If you search for translation length ratio QA, source target length check, or suspiciously short translation, the useful idea is not that translations should be the same length. Different languages expand, contract, segment and encode grammar differently. The useful check is whether a target is unusually short or long relative to what is normal for this language pair, content type and project.
A length-ratio QA rule can surface omissions, duplicated text, accidental copy-paste, unresolved placeholders, untranslated fragments or runaway machine output. It is a triage signal, not a verdict. A long German compound translated into several English words can be perfectly normal; a two-word target for a dense paragraph may deserve immediate inspection.
This guide explains how people translate quickly by using source–target length ratios as a review filter. It covers word, character and byte ratios, language expansion, scripts, UI strings, subtitles, CJK text, punctuation, repeated content, MT, threshold calibration, false positives, severity and batch review.
The owner job is distinct from character-limit QA. Character limits ask whether target text fits a UI container. Length-ratio QA asks whether the target’s size is suspicious compared with its source and expected language-pair behavior.
1. Length Difference Is Normal
Translation changes grammatical packaging, word count and character count even when meaning is preserved exactly. English, German, Japanese and Arabic do not encode the same information with the same number of tokens.
A useful QA rule therefore looks for outliers rather than exact equality. Treat length ratio as anomaly detection, not as a demand for visual symmetry. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
2. Choose the Measurement Unit Deliberately
Words, characters, bytes and tokens answer different questions. Character counts are often more comparable across CAT segments than word counts for scripts with different segmentation rules.
Bytes matter for technical storage constraints but are poor proxies for linguistic completeness. Use the unit that matches the QA purpose and language pair. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
3. Word Ratios Can Mislead Across Scripts
Word segmentation varies greatly across languages. Chinese, Japanese and Thai may be counted differently by different tools, while agglutinative languages can pack more meaning into fewer space-delimited words.
A word-based threshold copied from English–French work can create constant false alarms elsewhere. Calibrate with real project data. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
4. Character Ratios Are Useful but Not Universal
Character counts avoid some tokenization issues and can expose extreme omissions or duplication. They still behave differently across scripts because one character can encode very different amounts of information.
Full-width forms, combining marks and Unicode normalization can also affect counts. Use character ratio as a practical signal rather than a linguistic constant. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
5. Very Short Targets Can Reveal Omissions
A target that is dramatically shorter than comparable segments may be missing a clause, list item, qualifier or copied phrase. This is especially useful after heavy editing, MT post-editing or fuzzy-match reuse.
Inspect the source structure before concluding that concision is the problem. Some target languages legitimately compress information that the source expresses more explicitly. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
6. Very Long Targets Can Reveal Duplication
Accidental paste, duplicated clauses, repeated MT output or unresolved notes can make a target unexpectedly long. Length-ratio QA can surface these cases even when the repeated wording is fluent.
Search for duplicated spans, repeated sentences and imported comments. Do not assume that natural language expansion alone explains an extreme outlier. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
7. Short Source Strings Are Noisy
A one-word source translated into a two-word target produces a 2:1 word ratio that looks extreme but is completely ordinary. Ratios become less stable as segment length approaches zero.
Use minimum source-length thresholds before applying strict ratio warnings. Very short labels are better reviewed with context, terminology and character-limit rules. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
8. Long Segments Give Stronger Signals
In longer segments, extreme length differences are less likely to come from one article, auxiliary or inflection. A 70-word source and a 12-word target deserves different scrutiny from a 2-word source and 1-word target.
Thresholds can tighten as source length increases. Segment size should influence confidence in the warning. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
9. Language Pair Baselines Matter
Each language pair and direction develops a normal expansion or contraction range. Use completed, reviewed translations from the same domain to estimate typical ratios.
A baseline drawn from marketing text may not fit software UI or legal prose. Calibrate by pair, direction and content type when possible. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
10. Genre Changes Expected Length
Subtitles, technical tables, contracts, help articles and marketing headlines impose different target styles. A subtitle translation may deliberately compress spoken content; a legal translation may preserve more explicit structure.
Length-ratio thresholds should reflect genre expectations. One project-wide rule can be too crude for mixed content. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
11. UI Text Has Two Different Length Problems
UI strings can be suspiciously long relative to source and also exceed an absolute character limit. These are different checks.
Ratio QA identifies unusual translation behavior; length-constraint QA protects the visual container. Use both where the product needs them. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
12. Subtitles Often Compress by Design
Audiovisual translation may shorten speech to meet reading speed, timing and screen limits. A low target/source ratio can therefore be intentional.
