Country Adaptation in Neural Machine Transliteration of Person Names

Abstract

We have developed a neural machine transliteration system that can transliterate person names in many countries into Japanese. This system is constructed by two-step training. First, a base model is trained on a large country-unknown person-name dataset. Then, country-specific models are constructed from the base model by applying transfer learning with a set of small country-specific datasets. An experiment shows that our system outperforms the previous CRF-based system.