People searching post-edit distance, translation edit distance, post-edit rate, TER translation, HTER, review edit distance, machine translation post-editing metrics, or how to measure how much a reviewer changed a translation are trying to turn revision effort into evidence. They want to know whether a translation suggestion, MT engine, translator, workflow, or language pair is producing text that reviewers can accept with little intervention or text that has to be substantially rewritten.
Current localization analytics use practical terms such as post-edit rate, average edit distance, light edits, medium edits, heavy edits, new text, and retranslation. Translation Edit Rate, or TER, is a related metric that counts edit operations such as insertions, deletions, substitutions, and shifts relative to a reference. The search language sounds numerical, but the dominant reader job is not “produce one score.” It is: measure how much text changed, separate workload from quality, compare like with like, and use the result to tune MT routes, review practices, and translator support without turning every edit into an accusation.
This article has one dominant job: use post-edit distance and related review-effort metrics to understand how much human intervention translated text requires. It does not replace LQA, which classifies what kind of error occurred and how serious it was. It does not replace quality estimation, which predicts risk before human review. Post-edit distance analytics measure transformation effort after review: how much changed, how often, and where. The central challenge is interpreting those numbers correctly.
Quick answer
A reliable post-edit analytics workflow is:
- define exactly what text state counts as the “before” version;
- define exactly what reviewed state counts as the “after” version;
- calculate post-edit rate and edit distance consistently;
- separate no-edit, light-edit, medium-edit, heavy-edit, new-text, and retranslation cases;
- normalize by source words or reviewed segments when comparing workload;
- compare only comparable language pairs, content types, and review stages;
- pair edit-distance data with LQA categories and severity;
- investigate reviewer-style differences before blaming translation quality;
- use the data to compare MT routes, TM reuse, translators, and content classes cautiously;
- never treat low edit distance as proof of correctness or high edit distance as proof of poor translation.
The central rule is:
edit distance measures change, not truth.
What edit distance is measuring
At a simple level, edit distance asks:
How many operations are needed to transform Version A into Version B?
Common operations include:
- insertion;
- deletion;
- substitution.
Some metrics such as TER also include shifts of word sequences.
If a reviewer changes:
The user can delete the account.
to:
The user may delete the account.
one small substitution may be enough.
If the reviewer rewrites the entire sentence, distance rises.
Why this is useful
Post-edit effort can reveal:
- strong versus weak MT output;
- review-heavy language pairs;
- content that needs better terminology;
- translators whose work requires frequent revision;
- reviewers who rewrite unusually aggressively;
- source types that create repeated editing.
The metric is valuable because review effort costs time.
But the interpretation must be careful.
Post-edit rate versus edit distance
These are related but different.
Post-edit rate
The percentage of reviewed translations that required any edit.
If 100 segments are reviewed and 30 are changed:
post-edit rate = 30%.
Average edit distance
How much the edited translations changed relative to their length.
A project can therefore have:
- high post-edit rate but tiny edits;
- low post-edit rate but a few complete rewrites.
You need both views.
Example 1: many tiny corrections
100 segments reviewed.
80 edited.
Most edits are:
- punctuation;
- capitalization;
- one article;
- one terminology preference.
Post-edit rate: high.
Average edit distance: low.
Interpretation: quality may be close, but the workflow has many small corrections.
Possible root causes:
- style-guide mismatch;
- reviewer preference;
- punctuation settings;
- terminology configuration.
Example 2: few but heavy rewrites
100 segments reviewed.
20 edited.
10 of those are almost fully rewritten.
Post-edit rate: 20%.
Average edit distance among edited segments: high.
Interpretation: most output is accepted, but a subset fails badly.
Possible root causes:
- ambiguous short strings;
- one weak MT domain;
- high-risk content;
- source problems.
Step 1: define the baseline version
What does “before review” mean?
Possible baselines:
- raw MT output;
- translator-confirmed target;
- TM-pretranslated target;
- post-edited target before reviewer;
- imported vendor translation.
These are not interchangeable.
A reviewer editing raw MT is measuring MT-to-final effort.
A reviewer editing human post-edited MT is measuring post-editor-to-final effort.
Define the baseline.
Step 2: define the final version
Possible final states:
- first reviewer save;
- approved segment;
- client-reviewed segment;
- released text.
If one project measures translator → reviewer and another measures MT → final client approval, the numbers cannot be compared directly.
Step 3: choose the unit of analysis
You can measure by:
- character;
- token;
- word;
- segment.
Each produces different values.
Character distance is sensitive to spelling and morphology.
Word distance can better reflect lexical edits but depends on tokenization.
TER adds phrase shifts.
Do not compare scores from different implementations as if they are identical.
