English-Japanese Machine Translation for Lecture Subtitles using Back-Translation and Transfer Learning

In this study, we report our efforts to improve the performance of English-Japanese machine translation on the “Technology, Entertainment, Design (TED)” lecture data. We experimented with two methods to improve English-Japanese translation focusing on data augmentation. The first method is a data augmentation using Japanese monolingual corpus through back translation. The second method is transfer learning, in which a large out-of-domain English-Japanese parallel corpus was employed. We also investigated the combination of the two methods. The results indicated that the two methods produced better result than the baseline, which involved only the in-domain parallel corpus. There was an improvement in the BLEU score of 0.33 for the back-translation method and 1.37 for the transfer learning and 1.33 for the combined methods.