Intelligent assessment and personalized teaching strategies for English writing integrating large language model and text readability regression analysis

In intelligent assessment of English writing, a single model struggles to balance semantic depth and interpretability, easily overlooking individual learner differences and leading to feedback generalization. This article employs a method integrating the Large Language Model Meta AI-2 (LLaMA2) large language model and Extreme Gradient Boosting (XGBoost) text readability regression analysis to support the generation of personalized teaching strategies. LLaMA-2 is used to extract the contextual semantics, logical coherence, and pragmatic features of the essay; readability indices are calculated and input into XGBoost for nonlinear regression modeling; subsequently, a gating mechanism is used to fuse semantic vectors and readability prediction values to generate a comprehensive score and four-dimensional assessment dimensions; finally, based on the learner’s historical level and current weaknesses, rewriting suggestions, micro-skill training, or model essay recommendations are dynamically triggered from a predefined rule base, achieving a closed-loop linkage between assessment and teaching. Experiments show that the proposed method achieves a mean Quadratic Weighted Kappa (QWK) score of 0.88–0.93 and a mean mean absolute error (MAE) score of 0.80–1.15 across different language proficiency subsets, demonstrating both semantic depth and overall scoring accuracy. Across different error types of text, the mean feature attribution consistency score is 0.918–0.942, and the mean interpretability fidelity score is 0.872–0.903, improving feedback credibility. In low-scoring and high-scoring learner groups, the policy response rate and unit policy gain are 0.874 ± 0.032 and 2.31 ± 0.18, respectively, indicating effective matching of personalized strategies. This research effectively bridges the gap between deep semantic understanding and practical teaching applications, providing an interpretable and adaptable technical reference for intelligent writing feedback systems.