Research on Adaptive Optimization Algorithm for Japanese Machine Translation

The complexity of the Japanese language poses challenges for machine translation (MT), such as domain-specific terminology, corpus scarcity, and agglutinative characteristics. To address these issues, this study proposes an algorithm based on Dynamic Domain-Aware Hybrid Gated Network (DDA-HGN), which integrates domain feature extraction, context-aware weight allocation, and adversarial training mechanisms to enhance the model's domain adaptability. Additionally, a high-quality, multi-domain Japanese parallel corpus was constructed, covering fields such as medical, legal, and patent texts, with data augmentation and active learning techniques applied to improve corpus quality. Experimental results show that DDA-HGN outperforms baseline models like Transformer and mBART-50 in both general and medical domains, achieving excellent performance in BLEU-4, TER (Translation Edit Rate), DAS (Domain Term Score), and PER (Particle Error Rate) metrics. Ablation studies further demonstrate the importance of dynamic gating mechanisms and adversarial training in improving translation quality and term accuracy. This research provides an effective theoretical foundation and technical solution for the practical application of Japanese MT.