ChunkyBERT: a novel technique for multiclass political bias detection in news media

Abstract

Abstract With the increasing use of digital platforms for spread of information, political news has some of the most skewed sources which has confused people on what are facts and what are biased reportings. While most modern methods of finding bias in media use advanced Machine Learning algorithms and deep Learning techniques, such techniques heavily rely on manually generated features as sentiment analysis and lexical frequency which are tedious and time consuming. Hence we propose a novel method using Bidirectional Encoder Representations from Transformers (BERT) for ascertaining political media bias, mainly classifying articles into left-wing, centrist and right-wing inclinations via incorporation of the complete text. The suggested method includes dividing long political articles into segments of a set length. These are encoded separately using a pre-trained BERT model. A Transformer encoder then aggregates these segment-level embeddings, along with an attention pooling mechanism. This lets the model focus on and use the parts of the text that best show political bias for classification tasks. The experiments showed a highest validation accuracy of 86.22% and a validation AUC-ROC of 0.96, which is better than standard methods. This research gives a way to scale the detection of political bias, with potential uses in journalism, media monitoring, and improving digital literacy tools.