Researchers propose HierBias, a hierarchical model for context-aware media bias detection

Researchers at an unspecified institution introduce HierBias, a hierarchical context-conditioned media bias detector that explicitly models inter-sentence dependencies ignored by prior sentence-level approaches.

The paper formalizes a context-conditioned bias probability and proves that incorporating document context strictly reduces Bayes error for sentence-level classification when inter-sentence mutual information is non-zero.

HierBias combines a RoBERTa encoder at the sentence level with a cross-sentence Transformer aggregator and dual output heads for binary bias detection and four-class bias type classification.

On the BABE and BASIL benchmarks, HierBias reports 0.853 F1 and 0.723 MCC, surpassing the prior state-of-the-art bias detector by +2.6% F1 and +4.3% MCC, with statistical significance per McNemar’s test (p < 0.05).

Ablation experiments indicate that each theoretical component contributes independently and consistently to performance.

The authors also derive a multi-task generalization bound showing that jointly training binary bias detection and fine-grained type classification improves sample efficiency on small annotated corpora.