Text auto-correction is a crucial research direction in the field of Natural Language Processing (NLP), widely applied in intelligent writing, search engine optimization, educational assessment, and other domains. Traditional correction methods relying on rules and statistical models struggle to address complex semantic errors. In recent years, pre-trained language models based on deep learning (such as BERT and GPT) have significantly improved correction performance, yet challenges remain including high computational complexity and suboptimal performance on long texts. This paper proposes a DeepSeek architecture-based text auto-correction method, integrating Dynamic Masked Language Modeling (DMLM) and adversarial training strategies to enhance the model’s semantic understanding capability and robustness. Experimental results demonstrate that on NLPCC2018 and SIGHAN datasets, our method achieves a correction accuracy of 93.5% and an F1-score of 91.3%, outperforming existing mainstream models (e.g., BERT, T5). This research provides a novel technical solution for efficient and high-precision text correction.