RELEVANCE FEEDBACK AS AN INDICATOR TO SELECT THE BEST SEARCH ENGINE - Evaluation on TREC Data

This paper explores information retrieval system variability and takes advantage of the fact two systems can retrieve different documents for a given query. More precisely, our approach is based on data fusion (fusion of system results) by taking into account local performances of each system. Our method considers the relevance of the very first documents retrieved by different systems and from this information selects the system that will perform the retrieval for the user. We found that this principle improves the performances of about 9%. Evaluation is based on different years of TREC evaluation program (TREC 3, 5, 6 and 7), TREC-adhoc tracks. It considers the two and five best systems that participate to TREC the corresponding year.