WILDFRAME
收藏资源简介:
WILDFRAME数据集是由耶路撒冷希伯来大学的研究团队创建的,包含1000条经过精心挑选和重构的文本语句,旨在评估大型语言模型在自然发生文本上的框架效应,并与人类行为进行对比。数据集通过三步构建:首先从现实世界中选取具有明确情感倾向的语句;其次,通过添加前缀或后缀对这些语句进行正负情感的重构;最后,通过众包方式收集人类对这些重构语句的情感标注。该数据集可用于研究大型语言模型在情感分析任务中的框架效应,以及与人类行为的相似性。
WILDFRAME Dataset is developed by a research team from the Hebrew University of Jerusalem. It encompasses 1000 carefully selected and reconstructed textual statements, with the goal of evaluating framing effects of large language models (LLMs) on naturally occurring text and comparing their performance with human behaviors. The dataset is constructed in three stages: first, select statements with distinct emotional orientations from real-world contexts; second, reconstruct these statements to elicit either positive or negative sentiment by appending prefixes or suffixes; finally, collect human sentiment annotations for these reconstructed statements through crowdsourcing. This dataset can be utilized to study the framing effects of LLMs in sentiment analysis tasks, as well as the similarity between model behaviors and human responses.




