Selectively using linguistic resources throughout the question answering pipeline

It is generally believed that question answering can benefit from natural language processing methods. So far, however, there have been few systematic studies of this conjecture. We report on ongoing work that is aimed at understanding the contribution of linguisticallyinformed modules and resources to the overall performance of a generic question answering system. Specifically, we describe the ways in which currently we use linguistically motivated techniques, and demonstrate the impact of integrating, or not integrating, these techniques on the overall performance of our question answering system. Evaluation results are based on the TREC 2002 and TREC 2003 question sets.