Influence of Contrastive Learning on Source Code Plagiarism Detection through Recursive Neural Networks

Plagiarism during programming assignments is a problem in academia. It hinders the ability of academic instructors to truly judge students’ performance and thus, prevents students from receiving adequate help from their instructors. In cases where the number of code submissions for a particular assignment is relatively small, the instructor can inspect each code submission to determine whether they are similar. But as the number of code submissions grows, it becomes difficult to detect similarities between them. Therefore, this induces the need for an automatic source code plagiarism detector. Previous studies showed that we could use the abstract syntax tree (AST) of a source code to get an accurate representation of the source code for neural network computations. Although a study even presented a recursive artificial neural network named Abstract Syntax Tree-based Neural Network (ASTNN) that could represent source codes into vector embeddings using their ASTs, it does not use contrastive learning paradigms, shown to increase the performance of Siamese networks in similarity detection tasks. Therefore, this paper presents an improved version of the ASTNN for code clone detection, where we modify the original model for contrastive learning. Experiments demonstrated that we outperform the original ASTNN model in code clone detection tasks, with a+5% improvement in the F1-score of our model. This study aims at improving the way we perform similarity detection tasks involving programming languages.