Multilingual OCR quality assurance requires more than checking whether every word was recognized. This reusable checklist helps developers and QA teams test language detection, scripts, reading order, transliteration, dates, numbers, currencies, and mixed-language pages before releasing an OCR workflow or changing an OCR API.
Overview
OCR internationalization testing asks a simple question: can the system preserve the meaning and structure of a document across languages and locales? A result may look readable while still containing serious errors. For example, a comma may be a decimal separator, a date may be interpreted in the wrong order, or a right-to-left line may be returned in reverse reading order.
Start by separating three layers of quality:
- Recognition: Are the characters, words, accents, diacritics, and symbols transcribed correctly?
- Layout: Are columns, tables, labels, line order, and reading direction preserved?
- Interpretation: Are dates, numbers, currencies, addresses, and other fields normalized without changing their intended meaning?
Do not treat a provider's language list as a complete test plan. Confirm which languages are supported together, whether the model identifies scripts automatically, whether language hints are required, and whether handwriting, low-resolution scans, tables, or embedded images behave differently. Record these details in a language-and-locale coverage matrix that can be rerun after a model, SDK, preprocessing, or document-template change.
Reusable coverage matrix
Create one row for each document scenario rather than one row for each language alone. Useful columns include:
- Document or fixture ID
- Primary language and secondary language
- Script, such as Latin, Cyrillic, Arabic, Devanagari, Han, or a mixed script
- Expected reading direction
- Language hint supplied to the OCR API
- Image type, resolution, rotation, and scan quality
- Expected date, number, currency, and identifier formats
- Ground-truth text and structured fields
- Character, field, and layout results
- Confidence score, review status, and error notes
Keep representative samples for each market and document type. A synthetic test can verify a particular separator or character, but real documents are needed to expose typography, spacing, compression, stamps, folds, and layout interactions.
Checklist by scenario
1. Single-language documents
- Confirm the expected language or language hint is passed consistently.
- Test uppercase, lowercase, accents, ligatures, diacritics, and punctuation.
- Include common fonts as well as condensed, decorative, or faint fonts used in production.
- Compare extracted text with ground truth at character and word level.
- Check whether line breaks and paragraph boundaries remain usable for downstream processing.
Test more than clean digital text. Include scans with skew, shadows, uneven lighting, background patterns, and moderate compression. If quality is poor, isolate preprocessing from recognition so the team can identify whether the failure came from the image or the OCR model. The OCR image preprocessing guide covers deskewing, denoising, cropping, and related preparation steps.
2. Mixed-language pages
- Test pages where languages share a script, such as two Latin-script languages.
- Test pages where scripts differ, such as Latin and Arabic or Latin and Han characters.
- Place language changes within the same line, table, field, and paragraph.
- Check whether the API returns one language label for the page or language information by block, line, or word.
- Verify that language hints improve results rather than forcing the wrong model onto part of the page.
Mixed-language pages commonly fail at boundaries: product names, addresses, names, and technical terms may be treated as misspellings or transliterated unexpectedly. Preserve the original recognized text before applying translation, transliteration, spelling correction, or normalization. These later transformations should be separate, reversible steps.
3. Right-to-left and bidirectional text
- Test Arabic, Hebrew, and other right-to-left content where relevant to your users.
- Include lines containing Latin brand names, email addresses, URLs, or numbers inside right-to-left text.
- Check visual order, logical text order, punctuation placement, and digit grouping.
- Verify that exported JSON, plain text, searchable PDFs, and user-interface displays handle direction consistently.
- Test forms with labels on either side of the value.
Do not judge right-to-left output only by opening a text file in one editor. Compare the API response, rendered interface, and downstream parser. A string can be technically correct in logical order but appear confusing when a display layer applies bidirectional formatting incorrectly.
4. Dates and times
- Include day-month-year, month-day-year, and year-month-day examples.
- Test separators such as slashes, dots, hyphens, and spaces.
- Include month names, abbreviated months, local-language months, and dates without leading zeroes.
- Test ambiguous values such as 03/04/2026 against a document locale or explicit field label.
- Check calendar assumptions when the workflow supports regions using calendar systems other than the default application calendar.
- Include times with 12-hour and 24-hour notation, local time zones, and seconds where relevant.
OCR should first capture the printed value faithfully. A parser may then normalize it to an internal format, but it should not silently guess when the source is ambiguous. Store the raw value, the normalized value, the assumed locale, and a validation status.
