NER-CHANGED, ROLE-SWAPPED

本文介绍了两个专为自然语言推理（NLI）设计的新数据集：NER-CHANGED和ROLE-SWAPPED。NER-CHANGED数据集包含通过替换句子中的命名实体创建的中性或矛盾标注的句子对，而ROLE-SWAPPED数据集则通过交换句子中具有相同动词类型但不同角色的实体来生成。这两个数据集旨在帮助NLI系统更好地理解和区分实体与角色。创建过程中，研究者利用了如bAbI和AMR等现有语料库，并通过自动化的方式生成句子对。这些数据集主要应用于改进NLI模型的性能，特别是在理解实体和角色方面的能力。

This paper introduces two novel datasets designed specifically for Natural Language Inference (NLI): NER-CHANGED and ROLE-SWAPPED. The NER-CHANGED dataset consists of sentence pairs labeled as neutral or contradictory, which are created by replacing named entities within sentences. The ROLE-SWAPPED dataset, by contrast, is generated by swapping entities that share the same verb type but have distinct roles in the sentences. These two datasets are intended to help NLI systems better understand and distinguish between entities and their roles. During their development, researchers utilized existing corpora such as bAbI and AMR, and generated the sentence pairs through automated methods. These datasets are primarily applied to improve the performance of NLI models, particularly their ability to understand entities and their corresponding roles.