Step 4: understand Levenshtein-style distance
A simple word-level edit distance may count:
Insertion
Reviewer adds a word.
Deletion
Reviewer removes a word.
Substitution
Reviewer replaces one word with another.
If the original has ten words and two substitutions are needed, a normalized distance might be around 20%, depending on formula.
The exact implementation matters.
Step 5: understand TER
Translation Edit Rate is commonly defined around the number of edits needed to turn a candidate translation into a reference, normalized by reference length.
TER can include:
- insertions;
- deletions;
- substitutions;
- shifts.
The shift operation matters because moving a phrase can be one structural change rather than many delete-and-insert operations.
HTER in practice
Human-targeted TER, or HTER, uses a human post-edited version as the reference.
Conceptually:
raw system output → human post-edit.
It is often used to estimate human effort more directly than comparing against an unrelated reference translation.
Again, the exact implementation can vary.
Step 6: do not compare raw distance across segment lengths
Two edits in a five-word segment are huge.
Two edits in a fifty-word segment are small.
Normalize by length.
This is why percent edit distance is more interpretable than raw operation count for cross-segment comparison.
Step 7: build effort bands
A practical dashboard can group reviewed translations into:
- no edits;
- light edits;
- medium edits;
- heavy edits;
- new text;
- retranslation.
The threshold values should be documented.
Some current dashboards use bands around 15% and 30% for light/medium/heavy separation.
Those thresholds are useful operational conventions, not universal laws of translation quality.
Why effort bands help
Managers can scan:
65% no edit, 22% light, 8% medium, 5% heavy.
This communicates workload better than:
average edit distance = 11.4%.
Averages can hide distribution.
Step 8: separate new text from edits
Sometimes a reviewer creates a translation where no prior target existed.
That is not “post-editing.”
It is new translation.
Likewise, a complete retranslation should be separated from ordinary editing if the analytics platform supports it.
Otherwise the edit-distance average can become misleading.
Step 9: separate reviewer creation from reviewer revision
A reviewer may find an empty segment and translate it.
Do not count that as evidence that the original translator produced poor target text.
There was no target text.
Metrics need honest categories.
Step 10: compare language pairs separately
German, Japanese, Finnish, Arabic, and Spanish have different morphology, word order, and tokenization behavior.
The same underlying semantic change can produce different edit distances.
Do not rank language teams by one global raw threshold without calibration.
Step 11: compare content types separately
A marketing slogan can be heavily rewritten for voice even when accurate.
A technical instruction should often require much less stylistic rewriting.
If you combine both:
- marketing inflates edit distance;
- technical performance becomes hard to see.
Segment by content class.
Step 12: compare review goals separately
Review type matters.
Accuracy review
Minimal changes unless meaning/terminology is wrong.
Copyediting
May rewrite style extensively.
Brand review
May transform tone.
Legal review
May change wording for risk.
A heavy copyedit does not prove the translator failed.
It may be the planned workflow.
Step 13: reviewer behavior can dominate the metric
Reviewer A makes only necessary corrections.
Reviewer B rewrites every sentence into personal style.
Same translator.
Same source.
Different edit-distance result.
Therefore reviewer identity or reviewer-pairing can be an important analytics dimension.
Step 14: calibrate reviewers
Before using edit distance for performance decisions:
- give reviewers the same sample;
- define what changes are necessary;
- define style expectations;
- compare rewrite tendency.
Reviewer calibration improves metric comparability.
Step 15: pair edit distance with LQA
Suppose average edit distance rises from 10% to 20%.
Why?
LQA can tell you:
- terminology errors increased;
- accuracy stable;
- style edits increased.
Without LQA, the number says only:
more text changed.
With LQA, you know what changed and why.
Example 3: high edit distance, low error severity
Reviewer rewrites:
Start the device.
to:
Turn on the device.
Both are acceptable under project style.
Edit distance exists.
LQA defect may be none.
Do not convert the edit into an error automatically.
Example 4: low edit distance, critical error
Target:
Do not exceed 50 mg.
Reviewer changes:
Do not exceed 5 mg.
One small deletion.
Very low edit distance.
Potentially critical error.
This is the strongest reason not to treat edit distance as quality severity.
Step 16: use edit distance to compare MT engines cautiously
Same source sample.
Engine A average post-edit distance: 12%.
Engine B: 24%.
If review conditions are controlled, A probably requires less editing.
But also compare:
- critical errors;
- terminology;
- hallucinations;
- numbers;
- locale correctness.
Effort is one dimension.
Step 17: use edit distance to tune MT routing
If EN-DE technical output from Profile A consistently needs less editing than Profile B while quality is comparable, route technical German to A.
