Why do Successful Search Systems Fail for Some Topics

This paper describes and evaluates the vector-space and probabil-istic IR models used to retrieve news articles from a corpus writ-ten in the French language. Based on three CLEF test-collections and 151 queries, we classify the poor retrieval results of difficult topics under 6 categories. The explanations we obtain from this analysis differ from those suggested a priori by our students. We use the Web to manually or automatically find related search terms to the original query. We evaluate these two query expan-sion strategies in order to improve mean average precision (MAP) and to reduce the number of topics for which no pertinent re-sponses are listed among the top ten references returned.