Chinese Atomic Claim Decomposition Dataset (CACDD)
收藏资源简介:
Chinese Atomic Claim Decomposition Dataset (CACDD) 是一个用于长篇回答验证的声明分解基准数据集,由中国科学院计算技术研究所创建。该数据集基于WebCPM数据集,并增加了专家注释以确保高质量。CACDD包含500个人工注释的问题-回答对,共计4956个原子声明。数据集的创建过程包括句子分解、句子分类和原子声明分解三个步骤,旨在解决长篇回答中的事实验证问题,特别是在大语言模型生成的回答中识别和分解可验证的声明。
Chinese Atomic Claim Decomposition Dataset (CACDD) is a benchmark dataset for claim decomposition in long-form answer verification, developed by the Institute of Computing Technology, Chinese Academy of Sciences. This dataset is built upon the WebCPM dataset, with expert annotations added to ensure high data quality. CACDD contains 500 manually annotated question-answer pairs, totaling 4956 atomic claims. The dataset construction process includes three steps: sentence decomposition, sentence classification, and atomic claim decomposition, aiming to address the fact verification problem in long-form answers, particularly for identifying and decomposing verifiable claims from responses generated by large language models (LLMs).
CACDD 数据集概述
数据来源
- 数据集基于 WebCPM 数据集构建。
数据格式
- question: 开放领域的问题。
- answer: RAG 生成的长篇回答。
- sentence_classification: 从回答中提取的句子,并分类为事实(fact-1)、观点(opinion-2)、指令(instruction-3)和其他(other-4)。
- atomic_claim: 从上述事实句子中提取的原子声明。




