A missing value and an unchecked box are not the same state
Define a schema and state vocabulary before mapping OCR text. Plain OCR does not prove which answer belongs to which field.
The finished extraction should preserve field labels and explicit states, preventing an unreadable mark, empty response, unchecked box, and not-applicable answer from collapsing into one blank value.
Put the procedure into practice with Form to JSON. Continue with OCR Output Formats: TXT, JSON, TSV, hOCR, ALTO and PDF and Numeric OCR Validation for Dates, Totals and Identifiers.
Form-state review
- Write the destination schema first
- List every allowed field state
- Retain page and region identifiers
- Review handwriting separately
- Require human review for ambiguous marks
Define the form schema before mapping recognized text
Inventory field types
Separate printed labels, typed answers, handwriting, checkboxes, radio groups, signatures, dates, and continuation areas. Different field types need different recognition and review rules.
Define explicit states
For each field allow values such as present, blank, unchecked, checked, not applicable, and unreadable. Do not collapse these into an empty string because downstream users cannot recover the distinction.
Anchor repeated labels
Forms often repeat Name, Date, Yes, or No. Preserve section and page identifiers plus coordinates so each answer can be attached to the correct instance.
Separate extraction from interpretation
Recognizing a mark in a checkbox is not the same as deciding the legal or operational meaning of the form. Route ambiguous overwriting, correction fluid, and crossed-out answers to a reviewer.
Represent absence as an explicit reviewed state
A form can contain an unchecked box, an empty text field, an unreadable mark, a not-applicable response, and a field that is missing from the page. Those states must not collapse into one empty string. Define the permitted state for each field before mapping OCR text.
For repeated labels such as Name or Date, include the section, page, and region in the field key. Preserve the raw token and source location beside the reviewed value. Signatures, handwriting, and checkbox marks need visual review even when nearby printed labels are recognized perfectly.
- Define schema and states first
- Anchor repeated labels to their section
- Keep source location with every reviewed value
Form state vocabulary
| Visible source | Recorded state | Value |
|---|---|---|
| Empty box | unchecked | false |
| Marked box | checked | true |
| Field left empty | blank | null with reason |
| Mark cannot be resolved | unreadable | null plus review flag |
Absence is not one state
Plain OCR cannot reliably distinguish an unchecked box, an empty response, an unreadable mark, a hidden field, and a field that does not apply. Do not collapse them into one null value.