This study focuses on developing an automatic assessment model of English writing that integrates natural language processing (NLP) and grammatical analysis, aiming at solving the problems of inefficiency, subjectivity and delayed feedback in traditional manual correction of English writing. This paper studies and constructs a multi-dimensional evaluation framework covering grammar, semantics and text structure, and realizes grammatical constraints through semantic-grammar mutual reinforcement theory and pre-training model to extract semantic information and dependency syntax tree. Based on the hierarchical theory of text structure, the hierarchical features of the article are modeled with the help of graph neural network (GNN). The technical framework adopts three-stage pipeline architecture, including input processing layer, analysis fusion layer and evaluation feedback layer, which realizes grammar error detection, semantic coherence evaluation, structure analysis and personalized feedback generation. The experimental results show that the model is significantly superior to the existing benchmark model in grammar error detection performance, with an average F1 value of 89.1%, and the correlation coefficient with the median consistency of human assessors is 0.82, which is close to the consistency among human assessors. Personalized feedback is significantly better than general feedback in error correction, which effectively improves the error correction rate.