With the rapid advancement of scientific and technological innovation, the significance of intelligent translation services has grown considerably. This study leverages policy support to integrate AI technology into the domain of grammar error detection and correction. It utilizes a Long Short-Term Memory (LSTM) neural network and employs a Transformer model grounded in a self-attention mechanism to address the task of grammar error correction. This approach is designed to balance local contextual nuances and long-range dependencies within the text, culminating in the development of an AI-based grammar error correction model. A thorough evaluation of this model’s performance involved comparative analysis with alternative models, assessments across various types of syntactic errors, and detection of collocation errors. The findings indicate that our model exhibits superior grammatical error correction capabilities, outperforming comparative models. Specifically, it achieves Grammar Learning Evaluation Understudy (GLEU) scores that are 2.31% to 7.95% higher than those of its counterparts. Moreover, it demonstrates overall recognition rates for different grammatical and collocational errors between 18.88% to 58.98% and 50% to 70%, respectively, which underscore its practical applicability. This methodology not only enhances grammatical error detection but also holds promise for broader application in AI-driven translation services.