FAIR-PIVOT: Fairness-oriented Perspective-Inversion via Viewpoint-balanced Outline-then-generaTion for Mitigating Political Bias in Web Corpora
TL;DR
Fairness-oriented Perspective-Inversion via Viewpoint-balanced Outline-then-generaTion (simply, FAIR-PIVOT), a retrieval-augmented generation (RAG)-based two-level-agent framework that builds balanced perspective-inverted synthetic data for politically sensitive topics is proposed.
Abstract
Large-scale web corpora (e.g., C4) are severely imbalanced in terms of political perspective: on the death penalty issue, e.g., left-leaning documents outnumber right-leaning ones by more than 8-fold. Such imbalances can transfer to downstream large language models, reinforcing political bias. We propose Fairness-oriented Perspective-Inversion via Viewpoint-balanced Outline-then-generaTion (simply, FAIR-PIVOT), a retrieval-augmented generation (RAG)-based two-level-agent framework that builds balanced perspective-inverted synthetic data for politically sensitive topics. After annotating web texts with multi-persona LLMs, FAIR-PIVOT retrieves documents from contrasting orientations and inverts their viewpoint, preserving authentic discourse structure rather than rephrasing the same-view text. Across nine topics, FAIR-PIVOT generates 13,954 instances and reduces the total variation distance, a measure of how far the orientation distribution deviates from uniform balance, by up to 83%, while improving linguistic fluency as measured by Perplexity (25.5 → 21.8 ↓). Fine-tuning small language models on the balanced corpus further shifts their downstream behavior toward the intended orientation (Gemma (L--R)% +48 → +34, Qwen +55 → +50), confirming that data-level balancing can yield fairer model responses. Code and scripts are available at https://github.com/bigbases/FAIR-PIVOT.
