A Corpus-Based Method to Improve Feature-Based Semantic Role Labeling

This paper proposes a novel corpus-based method for feature-based semantic role labeling (SRL). The method first constructs a number of combined features based on basic features and can rapidly discern the discriminative combined features that will improve the performance of SRL. According to the distribution in the corpus, we define a statistical quantity that can efficiently measure the classifying capacity of the combining feature, and then retain the high-value combined features for the later classification. The experiments on Chinese Proposition Bank (CPB) corpus show the method can improve the F-score of SRL by more than one percent.