gesis.de). GIRT is an excellent example of a collection indexed by a multilingual thesaurus, The collection is managed and indexed by the GESIS organization (http://www.social-scienceThe GIRT collection consists of reports and papers (grey literature) in the social science domain.
we used translation (L & H Power and the Systran translator) combined with our normalized German bilingual runs with translation only (also: dieren t collection), one can see a 36% and allowed elds (including the controlled terms) in the document collection. Our nal Girt run was BKGREG1 (Berkeley Girt English to German automatic run 1) where elds for the query indexing was countered by the dieren t thesaurus matching techniques for the Comparing the Russian Girt runs (translation plus thesaurus matching) to the Russian to indexed the Title and Description query elds). The positive eect of including the narrative and narrative runs, respectively. technique resulted in a 3% drop in average precision compared to the BKGRRG2 run, which only results into one le. 70% improvement in average precision for the title and description only and the title, description average precision for the monolingual Girt runs BKGRGG1 compared to BKGRGG2 (which only Using all query elds and indexing the controlled terms resulted in a 45% improvement in Although the BKGRRG1 run used all query elds for searching, its fuzzy thesaurus matching used the title and description topic elds for searching but used a dieren t thesaurus matching technique. Both runs pooled 2 query translations (Systran and Promt) and the thesaurus matching results for the French bilingual runs were slightly better than those for the German runs. In both languages, adding the narrative to the query indexes improved average precision about 6% and topics (including all title, description and narrative) were searched against the French collection. BKMLFF1 (Berkeley Monolingual French against French Automatic Run 1). The original query We applied a blind feedback algorithm for performance improvement. For indexing the French 7% for the German and French runs, respectively. collection, we used a stopwordlist, the latin-to- lower normalizer and the Muscat French stemmer. For CLEF-2002, we submitted monolingual runs for the French and German collections. Our
BKBIEF1 (Berkeley Bilingual English against French Automatic Run 1). We translated the English queries with two translation programs: the Systran translator (Altavista Babelsh) and words occurring twice or more divided (to avoid overemphasis of terms that were translated the
words occurring twice or more divided (to avoid overemphasis of terms that were translated the English queries with two translation programs: the Systran translator (Altavista Babelsh) and BKBIEG1 (Berkeley Bilingual English against German Automatic Run 1). We translated the same by both programs). We used the German decompounding procedure to split compounds in