Transition-based Semantic Role Labeling Using Predicate Argument Clustering

This paper suggests two ways of improving semantic role labeling (SRL). First, we introduce a novel transition-based SRL algorithm that gives a quite different approach to SRL. Our algorithm is inspired by shift-reduce parsing and brings the advantages of the transitionbased approach to SRL. Second, we present a self-learning clustering technique that effectively improves labeling accuracy in the test domain. For better generalization of the statistical models, we cluster verb predicates by comparing their predicate argument structures and apply the clustering information to the final labeling decisions. All approaches are evaluated on the CoNLL’09 English data. The new algorithm shows comparable results to another state-of-the-art system. The clustering technique improves labeling accuracy for both in-domain and out-of-domain tasks.