NLVR2

NLVR2是一个用于研究自然语言和图像联合推理的数据集，特别关注语义多样性、组合性和视觉推理挑战。该数据集包含107,292个英文句子与网络照片配对，任务是确定自然语言标题是否与一对照片相符。数据集通过使用视觉丰富的图像集和比较与对比任务来诱导语言多样性。定性分析表明，数据需要组合性联合推理，包括数量、比较和关系。评估使用最先进的视觉推理方法显示，数据提出了强大的挑战。

NLVR2 is a dataset designed for investigating joint reasoning over natural language and images, with a particular focus on semantic diversity, compositionality and visual reasoning challenges. It comprises 107,292 English sentences paired with pairs of web photographs, and the core task is to determine whether a natural language caption matches the corresponding pair of photographs. The dataset induces linguistic diversity by leveraging visually rich image sets and comparative as well as contrastive tasks. Qualitative analysis reveals that the data demands compositional joint reasoning, including quantity, comparative and relational reasoning. Evaluations using state-of-the-art visual reasoning methods demonstrate that this dataset poses substantial challenges.