Side-by-Side Comparison Amplifies Dialect Bias in Language Models
A recent investigation published on arXiv (2605.24384) indicates that language models (LMs) display hidden dialect bias, linking African-American Vernacular English (AAVE) with more unfavorable stereotypes compared to Standard American English (SAE). Notably, this bias intensifies when SAE and AAVE tweet pairs are analyzed together, reflecting scenarios that influence critical decisions, such as candidate evaluations. Furthermore, the bias increases when dialect labels are clearly identified, despite ongoing attempts to address these biases. The research measures bias through stereotypical characteristics sourced from social psychology studies on racial prejudice.
Key facts
- Language models exhibit covert dialect bias against AAVE.
- Bias is amplified when SAE and AAVE tweets are compared side by side.
- Side-by-side comparison mirrors high-impact decision-making contexts.
- Explicit dialect labels worsen the bias.
- Study uses stereotypical traits from social psychology research on racial bias.
- Research published on arXiv with ID 2605.24384.
- Bias exists even without explicit dialect labels.
- Prior work showed bias in isolation, but side-by-side comparison increases it.
Entities
Institutions
- arXiv