A two-step interactive fusion of the improved BERT and attention mechanism is proposed as an intelligent assessment method for English translation teaching. The adaptive Fusion BERT Network (AFBNet) model is designed to address the difficulty of effectively utilizing complex linguistic knowledge in the English translation tasks using only the fine-tuning method. The specific procedures involve extracting BERT’s multilayer representation, building the mask knowledge matrix, and applying the pre-training knowledge contained in BERT to the encoding word embedding layer. Secondly, the attention mechanism module interactively extracts useful knowledge in the multilayer representation of BERT and adaptively fuses with our English translation teaching model. As a result, AFBNet improves BLEU by approximately 2.6 to 4.4 points over the vanilla Transformer baseline across the five translation tasks, which exact gains vary by task. AFBNet is outperformed by BERT-JAM on the German-related tasks in our experiments, while achieving stronger results on several other directions. In conclusion, this two-step method integrating improved BERT and attention mechanism model proposed in this paper reduces the differences between pre-trained models and English translation teaching models due to different training goals. AFBNet model can effectively use pre-trained linguistic knowledge to improve English translation teaching model. The proposed method is particularly suitable for English translation teaching scenarios, where fine-grained semantic assessment and robust evaluation under limited data conditions are required.
Wenting Duan (Mon,) studied this question.