摘要： 事件检测与类型识别是事件抽取的基础，具体实施分为触发词检测和事件类型识别2个阶段。分别对2个阶段进行研究，在前一阶段，针对词形特征过拟和问题，提出利用LDA模型对词语聚类的方法，考虑到中文自动分词与标注的触发词边界的不一致性，提出基于CRFs模型的触发词识别方法。在后一阶段，为提高事件类型识别的效果，将跨事件理论应用于中文事件类型识别。实验结果表明，该方法能提高系统性能，F值分别提高到66.3和62.0。

关键词: 事件抽取, 触发词检测, 事件类型识别, 跨事件, CRFs模型, LDA模型

Abstract: Event extraction is an important component of information extraction. Event detection and recognition is the basis of event extraction. It is implemented by two stages which are trigger word detection and event recognition. The two stages are studied. In first stage, words are clustered by LDA model to solve the words overfitting problem and a trigger word detection method is proposed based on character using CRFs model in view of the inconsistency between Chinese word segmentation and trigger word boundary. In the next stage, cross-event inference is applied in Chinese event recognition to enhance the result of event recognition. Experimental results show that the approach can significantly improve system performance, achieving the F-measure of 66.2 and 62.0 on the stage of trigger detection and event recognition respectively.

Key words: event extraction, trigger word detection, event type recognition, cross-event, CRFs model, LDA model

中图分类号: 

• TP391