Computational approaches to developing the implicit media bias dataset: Assessing political orientations of nonpolitical news articles

Abstract

Research on media bias has been primarily conducted a number of times of news outlets referred on political news articles, but nonpolitical articles can still convey media bias that indicates the political orientation of the news outlet. Using manual human evaluation and computational approaches, we developed and publicly released the Implicit Media Bias Dataset, which contains the political orientations of 24,576 news articles featuring nonpolitical events. News articles published in the information technology and science section of the two most biased Korean news outlets (the most conservative and the most progressive) were collected, and each article was manually evaluated by human annotators in terms of its objectiveness, fairness, unbiasedness, and political orientation. The results revealed significant differences between the articles from the conservative and progressive news outlets in these domains. Next, deep learning models trained with a large corpus of nonpolitical articles were used to identify the political orientations of the first set of articles. They achieved over 98% accuracy in classifying the articles as conservative or progressive. The findings of this study demonstrate the effectiveness of computational methods in identifying and analyzing diverse forms of polarization in society.