Advancing Arabic automated essay scoring through cross-encoder BERT models and interpretable explanations

Automated Essay Scoring (AES) in Arabic remains underdeveloped due to the language's diglossia, rich morphology, and limited annotated resources, reflecting inherent linguistic asymmetry. This study advances Arabic AES by introducing a BERT-based sentence-pair cross-encoder framework that leverages the structural symmetry of bidirectional attention and combines it with an Integrated Gradients (IG) explanation pipeline. Five widely used Arabic BERT variants were systematically evaluated: (1) asafaya/bert-base-arabic, (2) aubmindlab/arabertv02, (3) UBC-NLP/MARBERT,faisalq/ (4) SaudiBERT, and (5) CAMeL-Lab/CAMeLBERT-MSA. Across the experiments, the CAMeLBERT-MSA cross-encoder fine-tuned for ten epochs achieved high results, with an R2 of 98.47%, MAE of 0.07, and 98.32% accuracy within ± 0.5 points. This marks a significant improvement over both shallow similarity-based baselines and earlier BERT concatenation models, demonstrating the value of deeper task-specific adaptation on Modern Standard Arabic (MSA) corpora. Cross-prompt evaluations revealed both generalization symmetry-strong performance with limited prompts (R2 = 79.38%, 86.92% accuracy within ± 1.0 point at four prompts)-and generalization asymmetry, where performance declined as prompt heterogeneity increased. The interpretability pipeline enhanced pedagogical value by producing student-only, word-level rationales that identified rubric-aligned, score-raising, and score-lowering terms. These lexicons serve as lightweight concept inventories, reinforcing fairness and trust through explanatory symmetry between model reasoning and rubric expectations.