Semantic Role Lableing System using Maximum Entropy Classifier

摘要

A maximum entropy classifier is used inour semantic role labeling system, whichtakes syntactic constituents as the labelingunits. The maximum entropy classifier istrained to identify and classify the predicates'semantic arguments together. Onlythe constituents with the largest probabilityamong embedding ones are kept. Afterpredicting all arguments which havematching constituents in full parsing trees,a simple rule-based post-processing is appliedto correct the arguments which haveno matching constituents in these trees.Some useful features and their combinationsare evaluated.