Avaluació automàtica de codi font fent servir tècniques de deep learning

Abstract

[en] This degree thesis focuses on the potential automation in assessing algorithmic exercises in Python using "Code Embeddings" and Deep Learning with Neural Networks. Our hypothesis is based on the idea that the embedding generated from a student's exercise will have a distance from the embedding of the most efficient possible solution, and based on this distance, a grade can be generated for the exercise. By training this neural network with various exercises and expected grades, we hope to reach a point where the grades proposed by it are similar to those a teacher would assign when correcting exercises, thereby reducing the workload of grading numerous exercises for a human. One of the crucial stages in calculating this distance between the code embeddings is the generation of these embeddings, which have been created using a code transformer model called CodeT5. The research and tests conducted suggest a potential reduction in the grader's workload, albeit with the need to train the neural network with a substantial amount of data to enhance predictions and outcomes when employing this technique alongside others to refine the grading system for automation.