An Automatic Grading System for Arabic Language Short-Answer Questions Using Deep Learning

Assessing students' acquired knowledge is a core objective of the educational process. Automating this task can improve the quality of the evaluation and reduce the time, effort, and cost associated with manual grading. Automated Short Answer Grading (ASAG) systems aim to support this goal by providing accurate scoring and clear feedback to students. The ability to explain assigned scores is crucial for the real-world deployment of automated systems, ensuring transparency and trust. This work introduces an Arabic ASAG system designed to combine automated scoring with explainability. Several deep learning models and embedding techniques are evaluated, including BERT, LSTM, and attention mechanisms. Experiments are conducted on three datasets: AR-ASAG, a dedicated Arabic ASAG dataset; a manually translated Arabic version of the PT-ASAG dataset (originally in Portuguese); and a merged dataset combining both. The results highlight the effectiveness of the BERT model, which achieved strong performance with Pearson correlation coefficients from 0.811 to 0.923 and a minimum RMSE of 0.182. Prediction results are also interpreted to improve both the explainability and reliability of the system.