Winoground

Winodground数据集包含400项 (每个项由两个具有重叠词汇内容的图像文本对组成)。根据文本是否交换对象，关系或两者进行语言分类。这些项目根据是否涉及进一步分类: 语用学标签，指示所需的非文字/语用推理，符号标签，指示对符号空间中的事物进行推理 (例如，儿童绘画)，以及系列标签 (指示项目是否来自相同，而不是来自无关的照片)。

The Winoground dataset consists of 400 items, each composed of two image-text pairs with overlapping lexical content. Each item is linguistically classified based on whether the text swaps objects, relations, or both. These items are further categorized via three types of labels: pragmatic labels, which indicate the requirement for non-literal or pragmatic reasoning; symbolic labels, which denote reasoning about entities in symbolic spaces (e.g., children's drawings); and sequence labels, which indicate whether the items originate from the same set of photographs rather than unrelated ones.

搜集汇总

背景与挑战

背景概述

Winoground是一个由德克萨斯大学奥斯汀分校于2022年发布的数据集，包含400项图像文本对，每对具有重叠词汇内容，并根据文本是否交换对象、关系或两者进行分类。数据集还涉及语用学标签、符号标签和系列标签，用于指示推理类型和项目来源。