Researchers at the Massachusetts Institute of Technology have created a dataset of over 12,400 technology, trade, retail, and sports charts. For each chart, the dataset contains an image of the chart, its data table, a scene graph (a data structure depicting hierarchical relationships between data points), a caption describing its basic construction, and a caption describing trends and relationships in the chart’s data. Researchers can use the dataset to improve language models that generate chart captions.
Image credit: Flickr user Nikolas Techenburg