Duplicate bug report detection using pre - trained language models

Abstract

Software testing and defect reporting are significant factors of software development and maintenance. Defects are identified and reported in a bug tracking system like JIRA, or Bugzilla. Those reported defects are further triaged by an expert who has an understanding of the repository, system, and developers and assigns them to the developers to fix them. During this defect reporting there can be duplicate bugs reported and identifying duplicate bugs is a crucial task. Manual labeling of duplicate defects is time-consuming, may identify defects as duplicate bug reports, and also increases the cost of software maintenance. Therefore automated duplicate bug report detection is very significant. This research proposes a duplicate bug report classification methodology that leverages the Pre-trained language models BERT and XLNet with Multi-Layer Perceptron as the Deep Learning classifier for duplicate bug detection. We tested on publicly available datasets related to Eclipse, NetBeans, and OpenOffice bug reporting datasets. The selected models were shown to outperform the previously proposed systems for the same task. Among them, the approach used with BERT embeddings has shown the best results. Further experiments showed that BERT is capable of domain adaptation –meaning that even when the BERT was finetuned with different bug report datasets, it is still capable of detecting duplicate bugs

in

an unseen dataset. Finally, a multi-stage classification was done using a Convolutional Neural Network model and a BERT model using Eclipse and NetBeans datasets and a combined dataset of Eclipse and NetBeans. The approach used with the combined dataset has outperformed the baseline approach.