Automated Assessment of Java Code Using CodeT5+, Early Fusion, and CatBoost Regressor

With the increasing number of students pursuing education, the workload on educators with respect to time consumed on evaluation, has significantly risen. Manually assessing student-written code is a resource-intensive task. To address this issue, the study has developed a model which takes code as input and processes it. Further, regression models are trained to predict scores for the code, providing an automated and scalable solution for large-scale grading systems. The best model achieves an R² score of 0.56, demonstrating its improvement over previous methods and its potential for real-world educational applications. Additionally, an analyze offering insights into code’s contribution to the model’s performance is performed.