English translation teaching faces the challenge of insufficient semantic expression precision, and traditional teaching methods are difficult to capture cross-linguistic semantic nuances. In this study, a deep semantic space model is constructed based on the three principles of relevance, consensus and complementarity, which extracts image features through techniques such as transfer learning, feature adaptation and convolutional neural network, and utilizes bag-of-words model and recurrent neural network for text semantic learning. The experiments validate the model performance on two datasets, TGIF and MSVD, and the results show that on the TGIF dataset, the deep semantic spatial model (DSS) proposed in this paper achieves the R@1, R@5, and R@10 metrics of 9.97, 25.97, and 34.53, respectively, in the video dimension of text retrieval; and the corresponding metrics in the video retrieval of the video dimension are, in order, 15.06, 30.73, 41.22, significantly better than the comparison algorithm. Teaching application experiments show that the translation scores of students in the experimental class with the model-assisted teaching (12.14±1.76) are significantly higher than those of the control class with the traditional teaching method (9.73±2.46), and the difference is statistically significant (P<0.001). The study shows that the deep semantic space model based on multimodal learning can effectively improve the semantic expression accuracy in English translation teaching, which provides new technical support and methodological reference for the reform of English translation teaching.
