Do the Right Thing, Just Debias! Multi-Category Bias Mitigation Using LLMs

TL;DR

This paper introduces ANUBIS, a new dataset for nine social bias categories, and evaluates advanced LLM techniques for multi-category bias mitigation, emphasizing generalizability and societal impact.

Contribution

The paper presents ANUBIS, a comprehensive bias dataset, and assesses multiple LLM-based methods for effective multi-category bias reduction in language models.

Findings

ANUBIS contains 1507 curated sentence pairs across nine bias categories.

State-of-the-art models like T5 show varying effectiveness in bias mitigation.

Cross-dataset generalizability and environmental impacts are analyzed.

Abstract

This paper tackles the challenge of building robust and generalizable bias mitigation models for language. Recognizing the limitations of existing datasets, we introduce ANUBIS, a novel dataset with 1507 carefully curated sentence pairs encompassing nine social bias categories. We evaluate state-of-the-art models like T5, utilizing Supervised Fine-Tuning (SFT), Reinforcement Learning (PPO, DPO), and In-Context Learning (ICL) for effective bias mitigation. Our analysis focuses on multi-class social bias reduction, cross-dataset generalizability, and environmental impact of the trained models. ANUBIS and our findings offer valuable resources for building more equitable AI systems and contribute to the development of responsible and unbiased technologies with broad societal impact.