Research map / Learning theory and algorithms / Fairness
Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?
Yifan Wang, Mayank Jobanputra, Ji-Ung Lee, Soyoung Oh, Isabel Valera, Vera Demberg
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use L2-aggregated or Occlusion explanations to detect biased predictions | BERT and Qwen3-4B · Civil Comments / Jigsaw bias detection | fairness correlation | Improved |
| aggregate embedding attributions by mean instead of L2 norm (ablation) | explanation-based bias detection · Civil Comments bias detection | fairness correlation | Hurt |
| use validation-set sensitive-token reliance to select the fairest model | model selection · Civil Comments, 7 encoder and 4 decoder models | Spearman rank correlation / MRR@1 | Hurt |
| add sensitive-token reliance regularizer during training | BERT · Civil Comments race/gender/religion | FNR disparity and other fairness metrics | Improved |
| Grad L1 explanation regularization for gender bias | BERT · Civil Comments gender | accuracy and fairness | Improved |
| tune alpha on a fairness-balanced metric instead of task accuracy (ablation) | explanation-based debiasing · Civil Comments | fairness | Improved |
| combine traditional and explanation-based debiasing | BERT · Civil Comments | fairness | Improved |
| apply traditional debiasing methods (group balance, CDA, dropout, etc.) | BERT, Qwen3-4B · Civil Comments | fairness metrics: BERT race Avg_iu 3.17 -> 0.50 (CDA), FNR disp 10.04 -> 18.45 | Mixed |
| explanation-debias with attention-based methods then detect bias with attention explanations | explanation-debiased BERT · Civil Comments race | fairness correlation | Hurt |
| detect bias with Occlusion and L2 explanations in explanation-debiased models | explanation-debiased BERT · Civil Comments race | fairness correlation | No clear effect |
| replace explanation-based detection with LLM-as-a-judge (ablation) | Qwen3-4B predictions · Civil Comments race | fairness correlation: 0.194 -> 0.104 (Qwen3-4B, K=5) | Hurt |
| use larger GPT-OSS-120B as judge instead of Qwen3-4B (LLM decision) (ablation) | LLM-as-a-judge · Civil Comments race | Avg_iu biased/unbiased: 0.065/2.59 -> 4.42/2.35 | Improved |
Closest papers by meaning
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.87)
- Constructing Fair Latent Space for Intersection of Fairness and Explainability (AAAI 2025 · similarity 0.86)
- Collapsed Language Models Promote Fairness (ICLR 2025 · similarity 0.86)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.86)
- The Devil is in the Neurons: Interpreting and Mitigating Social Biases in Language Models (ICLR 2024 · similarity 0.86)
- All Should Be Equal in the Eyes of LMs: Counterfactually Aware Fair Text Generation (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
