Research on automatic Japanese text translation and error detection model based on deep learning

This paper studies an integrated model of Japanese text automatic translation and error detection based on deep learning, aiming at solving the limitations of traditional translation methods in dealing with Japanese complex grammatical structures and the lack of adaptability of existing models in specific fields. The model adopts the Transformer variant architecture of multi-task collaborative learning, and realizes the joint optimization of translation and error detection through shared encoder. The key modules of the model include semantic sharing encoder, dual-task decoder and domain adaptation module. Among them, the semantic sharing encoder combines Depthwise Separable Self-Attention (DSSA) and self-attention mechanism to reduce the computational complexity. Dual-task decoder introduces gated cross-language attention mechanism and error detection decoder integrating pointer network and conditional random field (CRF) respectively. The domain adaptation module enhances the generalization ability of the model in a specific domain through confrontation training and meta-learning. The experimental results show that the model performs well on the test sets in the legal and medical fields, with the score of BLEU-4 being 53.8, the TER being 24.1, the F1 value of false detection and the recall rate reaching 89.3 and 86.7, respectively, and the reasoning speed being 153 tokens/s, which completely surpasses many baseline models including NMT+Grammarly, JointNMT and LawsMT. In addition, the ablation experiment further verified the importance of each component of the model to improve the overall performance. This research provides an efficient and accurate solution for Japanese text automatic translation and error detection, which has important theoretical and application value.