This study firstly introduces the working principle of deep learning-based neural machine translation model (NMT) and its recurrent neural network translation backbone network, which enhances the semantic characterization capability through Glove word embedding layer. A tree-to-sequence based attention mechanism is innovatively introduced at the encoder side, and a tree-based encoder is appended to the traditional sequence encoder to construct syntax-aware context vectors. On the decoder side, the syntactic tree structure information is integrated into the sequence-to-sequence model (seq2seq), and this model is used to explore the knowledge transfer effect of the English translation teaching process. The results show that the accuracy rates of the neural machine English translation models incorporating syntactic information proposed in this paper are all above 90%. The experiment on the effect of English translation teaching shows that the mean values of students’ scores on the post-test of long sentence translation and composition translation in the reading section of the experimental class increased by 11.022 and 12.5388 points respectively compared with those of the control class, with significant differences between the scores of the two groups of students (p<0.05), and the same significant differences are presented between the scores on the pre-test and post-test of the students' scores on the long sentence translation and composition translation in the experimental class. It can be seen that the application of the model can effectively promote knowledge transfer and help students better understand and utilize translation skills.