<p>Sentiment analysis plays a vital role in understanding public opinions and emotions toward products, services, and events. Hence, this research aims to present an innovative approach to sentiment analysis that combines heuristic optimization with a hybrid model architecture. The proposed methodology consists of three key stages such as preprocessing, feature extraction, and classification. In the first stage, the input text is preprocessed using techniques such as tokenization, stemming, and removal of stop words. In the next stage, relevant features are extracted from the preprocessed outcomes, including Improved Lexicon-based Bidirectional Encoder Representations from Transformers embeddings, Term Frequency-Inverse Document Frequency, and aspect term extraction-based features. These features capture both the semantic and contextual information crucial for effective sentiment analysis. Subsequently, a hybrid model is introduced, combining the Modified LinkNet architecture with convolutional neural networks. This hybrid model receives the extracted features and ensures accurate sentiment detection through proper training. Additionally, the model’s weights are optimized using a novel heuristic method called the Self-Improved Pufferfish Optimization Algorithm, which iteratively refines the model’s parameters for improved performance in sentiment analysis tasks. The usefulness of the suggested model is confirmed by comprehensive evaluation on benchmark sentiment analysis datasets. The suggested method outperformed current approaches, achieving an accuracy of 0.952 and an F-measure of 0.928. Overall, this study advances sentiment analysis techniques and provides insightful information on textual sentiment analysis.</p>