CONV_GBERT: Leveraging BERT–GPT Fusion and IMP-FGSM for Adversarially Robust Sentiment Classification

This paper introduces CONV_GBERT, a hybrid deep learning framework for robust sentiment classification. The model integrates the bidirectional contextual understanding of BERT, the generative capabilities of GPT, and the local feature extraction strength of convolutional neural networks. By concatenating embeddings from BERT and GPT and passing them through a convolutional layer, CONV_GBERT captures both global semantic meaning and fine-grained sentiment patterns in text. To further enhance robustness and generalization, we incorporate Improved Fast Gradient Sign Method (IMP-FGSM) for adversarial training. IMP-FGSM introduces controlled perturbations to the input embeddings and iteratively refines them using gradient-based updates. This technique improves the model’s resistance to adversarial attacks while preserving valid token boundaries. Experiments on benchmark sentiment datasets demonstrate that CONV_GBERT significantly outperforms baseline models. It achieves an accuracy of 84.89%, precision of 83.67%, recall of 82.88%, and an F1 score of 83.37%. These results highlight the potential of combining multiple transformer-based architectures and adversarial training to advance sentiment analysis in complex, noisy, and dynamic textual environments.