Research map / Learning theory and algorithms / Fairness
Fair Text Classification via Transferable Representations
Thibaud Leteno, Michael Perrot, Charlotte Laclau, Antoine Gourru, Christophe Gravier
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 |
|---|---|---|---|
| add WFC fairness regularization to CE | CE · Moji | fairness: 61.2 -> 91.4 | Improved |
| add WFC fairness regularization to CE | CE · Bios | fairness: 85.1 -> 89.0 | Improved |
| increase beta from 0.1 to 100 (ablation) | WFC · Bios | leakage: 98.1 -> 77.7 | Improved |
| increase beta from 0.1 to 100 (ablation) | WFC · Bios | balanced accuracy: 82.8 -> 81.2 | Hurt |
| increase beta from 1.0 to 100 (ablation) | WFC · Moji | balanced accuracy: 75.2 -> 67.9 | Hurt |
| use hard predicted sensitive labels instead of demonic representations Z_a (ablation) | WFC · Moji | fairness: 91.4 -> 65.0 | Hurt |
| use hard predicted sensitive labels instead of Z_a (ablation) | WFC · Bios | fairness: 89.0 -> 87.5 | Hurt |
| train demonic zero-shot on EEC instead of Bios 1% (ablation) | WFC · Bios | fairness: 89.0 -> 88.9 | No clear effect |
| add domain adaptation (eta=1) to demonic training (ablation) | WFC · Bios | fairness: 86.8 -> 88.7 | Improved |
| add class rebalancing BTEO (ablation) | WFC · Bios | fairness: 89.0 -> 89.1 | No clear effect |
| replace BERT with SFR-Embedding-2_R representations (ablation) | WFC · Bios | balanced accuracy: 82.4 -> 85.2 | Improved |
| use last hidden layer rather than first layer or logits for Wasserstein distance (ablation) | WFC · Moji | fairness | Improved |
Closest papers by meaning
- Human-Guided Fair Classification for Natural Language Processing (ICLR 2023 · similarity 0.88)
- Simple and Effective Specialized Representations for Fair Classifiers (NeurIPS 2025 · similarity 0.86)
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.86)
- FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models (NeurIPS 2025 · similarity 0.86)
- Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference (NeurIPS 2025 · similarity 0.85)
- On Optimal Steering to Achieve Exact Fairness (NeurIPS 2025 · 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.