Review against subtitle rules, timing and preserved meaning rather than generic prose thresholds. Do not punish professional condensation as omission. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
13. Machine Translation Can Produce Runaway Output
LLM or MT systems can occasionally repeat, explain or expand beyond the source. Length outliers are a cheap way to surface suspicious generations.
The warning should send the segment to human review rather than automatically rejecting it. Some source sentences genuinely need longer target phrasing. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
14. Post-Editing Can Over-Compress
A reviewer may shorten machine output aggressively and accidentally drop nuance, conditions or attribution. Length-ratio QA can identify candidates for source-target comparison after heavy editing.
The check is particularly valuable when editing happened under time pressure. Use it as a second look, not a stylistic preference for longer text. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
15. Fuzzy Matches Can Create Hidden Ratio Changes
A translator may confirm an old target after the new source grows substantially. The target then looks unusually short because the fuzzy match was not fully repaired.
Length ratio can complement Unedited Fuzzy Match QA. Compare changed source material before rewriting the whole target. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
16. Repeated Content Can Distort Statistics
Boilerplate or repeated labels can dominate project averages and make a baseline look more stable than the truly novel text. Calculate representative baselines from reviewed material across several content types.
Do not let thousands of identical short strings define the expected ratio for long prose. Sampling matters. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
17. Punctuation and Formatting Can Inflate Counts
Bullets, numbering, inline codes and markup can alter character or token totals without changing linguistic content. Decide whether the measurement strips tags, placeholders and protected tokens.
Compare like with like. A ratio rule should count the material whose presence or absence it is actually trying to evaluate. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
18. Numbers and Codes Need Separate Handling
A dense technical string can contain many numbers, IDs or symbols that remain identical across languages. These tokens can make length ratios appear closer than the linguistic text really is.
Consider excluding protected tokens for some analyses. Keep number and code integrity in dedicated QA checks. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
19. Headings and Titles Can Legitimately Recast
Target-language headlines may be shorter, more explicit or differently structured for natural style. Do not use narrow prose thresholds on titles.
A heading ratio warning should be loose and contextual. The purpose is to catch obvious omissions or duplication, not to force headline syntax to mirror the source. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
20. Names Can Create Stable Length Islands
Segments dominated by names, product codes or URLs may have very similar lengths even across unrelated languages. Length ratio adds little value there.
Use entity and non-translatable checks instead. QA profiles should disable low-value warnings where other rules are stronger. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
21. Use Median and Percentile Thinking
Project averages can be distorted by extreme segments. A median and percentile range gives a more robust sense of normal behavior.
You do not need advanced statistics to benefit: reviewed examples can define a practical low and high band. The goal is a useful outlier queue, not mathematical elegance. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
22. Thresholds Should Be Tunable
A hard-coded global ratio such as 0.5–2.0 may be convenient but noisy. Allow different thresholds for language pair, segment size and content type.
Start broad, inspect false positives, then refine. A rule that reviewers trust is more valuable than a theoretically strict rule they ignore. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
23. Warnings Should Carry Context
Show source length, target length, ratio, file, segment ID and surrounding text. A bare ‘length mismatch’ forces the reviewer to reconstruct the problem.
Good QA output shortens the time from warning to decision. Context is part of speed. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
24. Batch Review Outliers
Length-ratio warnings are well suited to filtered batch review. Inspect the most extreme outliers first, then work toward the threshold boundary.
This prioritizes the candidates most likely to represent real omissions or duplication. Rerun after fixes to confirm the queue shrank for the right reasons. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
25. Do Not Auto-Fix Length
Software can detect an outlier but should not automatically add or delete language to force a ratio. Meaning, grammar and style determine the correct target length.
Automatic shortening or padding can create worse translation than the original issue. Use detection to route human attention. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
26. Compare Against Approved History
If the same segment type has been translated before, previous approved ratios can provide a local baseline. This is especially useful for recurring product releases or regulated templates.
Do not inherit old thresholds blindly when source style or target guidelines change. Version the QA profile with the project. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
27. Use Ratio QA After Major Edits
Large find-and-replace, MT retranslation, source update or style harmonization can create unintended length changes. Run the check after risky transformations.
It acts as a cheap regression detector. Pair it with backup/recovery and revision-history tools so suspicious changes can be inspected. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
28. Review False Positives as Training Data
Legitimate outliers reveal where the rule does not understand the language or content type. Collect recurring examples and refine thresholds or exceptions.
