TREC 2001 CROSS LANGUAGE DATASET

Ten groups participated in the TREC-2001 cross-language information retrieval track, which focussed on retrieving Arabic language documents based on 25 queries that were originally prepared in English. French and Arabic translations of the queries were also available. This was the first year in which a large Arabic test collection was available, so a variety of approaches were tried and a rich set of experiments performed using resources such as machine translation, parallel corpora, several approaches to stemming and/or morphology, and both pre-translation and post-translation blind relevance feedback. On average, forty percent of the relevant documents discovered by a participating team were found by no other team, a higher rate than normally observed at TREC. This raises some concern that the relevance judgment pools may be less complete than has historically been the case.