HierBias beats SOTA media bias detection by 4.3% with context-aware model

Researchers from an undisclosed institution (Li, Yan, Dong) have published HierBias, a new hierarchical model that tackles media bias detection by considering full document context rather than standalone sentences. Traditional sentence-level approaches classify each sentence independently, missing the contextual signals human annotators naturally use. HierBias formalizes a context-conditioned bias probability and proves theoretically that leveraging document context reduces Bayes error when inter-sentence mutual information is non-zero. The model architecture combines a sentence-level RoBERTa encoder with a cross-sentence Transformer aggregator, plus dual output heads for binary bias detection and four-category bias type classification.

Evaluated on the BABE and BASIL benchmarks, HierBias achieves 0.853 F1 and 0.723 Matthews Correlation Coefficient (MCC), surpassing the previous state-of-the-art bias detector by +2.6% F1 and +4.3% MCC—a statistically significant improvement per McNemar's test (p<0.05). Ablation experiments confirm each theoretical component contributes independently. A multi-task generalization bound also shows that jointly training binary and fine-grained type classification improves sample efficiency on small annotated datasets. This work provides both a theoretical framework and a practical tool for more accurate, context-aware media bias detection.