Compact and robust models for Japanese-English character-level machine translation

Abstract

Character-level translation has been proved to be able to achieve preferable translation quality without explicit segmentation, but training a character-level model needs a lot of hardware resources. In this paper, we introduced two character-level translation models which are mid-gated model and multi-attention model for Japanese-English translation. We showed that the mid-gated model achieved the better performance with respect to BLEU scores. We also showed that a relatively narrow beam of width 4 or 5 was sufficient for the mid-gated model. As for unknown words, we showed that the mid-gated model could somehow translate the one containing Katakana by coining out a close word. We also showed that the model managed to produce tolerable results for heavily noised sentences, even though the model was trained with the dataset without noise.