Improving Multilingual Neural Machine Translation with Artificial Labels

Abstract

Inspired by the work which uses Artificial Translation Units for generation of synthetic data in low-resource Neural Machine Translation systems [12], we propose using these translation units to enhance ability of sharing information between translation units in the multilingual Neural Machine Translation systems. In particular, we concentrate on improving the translation of rare-words. Our method also suggest a new idea about leveraging bilingual dictionaries in multilingual Neural Machine Translation systems which is still limited in prior works. Our experiments show improvements of up to +3.5 BLEU scores in the translation tasks between Chinese, Japanese and Vietnamese from the TED Talks domain. Our machine translation system outperforms the systems in [12] when translating from Chinese to Vietnamese although we do not use any additional techniques such as data argumentation or pre-trained model as shown in [12].