OCR Translation: Making a Scanned Document Translatable

The first step of the translation process is usually invisible. If the document in your hands is a scanned diploma, a photographed invoice or a locked PDF, no translation tool can read that text directly. OCR preparation for translation comes in between: the optical character recognition step that turns the letters in an image into editable text.

Contents

The step looks simple, but it has a direct bearing on the result. A misrecognised digit, a lost line or a confused letter can be carried into the translation exactly as it stands. In official documents an error of this kind can lead to an application being refused.

Below we set out how this step works, where it goes wrong and how the file you send determines the quality.

What does OCR do in the translation process?

Optical character recognition identifies the shapes of the letters and digits in an image and converts them into text characters. The software divides the page into regions, locates the lines, assesses the letters one by one and picks the most probable character. The output is text that can be searched and edited.

Without that conversion, translation tools cannot come into play. A translation memory looks for sentences; an image holds no sentences, only pixels. A file sent without recognition means the translator reading it off the screen and typing it out by hand, and that is both slow and expensive.

Its second function is preserving the layout. Good recognition rebuilds not only the letters but the table structure, the columns and the paragraph arrangement as well. That is what allows the translated document to be delivered looking like the original.

Its third function is archivability. Recognised text feeds the translation memory for later jobs. Similar documents from the same institution are processed faster the next time round, and because the templates of official documents repeat to a large extent, this accumulation makes a noticeable difference in speed over time.

One point deserves attention: recognition is not a translation step. It takes no interest in what the text means; it reads only its shape. Its output is therefore always treated as raw, and it passes under a human eye at the next step.

The difference between a scanned PDF and a text PDF

PDF files look the same from the outside, but there are two different kinds. A text PDF is a file produced on a computer that stores its letters as characters. A scanned PDF is a photograph of a page; what it holds is an image, not text.

The easiest way to tell them apart is this: open the file and try to select a word. If the word can be selected and copied, it is a text PDF. If you cannot select it, what you have is an image and the recognition step will be needed.

The distinction feeds straight through to the price. A text PDF usually needs no extra preparation. A scanned PDF brings the recognition, correction and layout steps with it, and those steps may appear as a separate item in the quotation.

There is a third kind as well: files that carry both an image and a hidden text layer. In files of this sort the hidden layer is usually the result of a recognition run carried out earlier, and its accuracy is not guaranteed. Even when the copied text looks correct on screen, it may carry faulty characters in the background; rather than trust the layer that is already there, recognition is therefore often run again.

Where does recognition go wrong?

The most common error involves characters of similar shape. The digit zero and the capital letter O, the digit one and the lower-case l, six and b are frequently confused. That confusion has serious consequences in fields such as identity numbers, dates and amounts.

The second source of error is the page layout. On a multi-column page, reading the lines in the wrong order causes sentences to run into one another. When cell boundaries are lost in tables, data can fall into the wrong column.

The third source is image quality. A page scanned at an angle, a photograph with shadows on it, low resolution and the reverse side showing through all lower the recognition rate. In documents that are photocopies of photocopies, the error rate rises noticeably.

The fourth source is the additional elements on the document itself. When a seal, a stamp, a wet signature or a watermark sits on top of the text, the letters underneath become unreadable. These areas are usually transcribed by hand.

The limits of recognition in official documents

In an official document, recognition is a convenience, not a source of accuracy. The translator who signs the translation is obliged to compare the text against the original document. A translation delivered without that comparison carries the recognition error into the document.

Critical fields are always verified by hand: the spelling of the given name and surname, the date of birth, identity and document numbers, institution names, amounts. A single character wrong in these fields can invalidate the document and add weeks to an application.

The second limit is the completeness of the document. In an official translation every element on the page must be accounted for: the text of the stamp, the signature note, the page number, the annexes. Recognition software usually skips these elements; adding them is work for a person.

The third limit is liability. An error in an automated step does not remove the responsibility of the person who signs. With a serious supplier, therefore, the recognition output always passes before a second pair of eyes.

What makes a good scan?

What determines the quality of the result more than anything else is the file you send. A good scan raises the recognition rate, shortens the preparation time and therefore brings the cost down. The points below are the ones that make the most difference in practice.

Place the document on a flat surface and fit the edges fully within the frame. An oblique angle makes it harder for the recognition software to find the lines. Use a scanner rather than a phone camera if you can; if you cannot, take the picture in even light with no shadows.

