Sentiment analysis is a key task in natural language processing (NLP) that focuses on identifying the emotional polarity of textual data. While transformer-based models such as BERT have achieved remarkable performance by generating contextualized word representations, they may not fully capture sequential dependencies in long text sequences. To address this limitation, this paper proposes a hybrid deep learning framework that combines a pre-trained BERT encoder with a Bidirectional Long Short-Term Memory (BiLSTM) network. The BERT component extracts rich contextual embeddings, which are further processed by the BiLSTM to model temporal dependencies. The model is trained using binary cross-entropy loss with label smoothing and optimized using the Adam optimizer with cosine annealing scheduling. Additional techniques such as dropout regularization and gradient clipping are applied to enhance generalization and training stability. Experimental results on the IMDB movie review dataset demonstrate that the proposed model achieves an accuracy of 88.98%, along with strong precision, recall, and F1-score. Further evaluation using ROC-AUC and Precision-Recall curves confirms the robustness and effectiveness of the approach.
Keywords: BERT, BiLSTM, IMDB dataset, deep learning, natural language processing, sentiment analysis
