BAAF: A Framework for Media Bias Detection

Soumyadeep Sar, Subinay Adhikary, Dwaipayan Roy

Abstract: Social media platforms are creeping with different types of bias, such as political, gender, cognitive, text-level context, and racial biases. Addressing these biases requires specialized methods tailored to the nuances of each type, making it a very complex problem. In this paper, we propose a lightweight, unified solution based on supervised learning to tackle multiple social media biases. We leverage AdapterFusion, a non-destructive way of combining knowledge of different domains for multi-task learning. Additionally, we also explore different frameworks for in-context learning (ICL) with Llama-3.1-70b, utilizing zero-shot, few-shot, and prompt optimization on zero-shot prompts for each bias type. Using the MBIB dataset, which contains multiple social media bias types, we demonstrate the effectiveness of our approaches in detecting these biases. Our findings indicate that Bias-Aware AdapterFusion (BAAF) presents a robust and scalable solution for bias detection across diverse social media content, outperforming competitive baselines that rely on ICL techniques. Our code and data are made publicly available for research purposes.