ARB-BERT: An Automatic Aging-Related Bug Report Classification Method based on BERT

Abstract

Software aging is a common phenomenon in most software systems. It refers to the increase of failure rates or the decline of performance in a long-running software system, mainly caused by Aging-Related Bugs (ARBs). Failure incurred by software aging may cause economic loss and may lead to casualties in security-critical systems. Automatic classification of ARB reports is an effective method to ensure the software system's quality by helping us detect and fix the bugs in software systems. There are two challenges in the automatic classification of ARB reports at present. Firstly, it is difficult to distinguish the ARB reports since the semantics of the text is ambiguous and hard to be recognized; secondly, the number of ARB reports is much smaller than other types of bugs, which may lead to class imbalance. Therefore, An ARB Report Automatic Classification Method based on BERT(ARB-BERT) is proposed in this paper to alleviate these problems. We combined back-translation, random under-sampling and random over-sampling to reduce the class imbalance problem in classification. By considering the ARB reports' characteristics, we utilize BERT as the semantic model, extracting feature vectors containing more accurate and sufficient information. The experimental results show that our method can improve the accuracy, precision, F-measure, and recall value compared with previous methods.