Improving Automatic Essay Assessment Through Cosine Similarity Leveraging a Semantic Corpus

Automatic Essay Scoring (AES) represents an effective solution for facilitating automated evaluation of written essays by mitigating evaluator subjectivity and accelerating the assessment process. Nonetheless, a persistent challenge lies in achieving high accuracy due to limitations in semantic understanding. This study employs Cosine Similarity as a baseline approach and further integrates a semantic data corpus to enhance the representation of textual meaning. Empirical results demonstrate that relying solely on Cosine Similarity captures predominantly lexical-level similarities, yielding limited correlation with human scoring. The incorporation of a semantic corpus substantially improves the system's capacity to recognize synonyms and linguistic variations, thereby enhancing the sensitivity and reliability of the scoring process. Despite these improvements, the findings underscore the necessity for further corpus refinement, evaluation on larger and more diverse datasets, and assessment using multiple correlation metrics. Overall, this study provides a substantive contribution to the development of AES systems that are more accurate, consistent, and closely aligned with human judgment, thereby advancing the field of automated educational assessment.