Clustering and Visualizing Programming Solutions: A Transformer-Based Analysis Using CodeBERT and CodeT5+

The evaluation of student code submissions is difficult due to the wide variety of syntax and logic applied to solve a similar problem. This research exploits transformer models, CodeBERT and Code'I'Ss-, to produce vector embeddings that capture semantic similarity while accounting for syntactic differences. By applying dimensionality reduction techniques such as UMAP and t-SNE, along with clustering methods like DBSCAN, the system can visualize and group similar solutions, facilitating easier evaluation of multiple submissions by educators. This approach supports the automation of feedback generation and enhances code evaluation, thereby making the grading process more scalable and insightful.