bjXnet: an improved bug localization model based on code property graph and attention mechanism

Abstract

Bug localization technologies and tools are widely used in software engineering. Although state-of-the-art methods have achieved great progress, they only consider the source code information at the text level, which may establish a wrong correlation between the source code and the bug report, affecting the localization accuracy and reliability. In this paper, we propose an improved bug localization model, which uses the semantics of source codes at the graph level to supplement its semantics at the text level, optimizing and adjusting the graph semantics in combination with the attention mechanism to obtain the code semantic feature including the shallow and deep semantics of the source code. Finally, the correlation between code semantic feature and report semantic feature is measured by cosine similarity. We conduct experiments on three open source Java projects to comprehensively evaluate the performance of proposed model. The experimental results show that the model is significantly better than state-of-the-art methods.