TSD_XAI: Integrating XAI with Transformer Networks for Interpretable Sarcasm Detection

Sarcasm, marked by a discrepancy between literal expression and intended meaning, poses a significant challenge for automated natural language understanding systems. Traditional sarcasm detection models often rely on surface-level textual features and struggle to capture the deeper contextual nuances required for accurate classification particularly in short texts such as news headlines. To address this limitation, we propose a Transformer-based deep learning model specifically designed for sarcasm detection. Experiments conducted on the News Headlines Dataset for Sarcasm Detection demonstrated that our model achieved an accuracy of 94%, outperforming conventional machine learning baselines, including Logistic Regression (83 %), Naive Bayes (81%), K-Nearest Neighbors (70%), Decision Tree (93%), and Support Vector Machine (93%). Model performance was comprehensively evaluated using ROC curves, confusion matrices, and accuracy-loss plots to ensure robustness. To enhance interpretability and user trust, we incorporated Explainable AI (XAI) techniques through LIME (Local Interpretable Model-agnostic Explanations). LIME provided token-level attribution insights, highlighting key textual features that influenced each prediction. This approach not only increased transparency but also offered valuable understanding of model behavior in am-biguous cases.