Abstract
News and media outlets are integral to shaping public opinion. Despite journalism's aim for impartial reporting, various biases can emerge during writing and publication phases. While some bias is to be expected, issues arise when it is systematically unrepresentative, potentially compromising the quality of public debate, causing polarization, and, in the worst case, undermining democracy.
Previous studies in social and computational sciences have attempted to identify news bias through methods that often prove laborious, require expert knowledge and produce inconsistent results, or, while being more efficient, focus on detection, while forgoing deeper understanding. Furthermore, news bias research has predominately focused on sentence- or article-level bias, omitting patterns that emerge from an outlet's overall publication choices. Our work aims to classify outlet bias on global data using machine learning, incorporating sources associated with multiple forms of bias. We demonstrate that our approach can detect political bias among news outlets. Furthermore, features related to under-explored forms of bias improved model performance, and yielded the best-performing model with a 75% accuracy and AUC of 81% (compared to a baseline accuracy of 45% and 50% AUC)
We include a quantitative study to provide model explanations, demonstrating underlying processes of biased news outlet behaviour. Our work thereby presents a novel, scalable and comprehensive approach to studying political bias in news.
Previous studies in social and computational sciences have attempted to identify news bias through methods that often prove laborious, require expert knowledge and produce inconsistent results, or, while being more efficient, focus on detection, while forgoing deeper understanding. Furthermore, news bias research has predominately focused on sentence- or article-level bias, omitting patterns that emerge from an outlet's overall publication choices. Our work aims to classify outlet bias on global data using machine learning, incorporating sources associated with multiple forms of bias. We demonstrate that our approach can detect political bias among news outlets. Furthermore, features related to under-explored forms of bias improved model performance, and yielded the best-performing model with a 75% accuracy and AUC of 81% (compared to a baseline accuracy of 45% and 50% AUC)
We include a quantitative study to provide model explanations, demonstrating underlying processes of biased news outlet behaviour. Our work thereby presents a novel, scalable and comprehensive approach to studying political bias in news.
Original language | English |
---|---|
Number of pages | 19 |
Journal | PLOS ONE |
Publication status | Submitted - 29 Nov 2024 |