An ensemble of transliteration models for information retrieval

Transliteration is used to phonetically translate proper names and technical terms especially from languages in Roman alphabets to languages in non-Roman alphabets such as from English to Korean, Japanese, and Chinese. Because transliterations are usually representative index terms for documents, proper handling of the transliterations is important for an effective information retrieval system. However, there are limitations on handling transliterations depending on dictionary lookup, because transliterations are usually not registered in the dictionary. For this reason, many researchers have been trying to overcome the problem using machine transliteration. In this paper, we propose a method for improving machine transliteration using an ensemble of three different transliteration models. Because one transliteration model alone has limitation on reflecting all possible transliteration behaviors, several transliteration models should be complementary used in order to achieve a high-performance machine transliteration system. This paper describes a method about transliteration production using the several machine transliteration models and transliteration ranking with web data and relevance scores given by each transliteration model. We report evaluation results for our ensemble transliteration model and experimental results for its impact on IR effectiveness. Machine transliteration tests on English-to-Korean transliteration and English-to-Japanese transliteration show that our proposed method achieves 78-80% word accuracy. Information retrieval tests on KTSET and NTCIR-1 test collection show that our transliteration model can improve the performance of an information retrieval system about 10-34%.