Employing Morphological Structures and Sememes for Chinese Event Extraction

Proceedings of COLING 2012: Technical Papers, pages 1619–1634, COLING 2012, Mumbai, December 2012. Employing Morphological Structures and Sememes for Chinese Event Extraction LI Pei Feng and ZHOU Guo Dong School of Computer Science & Technology, Soochow University, Suzhou, China, 215006 {pfli, gdzhou}@suda.edu.cn ABSTRACT Current Chinese event extraction systems suffer much from the low recall due to unknown triggers. To resolve this problem, this paper firstly introduces morphological structures to better represent the compositional semantics inside Chinese triggers and then proposes a mechanism to automatically identify the head morpheme (either verb or noun) as the governing sememe of a trigger. Finally, it proposes a mechanism of combining the morphological structures and sememes of Chinese words to infer unknown triggers to improve the recall of the Chinese event extraction system. Evaluation on the ACE 2005 Chinese corpus justifies the effectiveness of our approach over a state-of-the-art system. 形态结构和义原在中文事件抽取中的应用 由于存在大量未知的触发词，当前的中文事件抽取系统受限于它的低召回率。为了解决这 个问题，本文首先引入形态结构来更好地表示隐含在中文触发词内部的组合语义，然后提 出了一个自动识别触发词中作为支配义原的核心词素（动词或名词）的机制。最后，本文 提出了一个结合了中文词语的形态结构和义原去推测未知触发词的方法，用于提高中文事 件抽取系统的召回率。在ACE 2005中文语料上的实验验证了我们方法的性能超越了目前 最好的中文事件抽取系统。 KEYWORDS: Chinese event extraction; Morphological structure; Governing sememe; Trigger identification; Head morpheme. Keywords in L2: 中文事件抽取; 形态结构; 支配语义; 触发词识别; 核心语素  Corresponding author 1619 1 Introduction As a compromise to natural language understanding, Information Extraction (IE) aims to extract structured information (e.g., entities, relations and events) from a text. Event extraction, a classic subtask in IE, is to recognize event trigger mentions of a predefined event type and their participants and attributes. While most studies in the literature focus on English event extraction, there are few successful stories concerning Chinese event extraction due to the special characteristics and challenges in Chinese language. Even with ground truth entities, times and values, the performance of most Chinese event extraction systems is much lower than that of English ones. For Chinese event extraction, unknown triggers (a trigger in the test set doesn’t occur in the training set and otherwise, a known trigger.) and word segmentation errors are two major reasons for the low performance, particularly the recall. The statistics on the ACE 2005 Chinese and English corpora (Li et al., 2012) shows that these two cases cover almost 30% of Chinese trigger mentions while this figure reduces to only about 9% in English. Besides, given the same number of event mentions, there are about 30% more different triggers in Chinese than those in English. This amplifies the problem. Therefore, trigger identification becomes a key to the success of Chinese event extraction. Currently, there mainly exist two major mechanisms to solve this problem. The first one is to expand the triggers using predefined or automatically-clustered synsets, a common mechanism widely used in various NLP applications. The problem with this mechanism is that it fails to consider the sense shifting of a word in difficult contexts and thus may introduce too many pseudo triggers and harm the precision. This largely limits the contribution of this mechanism (Chen and Ji, 2009b; Ji, 2009; Qin et al. 2010). For example, as a trigger of the Start-position event, “担任” has more than five senses (e.g., serve as, bear, engage, do, etc.) and only one of …

Content truncated to 500 words