Differential Translation for Japanese Partially Amended Statutory Sentences

Takahiro Yamakoshi, Takahiro Komamizu, Yasuhiro Ogawa, Katsuhiko Toyama

Abstract: We propose a differential translation method that targets statutory sentences partially modified by amendments. After a statute is amended, we need to promptly update its translations for international readers. We must focus on the focality of translation. In other words, we should modify only the amended expressions in the translation and retain the others to avoid causing misunderstanding of the amendment’s contents. To generate focal, fluent, and adequate translations, our method incorporates neural machine translation (NMT) and template-aware statistical machine translation (SMT). In particular, our method generates n-best translations by an NMT model with Monte Carlo dropout and chooses the best one by comparing them with the SMT translation. This complements the weaknesses of each method: NMT translations are usually fluent but they often lack focality and adequacy, while template-aware SMT translations are rather focal and adequate but not fluent. In our experiments, we showed that our method outperformed both the NMT-only and SMT-only methods.