Context-Aware Bias Detection for Language Using Contrastive Learning

Abstract

Identifying implicit bias requires moving beyond word-level analysis to model deep contextual relationships across a corpus. We introduce a contrastive learning methodology specifically designed to enhance bias detection through the use of graph-based sentence embeddings. This approach effectively models multi-level contextual dependencies, leading to improved classification accuracy and a powerful capability to discern subtle framing techniques. Empirical validation demonstrates state-of-the-art performance against conventional NLP models. Our results affirm the efficacy of incorporating graph attention mechanisms for robust bias analysis, providing a key contribution to the development of reliable automated systems for implicit bias detection in media.