Enhancing LSTM performance in sentiment analysis through advanced data preprocessing and model optimization techniques

This study explores enhancing the performance of Long Short-Term Memory (LSTM) networks in sentiment analysis by integrating advanced data preprocessing techniques and hybrid model architectures. A robust preprocessing pipeline was implemented, involving tokenization, normalization, slang handling, and dataset balancing to improve data quality. A CNN-GloVe-LSTM hybrid model was developed, leveraging GloVe embeddings for semantic representation, CNN for local feature extraction, and LSTM for sequential dependency learning. The study also examined an ensemble of LSTM and Random Forest models. Performance metrics, including accuracy, precision, recall, F1-score, and AUC-ROC, were used for evaluation. Results indicate that the CNN-GloVe-LSTM model achieved the highest accuracy (92.05%) and computational efficiency, outperforming both the standalone LSTM and ensemble approaches. The hybrid model demonstrated a significant reduction in training time while maintaining robust classification capabilities, making it a superior choice for sentiment analysis tasks on social media data.