Fusion of Information Retrieval Engines (FIRE)

We examine the feasibility of fusing the outputs of multiple text retrieval engines to improve accuracy. We tested three Web-based search engines (Excite for Web Servers, Infoseek's Ultraseek Server, and Sony Search Engine) over a 74,520 document collection (TREC Wall Street Journal articles from 1990-92) with a set of 125 natural-language queries with relevance judgements (TREC topics 51-175). We show that a weighted combination of scores produces higher precision over the top 5, 105 20, and 30 documents than any single engine over the same data set. We also compare favorably against the state-of-the-art heuristics in merging search engines. Our results suggest that fusing the results from the most dissimilar engines (those with the least overlap in the retrieved sets) is a more effective strategy than simply weighting the best engines more heavily.