Multi-argument classification for semantic role labeling

This paper describes a Multi-Argument Classification (MAC) approach to Semantic Role Labeling. The goal is to exploit dependencies between semantic roles by simultaneously classifying all arguments as a pattern. Argument identification, as a pre-processing stage, is carried at using the improved Predicate-Argument Recognition Algorithm (PARA) developed by Lin and Smith (2006). Results using standard evaluation metrics show that multi-argument classification, archieving 76.60 in F₁ measurement on WSJ 23, outperforms existing systems that use a single parse tree for the CoNLL 2005 shared task data. This paper also describes ways to significantly increase the speed of multi-argument classification, making it suitable for real-time language processing tasks that require semantic role labelling