依存关系上的中文名词性谓词识别研究

摘要

首次实现了一个基于依存关系的中文名词性谓词识别平台,作为语义角色标注的前提,谓词识别的结果直接影响语义角色标注的性能.使用两种方法进行实验:一种是基于传统的特征向量的方法在Chinese Nombank的转换语料上进行了系统实验,对各种词法特征、结构特征及其组合进行了测试,标准语料上F1值达到89.65,自动语料上达到81.27.另一种是使用树核的方法进行探索性实验,在标准语料和自动语料上分别得到84.62和80.93的F1值.%This paper implements a Chinese dependency-based nominal predicate identification system.Being the premise of Semantic Role Labeling(SRL) ,predicate identification plays a critical role.Two different ways are used in this paper.One is the traditional feather vector based predicate identification.In particular,various kinds of lexical and structure features are incorporated to improve the performance with systematic evaluation on the transferred corpus from Chinese Nombank.The system achieves 89.65 in labeled F for golden corpus,81.27 for automatic corpus.In another way,a tree kemel-based method is proposed to explore structural information.It achieves 84.62 in labeled F for golden corpus,80.93 for automatic corpus.