Zovda
Charts are the densest common format in business documents and the least represented in training data. When they are included, they are usually an image with a question and an answer.
That is enough to test a model and not enough to teach one, because the answer came from numbers the model never sees.
Pixels are a lossy encoding
A bar that reaches the gridline between 40 and 50 could be 44 or 49. A model reading pixels is estimating, and an estimate that happens to be right on the test set is not a capability.
Pair the image with the table it was drawn from. Then the task has a ground truth that does not depend on how the renderer rounded.
The fields that make a chart record
- The image, at a known size
- The underlying table: series, categories, values
- The chart type and the axis ranges, including whether they start at zero
- The title and the axis labels, verbatim
- The question, and the operation it requires: read, compare, sum, or extrapolate
The last field separates a lookup from a computation, and the two should be counted separately.
Axis tricks are the test
Truncated axes, dual axes, and a line chart on a log scale are where a chart-reading model earns its keep. A dataset of well-formed zero-based bar charts teaches the easy half.
Collect the charts people actually make, with their habitual distortions, and label the ones that mislead.
Ask questions the table can answer
Write each question against the table first. If the table cannot answer it, the question is testing the renderer, not the reader.
Zovda builds these records through Datuva, where the table stays attached to the image. The image collection lists the fields a document image carries.
Tell us what your model has to read off a page. We reply within one business day.