English
Answers that stay attached to their questions
A form is two documents printed on one sheet: the questions, boxes and small-print instructions that were there before anyone touched it, and the answers somebody added afterwards. Read it as flat text and the two layers interleave into nonsense — a surname next to the words “Family name”, a tick with nothing to tick. Evidano keeps them apart, pairs each response with the field it belongs to, and tells you when a box was left empty rather than quietly skipping it.
Sample pages
Real pages from public collections, shown beside their transcriptions. Pages whose transcription is still being checked are marked.
Source · CC BY-SA 4.0 · Photograph by Admiral Rolland
Transcription being verified
A blank form, reproduced so that no patient data appears. No published transcription exists; the printed field labels are the text to be captured.
What the model is told to watch for: Separate preprinted labels and instructions from handwritten or typed responses, and associate each response with its field, checkbox, radio button, or rating scale. Represent checked, crossed, overwritten, blank, and conditionally skipped fields distinctly, including handwritten additions that extend outside their designated boxes.
National Archives and Records Administration, 5 June 1917 · Source · Public domain
Transcription being verified
The card is indexed in the NARA series of First World War draft registration cards; the index gives name, date of birth and registration date but not a field-by-field transcription of the sheet.
What the model is told to watch for: Separate preprinted labels and instructions from handwritten or typed responses, and associate each response with its field, checkbox, radio button, or rating scale. Represent checked, crossed, overwritten, blank, and conditionally skipped fields distinctly, including handwritten additions that extend outside their designated boxes.
Seattle Municipal Archives, 18 September 1894 · Source · CC BY 2.0 · Seattle Municipal Archives, via Flickr
Transcription being verified
Catalogued by the Seattle Municipal Archives as part of the Health Department licence applications; the catalogue record names the applicant but does not transcribe the sheet.
What the model is told to watch for: Separate preprinted labels and instructions from handwritten or typed responses, and associate each response with its field, checkbox, radio button, or rating scale. Represent checked, crossed, overwritten, blank, and conditionally skipped fields distinctly, including handwritten additions that extend outside their designated boxes.
Internal Revenue Service, tax year 1941 · Source · Public domain
Transcription being verified
A working copy of a federal form; the blank version is reprinted in the Internal Revenue Service’s annual instructions for 1941, but the pencilled figures on this sheet are unpublished.
What the model is told to watch for: Separate preprinted labels and instructions from handwritten or typed responses, and associate each response with its field, checkbox, radio button, or rating scale. Represent checked, crossed, overwritten, blank, and conditionally skipped fields distinctly, including handwritten additions that extend outside their designated boxes.
United States Census Bureau, 1940 · Source · Public domain
Transcription being verified
The wording of every column and of the supplementary questions is printed in the Census Bureau’s 1940 Instructions to Enumerators; a transcription keyed to this sheet has not been published.
What the model is told to watch for: Separate preprinted labels and instructions from handwritten or typed responses, and associate each response with its field, checkbox, radio button, or rating scale. Represent checked, crossed, overwritten, blank, and conditionally skipped fields distinctly, including handwritten additions that extend outside their designated boxes.
The two layers on a completed form
The printed layer is dense and highly structured. A form usually carries a title and a form number, a block of instructions in smaller type, numbered items, field labels with sub-labels in brackets underneath the rule (“(Given name)”, “(Street and number)”), comb boxes that take one character per cell, tick boxes and radio circles, rating scales, and one or more panels reserved for the office that will process the sheet. Some of the printed text is rotated ninety degrees along the edge; some of it is a conditional instruction that applies only to certain lines, as on a census schedule where supplementary questions are asked only of the people entered on two particular rows.
The response layer is whatever the person in front of the form did. Block capitals for a name, joined writing for the rest, a tick or a cross or a diagonal stroke or a circle round the chosen option, a figure written over another figure without the first being crossed out, a dash or “N/A” or nothing at all where a question did not apply. Answers routinely leave their boxes: a long address wraps into the field below, an explanation carries on up the margin with a bracket, an amount is written across the dollars-and-cents divider rather than either side of it.
Then there is everything added after the form was handed in. A rubber stamp with a date, a reference number written diagonally across the sheet in a different pen, an approval initialled in red, a coding mark made by a clerk preparing the data for punching, a note that the licence was issued and when. These belong to the file rather than to the respondent, and a transcription that treats them as answers will attach a clerk’s reference number to whichever question it happens to overlap.
