Ensemble Transformer Models for Github Issue Classification: a Bert-Electra Hybrid Approach

Managing GitHub issues such as bugs, feature requests, and questions is a critical yet time-consuming task when performed manually. To automate this process, an NLPdriven classification system is proposed, employing an ensemble of transformer models (BERT and ELECTRA) combined with a neural network meta-classifier. This hybrid approach addresses BERT's limitations in classifying 'Questions,' improving accuracy from 0.599 to 0.648. The proposed system achieves a precision of 0.888 and an F1-score of 0.888, with particularly strong performance in “Feature Requests” (precision: 0.913) and significant gains in “Questions” (F1score: 0.706). Notably, it preserves BERT's robustness for “Bugs” and “Feature Requests“ while ensuring balanced classification across all categories. The results demonstrate that combining transformer models with a meta-classifier optimizes accuracy and reliability for issue triage. By automating this process, the proposed solution enhances workflow efficiency, reduces manual effort, and improves resource allocation for development teams, offering a scalable approach to software maintenance challenges.