MBDef: A Media Bias Definitions Dataset

Definitions of media bias and related concepts extracted from 6,115 scientific papers: 34,916 terms with 103,476 definitions, with synonyms, descriptions, related terms, subtypes, supertypes, examples and structured citations. The zip keeps the Hugging Face repository's layout. Licence: our annotations are licensed CC BY 4.0. The definition texts are quotations; their copyright stays with the papers' authors and publishers (see docs/licence.md in the zip). Validation (added in 1.1.0): validation.csv, validation_agreement.py and validation_README.md hold the manual audit of MBDef reported in Step 5 of the paper: two random samples of 200 definitions each, labelled by three pseudonymised annotators per sample with the MBDef-Gold questions (Q1 definition? Q2 relevant to the media bias taxonomy?). 318 of 400 definitions (79.5%) are relevant; pooled Krippendorff's α Q1 0.17, Q2 0.30; Gwet's AC1 Q1 0.82, Q2 0.67. The dataset zip is unchanged from 1.0.0; the validation files are not on Hugging Face.