Source Code Classification using Neural Networks

Abstract

The purpose of this dissertation is to study the problem of source code classification using neural networks. More specifically, in this problem, a piece of code is classified into an algorithmic class based on the function it performs. Pre-existing research has shown that neural networks are an effective way of modeling source code and solving such classification problems. Although literature results are encouraging, there are limitations related not only to datasets and preprocessing techniques but also to machine learning models. To this end, we propose a system that initially builds quality datasets, which are free of biases and noise. It then uses compilers to process these sets and finally uses neural networks to classify them into an algorithmic class. In the context of optimizing the system above, we studied a variety of pre-processing techniques and machine learning models.