AUTOMATED ESSAY SCORING FOR STUDENT EXAMS USING DEEP NLP MODELS

Abstract

The increasing use of essay-based examinations in higher education has created significant challenges in maintaining efficient, objective, and consistent assessment processes. Manual essay grading is time-consuming and susceptible to subjective judgment, particularly when evaluating large numbers of student responses. Therefore, this study aims to develop and evaluate an Automated Essay Scoring (AES) system based on Bidirectional Encoder Representations from Transformers (BERT) to improve the accuracy and consistency of student essay assessment. This research employed an experimental quantitative approach using 2,500 student essay responses, of which 2,340 valid responses were retained after preprocessing and data cleaning. The dataset was divided into training, validation, and testing subsets using a 70:15:15 ratio. The proposed model was fine-tuned using the AdamW optimizer with a learning rate of 2 × 10⁻⁵, a batch size of 16, and 8 training epochs. Model performance was evaluated using Quadratic Weighted Kappa (QWK), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). The experimental results demonstrate that the proposed BERT model achieved a QWK score of 0.872, indicating strong agreement with human evaluators, while obtaining an MAE of 0.418 and an RMSE of 0.593, reflecting relatively low prediction errors. Comparative evaluation also showed that the proposed BERT model outperformed conventional baseline approaches in automated essay scoring, confirming the effectiveness of contextual language representations for understanding semantic information in student essays. These findings indicate that the proposed framework provides a reliable and efficient solution for automated essay assessment, offering practical benefits for improving scoring consistency, reducing lecturers' workload, and supporting the implementation of intelligent assessment systems in higher education.