Using Translation Memory to Improve Neural Machine Translations

Abstract

In this paper, we describe a way of using translation memory (TM) to improve the translation quality and stability of neural machine translation (NMT) systems, especially when the sentences to be translated have high similarity with sentences stored in the TM. The difference between the sentences to be translated and the sentences stored in the TM may only be in a few phrases. Our TM comprises not only paired sentences (i.e., a sentence in the source language paired with its translation in the target language) but also paired phrases. Translation quality is improved using good phrase translations for the differing phrases. The NMT system is used to assist phrase translation. We tested our TM on 3,000 English-Chinese paired sentences which were randomly picked from recent annual reports published and submitted to the Hong Kong Stock Exchange. Our TM translations achieved a significant BLEU improvement for high similar sentences compared with our NMT translations.