TURINGBENCH
收藏资源简介:
TURINGBENCH是由宾夕法尼亚州立大学创建的数据集,包含200,000个人工或机器生成的样本,涵盖20个不同的标签。数据集主要用于研究神经文本生成方法的“图灵测试”问题。数据集内容包括新闻文章,主要来源于CNN和华盛顿邮报等媒体。创建过程中,使用标题作为提示,让19个选定的神经文本生成器生成与人类编写相似的文章。TURINGBENCH的应用领域包括检测机器生成的文本,尤其是在假新闻检测和政治宣传分析中。
TURINGBENCH is a dataset created by Pennsylvania State University, comprising 200,000 human-written or machine-generated samples across 20 distinct labels. It is primarily intended for research on the "Turing test" problem in neural text generation methods. The dataset consists of news articles predominantly sourced from media outlets including CNN and The Washington Post. During the development of this dataset, headlines were used as prompts to enable 19 selected neural text generators to produce articles that mimic human-written works. Application areas of TURINGBENCH include the detection of machine-generated text, particularly in fake news detection and political propaganda analysis.



