Document Understanding
Long, semi-structured, and visual documents made usable.
Document understanding research makes long, visual, and semi-structured documents usable. Important context lives in scans, statements, filings, contracts, forms, and charts. Visually rich document understanding studies how to read that material while keeping structure, layout, and provenance intact.
Core questions
- How should a system represent pages, sections, tables, images, and extracted facts together, so that a downstream reader can trace a fact back to the page it came from?
- What should be preserved from layout and visual structure before a model summarizes or reasons over the content?
- How do you measure extraction quality on your own documents, cheaply enough to compare parsers and models as they change?
Where it is applied
This research is applied in Verdant, which digests images and PDFs into structured Markdown and HTML and scores each run against truth fixtures. The structured output feeds information extraction and, from there, GraphForge.