In this paper, we design and implement a model network for English writing style generation using UNet network as well as ViT for encoding and decoding, and PatchGAN to enhance the identification speed. Based on the CRF-NLG model to identify and extract professional English terms, and design a special loss function to optimize the quality of writing style generation. The F1 value is used to evaluate the model recognition ability, and the writing style generation effect is explored by controlled experiments of the proposed model and three baseline models. The practical application results of the proposed model are visualized from four perspectives: overall evaluation, style strength, content preservation, and fluency, to verify its practical application effect. The results show that the proposed model exhibits the strongest performance in the two levels of content preservation and fluency, which are improved by 12.71% and 39.11%, respectively, compared with the existing GAN-based style generation model. Of the 119 modifications 92 (77.3%) were better, 17 (14.3%) were average, and only 11 (9.2%) were worse.