Machine transliteration using multiple transliteration engines and hypothesis re-ranking
Jong–Hoon Oh, Hitoshi Isahara · 2007
This paper describes a novel method of improving machine transliteration by using multiple transliteration hypotheses and re-ranking them. We constructed seven machine-transliteration engines to produce a set of transliteration hypotheses. We then re-ranked the hypotheses to select the correct transliteration hypothesis. We propose a re-ranking method that makes use of confidence-score, languagemodel, and Web-frequency features and combines them with machine-learning algorithms including support vector machines and the maximum entropy model. Our testing of English-to-Japanese and English-to-Korean transliterations revealed that the individual transliteration engines used in our approach performed comparably to previous approaches and that re-ranking improved word accuracy compared to the best individual engine from about 65 to 88%. 1.