Towards Transparent and Context-Aware Automated Essay Scoring

The evaluation of descriptive answer scripts in examinations is both time-consuming and prone to human bias. Although advancements in Natural Language Processing (NLP) and Artificial Intelligence (AI) have enhanced Automated Essay Scoring (AES) systems, significant challenges remain, including inter-rater variability, intra-rater inconsistency, and the difficulty of assessing subjective responses. Existing AES models, developed over more than fifty years, predominantly rely on student-to-student answer comparisons, which can lead to biased outcomes such as over-scoring incorrect answers and undervaluing correct ones. In contrast, course-based examinations require student responses to align with teacher-defined expectations, ensuring fairness, efficiency, and clarity in grading. In response to this gap, we present Reference-based AES (RAES), a framework that assesses student digitalized answers (SDAessay) through direct comparison with teacher-provided reference answers. Through ablation studies, the framework demonstrates impartial and meaningful evaluation while reducing biases inherent in conventional AES approaches. The findings further reveal limitations of reference-free AES models, which can inaccurately assign high scores to incorrect responses, underscoring the need for expectation-driven evaluation in educational assessment.