A Deep Learning Approach to Locate Buggy Files

Bug localization, using a bug description to locate potential buggy source code files automatically, can assist developers to quick fix bugs and relieve maintaining costs for industry. Traditional models are mainly based on information retrieval (IR) techniques, employing the lexical similarity between code and text to locate potential buggy files. Recent researches involve deep learning techniques, utilizing a neural network based encoder to extract semantics from both code and text. However, rather than natural language, code contains not only semantic information, but also important relations among classes, functions, and variables. For this purpose, we introduce a deep learning approach with various embedding methods for bug localization, and it achieves competitive performances on two public datasets.