OCR Handwritten Forms: Intake Sheets and Questionnaires

OCR can read handwritten forms in part. The printed labels come back cleanly, block capitals written one per box give the software a fair chance, and free script on a line is a gamble. Process a stack by converting every form, then checking each handwritten answer against the page before it goes into a spreadsheet.

That split between the printed half and the handwritten half of the page shapes everything else. The form's designer decided, long before the client picked up a pen, how readable the answers would be. This post covers what the design decides, the forms a law office collects by hand, what to do with answers that leave their fields, and a procedure for getting a pile of completed sheets into one table.

Can OCR read a form filled in by hand?

OCR can read a form filled in by hand to the extent that the answers look like print. The pre-printed questions convert like any typed page. The handwritten answers convert according to the writer, the pen and the space the form allowed. Expect a draft of each answer and plan to check every one.

A completed form is two documents laid on top of each other. One was typeset: the firm's name, the questions, the consent wording, the field labels. The other was written by a person balancing a clipboard on one knee. Recognition software treats each mark it finds on its merits, so one page can produce a perfect line of label text followed by a garbled surname.

Confirm the client's name, date of birth, any amount, and signature on the intake sheet. On an accident questionnaire, confirm the date of the event and the statement the client wrote. If those answers later appear in a pleading or in discovery, the cell has to match the page. A statute of limitations count uses the date you verified. Build a deposition outline from that same checked date.

The general limits of handwriting recognition apply here without change, and the post on what machines can read in handwritten notes sets them out. What a form adds is constraint. A field tells the writer where to put the answer and how much room there is. A good field also slows the writer down. Both help.

Three results are possible for any single answer:

• Read correctly. The text matches the page.
• Read wrongly. The text is a plausible word, name or number that the client did not write. Nothing in the output distinguishes this from the first result.
• Left out. A faint or small answer is treated as a stray mark and no text appears for it.

The second and third results are why a converted form is never the record of what the client said. Schema Law's own homepage says recognition is strongest on clean print and weaker on handwriting, and that nothing marks a passage as uncertain. Read that as the working assumption for any tool you use on a stack of forms: the output will not tell you which answers to doubt, so the checking has to cover all of them.

Block capitals in boxes versus free handwriting on a line

Block capitals in boxes give software separate characters of known size in known places. Free handwriting on a line gives it a continuous stroke of unknown length. The form design decides which of the two the client produces, and so shapes the result before any scanning happens.

What a row of boxes does

A row of small boxes, sometimes called a comb field, forces one character per cell. The writer cannot join letters, because the box edges are in the way. The writer cannot shrink the writing to fit a long word, because each letter has its own cell. The answer comes out as spaced capitals, which is the handwriting closest to print.

Boxes also make a human check faster. A reader comparing the draft with the page can count cells. If the phone number field has ten boxes and the draft shows nine digits, a digit is missing, and you know it without reading a single character closely.

What an open line does

An open line invites the client's ordinary hand. That may be tidy print or it may be joined script, and the form has no say. Joined writing removes the gaps software uses to find where one letter stops. The post on whether OCR can read cursive covers that problem, so this one stops at the design point: an open line gives up control over the handwriting.

The same answer three ways

Take a fictional client, Marguerite Oyelaran, giving her address on three versions of an intake sheet.

None of the three is guaranteed. The table shows where the risk moves, and the only way to find out what your own forms produce is to convert a handful and compare.

Client intake sheets and questionnaires completed in a waiting room

Intake sheets and questionnaires completed in a waiting room are handwritten forms made under poor conditions: a clipboard, a borrowed pen, limited time and a client with other things on their mind. They carry names, dates, phone numbers and narrative answers that the firm then relies on for conflict checks, contact and the first draft of the facts.

The forms a law office collects this way include:

• New client intake sheets. Name, address, contact details, opposing party, how the client found the firm.
• Practice-specific questionnaires. An accident questionnaire, a family law financial questionnaire, an immigration history, a list of creditors.
• Authorizations and releases. Mostly printed text with a name, a date of birth and a signature added by hand.
• Forms the client completed elsewhere and brought in. A clinic registration sheet, an employer's incident report, an insurer's claim form.

Each has two kinds of answer, and they deserve different treatment.

Short identifying answers are names, dates, numbers and addresses. They are the answers a conflict check or a limitation calculation depends on, and they are the hardest for software to guess, because nothing in the surrounding words predicts a surname or a date of birth. Check each one character by character.

Narrative answers are the lines under "Describe what happened" or "List your injuries". They are longer, more hurried and more likely to be in the client's everyday script. A draft of a narrative answer is useful for search. It is not the client's statement until someone has read it against the page.

What to capture besides the answers

A paper intake form holds information that sits outside any field, and a text conversion can lose all of it:

• A date stamp or a receptionist's "rec'd" note in a corner.
• A staff member's initials and file number added after the client left.
• Blank fields. An empty "Prior attorney" line is an answer of a sort, and a spreadsheet cell should show that the field was blank on the paper instead of looking like a field nobody keyed.
• Which pen wrote which part. An answer added later in different ink is worth a note.

Keep the scan of every form with the matter. Whatever the spreadsheet says, the page image is what shows the client's own writing.

How should a firm design a form so handwritten answers convert?

A firm should design the form so that handwriting is needed as little as possible and is constrained where it is needed: boxes for identifiers, tick choices in place of written words, generous spacing, clear printed labels and dark ink. The best conversion result comes from the answer the client never had to write out.

