Use one pass only when it preserves every script
Use one combined pass only when it preserves both languages. Otherwise recognize script-specific regions separately and recombine them with page coordinates.
The method produces a documented choice between one combined pass and several script-specific passes, with page regions preserved so separately recognized text can be recombined safely.
Put the procedure into practice with Bilingual OCR. Continue with Right-to-Left OCR: Text Direction and Number Review and Vertical Text OCR for Chinese, Japanese and Mixed Layouts.
Mixed-language working plan
- Map every script region
- Test combined and separate models
- Retain coordinates across passes
- Document reading-order rules
- Assign reviewers by language
Separate script-specific regions when combined recognition fails
Create a page language map
Mark regions by script and language, including headings, footnotes, stamps, quotations, and handwritten additions. Note direction changes and whether translations appear side by side or line by line.
Compare combined and separate passes
A combined model may be convenient but can increase ambiguity between similar glyphs. Run a representative page both ways and score the difficult tokens in each script.
Recombine by coordinates
Keep page, region, and line identifiers when separate passes are used. Reassembly should follow the visible layout and documented reading order, not the order in which OCR jobs finished.
Use qualified language review
A fluent-looking secondary language can contain incorrect names or grammatical substitutions. Route each language to a reviewer who can compare the output with the source script.
Compare one combined pass with controlled regional passes
Mark each language, script, and writing direction on a representative page. Run a combined model first, then crop the regions where one script is weak and test the relevant model separately. Use the same source pixels so the comparison isolates model behavior.
Recombine output by page region rather than by sentence plausibility. Keep names, untranslated terms, numbers, and punctuation exactly as shown unless the task explicitly includes translation. Record which pass produced each region so later corrections remain traceable to the source and settings.
- Create a script map
- Compare combined and separate passes
- Recombine with coordinates and provenance
Create a page language map before choosing recognition passes
Mark where each script appears and what reading direction connects its regions. This exposes short passages that a dominant-language result might omit.
Main text
Record the dominant language, script, direction, and column structure.
Names and labels
Mark regions where spell-check or translation would be especially risky.
Numbers and punctuation
Check how direction changes interact with dates, prices, phone numbers, and Latin abbreviations.
Recombination
Retain page coordinates or region IDs so separately recognized output returns to the right place.
Choose one combined pass only when every mapped region survives. Otherwise, process regions separately and recombine by page evidence.
Pass strategy
| Layout | Suggested test | Main risk |
|---|---|---|
| Languages alternate by paragraph | Separate region passes | Paragraph order lost during merge |
| Translation in parallel columns | Column-specific passes | Lines paired incorrectly |
| One language with a few borrowed terms | Focused model plus token review | Borrowed terms normalized |
| Stamp or annotation in another script | Isolated crop | Small secondary region ignored |
One automatic pass may hide minority-script loss
A dominant language can look accurate while names, labels, or shorter passages in another script fail. Evaluate each script separately before accepting combined output.