Neural machine translation models have made significant progress in general-purpose domains, but there are still many challenges for the models to solve translation tasks in specialized domains with low resources, especially in how to make full use of terminology information. The study firstly expresses the fundamentals of neural machine translation from three perspectives: encoder-decoder framework, text feature representation and decoding search, respectively. Then, it introduces BLEU, the evaluation index at the core of neural machine translation, and proposes engineering domain adaptive techniques and data enhancement methods. Finally, improvements are made on the outside of the translation model, and a neural machine translation model fusing the underlying information is proposed. Aiming at the problem that the English-Chinese translation model in the field of electrical engineering does not fully utilize terminological information, the terminological vocabulary is taken as a priori knowledge. Through experimental validation, the method can significantly improve the quality of machine translation and show more satisfactory translation results in low-resource languages as well, and the method proposed in this paper improves the BLEU values by 1.25-1.82 on average on different datasets. In addition, the model proposed in this paper outperforms the translation performance of the traditional Transformer model in the Chinese-English translation task. In summary according to the BLEU value of the model, the training time rise reaches the enhancement and achieves the balance of translation quality and time cost.
