Automated Essay Scoring Using Large Language Models

Automated Essay Scoring (AES) faces challenges due to essay evaluation subjectivity. This study uses multi-task learning to predict six analytic metrics simultaneously, testing various deep learning models. RoBERTa performed best, with masked language models outperforming generative ones. Pre-trained RoBERTa encodings with a simple neural network yielded optimal results. Grammar and cohesion were the most difficult metrics to predict accurately across all models.