Chinese Semantic Role Labeling with Hierarchical Semantic Knowledge

This paper reports our work on Chinese semantic role labeling, which takes advantage of hierarchical semantic knowledge from a common sense knowledge base named HowNet. On one hand, the words in lexical features such as predicate and head word are generalized with their hypernyms in HowNet. On the other hand, the hypernym-hyponym relation between sememes is used to capture the semantic similarity between verbs. Experiment results show that both of the two methods can help our system achieve significant improvements on semantic role classification precision with golden parses as the input, by alleviating the problem of data sparseness. Further experiment indicates that by using fully automatic parses as the input, the accuracy of Chinese semantic role labeling can be close to the English state of the art.