This is a strong use of the metric.
The decision is operational, not moral.
Step 18: use edit distance to evaluate TM reuse
A fuzzy TM match may be 95% similar to source but require 30% target editing.
Another 85% match may need only 10%.
Match percentage and edit effort are different.
Analyzing both can improve TM threshold decisions.
Step 19: use edit distance to detect stale TM
If 100% or context matches are frequently edited heavily, investigate:
- stale terminology;
- wrong domain;
- context mismatch;
- poor TM governance.
High edit distance on high-confidence TM matches is a valuable warning.
Step 20: use edit distance to evaluate source changes
A source revision can cause large target edits.
Track whether:
- tiny source changes create heavy target rewrites;
- source style is unstable;
- segmentation changed.
The metric can reveal source-authoring cost.
Step 21: build a review-effort dashboard
A useful dashboard can show:
- translations reviewed;
- source words reviewed;
- post-edit rate;
- average edit distance;
- no-edit count;
- light-edit count;
- medium-edit count;
- heavy-edit count;
- new-text count;
- retranslation count;
- review turnaround time.
These metrics answer different questions.
Do not compress them too early into one score.
Step 22: watch the distribution, not only the average
Project A:
- 95% no edit;
- 5% complete rewrite.
Project B:
- every segment edited by 5%.
Both can have similar average edit distance.
The operational experience is very different.
Distribution matters.
Step 23: measure by translator-reviewer pairing
A high edit rate can come from a mismatch between:
- translator style;
- reviewer expectations.
If Translator X performs well with Reviewer A but poorly with Reviewer B, investigate calibration before declaring quality failure.
Pairing data can reveal collaboration problems.
Step 24: measure by source type
One translator may handle:
- simple help center;
- dense patents.
Raw edit distance across both is unfair.
Group by:
- content type;
- domain;
- source difficulty;
- project.
Compare like with like.
Step 25: measure by leverage source
Classify target baseline:
- new human translation;
- 100% TM;
- fuzzy TM;
- MT;
- LLM;
- auto-propagated repetition.
Then ask:
Which source of leverage creates the least downstream editing?
This can inform resource priorities.
Step 26: measure by match band
For TM output, compare:
- context/101%;
- 100%;
- 95–99%;
- 85–94%;
- 75–84%.
If 85–94% matches require nearly complete rewrites, lowering the insertion threshold may help.
If 95–99% matches are usually light edits, keep them prominent.
Step 27: measure pre-translation policy
A project may prefill:
- exact TM only;
- high fuzzy;
- MT for no-match.
Edit analytics can reveal which prefilled categories save real work.
A filled target is useful only if the human spends less time reaching final quality.
Step 28: watch “no edits” carefully
No edit can mean:
- excellent output;
- reviewer acceptance;
- reviewer did not examine carefully;
- trivial segment;
- protected term.
Do not treat no-edit rate as pure quality.
Pair with QA and sampling.
Step 29: watch “heavy edits” carefully
Heavy edits can mean:
- poor translation;
- creative transcreation;
- source rewrite;
- reviewer preference;
- outdated terminology.
Investigate category distribution.
Step 30: define a retranslation threshold
Some systems classify a target as retranslation when edit distance becomes very high or when a reviewer replaces the segment completely.
Define the rule.
Do not let “retranslation” mean different things across teams.
Step 31: use reviewer comments to explain heavy edits
For heavy-edit segments, sample comments.
Common reasons:
- wrong meaning;
- tone;
- terminology;
- grammar;
- source ambiguity.
This turns the metric into actionable diagnosis.
Step 32: create a light/medium/heavy calibration set
Take ten reviewed segments.
Ask the team:
- light?
- medium?
- heavy?
Compare with dashboard thresholds.
The numerical bands should roughly match human intuition about editing effort.
Step 33: beware morphology-heavy languages
A small source change can trigger several target inflection changes.
Edit distance may look high even though the translation task was conceptually small.
Compare within the same language pair.
Do not compare raw percentages across typologically different languages without calibration.
Step 34: beware word-order differences
A reviewer moving a phrase can generate many insert/delete operations under simple edit distance.
TER’s shift operation was designed partly to handle this more sensibly.
Know which metric your platform uses.
Step 35: beware punctuation and whitespace
If the metric counts punctuation or spaces heavily, a formatting cleanup can inflate editing.
Some analytics normalize these differences.
Others do not.
Document the calculation method.
Step 36: beware tokenization
What is a word?
Examples:
don't;state-of-the-art;- Chinese text without spaces;
- Japanese scripts.
Tokenization affects word-level distance.
Cross-language comparisons require caution.
Step 37: use character distance for some scripts cautiously
Character-level comparison can be useful where word tokenization is difficult.
