QA checklist for an Arabic OCR Word file before you deliver it

OCR has turned a scanned Arabic book, contract or report into an editable Word file, and someone else is waiting for it: a translator, a typesetter, an editor or a client's records team. The file opens, the text flows right to left and nothing looks obviously broken. That is exactly when errors slip through, because OCR mistakes in Arabic rarely look like mistakes. A missing dot turns one real word into another real word.

This checklist is for the person who signs off on the file. It assumes you already have a Word document and need to decide whether it is ready to leave your hands. If the text is reversed or its letters don't join, fix that first with our guide to reversed Arabic text after PDF to Word. If Find can't locate words you can see, read why Find can't locate Arabic words. Those are encoding problems. This article is about whether the content is right.

Pass 1: structure (ten minutes, every file)

Structural errors affect every page, so check them before you read any text closely.

1. Page count and coverage. Compare the first and last line of the Word file with the first and last page of the scan. Then pick three pages at random and confirm each one's opening words appear in the right place. A skipped page or a duplicated page shows up here.
2. Columns and reading order. On two-column pages, check that the right-hand column is read before the left-hand one and that lines from the two columns aren't interleaved. Commentary layouts, where a main text sits above or inside a commentary, need the same check.
3. Running heads and page numbers. Book scans carry a running title and a page number on every page. Decide with the recipient whether these should be removed, kept as page markers, or moved into a comment. Then search for one running title to see whether any copies are still sitting in the middle of paragraphs.
4. Footnotes. Check that footnote text is separated from the main text and that each marker in the body still matches its note. Footnotes merged into the running text are one of the most common structural problems in converted books.
5. Headings and styles. If the recipient will build a table of contents or typeset the file, chapter and section headings should use Word's heading styles rather than bold body text. The Navigation pane (View > Navigation Pane) shows at a glance which headings Word recognises.
6. Direction. Every Arabic paragraph should be set to right-to-left. Microsoft notes the direction buttons appear in Home > Paragraph only once a right-to-left language is enabled in Office.

Pass 2: a proofread sample against the scan

You will rarely have time to proofread a whole book before delivery, and it isn't always your job. What you can do is read enough of it, word by word against the image, to know what kind of errors the file contains and where they cluster.

Choose the sample on purpose rather than at random:

• the first page of the main text (often after a long introduction in a different font)
• the densest page: small type, heavy tashkeel or a page full of footnotes
• the worst scan: a dark, skewed, stained or curved page near the binding
• a page with tables, numbers or dates
• a page with proper names: people, places, book titles or the parties to a contract
• one ordinary page from the middle

Read each one against the scan, not from memory. Mark every difference, then sort them: a wrong letter, a dropped or added word, a missing line, a wrong number or a formatting problem. If a pattern repeats across pages, such as the same letter swapped or the last line of every page missing, tell the recipient or fix it throughout before you deliver. If the sample pages are clean, say that they were clean. Don't claim the whole file is.

Pass 3: hunt for the errors OCR makes in Arabic

Arabic separates many letters only by their dots, so a faint dot, a speck of dust or a worn typeface can change the letter while leaving a valid-looking word. W3C's Arabic layout requirements describe how these letters share base shapes, which is why the confusions below are predictable. Use Find on suspicious words, and pay most attention to names, terms and numbers, where a spell checker won't help.

Watch for	Typical confusion
Dot count and position	ب ت ث ن ي on the same tooth shape; ج ح خ; س ش; ص ض; ط ظ; ع غ; ف ق
One-dot pairs	د ذ; ر ز
Word endings	ة read as ه, and ى read as ي (or the reverse)
Hamza and alef	أ إ آ ا swapped or the hamza dropped
Digits	Arabic-Indic ٢ and ٣, ٠ (a small dot) lost or read as a full stop, ٥ read as ٠
Small marks	tashkeel dropped, moved to the wrong letter or turned into dots
Spacing	two words joined, or one word split around a non-joining letter such as و or ر

Then run Word's spell check in Arabic. Microsoft explains how to set the proofing language for selected text under Review > Language > Set Proofing Language. If Arabic paragraphs are tagged as English, Word either ignores them or marks every word. Spell check catches non-words, but it can't catch a dotted-letter error that happens to produce another real word. Your sample read is what catches those.

Illustrative example. Suppose the sample from a printed hadith collection shows the same pattern on three pages: the final ي is printed without dots in this edition, and the OCR has correctly kept it as ى, so في appears as فى. Nothing in the transcription is wrong, but a translator who searches for في with the dotted yeh will miss every occurrence. The right move is a note in the delivery message, not a silent bulk replacement that would also change real alif maqsura.

Pass 4: clean the file before it leaves

• Accept or reject tracked changes and resolve comments you added while checking, unless the recipient wants to see them.
• Run Document Inspector (File > Info > Check for Issues > Inspect Document). Microsoft lists comments, revisions, document properties and personal information, headers and footers, hidden text and invisible content among the things it checks. Run it on a copy: Microsoft warns that removed data can't always be restored. It also doesn't catch text hidden in other ways, such as white text on a white background.
• Fonts. Use a font the recipient has, or one that ships with Windows and macOS, so the file doesn't fall back to a font without full Arabic support.
• File name and format. Save as .docx with a name that identifies the source and the page range, such as source-title_pp001-120_ocr.docx.

Delivery checklist

• First and last lines match the first and last scanned pages
• No skipped, duplicated or out-of-order pages in a spot check
• Columns, main text and commentary are in reading order
• Running heads and page numbers handled the way the recipient asked
• Footnotes separated, and markers match their notes
• Headings use heading styles where the recipient needs them
• Every Arabic paragraph is right-to-left
• Six purposeful sample pages read word by word against the scan
• Repeated error patterns fixed throughout or reported
• Names, numbers and dates on the sample pages checked individually
• Arabic spell check run with the right proofing language
• Tracked changes and comments resolved; Document Inspector run on a copy
• A short delivery note listing what was checked and what wasn't

That last item matters most. A note such as "pages 1, 14, 37, 62, 88 and 119 proofread against the scan; footnotes separated; marginal notes not transcribed" tells the translator or editor exactly how much trust to place in the file and where to look first.

Where ScribeTools fits

ScribeTools produces editable Word files from scanned and photographed Arabic documents, with right-to-left paragraphs, and can also produce searchable PDF, tables as spreadsheets and translations. It does not replace this review: every OCR output, ours included, should be checked before anyone relies on it. If you are choosing a tool, see Arabic PDF to Word, and for long printed editions see digitizing Arabic books.

Limitations

• A sample proofread tells you about the pages you read. It is not proof the rest of the file is correct, so the delivery note should say which pages were checked.
• Dotted-letter errors that produce real words are invisible to spell check and to search. Only reading against the image finds them.
• Handwritten pages, manuscripts, marginal notes and heavily damaged scans need a full proofread, not a sample.
• Menu names in Word differ between Windows, Mac and the web version, and between Office releases. The steps above follow Microsoft's documentation for Word on Windows.
• This checklist covers transcription quality. It doesn't cover translation quality or legal review of a contract's content.