English
Get the text out of handouts, worksheets and textbooks
Teaching material is rarely one column of prose. A page carries a heading, a numbered exercise, an equation, a diagram with labels pointing into it, a box for the answer, and quite often a pupil’s working and a marker’s red pen on top of all that. Evidano reads the printed layer and the written one separately, keeps question numbers attached to their questions, and does not turn a subscript into an ordinary digit.
Sample pages
Real pages from public collections, shown beside their transcriptions. Pages whose transcription is still being checked are marked.
Source · Public domain · Digitised by Google, via the Internet Archive
Transcription
The lesson heading, the introductory sentence and all five questions, from the Wikisource transcription of this page; the caption inside the engraving reads DIVISION. and the bracketed page number at the foot is the book’s own.
DIVISION. LESSON LII. In the picture you can count five wild geese and twenty ducks. 1. If twenty ducks were divided into four flocks, each having the same number, how many ducks would there be in each flock? 2. How many times is 5 contained in 20? 3. If five wild geese have ten wings, how many wings has each goose? 4. How many times is 5 contained in 10? 5. There are four full-blown flowers on two branches: how many on each branch? (56)
Transcription from Ray’s New Primary Arithmetic for Young Learners, page 56, via Wikisource (Public domain).
Transcription being verified
A transcription exists at Page:Hill’s manual of social and business forms.djvu/69 on Wikisource, but it is still marked as not proofread — raw recognition with visible errors — so it is not reproduced here as a checked text.
What the model is told to watch for: Separate printed instructional content from student answers, teacher corrections, grading marks, and marginal annotations. Preserve mathematical operators, subscripts, superscripts, equation alignment, answer-box structure, diagram labels, and arrows connecting notes to specific passages or figures.
Source · CC BY-SA 4.0 · Photograph by Jonathan Schilling
Transcription being verified
Course papers of this kind are not published; the sheets carry only the instructor’s name and the course details, and no student information.
What the model is told to watch for: Separate printed instructional content from student answers, teacher corrections, grading marks, and marginal annotations. Preserve mathematical operators, subscripts, superscripts, equation alignment, answer-box structure, diagram labels, and arrows connecting notes to specific passages or figures.
NASA Johnson Space Center, 2002 · Source · Public domain
Transcription being verified
The guide is published as a PDF with a text layer; this page is used here to test two-column reading order and the separation of the running foot from the lesson.
What the model is told to watch for: Separate printed instructional content from student answers, teacher corrections, grading marks, and marginal annotations. Preserve mathematical operators, subscripts, superscripts, equation alignment, answer-box structure, diagram labels, and arrows connecting notes to specific passages or figures.
What a page of teaching material contains
The genre covers school textbooks and readers, workbooks and photocopied worksheets, lecture handouts and slide printouts, past papers and mark schemes, laboratory sheets, revision cards and the notes students take from all of them. What they have in common is structure that carries meaning. A number at the start of a line is a question number, not a bullet; an indented paragraph in smaller type is a worked solution; a rule across the page divides one exercise from the next; and a word in bold is usually the term the lesson is about. Older schoolbooks add a running head and a page number in brackets at the foot, and set their exercises in the same size as the explanation, so the two run together in a plain reading.
Then there is everything that is not prose. Mathematics brings fractions written on two lines, exponents, subscripts, roots and aligned equals signs, none of which survive a straight left-to-right reading. Science brings units, degree signs and formulae such as H₂SO₄ where the digits are part of the name. Diagrams carry labels connected by leader lines to the thing they name, and a labelling exercise may be nothing but blank rules waiting to be filled in. Language sheets use gapped sentences, tick boxes and matching columns joined by lines drawn across the page.
Most real pages are also layered. A worksheet is printed, filled in by a pupil in pencil, marked by a teacher in red, and then photocopied for a file, so three inks and a photocopier’s shadow arrive together. Handouts get stapled, folded, annotated in the margin, and photographed at an angle on a desk with a second sheet showing underneath. A slide deck printed six to a page shrinks every line to a few millimetres. Anything you send is likely to have at least two of these problems at once.
