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 HyperGraph Attention Network (HyperGAT) with BERT for enhanced semantic representation; (2) a Reference Answer Set (RAS) constructed via clustering of full-score answers; and (3) Siamese Neural Networks (SNNs) for similarity-based scoring. Experiments on the Ohsumed and ASAP-5 datasets demonstrate that (i) HyperGAT-BERT achieves 72.95% accuracy on Ohsumed text classification, outperforming baseline HyperGAT by 3.28%, and (ii) the full HyperGAT-BERT-RAS achieves 78.66% accuracy and 0.7806 F1-score, with RAS contributing the most to performance gains. These improvements suggest the potential for more reliable scoring of diverse student answers, reduced teacher grading burden, and enhanced feasibility of AI-assisted formative assessment in real classrooms, although empirical validation with teachers and students is needed.