NaN-NLI

NaN-NLI是由墨尔本大学创建的一个自然语言推理测试套件，专注于子句级否定。该数据集包含258对前提-假设，其中前提包含子句级否定，而假设通过最小修改反映不同可能的解释。数据集的构建遵循严格的语言学框架，包括否定类型和构造的注释，以及用于构建假设的操作。NaN-NLI旨在通过精细分析模型性能，解决当前语言模型在处理否定方面的局限性，特别是在量化和否定结合的复杂情况下。

NaN-NLI is a natural language inference test suite created by the University of Melbourne, focusing on clause-level negation. This dataset contains 258 premise-hypothesis pairs, where the premise includes clause-level negation, and the hypothesis reflects various plausible interpretations through minimal modifications. The construction of the dataset follows a rigorous linguistic framework, covering annotations of negation types and constructions, as well as the operations used to generate the hypotheses. NaN-NLI aims to address the limitations of current language models in handling negation, particularly in complex scenarios where quantification and negation are combined, via fine-grained analysis of model performance.