Explainability in Transformer-Based Sentiment Analysis: A Survey of Methods, Applications, and Open Challenges

Transformer-based models like BERT have pushed sentiment analysis to impressive levels of accuracy. But accuracy alone is not enough especially when these models are deployed in hospitals, content moderation teams, and mental health platforms where a wrong prediction can cause real harm. The core problem is straightforward: these models do not explain themselves. They give you an answer but not a reason. This survey examines how Explainable AI (XAI) techniques are being applied to transformer-based sentiment analysis across three domains social media, product reviews, and healthcare asking whether current methods are adequate for the stakes involved. We organize existing XAI approaches into a structured taxonomy covering attention visualization, gradient-based methods, perturbation approaches (LIME and SHAP), rationale extraction, counterfactual explanations, and natural language justifications. To support these observations empirically, we present an original experiment applying LIME to both a TextBlob classifier and a DistilBERT model on fifteen challenging social media posts covering sarcasm, slang, negation, mixed sentiment, and masked distress. Five open challenges are identified and directions for future research proposed