To address the problems of poor translation quality, poor domain adaptability, and insufficient computational efficiency faced by existing English Neural Machine Translation (NMT) systems when processing complex syntactic structures and specialized domain texts, this paper conducts an optimization study of English NMT systems based on deep learning. First, a hierarchical fusion attention mechanism is introduced, combining standard self-attention and graph attention networks (GAT) in parallel, dynamically integrating syntactic dependencies and contextual information through gating fusion. Then, a joint strategy of dynamic vocabulary generation and domain-adaptive fine-tuning is proposed, enabling the model to generate out-of-vocabulary words based on sub-word units, and using course learning to adjust the data mixing ratio for a smooth transition. Next, to improve model inference efficiency and facilitate deployment while maintaining translation quality, a progressive knowledge distillation technique is adopted, using a teacher-student framework to hierarchically transfer knowledge to achieve model lightweighting. Experimental results show that the proposed method achieves BLEU (Bilingual Evaluation Understudy) scores of 38.4% and 40.2% on the WMT22 (Workshop on Machine Translation 2022) English-German dataset and the WMT23 (Workshop on Machine Translation 2023) English-Chinese dataset, respectively. In the medical field, the lexical, syntactic, and semantic error rates are as low as 4.27%, 6.88%, and 8.98%, respectively. Furthermore, in terms of translation quality and inference efficiency, the proposed method achieves an inference speed of 129.2 sentences/second with a translation editing rate of only 32.2%, achieving a balance between quality and efficiency. The research method in this paper optimizes the accuracy, robustness and efficiency of the NMT system through collaboration.
Zhu et al. (Tue,) studied this question.