Select the model for the language printed on the page
Choose the model for the language actually printed on the page, confirm it is installed, and test names, numbers, punctuation, and borrowed words.
The decision should rest on installed model coverage and a representative test of the page’s names, punctuation, numerals, borrowed words, and smallest important text.
Apply the decision with Bilingual OCR. Then test the tradeoff against A Reviewable Workflow for Mixed-Language OCR and OCR Review for Diacritics, Names and Punctuation.
The multilingual OCR review collection adds mixed-language, right-to-left, vertical-text, and diacritic checks.
Language choice
| Document pattern | Starting approach | Review focus |
|---|---|---|
| One language, one script | Focused language model | Names and numbers |
| Two languages, same script | Combined or separate comparative test | Accents and dictionary substitutions |
| Two different scripts | Script-specific region passes | Boundary and reading order |
| Unknown language | Identify script and obtain a qualified reviewer | Do not infer meaning from OCR fluency |
Test the language model on the page’s hardest tokens
Inventory the languages and scripts actually printed on the page. Select a passage containing names, punctuation, numerals, borrowed words, and the smallest important text. Run the available candidate models on that same region without changing the image.
Compare exact corrections, not sentence fluency. A dictionary can improve common words while damaging names or technical terms. Confirm that the required language data is installed and record the model used. If one model improves a script but harms another, use region-specific passes and preserve coordinates for recombination.
- Inventory before selecting
- Test names and borrowed words
- Record model and region decisions
Test names, punctuation, numbers, and borrowed words
Inventory script and language separately
Latin script does not identify whether the page is English, French, Vietnamese, or another language. Record every script, language, and direction present, including short headings and annotations.
Confirm available trained data
An interface should offer only models installed on the recognition server. OCR.now follows that rule. A selected but unavailable language should fail clearly rather than silently producing output with another model.
Test discriminating features
Choose samples containing diacritics, language-specific punctuation, uncommon letter combinations, names, and abbreviations. Ordinary words may look correct even when the wrong model is damaging the difficult parts.
Avoid dictionary overcorrection
Codes, identifiers, historical spellings, and names may be changed toward familiar words. Preserve the image location and review these tokens independently from spell-check suggestions.
Language-model selection check
- List all scripts and languages
- Verify trained data is installed
- Test language-specific characters
- Review names and codes separately
- Record model versions with the output
Language selection cannot repair the source
The correct model may reduce substitutions but cannot restore blurred strokes, missing diacritics, clipped text, or unsupported historical forms. Mixed pages may still require region-specific passes.
What model documentation should confirm
Use engine documentation to verify installed languages and script coverage, then test names, punctuation, numerals, and borrowed words from the real document.