BERT-Enhanced HyperGAT with Siamese Networks and Reference Answer Set for Automated Short-Answer Scoring

This paper proposes a novel framework, HyperGAT-BERT-RAS, that integrates: (1) a Hy perGraph Attention Network (HyperGAT) with BERT for enhanced semantic representa-tion; (2) a Reference Answer Set (RAS) constructed via clustering of full-score answers; (3) Siamese Neural Networks (SNNs) for similarity-based scoring; and (4) GPT-4-based data augmentation to address class imbalance. Experiments on the Ohsumed and ASAP-5 da-tasets demonstrate that: (i) HyperGAT-BERT achieves 0.7317 accuracy on Ohsumed text classification, outperforming baseline HyperGAT by 2.69%; (ii) the full Hyper-GAT-BERT-RAS achieves 0.7991 accuracy and 0.7956 F1-score, with RAS contributing the most to performance gains (4.34% accuracy drop when removed); (iii) GPT-4 augmentation improves Quadratic Weighted Kappa from 0.584 to 0.880 and minority-class (scores 2–3) F1 by 15.3%. These improvements translate into more reliable scoring of diverse student answers, reduced teacher grading burden, and enhanced feasibility of AI-assisted forma-tive assessment in real classrooms. Ablation and error analyses confirm the contribution of each component. The framework advances ASAG by synergizing graph-based relational modeling, pretrained language understanding, and knowledge-guided scoring.