Scholarlink

This study presents a methodology to mitigate biases in text generation across multiple demographic attributes, including gender, profession, religion, and race. By fine-tuning the GPT-2 model, we integrate dual-embedding space contrastive learning, optimal transport for embedding alignment, differential privacy for secure updates, and fairness-constrained optimization. Experiments using the StereoSet dataset achieve a balanced Stereotype Score (SS) of 51.24, outperforming most existing models.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Bias Busters: Fair Text Generation with GPT-2

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Md. Nur Amin,
  • Fatih S. Bayram,
  • Alexander Jesser

摘要

This study presents a methodology to mitigate biases in text generation across multiple demographic attributes, including gender, profession, religion, and race. By fine-tuning the GPT-2 model, we integrate dual-embedding space contrastive learning, optimal transport for embedding alignment, differential privacy for secure updates, and fairness-constrained optimization. Experiments using the StereoSet dataset achieve a balanced Stereotype Score (SS) of 51.24, outperforming most existing models.