PERSUASIVE-PAIRS
收藏资源简介:
数据集PERSUASIVE-PAIRS由奥胡斯大学和哥本哈根大学的研究团队创建,包含2697对短文本,旨在评估大型语言模型生成说服性语言的能力。该数据集通过多种语言模型对原始文本进行重写,以增强或减弱其说服力,并通过多重注释在说服力上进行相对评分。数据集内容涵盖新闻摘录和聊天或辩论中的语句,来源包括PT-Corpus、Webis-Clickbait-17等。创建过程中,研究团队使用了多种指令调整的语言模型,如GPT-4和LLaMA3,以确保数据集的多样性和广泛性。该数据集的应用领域主要集中在语言模型的评估和比较,特别是在生成说服性文本的能力上,有助于解决语言模型在不同领域和情境下的应用问题。
The dataset PERSUASIVE-PAIRS was developed by a research team from Aarhus University and the University of Copenhagen. It comprises 2697 pairs of short texts, and is designed to evaluate the ability of large language models (LLMs) to generate persuasive language. During its construction, original texts were rewritten using multiple language models to either enhance or reduce their persuasiveness, and relative scoring on persuasiveness was performed via multiple annotations. The dataset covers news excerpts and statements from chats or debates, with sources including PT-Corpus, Webis-Clickbait-17, and other similar corpora. The research team adopted various instruction-tuned language models such as GPT-4 and LLaMA3 during the creation process to ensure the diversity and broad coverage of the dataset. The main application scenarios of this dataset focus on the evaluation and comparison of language models, particularly their capability to generate persuasive texts, which helps address the application issues of language models across different domains and contexts.
Persuasive-Pairs 数据集
概述
Persuasive-Pairs 数据集包含成对短文本,每对文本中一个是来自新闻、辩论或聊天(通过字段 source 查看文本来源),另一个由语言模型(LLM)重写,以包含更多或更少的说服性语言。
标注
每对文本由三名标注者根据说服性语言的程度进行评判:任务是选择哪个文本包含更多的说服性语言,并在普通尺度上选择“略微”、“适度”或“非常”更多。评分是6点制,负分表示文本1比文本2更具说服力,反之亦然。字段 flip 标记LLM(模型)是否被提示生成更多/更少说服性语言,而 gen_place 指示生成文本在每对中的位置。
引用
使用数据时,请引用相关论文。




