layoutlm

Definition

A family of document-understanding models from Microsoft that pre-trains text jointly with its 2D position on the page — token bounding boxes — and, in later versions, with visual features as well. On forms, invoices, and receipts, where meaning depends on where something sits, this clearly beats text-only encoders like BERT. It consumes OCR output as input, which is what separates it from OCR-free models such as Donut.

Explainers using this term

• Document AI and OCR Today — How an LLM Ends Up Reading Your InvoicesLayoutLM: Pre-training of Text and Layout for Document Image Understanding