Data Driven Model to Investigate Political Bias in Mainstream Media

Abstract

Media bias refers to the tendency of mainstream media outlets to report news in a way that reflects their own political, social, or ideological beliefs or preferences. Such bias may obfuscate facts, manipulate public beliefs, misinform readers, narrow perspectives and viewpoints, and result in greater polarization and division. To counter this issue, this study presents a model for quantifying media bias, aimed at enabling individuals to make more informed media choices. The proposed media analysis model includes a pipeline that gathers articles from three distinct sources: mainstream media news outlets, known conservative outlets, and known liberal media outlets. The collected articles were subjected to a range of text pre-processing operations and subsequently, curated n-gram and topic lists were generated. Several classification models including BERT, logistic regression, random forest, multinomial, and long short-term memory (LSTM) were created and fine-tuned on polarized news sources and used for analyzing news articles from the mainstream media. Among the various classification models that we investigated in this study, BERT achieved overall higher accuracy across the majority of topics. The analysis of mainstream media on various topics yielded different results, with some being balanced and others leaning left or right, depending on the topic. The research also suggests the effectiveness of using highly polarized news sources for developing models to predict media bias.

Keywords

• Classification
• deep learning
• mainstream media
• media bias
• natural language processing
• sentiment analysis