Religious Bias Landscape in Language and Text-to-Image Models: Analysis, Detection, and Debiasing Strategies
Abstract: The increasing utilization of language models necessitates critical examinations of their inherent biases, particularly concerning religion. This study thoroughly examines religious bias across a wide range of language models, encompassing pre-trained models like BERT, RoBERTa, ALBERT, and DistilBERT, alongside diverse open-source large language models such as GPT-2, Llama-3, Mixtral-8x7B, Vicuna-13B, and closed-source models including GPT-3.5 and GPT-4, along with DALL·E 3 for image generation. Using diverse methodologies like mask filling, prompt completion, and image generation, we assess each model’s handling of content related to different religions to uncover any underlying biases. We also investigate cross-domain bias concerning gender, age, and nationality within the context of religious content. Furthermore, this paper explores the effectiveness of targeted debiasing techniques, employing corrective prompts to mitigate identified biases. Our findings indicate that language models continue to exhibit biases in both text and image generation. However, the use of debiasing prompts has proven effective in mitigating these biases.
Paper Type: Long
Research Area: Ethics, Bias, and Fairness
Research Area Keywords: Computational Social Science and Cultural Analytics, Ethics, Bias, and Fairness, Generation
Contribution Types: Model analysis & interpretability, Data analysis
Languages Studied: English
Submission Number: 5229
Loading