ChartQA 使用视觉和逻辑推理回答图表问题的基准数据集

图表在数据分析中非常受欢迎。在探索图表时，人们经常提出各种涉及多个逻辑和算术操作的复杂推理问题。他们在问题中通常还常常提到图表的视觉特征。然而，大多数现有数据集并未专注于此类复杂推理问题，因为它们的问题是基于模板的，答案来自固定的词汇。

Charts are extremely popular in data analysis. During chart exploration, people often raise a variety of complex reasoning questions involving multiple logical and arithmetic operations. These questions frequently also reference the visual characteristics of the charts. However, most existing datasets do not focus on such complex reasoning questions, as their questions are template-based and their answers are derived from a fixed vocabulary.

搜集汇总

背景与挑战

背景概述

ChartQA是一个大规模基准数据集，专注于通过视觉和逻辑推理回答图表中的复杂问题。它包含9.6K个人工编写问题和23.1K个生成问题，基于真实世界图表和人工标注，区别于以往基于模板的数据集。