Do not suppress the entire check because one genre behaves differently. Tune narrowly so real anomalies remain visible. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
29. Length Ratio Complements Empty-Target QA
Empty targets sit at the extreme end of shortness but deserve their own deterministic rule. Length-ratio QA is for non-empty segments that still look suspicious.
Keeping both owners separate produces clearer remediation. A zero-length target should not wait inside a statistical outlier report. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
30. Length Ratio Complements UI Limits
A target can be perfectly proportional to source and still exceed a button limit, or be wildly disproportionate and still fit. These checks answer different questions.
Use ratio QA for completeness and anomaly triage; use absolute length constraints for layout. Do not merge semantics and design into one threshold. The warning is useful only when it reduces the search space for the reviewer instead of dictating how long a good translation must be.
A practical check is to compare the outlier with nearby reviewed segments, inspect whether protected tokens were counted, and verify whether the source–target meaning is complete. If the ratio is legitimate, document the pattern or widen the profile narrowly rather than disabling the rule globally.
Practical Length-Ratio QA Workflow
Choose the counting unit that fits the language pair and file type.
Exclude or consistently count tags, placeholders and protected tokens.
Set broad thresholds from reviewed project data rather than generic internet rules.
Use looser handling for very short segments and tighter confidence for longer text.
Sort warnings by extremity so the most suspicious outliers are reviewed first.
Check omission, duplication, runaway MT and unrepaired fuzzy matches before editing style.
Record recurring legitimate outliers and tune the profile narrowly.
Applied Calibration Lab
Take 100 reviewed segments from one language pair and content type. Record source and target character counts, then sort the ratios from shortest to longest. The point is not to discover a universal linguistic law; it is to see what ordinary approved work looks like inside this exact project.
Inspect the ten shortest and ten longest targets. Mark why each is legitimate or suspicious: language-pair expansion, subtitle compression, list formatting, repeated content, omitted clause, duplicated sentence, long proper name, or another visible cause. These notes become the first calibration data for the rule.
Now choose a broad warning band that catches the suspicious examples without flooding reviewers with ordinary ones. Re-run the same rule on a different reviewed sample. If the second sample behaves differently, investigate whether genre, segment length or token handling changed before tightening the threshold.
Next test the profile on a deliberately damaged copy: delete one clause, duplicate one sentence, leave one fuzzy match unrepaired and add one runaway MT explanation. The QA rule should surface several of those defects. If it misses them, adjust the rule or combine it with more specific checks rather than expecting one ratio threshold to detect every failure.
Finally document the scope: language pair, direction, content type, unit, excluded tokens, minimum source length and warning band. A calibrated QA profile is reusable only when another reviewer can understand what the numbers mean and when they should be ignored.
Revisit the calibration after a major style-guide change, new MT provider, different content type, or substantial source-authoring shift. Ratios are properties of a workflow and corpus, not permanent facts about a language pair. Periodic recalibration keeps the warning useful instead of allowing old thresholds to become noise.
Frequently Asked Questions
What is a good source–target length ratio?
There is no universal ratio. It varies by language pair, direction, script, genre and segment length. Use reviewed project data to establish a practical range.
Should I use words or characters?
Either can work, but character ratios are often easier across languages with different tokenization. The important rule is consistency and pair-specific calibration.
Is a very short target always an omission?
No. Subtitles, headings and some languages legitimately compress information. The check flags a candidate for review; it does not decide correctness.
How is this different from UI character limits?
UI limits compare target length with an absolute container requirement. Length-ratio QA compares target length with source length to surface suspicious translation behavior.
Can MT be checked this way automatically?
Yes. Ratio outliers are a useful deterministic signal for runaway or over-compressed output, but a human should still decide whether the target is actually wrong.
Current QA Practice
Current CAT/TMS QA systems commonly include checks for empty targets, identical source/target, numbers, punctuation, whitespace and length-related issues. The broader QA principle is to automate deterministic anomalies while leaving linguistic judgment to humans.
Recent localization QA guidance also stresses that mechanical checks should produce candidate findings rather than rewrite translations automatically. Length ratio fits that model: deterministic detection, contextual human decision.
Conclusion
Length-Ratio QA is useful because omissions and duplication often change the size of a segment before they become obvious in a linear reread.
The rule works best as a calibrated outlier detector, not as a belief that good translations have matching lengths.
Measure consistently, calibrate by language pair and content type, review the extremes first, and let meaning—not ratio—decide the final target.
Continue the Translation Series
Read How People Translate Quickly | Empty Target QA.
Read How People Translate Quickly | Length Constraints Pass.
Read How People Translate Quickly | Unedited Fuzzy Match QA.