Multi-Strategy and Multi-Source Approach to Question Answering

摘要

We are interested in improving the performance of QA systems by breaking away from the strict pipeline architecture. In addition, we require an architecture that allows for hybridization at low development cost and facilitates experimentation with different instantiations of system components. Our resulting architecture is one that is modular and easily extensible, and allows for multiple answering agents to address the same question in parallel and for their results to be combined. Our new question answering system, PIQUANT, adopts this flexible architecture. The answering agents currently implemented in PIQUANT vary both in terms of the strategies used and the knowledge sources consulted. For example, an answering agent may employ statistical methods for extracting answers to questions from a large corpus, while another answering agent may transform select natural language questions into logical forms and query structured knowledge sources for answers. In this paper, we first describe the architecture on which PIQUANT is based. We then describe the answering agents currently implemented within the PIQUANT system, and how they were configured for our TREC2002 runs. Finally, we show that significant performance improvement was achieved by our multi-agent architecture by comparing our TREC2002 results against individual answering agent performance.