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

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Educational Technology

Background:

• Automated grading systems require robust methods for semantic understanding and similarity assessment.
• Existing approaches may struggle with the diversity of student responses and the nuances of grading.

Purpose of the Study:

• To propose and evaluate the HyperGAT-BERT-RAS framework for enhanced automated scoring of student answers.
• To improve the accuracy and reliability of AI-assisted grading systems.

Main Methods:

• Integration of HyperGraph Attention Network (HyperGAT) with BERT for superior semantic representation.
• Development of a Reference Answer Set (RAS) using clustering of high-scoring answers.
• Application of Siamese Neural Networks (SNNs) for similarity-based scoring.

Main Results:

• HyperGAT-BERT achieved 72.95% accuracy in Ohsumed text classification, surpassing baseline HyperGAT by 3.28%.
• The complete HyperGAT-BERT-RAS framework reached 78.66% accuracy and a 0.7806 F1-score on ASAP-5.
• The Reference Answer Set (RAS) was identified as a key contributor to performance improvements.

Conclusions:

• The HyperGAT-BERT-RAS framework shows significant potential for reliable scoring of diverse student answers.
• This AI-driven approach can reduce teacher grading burden and enhance formative assessment feasibility.
• Further empirical validation with educators and students is recommended.