Passage-level Evidence for Cross-Language Information Retrieval

Machine translation (MT) techniques can be used to generate a query in a target language from a query in a source language for the cross-language information retrieval (CLIR). Recent MT systems have advanced enough to generate translations which are human-readable, However, translation error is still a serious impediment which hurts the effectiveness of a CLIR system. To compensate for defects in a machinetranslation result, we propose a method using passage-level evidence. By combining a document retrieval model with a passage retrieval model, we prevent the retrieval model from assigning a high score to a non-relevant document because of translation error. The retrieval model incorporating passage retrieval shows better results than a document retrieval model. In particular, the passage retrieval model achieves more improvement when the translation quality of queries is relatively low.