NEUTRALYZE: Automated Media Bias Detection and Neutralization in News Articles

Abstract

Media bias is a pervasive problem that can significantly impact public discourse and decision-making processes. With the growth of digital media, the potential for bias to spread quickly and widely has become even more pronounced leading to harmful consequences, such as social polarization, misinformation, and distrust in institutions. Manual methods for detecting and neutralizing bias are time-consuming and expensive. Though significant progress has been made in the field of automated media bias detection, the majority of the work has been focused on sentence level classification. Automated bias neutralization is a relatively new domain with limited research contribution. As a result, based on the limitations identified in existing research in related domains, this work proposes a token level classification task that operates at the level of individual words, allowing for more granular and accurate analysis of media bias. The research also contributes to the limited body of work on text neutralization, by utilizing a sequence-to-sequence models for conditional generation of neutralized texts.