Multilingual Neural Machine Transliteration with Adaptive Segmentation Schemes

Haoran Sun, Yikun Lei, Deyi Xiong

Abstract: The lack of training data and divergences in the phonological structures of languages pose challenges to machine transliteration. In this paper, we present a multilingual neu-ral machine transliteration framework to encourage knowledge transfer from high-resource languages and low-resource languages. In order to mitigate the phonological divergence issue, we propose to use different segmentation schemes that are adaptive to the source or the target language. Experiment results on public datasets demonstrate that multilingual neural machine transliteration significantly outperforms bilingual transliteration on both mid- and low-resource languages and that segmentation schemes have a great impact on transliteration quality.