cyberagent/AdParaphrase

AdParaphrase是一个包含人类偏好判断的语义等价但表述和风格不同的广告文本对的新型改写数据集。该数据集通过收集语义相似的广告文本，并由五名评委对每对文本进行改写识别，再由十名评委对每对改写文本进行吸引力评价构建而成。数据集允许我们专注于单个广告文本之间的语言特征差异，同时最小化语义内容差异的影响，从而分析以这些特征为中心的人类偏好。

AdParaphrase is a novel paraphrase dataset containing human preferences for pairs of ad texts that are semantically equivalent but differ in wording and style. The dataset was carefully constructed by collecting semantically similar ad texts, performing paraphrase identification with five judges per pair, and collecting human preference judgments with ten judges per pair. The dataset allows us to focus on differences in linguistic features between individual ad texts while minimizing the impact of differences in semantic content, thus enabling the analysis of human preferences centered on these features.