Leveraging Distance-Optimized Transformers for High-Performance Arabic Short Answers Grading

This study presents comprehensive distance-optimized transformer architecture for Automated Arabic Short Answers Grading (AASAG) that systematically evaluates multiple semantic similarity measures. Short answer grading—assessment of responses typically 1-3 sentences long requiring conceptual understanding rather than factual recall—poses significant challenges in Arabic due to morphological complexity and limited computational resources. Our approach integrates pre-trained Arabic transformer models (AraBERT v02) with four distinct distance algorithms: cosine similarity, Manhattan distance, Euclidean distance, and dot-product calculations within a Siamese network architecture. Through systematic evaluation across three progressively enhanced datasets (original AR-ASAG, SemEvalaugmented, and reference-integrated versions), our distance-optimized approach achieves state-of-the-art performance with correlation coefficients of 0.7998, representing a 5.5% improvement over existing methods. This advancement significantly outperforms traditional vector space models (0.7037 correlation), BERT-based approaches (0.7616), and hybrid semantic analysis methods (0.745), establishing new benchmarks for Arabic educational assessment technology.