UC Berkeley at CLEF 2003 - Russian Language Experiments and Domain-Specific Cross-Language Retrieval

As in the previous years, Berkeley’s group 1 experimented with the domain-specific CLEF collection GIRT as well as with Russian as query and document language. The GIRT collection was substantially extended this year and we were able to improve our retrieval results for the query languages German, English and Russian. For the GIRT retrieval experiments, we utilized our previous experiences by combining different translations, thesaurus matching, decompounding for German compounds and a blind feedback algorithm. We find that our thesaurus matching technique compares to conventional machine translation for Russian and German against English retrieval and outperforms machine translation for English to German retrieval. With the introduction in CLEF 2003 of a Russian document collection, we participated in the CLEF main task with monolingual and bilingual runs for the Russian collection. For bilingual retrieval our approaches were query translation (for German or English as topic languages) and ‘fast’ document translation (for English as the topic language). Document translation significantly underperformed query translation (using the PROMPT translation system).