The Impact of Transformer Models and Regression Algorithms on Automated Short Answer Grading

Grading short answer questions in higher education often poses significant challenges due to its subjective nature and time-consuming process. As educational institutions expand, the need for efficient and objective grading solutions becomes critical. This paper investigates the use of machine learning and natural language processing to automate the evaluation of short answer responses. Our study employs a two-phase methodology: first, several transformer models—including BERT, Sentence-BERT, T5, Genism, and XLNet—are used to transform text into numerical representations. In the second phase, various regression techniques, such as Linear Regression, Support Vector Machine (SVM), Random Forest, and AdaBoost, are applied to these representations to estimate scores. Our findings indicate that the combination of the BERT model with Random Forest performs the best, demonstrating superior accuracy with the lowest error metrics. This optimal model pairing is proposed for integration into the Ample Learning Management System (LMS) to enhance grading efficiency and provide quicker, more reliable feedback. The significant implications of this research provides teachers with a robust tool for automated grading, which can enhance the educational experience by delivering timely and accurate evaluations.