Joint Modeling for Chinese Event Extraction with Rich Linguistic Features

Proceedings of COLING 2012: Technical Papers, pages 529–544, COLING 2012, Mumbai, December 2012. Joint Modeling for Chinese Event Extraction with Rich Linguistic Features Chen CHEN Vincent NG Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688, USA {yzcchen,vince}@hlt.utdallas.edu Abstract Compared to the amount of research that has been done on English event extraction, there exists relatively little work on Chinese event extraction. We seek to push the frontiers of supervised Chinese event extraction research by proposing two extension to Li et al.'s (2012) state-of-the-art event extraction system. First, we employ a joint modeling approach to event extraction, aiming to address the error propagation problem inherent in Li et al.'s pipeline system architecture. Second, we investigate a variety of rich knowledge sources for Chinese event extraction that encode knowledge ranging from the character level to the discourse level. Experimental results on the ACE 2005 dataset show that our joint-modeling, knowledge-rich approach significantly outperforms Li et al.'s approach. Title and Abstract in Chinese 运用丰富语言学特征的中文事件抽取联合模型 ��文的事件抽取�������中文的事件抽取���������� Li et al.(2012) 的 ����学�的事件抽取���������������������中文事件抽取的 ���������用���联合模型���� Li et al. �����中的������� �����中文��抽取���������������文���的特征�� ACE2005 ���的���������运用丰富语言学特征的联合模型����� Li et al. 的��� Keywords: event extraction, Chinese language processing. Keywords in Chinese: ��抽取, 中文��语言��. 529 1 Introduction Recent years have seen a surge of interest in automatically extracting events from textual docu- ments. While diverse types of event extraction have been examined in the literature, the one we will focus on in this paper is ACE event extraction, which involves extracting instances of a prede- fined event type from documents. For example, consider the following Chinese text segment: Resneft ���������� 93.5 ��� (Resneft acquired Yugansk, paying only 9.35 billion US dollars) When applied to this example, an ACE event extraction system for Chinese should identify one event instance, which (1) is triggered by the verb �� [acquired] whose type is Transfer- Money, and (2) has three arguments, Resneft, ���� [Yugansk] and 93.5 ��� [9.35 bil- lion], which fulfill the roles of Buyer, Artifact and Price, respectively. This example illustrates the four primary subtasks of an ACE event extraction system, namely, (1) trigger identification (e.g., �� [acquired] should be identified as the trigger of an event); (2) trigger type determination (e.g., �� [acquired] should be assigned the type Transfer-Money; (3) argument identification (e.g., Resneft, ���� [Yugansk] and 93.5 ��� [9.35 billion] are the arguments of this Transfer-Money event); and (4) argument role determination (e.g., Resneft, ���� [Yugansk] and 93.5 ��� [9.35 billion] play the roles of Buyer, Artifact and Price respectively in this event). Compared to the amount of research on English event extraction (e.g., Finkel et al. (2005), Grish- man et al. (2005), Ahn (2006), Hardy et al. (2006), Maslennikov and Chua (2007), Ji and Grishman (2008), Patwardhan and Riloff (2009), Liao and Grishman (2010), Hong et al. (2011)), there is considerably less work on Chinese event extraction. Work on end-to-end Chinese event extraction was pioneered by Chen and Ji (2009b), who adopt a pipeline system architecture composed of four components that correspond to the four major subtasks mentioned …

Content truncated to 500 words