arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2407.02066cs.CL
收藏

BiasDora:探索视觉-语言模型中隐藏的偏见关联

BiasDora: Exploring Hidden Biased Associations in Vision-Language Models

George Mason University(乔治梅森大学);Archimedes AI Research Unit, Athena RC, Greece(希腊雅典娜研究中心阿基米德人工智能研究单元);University of Washington(华盛顿大学)

Chahat Raj, Anjishnu Mukherjee, Aylin Caliskan, Antonios Anastasopoulos, Ziwei Zhu

AI总结:

BiasDora通过探测视觉-语言模型,系统揭示9个偏见维度中隐藏的隐式关联,分析其负面性、毒性和极端性,并公开Dora数据集以识别现有方法遗漏的微妙与极端偏见。

摘要:

现有研究视觉-语言模型(VLMs)社会偏见的工作主要关注有限的已记录偏见关联,例如性别:职业或种族:犯罪。这种狭窄的范围往往忽略了大量未经检验的隐式关联,限制了对这些偏见的识别,从而也限制了缓解。我们通过探测VLMs来填补这一空白:(1)揭示9个偏见维度中隐藏的隐式关联。我们系统地探索多样的输入和输出模态,并(2)展示偏见关联在负面性、毒性和极端性方面的差异。我们的工作(3)识别出现有方法通常无法识别的微妙和极端偏见。我们将检索到的关联数据集(Dora)公开发布在https://github.com/chahatraj/BiasDora。

英文摘要:

Existing works examining Vision-Language Models (VLMs) for social biases predominantly focus on a limited set of documented bias associations, such as gender:profession or race:crime. This narrow scope often overlooks a vast range of unexamined implicit associations, restricting the identification and, hence, mitigation of such biases. We address this gap by probing VLMs to (1) uncover hidden, implicit associations across 9 bias dimensions. We systematically explore diverse input and output modalities and (2) demonstrate how biased associations vary in their negativity, toxicity, and extremity. Our work (3) identifies subtle and extreme biases that are typically not recognized by existing methodologies. We make the Dataset of retrieved associations, (Dora), publicly available here https://github.com/chahatraj/BiasDora.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