Our system for semantic role labeling ismulti-stage in nature, being based on treepruning techniques, statistical methods forlexicalised feature encoding, and a C4.5decision tree classifier. We use both shallowand deep syntactic information fromautomatically generated chunks and parsetrees, and develop a model for learningthe semantic arguments of predicates as amulti-class decision problem. We evaluatethe performance on a set of relatively‘cheap’ features and report an F1 score of68.13% on the overall test set.
展开▼
