Document Translation vs. Query Translation for Cross-Lingual Information Retrieval in the Medical Domain

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6849–6860 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6849 Document Translation vs. Query Translation for Cross-Lingual Information Retrieval in the Medical Domain Shadi Saleh and Pavel Pecina Charles University, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´am. 25, 118 00 Prague, Czech Republic {saleh,pecina}@ufal.mff.cuni.cz Abstract We present a thorough comparison of two prin- cipal approaches to Cross-Lingual Informa- tion Retrieval: document translation (DT) and query translation (QT). Our experiments are conducted using the cross-lingual test collec- tion produced within the CLEF eHealth infor- mation retrieval tasks in 2013–2015 contain- ing English documents and queries in several European languages. We exploit the Statistical Machine Translation (SMT) and Neural Ma- chine Translation (NMT) paradigms and train several domain-specific and task-specific ma- chine translation systems to translate the non- English queries into English (for the QT ap- proach) and the English documents to all the query languages (for the DT approach). The results show that the quality of QT by SMT is sufficient enough to outperform the retrieval results of the DT approach for all the lan- guages. NMT then further boosts translation quality and retrieval quality for both QT and DT for most languages, but still, QT provides generally better retrieval results than DT. 1 Introduction Multilingual content has been growing significantly in the last few years simultaneously with rapid in- ternet access growth all over the world. Monolin- gual information retrieval task allows users to find information in documents that are written in the language that they use to write their queries. This ignores a vast amount of information that is repre- sented in other languages. Cross-Lingual Informa- tion Retrieval (CLIR) breaks this language barrier by allowing users to look up information that is represented in documents written in languages dif- ferent from the language of the query. We reinvestigate the effectiveness of two princi- pal approaches to CLIR: document translation (DT) and query translation (QT). The existing compari- son studies of the two approaches are outdated (e.g. Oard, 1998) and do not reflect the current advances in Machine Translation (MT). Even in very recent works, the authors have blindly assumed that DT is superior to QT (Khiroun et al., 2018), giving the ar- gument that in DT, the text is translated in a larger context compared to the translation of short isolated queries in QT. The larger context should help in translation disambiguation and better lexical selec- tion during translation, which should subsequently lead to better retrieval results. This hypothesis needs to be revised, taking into consideration the significant improvement of ma- chine translation quality in recent years, despite the strong practical disadvantages of DT over QT: DT is computationally expensive and hard to scale (every document needs to be translated into each supported language and then indexed) while QT is performed in query time and only a short text (the query) is translated into the document language. In …

Content truncated to 500 words