For documents with seals and stamps on them, a colour scan gives a better result than black and white, because the difference in colour makes the writing easier to separate out. With multi-page documents, combine the pages into a single file in the correct order.

Finally, if you have the digital original of the document, do not bother with scanning at all. An electronically signed document obtained from the institution is always a better starting point than a scan of its printout.

How does the cost of preparation show up in the quotation?

Document preparation does not appear separately in most quotations, but it is always inside the price. The table below shows how different file types are classified according to the preparation load they carry.

File type	Preparation load	Typical outcome
Word or editable text	Next to none	Goes straight into translation
PDF with a text layer	Low	A short layout correction
Clean scan	Medium	Recognition and checking step
Photograph or faded photocopy	High	Manual transcription may be needed
Handwriting or text over a seal	Very high	Worked out entirely by hand

The conclusion from the table is clear: the better the quality of the file, the smaller the preparation item. That is why it is quite normal to be given two different prices for the same document from two different scans. The difference in price comes not from the translation work itself but from the time spent preparing for that work.

Whose job is the check after recognition?

When the automatic recognition is finished, two texts are on the table: the original image and the text the software has produced. Comparing the two is a job in its own right, and most delays occur at exactly this point. If translation begins without the comparison, the error travels on silently and is only noticed after delivery.

The responsibility lies with the supplier, but your contribution shortens the job. If you tell us the correct spelling of the proper names, institutional titles and numbers that appear in the document, the checker does not have to guess. In documents coming into Turkish from a foreign language in particular, the spelling of names is the field corrected most often.

The second leg of the check is the layout. Whether the table structure, the column order and the page sections have been set up correctly is inspected by eye. If this step is skipped, the document can become unreadable even when the translation is correct; the risk is high with multi-column official forms in particular.

The third leg is the record. If a note is kept of which sections were transcribed by hand, the same areas are handled by hand from the outset in the next document from the same institution, and the time is not lost twice.

The Türkiye context

In Turkish, part of the recognition error comes straight from the alphabet. The dotted and dotless i, and the characters ç, ğ, ö, ş and ü, can turn into one another or into their unaccented counterparts in low-resolution scans. In the spelling of a given name or a surname, an error of this kind leads to the document being refused in an official application.

The second particularity is the age of the document archive. Old civil registry extracts, title deeds typed on a typewriter and military records filled in by hand are still in circulation. The recognition rate falls with these documents and the process runs largely by hand. Where a document carries wording in the old script, a separate expertise is required.

The third is notarial practice. In a notarised translation the translator is expected to see the document in the original or in a certified copy. For that reason the recognition output is never treated as sufficient on its own; producing the original document is a rule of the process.

This article is for general information and is not legal advice.

What does this mean for you?

The practical value of knowing about this step is this: the quality of the file you send has a direct effect on the price you are quoted and on the delivery time. Five minutes of preparation can head off hours of transcription. The sequence below sets out the steps that help most.

1. Look for the digital original of the document; if there is one, send that rather than the scanned version.
2. Scan on a flat surface, fully within the frame and free of shadows; scan stamped documents in colour.
3. Combine a multi-page document into a single file in the correct order.
4. Set out the spelling of given names, surnames and numbers separately in your message; that is how critical fields are verified.
5. State which institution the document will be submitted to; the form of delivery is decided accordingly.

To see the price of your document according to its preparation load, you can send the file through the quotation form, and you can reach the guides written for particular document types through the knowledge centre.

Frequently asked questions

Is it enough if I photograph my document and send it?

In most cases it is enough, but a scan is safer. An oblique angle, shadows and low resolution increase recognition errors noticeably.

Do OCR errors carry through into the translation?

They can if the checking step is skipped. That is why the recognised text is compared with the original image before the translation is started.

Can OCR be used in a notarised translation?

It can, but only at the preparation stage. The translator whose signature is on the work is responsible for its accuracy and is obliged to compare the document with the original.

Can handwritten documents be recognised?

Only to a limited extent. Handwriting, writing over a seal and a faded stamp are usually transcribed by hand, and that takes extra time.

Sources

• Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation. Routledge Focus, Chapter 2.
• Notary Public Law No. 1512, provisions on translation and certification.
• Garcia, I. (2015). Computer-aided translation: systems. Routledge Encyclopedia of Translation Technology.