DIALFACT
收藏资源简介:
DIALFACT是一个基于Wizard-of-Wikipedia数据集构建的对话事实核查基准数据集,包含22,245条人工标注的对话声明,每条声明都配有来自维基百科的证据片段。该数据集旨在解决对话中事实核查的挑战,特别是处理非正式语言、指代和检索歧义等问题。DIALFACT包含三个子任务:可验证声明检测、证据检索和声明验证,旨在预测对话响应是否应被视为可验证声明,并找到相关证据,最终预测声明是被支持、反驳还是信息不足。此数据集不仅包含人工编写的声明,还通过操作如矛盾、填充和替换创建了合成声明,由合格的众包工作者进行标注,以提高对话系统的事实准确性和可信度。
DIALFACT is a conversational fact-checking benchmark dataset built upon the Wizard-of-Wikipedia dataset. It contains 22,245 manually annotated conversational claims, with each claim paired with evidence snippets sourced from Wikipedia. This dataset aims to address the challenges of fact-checking in conversations, particularly those involving informal language, reference resolution and retrieval ambiguity. DIALFACT includes three subtasks: verifiable claim detection, evidence retrieval and claim verification, which are designed to predict whether a conversational response should be regarded as a verifiable claim, retrieve relevant evidence, and finally determine whether a claim is supported, refuted or lacks sufficient information. In addition to manually written claims, this dataset also includes synthetic claims created through operations such as contradiction, padding and substitution, which are annotated by qualified crowdworkers to enhance the factual accuracy and credibility of conversational systems.

- 1DialFact: A Benchmark for Fact-Checking in Dialogue语言技术研究所,卡内基梅隆大学† · 2022年



