Juru at TREC 2005: Query prediction in the terabyte and the robust tracks

Abstract

Our experiments focus this year on the ad-hock tasks of the Terabyte and the Robust tracks. In both tracks we experimented with the query prediction technology we developed recently. In the Terabyte track, we investigated how query prediction can be used to improve federation of search results extracted from several indices. We show that federated search based on query prediction can achieve comparable results to single-index search. In the Robust track we trained a predictor over one collection (TREC-8) for predicting query difficulty over another collection (AQUAINT). The experimental results show that difficult topics on the TREC-8 collection were not found to be consistently difficult on the AQUAINT collection.