Side-by-side Comparison Amplifies Dialect Bias in Language Models

Technical Summary: Side-by-Side Comparison Amplifies Dialect Bias in Language Models

Problem Statement
Large Language Models (LMs) exhibit systematic biases against speakers based on dialectal variations, a phenomenon known as covert dialect bias. This occurs even when explicit dialect labels are absent, as models associate negative stereotypes with African-American Vernacular English (AAVE) relative to Standard American English (SAE). While prior research has established the existence of this bias in isolation, it remains unclear how bias manifests in high-stakes decision-making contexts where models are required to rank or compare candidates side-by-side. Furthermore, it is unknown whether explicitly specifying dialect labels (overt bias) mitigates or exacerbates these disparities compared to implicit inference (covert bias). Existing evaluation frameworks often rely on absolute prompting (evaluating tweets in isolation), which may underestimate the severity of bias in comparative settings.

Methodology
The authors employ a matched-guise probing technique using a dataset of 2,019 intent-equivalent tweet pairs (one SAE, one AAVE) derived from Groenwold et al. [20]. The study evaluates three models: LLaMA-3.1-8B, DeepSeek-V3, and GPT-4.0-mini.

The evaluation is conducted across four distinct conditions:

1. Absolute Prompting: SAE and AAVE tweets are rated independently.
2. Contrastive Prompting: SAE and AAVE tweets are presented side-by-side for direct comparison.
3. Covert Setting: No explicit dialect labels are provided; the model infers dialect from linguistic features.
4. Overt Setting: Explicit dialect labels (e.g., "This tweet is written in AAVE") are included in the prompt.

Models are prompted to rate each tweet on a 1–5 Likert scale across 12 stereotypical traits grouped into six valence pairs (e.g., Intelligence/Stupidity, Politeness/Rudeness), informed by the Princeton Trilogies [18, 25, 26].

To quantify bias, the study utilizes:

• Cohen's $d$: To measure the effect size of score disparities between SAE and AAVE.
• Counterfactual Fairness Gap (CF Gap): To measure the normalized mean absolute error between scores assigned to intent-equivalent pairs.
• $Q$-Value: A log-likelihood ratio metric to assess model confidence in assigning specific scores to different dialects.

For mitigation, the authors apply counterfactual fairness finetuning to LLaMA-3.1-8B. This involves training the model to assign identical trait scores to SAE and AAVE tweet pairs, using the model's own SAE scores from the absolute setting as ground truth. A prompting-based debiasing approach is also tested as a baseline.

Key Results

1. Amplification in Contrastive Settings: The most significant finding is that side-by-side comparison significantly amplifies covert dialect bias. Across all models, the effect sizes (Cohen's $d$) and CF gaps are substantially larger in the contrastive setting than in the absolute setting. For instance, DeepSeek-V3 exhibited a 51.78% increase in Cohen's $d$ values when moving from absolute to contrastive prompting. This suggests that comparative contexts, common in hiring or content moderation, exacerbate existing biases beyond what is observed in isolation.
2. Overt Bias Exacerbation: Contrary to the assumption that explicit labeling might reduce bias by making the model aware of the context, explicitly specifying dialect labels (overt setting) often amplifies bias further, particularly in contrastive scenarios. In the overt contrastive setting, disparities for traits like Articulation and Sophistication reached large effect sizes.
3. Model Sensitivity: LLaMA-3.1-8B demonstrated the highest sensitivity to dialectal variation, showing the largest CF gaps and the most inconsistent self-consistency (stability of scores across repeated prompts) for AAVE inputs compared to SAE.
4. Mitigation Limitations: Counterfactual fairness finetuning successfully reduced bias for a subset of traits (e.g., Unsophistication, Stupidity, Incoherence) in the absolute setting. However, these improvements did not consistently hold in the contrastive setting. In some cases, finetuning increased disparities for other traits (e.g., Intelligence, Politeness) when models were forced to compare dialects directly. Prompt-based debiasing showed mixed results and was less reliable due to sensitivity to prompt formulation.

Significance and Contributions
The paper makes four primary contributions:

1. Contrastive Amplification: It demonstrates that covert dialect bias is not static but is dynamically amplified when models perform comparative judgments, a critical insight for real-world applications involving ranking or selection.
2. Overt vs. Covert Comparison: It provides a direct comparison showing that overt dialect cues do not mitigate bias but often intensify it, challenging the efficacy of current safety alignment strategies in handling explicit identity cues.
3. Evaluation Framework: It highlights that standard evaluation settings (absolute prompting) may significantly underestimate the severity of dialect bias, necessitating the inclusion of contrastive testing in fairness audits.
4. Mitigation Challenges: It reveals that while counterfactual fairness finetuning can reduce aggregate bias, it is insufficient for ensuring consistent fairness across all traits and evaluation contexts, particularly in high-stakes comparative scenarios.

The authors conclude that dialect bias remains an unresolved problem in LMs, with existing mitigation strategies failing to address the amplification effects inherent in contrastive decision-making. They argue for more robust evaluation frameworks that account for these comparative dynamics to prevent the reinforcement of structural inequalities in automated systems.