Analyzing Translation Accuracy of Concatenation-based Multi-source Neural Machine Translation

Abstract

In this paper, we study a concatenation-based multi-source neural machine translation (NMT) model trained with three-language parallel corpus. We show that the concatenation-based multi-source NMT model where a parallel English and Chinese sentences are input to the model as the source sentences improves the BLEU score of the single-source NMT where only English or Chinese source sentence is input to the model. Among major phenomena where the BLEU improves when translating from the source English sentence than from the source Chinese sentence are translation of Katakana loanwords, tense, and particles, etc., while, in the translation of Chinese words when the Chinese and Japanese words share an identical Chinese character, the BLEU improves when translating from the source Chinese sentence than from the source English sentence. We then show that, in the translation by the concatenation-based multi-source NMT model, the BLEU improves the most by correctly incorporating translation of both types of phenomena in a complementary style.