Document annotation and OCR training data
Invoices, forms, IDs, contracts and handwritten pages labeled for document AI: fields boxed and named, tables mapped, reading order set and text transcribed exactly as written, in any script. Built for key-value extraction, OCR and layout models.
Example invoice with sample data. The date is labeled day first, as written in the UK, and normalized in the output.
- Documents
- Invoices, receipts, forms, IDs, contracts, bank statements and handwriting
- Labels
- Field boxes, key-value pairs, tables, reading order and transcription
- Scripts
- Latin, Arabic, Cyrillic, CJK, Devanagari and more
- Quality
- Character-level checks and a reviewed sample on every batch
What document annotation is
Document annotation is labeling scanned or digital documents so a model can learn to read them: drawing boxes around fields, naming each one, mapping table rows and columns, and transcribing the text exactly. The labels teach OCR and document AI models where information sits and what it means.
Documents from different countries follow different conventions for dates, addresses, names and numbers. Our annotators read the language and script on the page, so a date written 03/09 is labeled the way it was meant.
Document annotation tasks
For extraction, OCR and layout models.
Key-value extraction
Fields boxed and named, with values normalized to your schema.
Table structure
Rows, columns, headers and merged cells mapped in statements and reports.
Handwriting transcription
Handwritten forms, notes and letters transcribed exactly, with uncertain characters flagged.
OCR ground truth
Line and word boxes with exact text, for training and testing OCR engines.
Document classification
Document type, language and quality labeled before extraction.
Redaction
Personal and sensitive fields marked or masked before documents are used for training.
How a document annotation project runs
Field definitions come first, with examples of every document type.
Schema
Your fields, value formats and rules for missing or unclear values.
Pilot
A set of documents per type and language labeled by several annotators.
Accuracy check
Field and character accuracy measured, and the schema updated with you.
Accuracy report QAProduction
Batches labeled to your schedule, with a reviewer checking a sample.
Delivery
JSON, COCO-style boxes, PAGE XML or your tool's export.
Related services
Questions about document annotation
What is document annotation?
Document annotation is labeling scanned or digital documents for machine learning: boxing and naming fields, mapping tables, setting reading order and transcribing text, so models can learn to extract information from similar documents.
Can you annotate handwritten documents?
Yes. Annotators transcribe handwriting exactly as written and flag characters they can't read with confidence, so your model isn't trained on guesses.
Can you work with documents in Arabic, Chinese or Hindi?
Yes. Documents are labeled by native readers of the script, including right-to-left and CJK layouts, and values are normalized to your schema.
How do you protect personal data in documents?
We sign your NDA, limit access to the project team and can redact personal fields before or during labeling. See data security.
Send us sample documents
Send a few documents and your field list, or book a call. We label a pilot set and report field accuracy.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time