<p>Classification of bug report is of critical importance in software maintenance since it helps one to triage, prioritize and perform the task of fixing the bug in software. Typical bug handling methods are heavily labor intensive, and therefore, slow, error prone and near impossible to scale up in modern development environments with the creation of huge volume of reports. However, many existing solutions have inherent drawbacks, such as based on shallow textual features, ineffective dealing with class imbalance, and weak generalization ability in cold start. To address these challenges in this paper, an integration of the strengths of two advanced Transformer architectures, namely, RoBERTa (Robustly Optimized BERT Pre-training Approach) and XLNet, is proposed as hybrid deep learning model for feature extraction followed by Random Forest (RF) classifier for robust prediction. Thus, the dual-transformer design allows the model to learn richer semantic and syntactic features. RF helps stabilize the decision and makes the system more interpretable. In order to evaluate experimentally, we culled over 2000 project bug reports from the Mozilla and Eclipse projects through Bugzilla with varying category and priority level. In order to tackle the class imbalance of the dataset, it was first balanced using SMOTE (Synthetic minority oversampling technique). Retailers can predict the next Bug category or Bug priority with an F1 score of up to 98.82% for prioritization and 98.40% for categorization, which is an order of magnitude higher than classical machine learning baseline methods.</p>