Improving neural machine translation by word translations

Abstract

Transformer-based neural machine translation (NMT) models have achieved state-of-the-art performance in the machine translation paradigm. These models learn the translation knowledge from the bilingual corpus through the attention mechanism automatically. This differs from the way human translators approach sentence translation, where prior knowledge plays a significant role. Inspired by this, a word translation augmentation (WTA) method is proposed to improve the Transformer-based NMT model. The main steps are as follows: Firstly, constructing the word alignment rules based on the training set. Next, generating the translation rules for source words according to the word alignment rules. Lastly, incorporating the potential translation candidates for each source word into the NMT model during the training and testing procedure. In addition, the WTA method introduces the idea of Mixup for translation candidates of a source word and employs two augmentation strategies to augment the encoder. The results of experiments on several translation tasks with high-resource and low-resource indicate the effectiveness of the proposed method compared with the corresponding strong baseline, and the improvement in BLEU score achieved ranges from 0.42 to 0.63.