# How OCR Errors Become Translation Errors—and What to Check First

Trace mistranslations back to scanned text, inspect names, numbers, symbols, and complex pages, and decide when to redo OCR instead of translating again.

A translated sentence can be grammatical and completely wrong. If OCR read “not permitted” as “now permitted,” translation may confidently carry the wrong instruction into another language. The translator never saw the word you saw on the scanned page.

When a result looks suspicious, compare three things: the page image, the recognized source text, and the translation. This separates recognition mistakes from translation decisions and keeps you from paying to repeat the wrong stage.

## Follow the error through the workflow

OCR recovers characters and document structure from page images. Translation then works with that recovered material. A mistake can enter at either stage, and layout errors matter as much as misspelled words.

Imagine a two-column page. Every word is recognized correctly, but sentences from the right column are inserted into the left column's paragraph. A translator receives an argument that no longer makes sense. Polishing the English will not restore the lost order.

Use a short diagnosis before making changes:

| What you find                                       | Likely place to investigate   |
| --------------------------------------------------- | ----------------------------- |
| A name differs from the visible scan                | Character recognition         |
| Correct words appear in the wrong sequence          | Layout and reading order      |
| Source text is correct but the meaning changes      | Translation                   |
| A correct translation loses a note marker on export | Output structure or rendering |

More than one problem can occur in the same passage. Fixing spelling does not prove that the paragraph order is correct.

## Give names, numbers, and symbols their own pass

Ordinary prose sometimes survives a missing letter. A measurement, identifier, or personal name may not. Look closely at `0/O`, `1/l/I`, minus signs, decimal separators, superscripts, and accented letters. Search both the expected spelling and plausible misreadings of a recurring name.

Do not judge a table only by its visible values. A correctly recognized number under the wrong heading is still wrong. Confirm the row label, column label, unit, and any qualifying note together. For formulas, compare subscripts, exponents, brackets, and operators against the original; a fluent surrounding paragraph is no evidence that the notation survived.

Keep a small issue sheet with page number, source fragment, observed error, and correction. For example: “PDF page 38, table 2: minus sign missing before 0.4.” This is more useful than a general note that the chapter needs work.

## Sample the difficult pages, not just the opening

Review at least one ordinary page and each layout type that appears in the document: columns, dense notes, a table, a diagram with labels, and a page with faint or damaged print. Include a page transition where a sentence continues onto the next image.

Read a complete paragraph across each transition. Verify that a running header has not become part of a sentence and that a footnote has not interrupted the main text. Then compare the translation of that same passage. If it introduces an actor, conclusion, or qualification absent from the source, investigate further.

Sampling is triage, not a guarantee for the rest of the book. If the same error repeats, expand the check to all pages with that feature. A repeated heading in the body is a reason to search the whole output, not merely delete the first example.

## Decide whether to redo OCR or correct extracted text

If many pages are blurred, skewed, cut off, or in the wrong order, start with a better source where possible. Redoing recognition on the same unreadable image may reproduce the same problem. Preserve the old result and give the new run a distinct filename so you can compare them.

If recognition is mostly sound and errors are isolated, an editable Markdown copy may be easier to correct in your own editor. Keep corrections tied to source pages. This is a manual workflow; do not assume PDF Craft supplies an OCR correction screen or automatically reuses your edits in an existing task.

[PDF Craft's PDF conversion workspace](https://pdfcraft.ai/en/pdf-craft/) can produce EPUB for reading or Markdown for editing. Conversion does not require translation; enable translation only when you want it. If you need to inspect recognition independently, first obtain an untranslated output. A later translation or rerun is a separate processing decision; consult [Pricing](https://pdfcraft.ai/en/pricing/) before repeating long tasks.

After correcting the source, compare the new translation with the earlier one at the known error locations. Then check the surrounding sentences: a corrected name or negation may affect pronouns, agreement, or an entire explanation. Save your issue sheet with the source and finished result so the next review starts with evidence rather than memory.