Aiming at the limitations of traditional scoring model, this paper proposes a Transformer-Enhanced BiLSTM model based on multi-head self-attention mechanism, which can effectively capture the high-level writing skills of texts. At the same time, a multi-modal feature fusion framework is constructed, which integrates linguistic, semantic and textual features, and the dimension of features is reduced by improved mRMR (minimum Redundancy Maximum Relevance) algorithm, and the contribution of features is analyzed by SHAP value. The experiment uses three data sets covering academic writing, social topic discussion and business e-mail. The results show that the proposed model outperforms the baseline model in two indicators: Quadratic Weighted Kappa (QWK) and Root Mean Square Error (RMSE), with QWK reaching 0.862 and RMSE 0.127. The feature ablation experiment shows that multi-feature fusion significantly improves the scoring accuracy. The interpretable design goal of the model is verified by attention weight heat map and SHAP value visualization, which provides concrete and operable suggestions for teaching feedback.