<![CDATA[Essay assessment remains a persistent challenge in many schools due to the time-consuming nature of manual grading and the variability that arises from assessor subjectivity. This situation highlights the need for an automated scoring system capable of producing fast, consistent, and teacher-comparable evaluations. This study proposes a hybrid approach for automatic essay scoring by combining semantic similarity from Sentence-BERT with lexical features derived from Jaccard Similarity, Levenshtein Similarity, Keyword Coverage, and Length Penalty. The five similarity components are integrated using a weighted aggregation scheme and calibrated to the 0–100 scoring scale through linear regression. The model was tested on a dataset of high-school essay responses accompanied by manual teacher scores. Experimental results indicate that the proposed system performs reliably, achieving a Mean Absolute Error (MAE) of 3.58 and a Root Mean Square Error (RMSE) of 4.48 on the test set. The model also demonstrates strong practical alignment with teacher scoring, reaching an agreement rate of 87.30% within a tolerance of ±7 points. These findings suggest that the hybrid method can approximate human scoring patterns with a high degree of consistency, providing a promising tool to support objective and efficient assessment processes in educational settings.]]>
Copyrights © 2026
