A Novel Semantic Role Labeling model based on Statistical Semi-Supervised Learning

Abstract

Semantic role labeling is an important task in natural language processing to determine the relation between the predicate in a sentence and its corresponding argument. At present, most of supervised learning methods can achieve better performance in large-scale labeled training data application. But it is difficult to obtain high quality labeled data. To solve the problem, we proposed a novel semantic role labeling algorithm based on semi-supervised statistical learning in this paper. Firstly, an initial classifier was constructed based on small-scale labeled data. Secondly, a noise filter is employed to reduce the noise in the train data. The experiment results show the proposed algorithm has a higher performance for semantic role labeling, and provide a good practicability for other natural language processing tasks.