摘要
传统的分步骤事件抽取方法中,事件元素识别的结果无法指导事件类型识别,而事件类型识别的效果在很大程度上决定了事件抽取系统的整体性能.文中为解决事件类型识别对元素识别的后向依赖问题,将事件抽取看作序列标注,构建一个改进的条件随机域联合标注模型,将事件类型和事件元素在图模型中同时进行标注.同时,通过触发词嵌入试图解决事件抽取中的数据不平衡问题.ACE2005中文语料上的实验表明,基于该模型的方法提高了事件类型识别的性能,最终F值达到63.53%.
The result of event argument recognition cannot guide event type recognition in the traditional multi - step event extraction methods. Nevertheless the performance of event extraction system largely depends on event type recognition. In order to address the backward dependency of event type recognition on event argument recognition, event extraction is considered as a sequence labeling. In this paper, an improved conditional random field joint labeling model is proposed. The event type and event argument are labelled simultaneously in the graph model. The solution of the unbalanced data problem is discussed through embedding trigger word. The experiments on ACE 2005 Chinese corpus show that the performance of event type recognition is improved by the proposed method and F-score achieves 63.53%.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2012年第3期445-449,共5页
Pattern Recognition and Artificial Intelligence
基金
国家自然科学基金项目(No.61100123)
教育部博士点基金项目(No.20110032120040)
天津市科技支撑计划重点项目(No.08ZCKFGX0180)资助
关键词
事件抽取
事件类型识别
条件随机域
Event Extraction, Event Type Recognition, Conditional Random Fields
