Outlet-Level Racial Bias Analysis in South African COVID-19 News

27,140 COVID-19 vaccination news articles from 39 South African outlets form the corpus for this racial bias study. Researcher Nnaemeka Ohamadike trained an ensemble of Word2Vec models to embed each outlet's language and measure its association with racial stereotype vocabularies. The dataset was published on April 5, 2026.

Use Cases

• Analyze associations between outlet-level word embeddings and validated racial stereotype vocabularies.
• Compare racial bias patterns across 39 specific news outlets, such as business/finance versus community newspapers.
• Replicate the finding that socioeconomic race bias correlates with South African human judgments using the provided ensemble models.
• Apply the scalable framework to audit institutional racial bias in other news corpora or time periods.

Strengths

• Corpus contains 27,140 articles, providing a substantial text base for analysis.
• Analysis covers 39 distinct South African news outlets, enabling granular institutional comparison.
• Method employs an ensemble of 10 Word2Vec models trained on resampled data for robustness.

Limitations

• Sample size per outlet is not explicitly detailed beyond the resample size of 3,900 articles total.
• Analysis is temporally limited to COVID-19 vaccination news, which may not represent general media bias.
• The primary data (word embeddings, model outputs) are derived, not the raw article text.

Provenance

Source

figshare, authored by Nnaemeka Ohamadike.

Collection Method

Word embeddings (Word2Vec) trained on a corpus of news articles, with outlet-level vectors calculated from their language.

Time Range

Covers the period of COVID-19 vaccination news; specific dates not provided.

Freshness

Data was last updated in April 2026.

Geography

South Africa.