Memory-based semantic role labeling: Optimizing features, algorithm, and output

In this paper we interpret the semantic role labeling problem as a classification task, and apply memory-based learning to it in an approach similar to Buchholz et al. (1999) and Buchholz (2002) for grammatical relation labeling. We apply feature selection and algorithm parameter optimization strategies to our learner. In addition, we investigate the effect of two innovations: (i) the use of sequences of classes as classification output, combined with a simple voting mechanism, and (ii) the use of iterative classifier stacking which takes as input the original features and a pattern of outputs of a first-stage classifier. Our claim is that both methods avoid errors in sequences of predictions typically made by simple classifiers that are unaware of their previous or subsequent decisions in a sequence.