Multilingual Query Expansion for CLEF Adhoc-TEL

In this paper we will briefly describe the approaches taken by the Cheshire (Berkeley) Group for the CLEF Adhoc-TEL 2009 tasks (Mono and Bilingual retrieval). Recognizing that many potentially relevant documents in each of the TEL sub-collections are in other languages, we tried to use multiple translations of the topics for searching each subcollection, combined into a single query. Overall this strategy performed very poorly compared to the the basic monolingual approach used last year (and repeated for one run in each language this year). We haven’t yet completed our analysis of the reasons for this (we suspect that results were evaluated expecting the retrieved items to also be in the same language as the topic). Once again this year we used probabilistic text retrieval based on logistic regression and incorporating blind relevance feedback for all of the runs. All translation for bilingual tasks was performed using the LEC Power Translator PC-based MT system. Our results this year, however, were surprising poor compared to last year’s results. Some testing has shown that, for some cases, unexpected hyphenations in the machine translation and untranslated words were to blame. It may also be the case that others have significantly improved their approaches for this task.