Inducing Romanization Systems

We propose a method for inducing romanization systems directly from a bilingual alignment at the grapheme level. First, transliteration word pairs are aligned using a non-parametric Bayesian approach, and then for each grapheme sequence to be romanized, a particular romanization is selected according to a user-specified criterium. We apply our approach to the task of transliteration mining, and used Levenshtein distance as the selection criterium. We performed experiments on three languages with differing characteristics: Japanese, Russian and Chinese. Our experiments show that the mining system built from the induced romanization system is able to outperform existing baseline romanization systems. By extending our approach to induce romanization systems based on other criteria we expect our technique may find more general application in the future.