CB2VS-Net: A Transformer-Augmented CNN for Sentiment Analysis

Sentiment Analysis helps organizations and businesses in understanding public opinion and feeling towards their brand, services, and so on. This paper introduces "CB²VS-Net" - a streamlined modern hybrid model that integrates Transformer-augmented Convolutional Networks, Deep Vector Quantization, and Polynomial Kernel SVM for enhanced sentiment analysis. The model begins by converting text into subword tokens using a BERT tokenizer and these are passed into the ConvBERT module for extraction of both local semantic features and global contextual patterns. The high-dimensional embeddings are then encoded into compact, fixed-length vectors through a VG-VAE encoder effectively for upgrading the traditional Bag of Visual Words representation. The resulting discrete histograms are classified using a Polynomial SVM for capturing nonlinear sentiment boundaries. This simple design avoids complex architectural stacking while ensuring high fidelity and contextual depth. The proposed CB²VS-Net is evaluated on the IMDb dataset from Kaggle showing improved accuracy of 91.3% over baseline CNN, BoVW, and standalone BERT-SVM models. Further, the approach achieved 90.2% in precision, 88.5% in recall, and 89.3% in F1-score. Overall, the proposed framework is robust, modular, and suitable for real-world NLP applications under varying domain conditions.