FigCaps-HF

FigCaps-HF是一个大规模的图表-标题对数据集，由马萨诸塞大学阿默斯特分校和Adobe Research共同创建。该数据集包含133,543对图表及其对应的标题，每对都附有人类反馈评分，用于评估标题的有用性、视觉描述性和文本内容的相关性。创建过程中，研究团队首先从公开的科学文章中提取图表和标题，然后通过领域专家对这些对进行评分，以确保标题的质量和相关性。该数据集主要用于开发和评估基于人类反馈的图表到标题生成模型，旨在提高科学文档的可理解性和可访问性，特别是帮助视觉障碍读者更好地理解科学内容。

FigCaps-HF is a large-scale chart-title pair dataset co-created by the University of Massachusetts Amherst and Adobe Research. It contains 133,543 chart-title pairs, each accompanied by human feedback scores for evaluating the title's usefulness, visual descriptiveness and textual relevance. During the dataset creation, the research team first extracted charts and their corresponding titles from publicly available scientific articles, then had domain experts score these pairs to ensure the quality and relevance of the titles. This dataset is primarily used to develop and evaluate human feedback-based chart-to-title generation models, aiming to improve the comprehensibility and accessibility of scientific documents, particularly to help visually impaired readers better understand scientific content.