With the deep integration of big data technology and artificial intelligence, the field of English translation is experiencing a paradigm shift from single text processing to multimodal information fusion. In this paper, we design a multimodal English translation model based on Transformer, which fuses cross-modal information through the source language sequence encoding layer. Using the pre-trained cross-modal model CLIP to extract graphic features, combined with the dynamic contextual visual guidance vector mechanism, the adaptive fusion of text and image information is realized. In English-Chinese translation, the perplexity of the model in this paper is only 1.36 after 100 rounds of iterations, and the correct rate reaches 99.80%. The control model 1 still has a perplexity of more than 10, and the correct rate is only 94.49%. In Chinese-English translation, the model in this paper also achieves optimal results. The model’s translation performance on the six sets of parallel corpus, including EnglishGerman, English-French, and Chinese-English, is significantly better than that of the baseline model, with the highest improvement of BLEU value of 16.89, which verifies the effect of multimodal information enhancement on the improvement of translation quality and context adaptation.
