VISOGENDER

VISOGENDER是由牛津人工智能协会和牛津大学创建的一个新颖数据集，旨在评估视觉语言模型中的性别偏见。该数据集专注于职业相关的偏见，每张图像都与包含场景中主体和对象代词关系的标题相关联。VISOGENDER通过平衡职业角色中的性别代表性，支持两种偏见评估方式：代词解析偏见和检索偏见。数据集涵盖23种独特的职业，每种职业在数据集中以两种模板形式出现，用于测试模型在复杂场景中解析二元性别的能力。VISOGENDER的应用领域包括评估和改进视觉语言模型在处理性别相关任务时的性能，以减少性别偏见对模型输出的影响。VISOGENDER is a novel dataset developed by the Oxford Artificial Intelligence Society and the University of Oxford, aimed at evaluating gender bias in vision-language models. This dataset focuses on occupation-related bias, where each image is paired with a caption that depicts the pronominal relationships between the subject and objects in the scene. By balancing gender representation across occupational roles, VISOGENDER supports two bias evaluation paradigms: pronoun resolution bias and retrieval bias. The dataset encompasses 23 unique occupations, each presented in two template formats within the corpus to test models' capability to resolve binary gender cues in complex scenarios. Application scenarios of VISOGENDER cover evaluating and optimizing the performance of vision-language models when handling gender-related tasks, so as to alleviate the impact of gender bias on model outputs.