未明确提供

本研究构建了一个用于评估视觉语言模型中性别职业偏见的统一框架，并创建了一个合成的高质量文本和图像数据集，该数据集模糊了性别差异，用于基准测试性别偏见。数据集通过使用性别中性语言和避免特定性别形容词的文本，以及使用机器人代替人类专业人士的图像，来生成性别模糊的输入。此数据集旨在帮助未来改进视觉语言模型，以学习社会无偏见的表示，并系统地评估模型在不同输入输出模式下的性别偏见。

This study constructed a unified framework for evaluating gender-occupation bias in vision-language models, and created a high-quality synthetic text and image dataset that blurs gender differences for benchmarking gender bias. The dataset generates gender-ambiguous inputs by adopting gender-neutral language in text that avoids gender-specific adjectives, and by using images of robots instead of human professionals. This dataset aims to facilitate future improvements of vision-language models to learn socially unbiased representations, and to systematically evaluate the gender bias of models across different input-output modalities.