HERO: a hierarchical attention and multi-feature fusion framework for automated English essay scoring

Automated essay scoring (AES) has emerged as a crucial task in educational technology, aiming to replicate human evaluation of writing with speed, fairness, and accuracy. This study proposes an advanced deep learning framework, HERO (Hierarchical Essay Representation Optimizer), that uses hierarchical attention mechanisms and multi-feature fusion for holistic and grammar-based essay scoring. The proposed model employs a two-level bidirectional LSTM with attention to capture contextual information at both word and sentence levels, while integrating auxiliary metadata features such as prompt ID, gender, and grade level to enhance prediction performance. Experiments are conducted on the ASAP 2.0 dataset containing nearly 12,000 human-scored essays. The proposed framework demonstrates superior performance over baseline models, including MLP, KNN, and XGBoost, achieving a Pearson correlation coefficient of 0.92, MAE of 0.019, and CCC of 0.91. Comprehensive evaluation using error metrics, correlation measures, and statistical validation confirms the robustness and reliability of the model. The findings indicate that the integration of hierarchical representation learning with feature-level fusion significantly enhances essay scoring accuracy while supporting interpretability. However, the fairness analysis was limited to aggregate evaluation on the available dataset, and subgroup-level calibration was not comprehensively validated.