Why ordinary OCR struggles here
Printed text and pupil’s answers mixed together
A completed worksheet holds the question, the answer written into the gap and the marker’s comment on the same line. Read as one stream they become nonsense. They have to be separated by hand or by colour and shape, and each part labelled, before the sheet can be marked up or analysed.
Mathematics is not a line of characters
A fraction occupies two lines, an index sits half a line up, and a long division is a spatial arrangement. Flattened into a sentence, x² becomes x2 and ½ becomes 1 2. Getting this right means recognising the notation as notation and writing it out in a form — LaTeX, MathML or plain words — that can be read back.
Two columns, and the order they run in
Teacher’s guides and textbook pages are commonly set in two columns with headings inside them. A recogniser that reads straight across the page produces alternating half-sentences from the left column and the right, which is worse than useless for a lesson plan whose steps are numbered.
Photocopies, angles and show-through
Class material is copied until the small print greys out, the staple shadow falls across the first line, and the sheet is photographed on a desk where the page below shows at the edge of the frame. Text bleeding through thin paper is often as dark as the text on the front.
Who works with this material
Teachers rebuilding old resources
A worksheet that exists only as a photocopy of a photocopy cannot be edited, translated or made larger for a pupil who needs it. Getting the text back gives you something you can change, and the numbering and answer spaces come with it.
Accessibility and inclusion teams
A scanned handout is invisible to a screen reader. Converting it properly means reading order, headings, alternative text for diagrams and mathematics written so it can be spoken — not a picture of a page with a caption stuck under it.
Publishers and course teams
Moving a print course into a virtual learning environment means recovering thousands of exercises with their numbering and answers intact. A first pass that keeps question and answer keyed to each other saves most of the retyping.
Education researchers
Studies of curriculum change, marking practice or textbook language need the text of the materials, not scans. Where pupils’ work is involved, names and other identifying details are best removed or masked before the sheets go anywhere.
Getting the best transcription
Say which layer you want
Ask for the printed question paper, the handwritten answers, or both with each labelled. On a marked script, ask for the marker’s annotations and any grade as a separate field; it is the difference between a usable transcript and a tangle.
Choose a notation for mathematics
Name the format — LaTeX, MathML, Unicode or spoken words — before you start. LaTeX suits anything that will be typeset again, Unicode suits simple expressions in a word processor, and spoken forms suit accessibility work. Mixing them across a set of papers makes the results hard to use.
Flatten the sheet and shoot it square
Photograph one page at a time, from directly above, with the neighbouring sheet moved out of the frame. If a page is stapled, unfold it; if it curls, weight the corners. This costs seconds and removes most of the errors that come from a warped or skewed line.
Handle pupils’ names before you share
Class material often carries a child’s name, a class list or a candidate number in the header. Decide whether those fields should be dropped or replaced with a code as part of the transcription, so that what leaves your machine is already anonymous.
Frequently asked questions
- Can it keep question numbers with their questions?
- Yes. Numbered and lettered items — 1, 2, 3, then (a), (b), (c) and roman sub-parts — can be returned as a structure rather than as a run of text, which means a question and its parts stay together even when they break across a column or a page.
- How are equations returned?
- In the notation you ask for. LaTeX is the usual choice because it can be typeset or converted onward; simple expressions can come back as Unicode with proper superscripts and subscripts. Alignment of a multi-line calculation is preserved as far as the notation allows.
- What happens to a diagram with labels?
- The labels are transcribed and can be listed with the part of the figure each one points to, so a labelling exercise or a circuit diagram keeps its sense. The drawing itself is described briefly rather than redrawn.
- Can a printed answer key be matched to the questions?
- If the key is in the same image, the answers can be returned in the same numbering as the questions. If it is on a separate sheet, send both and say which is which; the numbering is what joins them.
- Is it safe to send marked work with pupils’ names on it?
- Treat it as personal data. The safest route is to crop or cover the name and any candidate number before uploading, or to ask for those fields to be replaced with a code in the output. Nothing on this page uses a real pupil’s work for that reason.
Related scripts and pages
Turn your teaching materials into text
Upload a worksheet, a past paper, a handout or a page of class notes, choose Educational Materials, and say whether you want the printed layer, the written one, or both.