基于特征组合的中文语义角色标注

摘要

This paper proposes a semantic role labeling (SRL) approach for the Chinese, based on feature combination and support vector machine (SVM). The approach takes the constituent as the labeling unit. First, this paper defines the basic feature set by selecting the high-performance features of existing parsing-based SRL systems. Then, a statistics-based method is proposed to construct a combined feature set derived from the basic feature set. According to the distribution of combining features in both positive and negative instances, the ratio of between-class to within-class distance is utilized as the measurement of classifying the performance the feature, and then choosing the combining features with high ratios into the combining feature set. Finally, the experimental results show that the feature combination method-based SRL achieved 91.81％ F-score on Chinese PropBank (CPB) corpus, nearly 2％ higher than the traditional method.%提出一种基于特征组合和支持向量机(support vector machine,简称SVM)的语义角色标注(semantic role labeling,简称SRL)方法.该方法以句法成分作为基本标注单元,首先从当前基于句法分析的语义角色标注系统中选出高效特征,构成基本特征集合.然后提出一种基于统计的特征组合方法.该方法能够根据正反例中组合特征的分布状况,以类间距离和类内距离之比作为统计量来衡量组合特征对分类所产生的效果,保留分类效果较好的组合特征.最后,在Chinese PropBank(CPB)语料上利用支持向量机进行分类实验,结果表明,引入该特征组合方法后,语义角色标注整体F值达91.81%,提高了近2%.