5. Numbers, decimals, and currencies
- Test comma and period decimal separators in both directions.
- Include thousands separators, spaces, apostrophes, and numbers without grouping.
- Test negative values, percentages, fractions, ranges, and leading or trailing symbols.
- Include currency symbols, currency codes, and currency names in multiple positions.
- Check whether a symbol is confused with a letter or digit in the selected script.
- Compare extracted totals, tax amounts, quantities, and unit prices independently.
Never infer a decimal convention from one field when the document contains a locale, currency, or explicit formatting clue. For invoices and receipts, validate arithmetic relationships where appropriate: line items, subtotal, tax, discount, and total should be compared with tolerances defined by the product team. A mathematically plausible value can still be an OCR error.
6. IDs, passports, and structured fields
- Test letters and digits that are easily confused, such as O and 0 or I, l, and 1.
- Check fixed-length identifiers, check digits, separators, and permitted character sets.
- Test machine-readable zones separately from the visual document area.
- Verify that names, dates, nationality codes, and document numbers are mapped to the correct fields.
- Route uncertain or conflicting fields to review rather than silently correcting them.
For sensitive documents, limit fixtures to approved test data and follow the workflow's retention and access controls. See the OCR API security review checklist and PII redaction workflow when designing test handling.
What to double-check
Script detection and language behavior
Record whether script detection is automatic, configured per request, or returned as metadata. Test what happens when the detected language is absent, incorrect, or ambiguous. Also check whether a multilingual request supports simultaneous languages or merely selects one model. Provider documentation and supported-language behavior can change, so treat both as versioned test inputs.
Reading order and layout
Compare single-column pages, multi-column articles, sidebars, tables, captions, footnotes, stamps, and rotated labels. A text-only accuracy score may hide a reading-order defect that breaks search, indexing, or structured extraction. Keep layout assertions separate from character assertions.
Confidence and human review
Define thresholds by field risk, not only by a global score. A slightly uncertain street name may need different handling from an uncertain passport number or payment total. Test low-confidence, missing, conflicting, and out-of-vocabulary cases. The OCR confidence scores guide can help structure review queues.
Regression measurement
Use a fixed golden set and report results by language, script, document type, and field. Useful measures include character error rate for text, word error rate for readable content, exact-match rate for identifiers, field accuracy for structured extraction, and reading-order pass rate. Do not hide a regression in an overall average: a small language segment can be important even when it represents few documents.
For a repeatable benchmark process, maintain the test fixtures, expected outputs, scoring rules, and release comparison in one place. The OCR performance benchmarking playbook provides a practical structure for this work.
Common mistakes
- Testing only clean samples: Production scans include rotation, blur, compression, shadows, and unusual layouts.
- Using one locale for every parser: Language, currency, document origin, and user settings may not agree.
- Normalizing too early: Replacing separators or transliterating text before preserving the raw OCR result makes debugging difficult.
- Scoring only text: Correct characters in the wrong column or field can still break automation.
- Assuming language support is additive: Two individually supported languages may behave differently when mixed on one page.
- Accepting automatic correction without evidence: Corrections can make an output look cleaner while changing names, identifiers, or amounts.
- Ignoring downstream systems: Character encoding, database collation, JSON serialization, fonts, and search indexes can introduce failures after OCR.
- Changing the fixture set after a failure: Preserve the failing sample and add it to the regression suite.
When to revisit
Run the full international OCR QA checklist before launching a new market, adding a language or script, introducing a new document type, changing an OCR API or model, modifying preprocessing, or altering date and number parsers. Run a focused regression whenever a template, SDK, output schema, or downstream parser changes.
Review the coverage matrix at least twice a year and before major seasonal planning cycles. Refresh the provider's supported-language documentation, script limitations, model behavior, file-format assumptions, and locale examples rather than relying on an old implementation note. Recheck samples after changes to cameras, scanners, compression settings, or document suppliers.
To make the next review efficient, finish each test cycle with four actions:
- Save raw inputs, raw OCR output, normalized output, and version information.
- Classify failures as recognition, layout, parsing, encoding, or workflow errors.
- Add representative failures to the permanent golden set, subject to privacy controls.
- Update thresholds, review rules, and release notes before production rollout.
A multilingual OCR workflow is ready for release when its supported scenarios are explicit, ambiguous values are handled deliberately, high-risk fields have a review path, and regressions can be detected by language, script, locale, and document type—not just by one overall accuracy number.