The Research On Chinese Semantic Role Labeling Based On A Combination Strategy

Semantic role labeling is an implement form of shallow semantic parsing. The task does not perform semantic parsing deeply. It only labels the constituents with semantic roles which have direct relation with the predicate in a sentence. The typical semantic roles include Agent, Patient, Source and Goal and so on. These semantic informtion can give great support to many NLP applications, such as question and answering, information extraction, machine translation and so on.At present, Chinese semantic role labeling (SRL) systems based on machine learning can be classified to full parsing based SRL and shallow parsing based SRL, generally. This paper firstly explores more features on the basic feature set of SRL based on full parsing, and then compares SRL systems based on these two methods in terms of performance. At last, Considering that these two systems have better performance on identifying different roles repectively, this paper attempts to combine the outputs of these two systems, uses some global features, such as how much individual systems generate the same argument, to train combination model for filtering candidate arguments in the pool, then resolves possible conflicts with domain knowledge constraints to obtain the final solution to improve performance of labeling. We report the experiments on the dataset from Chinese PropBank (CPB) 1.0. the F-Score of our combination system reaches 78.41%, Significantly improves performance on the basis of the F-Score(67.34%) of full parsing based system and the F-Score(71.67%) of shallow parsing based system. it is proved that our combination method is effective.