Sentiment analysis has become an invaluable exercise on social media addressing the understanding of opinion and consumer behavior and trends. The system of informality and context richness against platforms, such as Twitter, raises serious issues to current sentiment classification models. It is hypothesized that integration of the contextual feature representation capacity of embeddings in transformers with the sequential learning ability of BiLSTM networks in DitilBERT-BiLSTM can help to improve accuracy in sentiment classification. The Sentiment140 dataset is used as a training and testing set with 1.6 million small tweets among which there are equal numbers of positive and negative tweets. Experiments demonstrate that the proposed model obtains superior performance compared to traditional deep learning architectures, 81% accuracy, 82% precision, recall of 80% and F1-score of 81%. DistilBERT-BiLSTM shows good generalization and learns efficiently using less computer power than full-scale transformer models, indicating that it is more suitable to applications of real-world social media sentiment analysis. This study points out the success of hybrid deep learning networks in modelling both global semantic structure and local sequential relations as a strong remedy to resolving opinion mining in dynamic and noisy digital situation.