A Two-Phase Bug Localization Approach Based on Multi-layer Perceptrons and Distributional Features

Bug localization is a challenging and time-consuming task of the process of bug fixing and, more in general, of software maintenance. Several approaches have been proposed in the literature which support developers in this task by identifying source code files in which the bug is likely to be located. However, the research on this topic never stopped, looking for new methods providing better accuracy and/or better efficiency. In this paper, we propose a two-phase bug localization approach which leverages multi-layer neural networks and distributional features. First phase locations are obtained thanks to a neural network trained on word embeddings representations of fixed bug reports. The second phase refines bug locations taking into account the number of times source code files co-occur in fixed bug locations. To evaluate the approach, we conducted a large-scale experiment on five open source projects, namely Mozilla, Eclipse, Dolphin, httpd, and gcc. Results show that, thanks to pre-trained word embeddings, we were able to implement a scalable approach with a training running time of few hours on large datasets. Performances are comparable to other existing deep learning approaches.