An Automated Ensemble Learning Framework using BERT for Critical Software Bug Reports

Abstract

In order to prioritize important concerns in large-scale software development and increase the effectiveness of bug triage, automated classification of software bug reports is essential. Though there are still issues with robustness, class imbalance, and adaptability to various software settings, recent research has shown that transformer-based language models like BERT are excellent at capturing the contextual semantics of bug report content. This paper suggests an automated ensemble learning framework using BERT for significant software issue reports, motivated by the future research directions found in earlier BERT-based bug classification studies. In order to improve prediction stability and lessen sensitivity to noisy and unbalanced data, the suggested method uses BERT to create deep contextual representations from bug report titles and descriptions. These representations are then processed by an ensemble of machine learning classifiers. The system enhances recall and reliability in identifying important bug categories including security and performance issues by combining ensemble learning with transformer-based feature extraction. Since the system is completely automated, human involvement in prediction is not necessary. When compared to single-model techniques, the suggested framework provides better and consistent classification performance, according to experimental results on real-world bug report datasets. This study allows additional extensions like multimodal inputs and domain-specific adaption and shows a scalable and extensible foundation for future automated bug triage systems.