The exponential growth of social media platforms has made sentiment analysis a crucial tool for understanding public opinion, enhancing customer experience, and guiding organizational strategies. While transformer-based models such as BERT have set new benchmarks in sentiment classification, their lack of interpretability limits their applicability in domains requiring transparent decision-making. This study addresses this challenge by proposing a novel sentiment analysis framework that combines the predictive power of BERT with a dual-explainability mechanism integrating LIME and attention mechanisms. The proposed framework provides both localized and global explanations, enabling token-level interpretability and offering actionable insights into sentiment trends. Evaluated on the Twitter US Airline Sentiment dataset, the model achieves a classification accuracy of 92.0%, demonstrating its ability to effectively analyze noisy and unstructured social media data. The dual-explainability approach ensures transparency by highlighting influential tokens in individual predictions while identifying broader sentiment patterns across datasets. This combination of high accuracy and interpretability makes the framework particularly suited for real-world applications such as customer feedback analysis and brand monitoring. Future extensions of this work include addressing dataset imbalances, optimizing for real-time sentiment monitoring, and exploring multilingual applications to further broaden the framework’s impact.