Topic-level fairness improves fairness in retrieval based text generation
Topic-Level Fairness in Retrieval-Augmented Generation
Artificial Intelligence
Summary
Retrieval-augmented generation systems use information retrieval to help generate responses, but they can treat different groups unfairly depending on the topic. The authors found that measuring fairness overall can hide unfairness within specific topics. They improved fairness by adjusting the system separately for each topic using a ranking method guided by the system’s own generated preferences. This approach reduced unfairness significantly without losing accuracy.
What this means in practice
- •For search engine developers: Reduce topic-specific group biases in retrieval-augmented generation results using adapted reranking without losing accuracy.
- •For social media platform engineers: Improve topic-level fairness in automatic content recommendation and generation systems by applying topic-specific reranking methods.
Authors
Jaeyoung Park, Hantian Zhang, Kexin Rong, Steven Euijong Whang
Abstract
Group fairness in retrieval-augmented generation (RAG) systems is typically evaluated using metrics aggregated across queries. However, such metrics can conceal topic-level disparities: a system may favor women on questions about human rights and men on questions about actors and filmmakers, appearing nearly neutral overall even though neither topic is fair. We therefore study topic-level fairness, which requires group fairness to hold within each topic rather than only in aggregate. To reduce the topic-level disparity, we adapt the reranker separately for each topic with a Bradley-Terry objective under direct supervision from the generated output: preference pairs built from the generator's own output carry the direction that topic requires. With preference pairs built from RoMQA alone, the adaptation reduces the mean absolute per-topic disparity by 51% on RoMQA and by 40% on TREC Fair Ranking while retaining accuracy. No baseline reaches the reduction, even when modified to treat each topic separately rather than apply one correction to every topic. Neither does our method with one reranker for all topics, which follows the direction that most topics require.