Abstract:
"In this thesis, we present a novel deep-learning approach to bug localization using neural
networks, by employing the pre-trained model, GraphCodeBERT and training it with the
XLCoST dataset. The primary goal of this research is to improve the efficiency and accuracy of bug localization in software development, which is a critical aspect of the software maintenance process. The suggested method aims to address the drawbacks of conventional Information Retrieval (IR) and Convolutional Neural Network (CNN) based approaches, which often struggle to accurately identify the location of bugs within a large codebase.
To accurately capture the complex links between code snippets and their related descriptions, our model combines the GraphCodeBERT transformer, a cutting-edge pre-trained model, with cross attention mechanisms. In addition, we incorporate a customized weighting scheme that further refines the model's ability to identify relevant code components associated with specific bugs. The model's ability to concentrate on the most important features of the code is improved by this weighting method which enhances the model's capacity to focus on the most pertinent aspects of the code.
"
