À la recherche des paramètres des modèles de RI

We address in this study the problem of estimating standard IR model parameters on new collections for which no relevance judgements are available. To do so, we rely on past collections with relevance judgements and introduce a new, collection independent query representation. From this representation as well as from past collections, we learn a regression function linking queries to parameter values. This regression function is then applied to each new query so as to determine the value of the parameter of the IR model considered. The experiments, conducted on standard IR collections, reveal that this approach significantly outperforms the approach based on default parameter values as well as the one making use of relevance judgements on the new collection to find global values for the IR model parameters. MOTS-CLÉS : Modèles de RI, estimation des paramètres.