Selection guide / editorial
Extract invoices and forms
The useful outcome is correct fields with evidence: supplier, invoice number, dates, currency, totals and line items.
Define a schema
Specify required fields, null for missing information, currency handling and how to represent line items. Keep source page and region references.
Separate recognition and extraction
Measure OCR mistakes separately from field mapping. Correct text can be assigned to the wrong field, and valid JSON can invent a missing number.
Use held-out layouts
Hold out suppliers and document templates. Measure exact field match, numeric tolerances fixed in advance, missing-field abstention and line-item structure.
Review exceptions
Track how often a person must intervene and which errors are costly. FUNSD offers form-understanding context; it is not a substitute for your invoice-specific hold-out set.