We Study a cross-language IR approach using the NEAT information filtering system and the ASTRANSAC machine translation system. The BMIR-J2 standard Japanese test collection and our own translated data are used for evaluation. In our English-Japanese experiments, we compare the retrieval performance of the document translation approach and the query translation approach, and show that they ashieve over 90/100 and 80/100 of monolingual performance, respectively. Moreover, in our query translation experiments, we consider the case where the Japanese queries are translated into English by two different translators, and show that the manual translation can greatly affect the performance in traditional evaluation us- ing a monolingual baseline. In our Japanese-pseudo-English experiments, we perform local feedback both before and after query translation, and achieve retrieval performance of the monolingual standard. This can be regarded as an upperbound of performance in Japanese- English cross-language IR using real English documents. Finnally we study the relationship between the quality of machine translation and retrieval performance.
展开▼