A paper form the firm controls is the one place in a handwriting workflow where you set the conditions. Use that.

A design checklist

1. Replace written answers with choices. "Marital status" with four tick boxes produces a mark to check. The same question with a line produces a handwritten word to read.
2. Use one-character boxes for anything you will search or calculate on. Names, dates of birth, phone numbers, postcodes, policy and claim numbers.
3. Print the date format inside the field. Three small groups of boxes labelled MM, DD and YYYY remove the question of which number is the month.
4. Ask for capitals in words, next to the field. An instruction at the top of the page is forgotten by the second question.
5. Make boxes large enough for an adult hand. Cramped boxes push strokes onto the borders, where they merge with the printed line.
6. Leave white space between fields. Answers that touch each other are hard for software and for the person checking.
7. Print labels in a plain typeface, to the left of or directly over the field, never inside the writing area. Ink written across a printed label damages both.
8. Give narrative questions real room. Four ruled lines, widely spaced, with a printed "continue on the back of this sheet" beneath them.
9. Number every question. A number gives you a stable column name for the spreadsheet and a way to say which answer you mean.
10. Print a form version and page number in the footer. When the form changes next year, you can tell which layout a scanned sheet follows.
11. Hand out black pens. The effect of pencil and pale ink on a scan is covered in the post on scanning handwriting, under pencil, blue ink and felt tip under the scanner.

Design choices to avoid

• Grey shading behind fields, which a scanner can turn into speckle around the writing.
• Two-column layouts where an answer on the left can run into a question on the right.
• A signature line placed so close to a field that the signature crosses it.
• Double-sided printing on thin paper, where writing on the back shows through.

Test the form before you print a ream

Ask three colleagues with different handwriting to complete the draft form with a pen from reception. Scan the three sheets the way your office scans everything else, convert them, and compare each answer with the page. Fields where all three went wrong are design problems. Fields where one went wrong are handwriting problems you will meet again with clients. Change the first kind before printing.

Handwritten answers that spill outside the field or into the margin

Handwritten answers that spill outside the field are the answers most likely to be attached to the wrong question, cut in half or dropped. Software sees ink in a place the layout did not predict. Handle overflow by finding it on the image first and recording where each stray piece belongs.

Overflow takes a few recognisable forms:

Matching each printed label to the right answer is its own subject. The post on scanned forms explains keeping each answer beside its field label, and that method applies to typed and handwritten entries alike. Tick boxes are treated there too, under whether OCR can tell a checked box from an empty one. A client who ticks beside a box, circles the printed word or writes "yes" across two boxes has produced an overflow problem and a checkbox problem at once, so look at those marks on the image yourself.

Find overflow before you read the draft

Go through the stack of page images once, looking only at margins, the foot of each page and the reverse side. Put a flag on every sheet with writing outside a field. This pass takes seconds per sheet because you are looking at shapes, not reading. The flagged sheets are the ones where you compare the draft with the image line by line.

Scan both sides of every form, including sides that look blank. A blank reverse is evidence that nothing was written there. A reverse that was never scanned is a question you cannot answer later.

From a stack of handwritten forms to one spreadsheet row per client

Getting a stack of handwritten forms into one spreadsheet row per client takes seven steps: fix the columns, scan, convert, key each answer with the image open, mark doubtful cells, have a second person check, and store the scans with the sheet. The conversion supplies a draft. The person keying supplies the accuracy.

1. Fix the columns from the blank form. One column per numbered question, named with the question number and a short label, such as Q07_date_of_birth. Add four columns the form does not have: source_file, pages, overflow_note and checked_by.
2. Sort the stack by form version. Read the footer. Forms with different layouts go in different batches, because question 12 on one may be question 14 on another.
3. Scan each form as its own file, both sides. Name the file so the client or matter is identifiable. One form per file means one row per file, and no splitting later.
4. Convert a sample first. Take five forms from different writers and run them. Compare the handwritten answers with the pages and decide whether correcting the draft or typing from the image is faster for this form. The post on choosing a converter describes how to score a handwriting converter's output.
5. Key each row with the page image beside the sheet. Paste or type the answers into their columns. For names, dates and numbers, read the image and ignore the draft until you have your own reading, then compare the two.
6. Mark what you could not read. Use one fixed marker for an unreadable answer and another for a field left blank on the paper, for example [illegible] and [blank]. Never fill a cell with a guess. Conventions for doubtful words are set out under how a transcript should mark words nobody can read.
7. Have a second person check the identifiers. They read the name, date of birth, phone number and any claim or policy number from the image aloud while the first person watches the row.

What the tool contributes

A general conversion tool gives you the printed text and a first reading of the handwriting. In Schema Law you sign in, upload a PDF, PNG or JPEG of up to 20 MB, and view the parsed pages. It reads pages into text and tables. It does not hand back a row with one named field per answer, so the step from page text to columns is yours. The free plan takes one document a day of up to three pages, which is enough to try one completed form, and every plan includes all eight export formats, Excel and CSV among them.

A worked row

A fictional firm, Thackeray & Lund, has 40 completed accident questionnaires. The row for one client ends up like this:

The cell for question 11 is the valuable one. It tells whoever opens the sheet that the accident date on this form has to be confirmed with the client. A confident wrong date in that cell would have gone unnoticed until it mattered.