Error analysis in anaphora resolution

Proceedings of the Third International Conference on Language Resources and Evaluation (LREC 2002)

Abstract

This paper deals with error analysis and their influence in comparative and qualitative evaluation of systems performing anaphora resolution. It presents a corpus-based analysis of errors reported by four anaphora resolution systems, leading to an investigation of the type and source of errors; as a direct application of the investigation's results, a simple probabilistic hybrid method is described, that takes advantage of the strong points of each of the methods analysed, while trying to avoid their weak points.