Scholarlink

Bias detection in educational materials is crucial for higher education, where academic texts significantly shape students’ understanding of societal narratives. While bias detection has been studied in public discourse, identifying bias in higher educational resources remains underexplored. This study fine-tunes GPT-4o-mini to detect gender and ethnic biases in academic texts and benchmarks its performance against zero-shot large language model (LLM) approaches. The fine-tuned model achieved 81.08% accuracy, representing a significant 13.5% improvement over the zero-shot GPT-4o mini baseline (67.57%). Through structured annotation involving human reviewers, the study highlights the subjectivity inherent in bias classification. These findings contribute to advancing automated bias detection in higher education, promoting more inclusive academic content.

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 Detection in Online Higher Education Texts: Comparing Fine-Tuned and Zero-Shot LLM Approaches

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Rawan Bin Shiha,
  • Eric Atwell,
  • Noorhan Abbas

摘要

Bias detection in educational materials is crucial for higher education, where academic texts significantly shape students’ understanding of societal narratives. While bias detection has been studied in public discourse, identifying bias in higher educational resources remains underexplored. This study fine-tunes GPT-4o-mini to detect gender and ethnic biases in academic texts and benchmarks its performance against zero-shot large language model (LLM) approaches. The fine-tuned model achieved 81.08% accuracy, representing a significant 13.5% improvement over the zero-shot GPT-4o mini baseline (67.57%). Through structured annotation involving human reviewers, the study highlights the subjectivity inherent in bias classification. These findings contribute to advancing automated bias detection in higher education, promoting more inclusive academic content.