Bug Triaging Based on Transformer Models Utilizing Commit Messages

Abstract

Bug triaging is a crucial process in software maintenance involving assigning bug reports to the appropriate developers for resolution. Automated bug triaging eliminates the need for manual assignment by leveraging machine learning (ML), information retrieval (IR), and graph neural networks (GNN) techniques to classify bug reports based on extracted features. These approaches typically utilize bug report fields and metadata as training features. However, models often underperform when relying heavily on a few specific fields, such as the bug report summary and description. To address this issue, we integrated bug commit messages into the feature set to enhance the efficiency of bug-triaging models. Our study demonstrates significant improvements in triaging accuracy by incorporating commit messages. In the Firefox dataset, prediction accuracy increases from 23.28% to 69.45%. Similarly, in the Eclipse dataset, accuracy rises from 70.01% to 89.28%. These results prove that including commit messages can significantly improve bug-triaging accuracy.