Task-Based Evaluation of Anaphora Resolution: The Case of Summarization

One of the types of semantic interpretation processes that may help ‘crossing the barriers’ in text summarization is anaphora resolution. In this paper, we show that summarization is a good task for evaluating the performance of an anaphoric resolver, in the sense that it encourages developing anaphoric resolvers that build good-quality discourse models, and the performance of the anaphoric resolver correlates well with the performance on the task. Two versions of the GUITAR anaphora resolution system, 1.1 and 2.1, were used in conjunction with a LSA-based summarizer; we demonstrate that whereas both versions of the system result in improvements over the pure LSA system, only the latest version, 2.1, leads to significant improvements.