Why ordinary OCR struggles here
The label and the answer are separate things
Plain text extraction returns the words in the order they appear on the page, which mixes the questionnaire with its responses. What you actually want is a pairing: item 7 asks for the present trade or occupation, the answer is “Motion Picture Perf”. Without the pairing you cannot load the sheet into a spreadsheet, and you cannot tell a printed example answer from a real one.
Empty boxes carry meaning
A blank field can mean the question did not apply, that the respondent refused, that the enumerator skipped it under an instruction, or that the form was never finished. Silently dropping empty fields destroys all four possibilities, and makes an incomplete return look like a complete one. Blanks have to be reported as blanks, and distinguished from fields the model could not read.
Marks are not characters
Ticks, crosses, dots, filled bubbles and circles drawn round a printed word are answers, but no letter corresponds to them. Add the human tendency to put the mark between two boxes, or to tick and then cross out, and the interesting cases are exactly the ones a character recogniser cannot represent at all.
Writing that will not stay in its box
Comb boxes assume one letter per cell and long names ignore them. An explanation runs past the ruled line into the next question. A pencilled correction sits above the figure it replaces. Deciding which field a stray line of writing belongs to needs the geometry of the blank form, not just the ink on this copy.
Who works with this material
Records teams digitising a back-file
Licence applications, permits, membership forms and registration cards sit in boxes because keying them costs more than they are worth one at a time. Field-and-value output turns a run of identical sheets into rows, and the exceptions — a form filled in wrongly, an unsigned declaration — surface instead of being averaged away.
Survey and evaluation teams
Paper questionnaires still come back from clinics, classrooms and doorsteps. What matters is that a rating is read as the point on the scale that was marked, that free-text comments stay attached to the question that prompted them, and that partially completed returns are visibly partial when they reach the analysis.
Genealogists and archive indexers
Census schedules, enlistment papers and registration cards are the backbone of family history, and their value depends on names, ages and places being read exactly. An indexer wants the enumerator’s own spelling, the crossings-out, and a flag on anything doubtful, not a tidied version.
Administrators handling sensitive returns
Medical request forms, benefit claims and consent forms carry identifiers that most downstream uses do not need. Being able to capture the clinical or operational fields while holding back names, numbers and dates of birth is often the difference between a usable workflow and one that cannot be approved.
Getting the best transcription
Show the blank form first
If you have an unused copy of the same form, transcribe that before the completed ones. It gives you the printed layer on its own — every label, every instruction, the position of every box — so the completed sheets can be read as answers against a known template rather than deciphered from scratch each time.
Ask for pairs, not paragraphs
Request the output as field and value, one per line or as JSON, and say what an untouched field should look like: an empty string, the word “blank”, or an explicit note that the enumerator was told to skip it. That single decision is what makes a stack of forms loadable into a spreadsheet.
Spell out how marks should be reported
Say whether a tick, a cross and a circled option should all come back as “checked”, and what should happen when a mark straddles two boxes or an option is ticked and then struck out. Ambiguous marks are better reported as ambiguous than resolved by a coin toss.
Decide about personal data before you start
Forms are the most identity-dense documents most people scan. Say in the prompt which fields should be redacted or replaced by a placeholder — name, address, date of birth, reference numbers — and check that they really are absent from the first few outputs before you run the batch.
Frequently asked questions
- How does it know which answer belongs to which question?
- From the geometry of the sheet: the position of the writing relative to the rule, the box and the label, and the numbering of the items. On a form where the questions run on as sentences, the answer is the writing inside the sentence, and it is returned against the printed wording that surrounds it.
- Are ticks, crosses and circled options recognised?
- Yes, and they are reported as the state of the field rather than as characters. A ticked box, a crossed box, a filled circle and a ring drawn round a printed word are all treated as a selection; a mark that sits between two options or that has been struck out again is flagged rather than assigned.
- Can it tell an empty field from one it could not read?
- These are reported differently. An empty field comes back as empty; writing that is present but illegible is marked as unread, so a person can go back to the image. Conflating the two is how partly completed forms turn into confident, wrong data.
- What happens when the answer runs outside its box?
- The writing is assigned to the field it started in, and the overflow is kept with it rather than split between two fields or moved into the answer below. Where a comb box has been ignored and a name written straight across the cells, the name is read as a word.
- Can names and identifiers be kept out of the transcription?
- Yes. You can ask for named fields to be redacted or replaced with a placeholder so that the structure and the substantive answers survive without the identity. This is worth doing on any batch of medical, benefit or membership forms before the text goes anywhere else.
Related scripts and pages
Turn a box of completed forms into rows
Upload a blank form and a filled one, choose Forms and Questionnaires, and check the field pairing before you run the rest of the batch.