Factors Influence Cross-Prompt Scoring of Arabic Essays

الملخص

Cross-prompt automated essay scoring (AES), which involves training on specific prompts and testing on unseen cases, represents a practical application scenario; however, this area remains underexplored for Arabic. Progress in AES is typically characterized by a succession of neural architectures that exhibit escalating levels of complexity. However, such progress is seldom evaluated against more basic choices such as tokenization, model-selection criteria, and text segmentation. We present the first controlled factorial study of cross-prompt Arabic AES on the LAILA dataset, crossing three architectures of increasing complexity (flat, hierarchical, multi-trait), three tokenizer–model configurations, and three random seeds, and benchmarking against a strong prompt-agnostic feature baseline. Our central finding is that tokenizer configuration drives performance far more than architecture: it outperforms each architectural step by approximately four times, covering a range of 0.16 to 0.19 in the overall Quadratic Weighted Kappa (QWK), the agreement metric used throughout. By contrast, the move from a flat to a hierarchical encoder yields only a small and seed-fragile gain (+0.045), multi-trait attention is mildly harmful (−0.02), and flat AraBERT matches the feature baseline (0.621 QWK) — which itself outperforms the hierarchical and multi-trait models in four of six configurations. We additionally provide evidence on how a seemingly straightforward "more-complex-is-better" result (M2 = 0.637 vs. M1 = 0.556 in our own early runs) is an artifact that dissolves once tokenization, selection criterion, and segmentation are each controlled. We release a corrected evaluation protocol — mean-per-prompt checkpoint selection and a sentence-based segmenter suited to Arabic — and argue that such controls are prerequisites for credible architecture claims in low-resource cross-prompt AES.