Semantic Role Labeling Based on Valence Structure and Deep Neural Network

Abstract

Semantic roles labeling is a kind of shallow semantic analysis. Existing Chinese semantic analysis methods and semantic roles labeling systems do not effectively characterize Chinese essential features, and it causes the currently larger difference between Chinese SRL systems and English SRL systems. Valence structures can better characterize syntactic structures and semantic constitution relations of Chinese sentences, so we incorporated the valence information of predicates into semantic roles labeling. Experimental results show that proper use of valence information significantly improves the performance of semantic roles labeling system: the verbal SRL approach achieves the performance of 93.69% in F1-measure and the nominal SRL approach achieves the performance of 79.23% in F1-measure on golden parse trees and golden predicates, and all outperform the state-of-the-art SRL systems. In recent years, end-to-end semantic role labeling based on deep neural networks has attracted more and more attention. However, current semantic role labeling methods use deep neural networks without language features. We propose a deep neural network model that integrates valence information, which was evaluated on the CoNLL-2005, CoNLL-2012 shared task datasets, and achieved better results than previous work: the performance of the CoNLL-2005 shared task dataset is improved 0.74 percentage points; the semantic role labeling methods proposed in this paper achieved an F1 value of 84.80% in CoNLL-2012 shared task datasets.