Publication

Abstract : The identification and mitigation of biases in textual data are essential for ensuring fairness, transparency, and equity across diverse Natural Language Processing (NLP) applications. To address these challenges, we propose a scalable framework for bias detection using a Transformer-based Generative Adversarial Network (TransGAN), which uses a hierarchical structure for capturing explicit and implicit biases. The framework integrates Swish activation function for improved gradient flow, and Explainable Artificial Intelligence (XAI) techniques, such as attention visualization and SHapley Additive exPlanations (SHAP) values to provide interpretable decision-making insights. The adversarial training enhances discriminator performance in identifying nuanced bias patterns through deep contextual modeling of textual relationships, while the generator synthesizes contextually-aware biased samples for robust decision boundary learning for bias classification. The proposed model is validated through extensive experiments, which demonstrate better performance with an accuracy of 91.2% for unbalanced data and 95.3% for balanced data, compared to traditional machine learning and transformer-based methods, while providing transparent explanations crucial for ethical AI deployment in various NLP applications.