features. Lexical features contain part-of-speech tags (POS), entity information, and morphology fea- tures (e.g., token, lemma, etc.), which aim to cap- ture semantics or the background knowledge of words. For example, consider the following sen- tence with an ambiguous word beats: S1: Obama beats McCain. S2: Tyson beats his opponent . In S1, beats is a trigger of type Elect. However, in S2, beats is a trigger of type Attack, which is more common than type Elect. Because of the am- biguity, a traditional approach may mislabel beats in S1 as a trigger of Attack. However, if we have the priori knowledge that Obama and McCain are presidential contenders, we have ample evidence to predict that beats is a trigger of type Elect. We call these knowledge lexical-level clues. To repre- sent such features, the existing methods (Hong et al., 2011) often rely on human ingenuity, which is a time-consuming process and lacks general- ization. Furthermore, traditional lexical features in previous methods are a one-hot representation, which may suffer from the data sparsity problem and may not be able to adequately capture the se- mantics of the words …

Content truncated to 500 words