DCU@FIRE2010: term conflation, blind relevance feedback, and cross-language IR with manual and automatic query translation

摘要

For the first participation of Dublin City University (DCU)udin the FIRE 2010 evaluation campaign, information retrievalud(IR) experiments on English, Bengali, Hindi, and Marathiuddocuments were performed to investigate term conationud(different stemming approaches and indexing word prefixes),udblind relevance feedback, and manual and automatic queryudtranslation. The experiments are based on BM25 and onudlanguage modeling (LM) for IR. Results show that term conation always improves mean average precision (MAP)udcompared to indexing unprocessed word forms, but different approaches seem to work best for different languages. For example, in monolingual Marathi experiments indexing 5-prefixes outperforms our corpus-based stemmer; in Hindi,udthe corpus-based stemmer achieves a higher MAP. For Bengali, the LM retrieval model achieves a much higher MAPudthan BM25 (0.4944 vs. 0.4526). In all experiments usingudBM25, blind relevance feedback yields considerably higherudMAP in comparison to experiments without it. Bilingual IR experiments (English!Bengali and English!Hindi) areudbased on query translations obtained from native speakersudand the Google translate web service. For the automaticallyudtranslated queries, MAP is slightly (but not significantly)udlower compared to experiments with manual query translations. The bilingual English!Bengali (English!Hindi)udexperiments achieve 81.7%-83.3% (78.0%-80.6%) of the bestudcorresponding monolingual experiments.