BTAL: An imbalance software bug report triage approach based on BERT-TextCNN

Abstract

With the expansion of software project scale, a large number of software bug reports have been generated. Bug triage is an indispensable task in software development and maintenance, which directly affects the efficiency of software bug fixing and maintenance cost. Existing bug triage methods often fail to make full use of the useful information in defect reports, resulting in the neglect of some auxiliary information that is critical to the defect triage task. Meanwhile, they also fail to take into account the uncertainty of developers’ work and the gap in their activity, thus leading to the imbalance in software defect report datasets. To address this issue, we propose an imbalance software bug report triage method, BTAL, based on BERT-TextCNN. Firstly, the method utilizes multiple information from software bug reports and employs BERT and TextCNN models for vector representation and feature extraction. Then, the output feature vectors are input into the softmax function to obtain the probabilities of bug reports being triaged to developers. To solve the problem of imbalanced datasets, we propose an adaptive loss function that can adaptively adjust the loss weights based on different categories of samples. This helps the network reduce its focus on majority classes and increase its focus on minority classes, thereby improving triage accuracy. Experimental results on five large-scale open-source software projects, namely GCC, NetBeans, Eclipse, Mozilla, and OpenOffice, demonstrated the effectiveness of the BTAL method in solving software bug triage problems, outperforming current state-of-the-art models.