半监督中文事件抽取中的模板过滤和转换方法

Pattern Filtering and Conversion Methods for Semi-supervised Chinese Event Extraction

事件模板是指导事件抽取工作的依据,半监督方法下模板的准确性显得尤为重要。目前,基于双视图的"触发词-论元"模板的中文信息事件抽取系统不能有效地解决触发词一词多义的现象和模板稀疏现象。提出了一种借助论元进行触发词语义消歧的方法,并利用该方法进行模板过滤以消除无效模板的影响。另外,针对几种特殊的中文句型,根据句法结构提出了模板转换规则,从而提高了模板的适用性。在ACE2005中文语料上的测试表明,该方法可有效地提高半监督中文信息事件抽取系统的性能。

The accuracy of event patterns is very important in semi-supervised event extraction.Currently,semi-supervised Chinese event extraction system based on the pairwise pattern（e.g.,Trigger-Argument）suffers much from the issues of polysemy of triggers and sparse patterns.This paper put forward a argument-based mechanism to solve trigger sense disambiguation,and then applied it to pattern filtering to eliminate invalid patterns.In addition,for several special Chinese sentence structures,this paper proposed a pattern conversion method based on syntactic structure to enhance the applicability of the pattern.The experimental results on the ACE 2005 Chinese data show that our methods can effectively improve performance of semi-supervised Chinese event extraction system.