Joint Modeling of Trigger Identification and Event Type Determination in Chinese Event Extraction

Proceedings of COLING 2012: Technical Papers, pages 1635–1652, COLING 2012, Mumbai, December 2012. Joint Modeling of Trigger Identification and Event Type Determination in Chinese Event Extraction LI Pei Feng, ZHU Qiao Ming, DIAO Hong Jun and ZHOU Guo Dong School of Computer Science & Technology, Soochow University, Suzhou, China, 215006 {pfli, qmzhu, hjdiao, gdzhou}@suda.edu.cn ABSTRACT Currently, Chinese event extraction systems suffer much from the low quality of annotated event corpora and the high ratio of pseudo trigger mentions to true ones. To resolve these two issues, this paper proposes a joint model of trigger identification and event type determination. Besides, several trigger filtering schemas are introduced to filter out those pseudo trigger mentions as many as possible. Evaluation on the ACE 2005 Chinese corpus justifies the effectiveness of our approach over a strong baseline. 一个应用于中文事件抽取的事件触发词识别和类型判别联合模 型 当前，有2个问题困扰着中文事件抽取系统：低质量的事件标记语料库和假事件触发词相 对于真事件触发词的高比例。为了解决以上2个问题，本文提出了一个结合事件触发词识 别和事件类型判别的联合模型。另外，几个触发词过滤模式同样被引入本系统用于过滤掉 尽可能多的假触发词实例。在ACE2005中文语料上的测试结果表明，本文的方法和基准系 统相比具有更高的性能。 KEYWORDS: Joint modeling, Event type determination, Trigger identification, Trigger filtering. KEYWORDS IN L2: 联合模型, 事件类型判别, 触发词识别, 触发词过滤 Corresponding author 1635 1 Introduction Information extraction (IE) is a task of extracting structured information (e.g. entities, relations and events) from the text. As a critical part of IE, event extraction is to identify trigger mentions of a predefined event type, and their participants and attributes. It can be typically divided into four components: trigger identification, event type determination, argument identification and argument role determination. Due to the central role of the contained events in a text, it is critical to mine their semantics in order to understand a text. Unfortunately, event extraction has been proven its performance is still very low. In the literature, most studies focus on English event extraction and have achieved certain success (e.g., Grishman et al., 2005; Ahn, 2006; Patwardhan and Riloff, 2009; Hong et al., 2011; Lu and Roth, 2012; Llorens et al., 2012). However, there are few successful stories regarding Chinese event extraction due to the special characteristics in Chinese trigger identification. Besides unknown triggers1 and word segmentation errors (Li et al., 2012), the low quality of annotated corpora and the high ratio of pseudo trigger mentions to true ones are also blamed for the low performance of Chinese event extraction. To examine the low quality of annotated corpora in Chinese event extraction, we take the ACE (Automatic Content Extraction) 2005 Chinese corpus (with 8 types and 33 subtypes of events), one of the most popular corpora in event extraction, as an example. In particular, we randomly select 33 documents from the training set and ask two human annotators to manually tag event mentions and their types following the definition of the ACE 2005 corpus. Here, human annotator 1 is a first year postgraduate student with no background in Chinese event extraction while human annotator 2 is a third year postgraduate student working on Chinese event extraction. Table 1 justifies the difficulty of Chinese event extraction, particularly for trigger identification and event type determination, even for a well-educated human being. As shown in Table 1, the IAA (Inter-Annotator Agreement) on both trigger identification and event …

Content truncated to 500 words