From Detection to Characterization: A Holistic and Perspectivist Approach to Media Bias Analysis

Abstract

Doctoral thesis on automated media bias detection, supervised by Jorge Carrillo-de-Albornoz and Laura Plaza. Current bias-detection systems do not generalize reliably beyond the linguistic, thematic, or ideological contexts they were trained on, and this thesis argues the cause is structural rather than technical. It diagnoses five open problems in the field (conceptual fragmentation, weak cross-corpus generalization, the ground-truth fallacy, anglocentric resources, and opaque evaluation) and contributes four interlocking artifacts to address them: (1) a hierarchical taxonomy of seventeen manifestations of media bias distilled from a systematic review of 118 studies; (2) MBBMD, a hierarchical and perspectivist Spanish corpus that preserves annotator disagreement as informative signal following the Learning with Disagreements paradigm; (3) a rigorous, reproducible cross-dataset diagnosis across five reference corpora (MBIC, Crisis in Ukraine, FIGNEWS, BEADs, MBBMD); and (4) a hierarchical, perspectivist detection framework that, trained only on MBBMD, reaches a cross-dataset macro-F1 of 0.77 (matching in-domain performance and improving +0.30 over the strongest flat baseline). The central contribution is a reframing: from treating bias as an objective property of the document to modeling it as reasoning under uncertainty, grounded in plurality and context.

Keywords