It can also overstate small orthographic changes.
Again:
metric choice should match the question.
Step 38: distinguish human post-edit distance from automatic MT metrics
BLEU, COMET, chrF, TER, and human edit distance are related evaluation tools but answer different questions.
Human post-edit distance asks:
How much did the human finalizer change this output?
An automatic metric asks:
How similar or semantically aligned is this output to references or learned quality criteria?
Do not mix them casually.
Step 39: HTER needs a genuine post-edited reference
If the “reference” is an independent translation written from scratch, the distance can reflect stylistic variation rather than editing effort.
HTER is most interpretable when the human reference represents the actual edits needed to make the system output acceptable.
Step 40: use a controlled post-editing brief
Tell reviewers what level of editing is expected.
For example:
Full post-edit
Correct meaning, grammar, terminology, and style to publication quality.
Light post-edit
Correct serious errors while allowing less-polished style.
The same MT output can produce different edit distance under different briefs.
Step 41: do not compare light and full post-edit projects directly
A full-post-edit workflow should create more changes.
That is planned.
The metric must be interpreted against the quality target.
Step 42: measure time as well as edits
Two segments can have the same edit distance but different cognitive effort.
Example:
Segment A: replace two obvious words.
Segment B: research one technical term for ten minutes, then change two words.
Edit distance is identical.
Time is not.
Combine edit metrics with:
- review duration;
- words per hour;
- query count.
Step 43: editing less can take longer
A reviewer may spend several minutes confirming that a difficult translation is correct and make no edits.
No-edit does not mean zero effort.
This is another reason to avoid using edit distance as a complete productivity measure.
Step 44: edit distance and translator compensation
Be cautious about using raw edit metrics for pay or ranking.
Factors include:
- content difficulty;
- reviewer style;
- source quality;
- language pair;
- project instructions.
If used commercially, normalize carefully and allow dispute review.
Step 45: edit distance and MT discounts
Some organizations connect MT or TM leverage to pricing.
Edit analytics can inform whether assumptions are realistic.
If 95% fuzzy matches consistently require 30% editing, a pricing model treating them as almost finished may be unfair and operationally unsound.
Data can improve negotiation.
Step 46: use edit distance to tune fuzzy thresholds
Suppose:
95–99% TM: 8% average edit distance.
85–94%: 21%.
75–84%: 43%.
Below 75%: 60%.
This suggests a point where inserting the fuzzy target may stop saving much effort.
Test by language/domain.
Step 47: use edit distance to tune auto-confirm rules
If supposedly high-confidence matches are frequently edited, do not auto-confirm them.
A context match with stable history may be safer than a 100% match from a low-trust TM.
Use metadata and edit evidence together.
Step 48: use edit distance to compare reviewers
Reviewer A average edit distance: 8%.
Reviewer B: 24%.
Before concluding B is stricter or better, sample the edits.
Maybe B handles marketing projects.
Maybe B rewrites stylistically.
Maybe A misses errors.
The metric starts the question.
It does not answer it.
Step 49: use edit distance to detect reviewer drift
If the same reviewer’s average rises sharply over time with similar content:
- project instructions changed?
- source quality changed?
- new terminology migration?
- reviewer style drift?
Investigate.
Trend data is more useful than one isolated score.
Step 50: build a root-cause review for heavy-edit segments
Every month, sample the top 20 heavy-edit segments.
Classify root cause:
- MT;
- TM;
- source;
- terminology;
- translator;
- reviewer preference;
- context missing.
Then fix the most common system cause.
Failure mode 1: low edit distance treated as high quality
Result: critical one-word errors are underestimated.
Repair: LQA severity.
Failure mode 2: high edit distance treated as poor translator quality
Result: creative or difficult work is penalized.
Repair: content/reviewer context.
Failure mode 3: different baselines compared
Result: metrics meaningless.
Repair: define before/after states.
Failure mode 4: language pairs compared raw
Result: tokenization/morphology distort ranking.
Repair: within-pair calibration.
Failure mode 5: average hides heavy tail
Result: rare rewrites invisible.
Repair: effort bands.
Failure mode 6: new text counted as post-editing
Result: edit burden inflated.
Repair: separate new/retranslation.
Failure mode 7: reviewer preference counted as defect
Result: quality inference distorted.
Repair: LQA categories.
Failure mode 8: time ignored
Result: research-heavy no-edit segments look free.
Repair: turnaround metrics.
Failure mode 9: one month triggers route change
Result: noise causes configuration churn.
Repair: trend analysis.
Failure mode 10: metric becomes performance weapon
Result: people optimize edits rather than quality.
Repair: use as diagnostic evidence.
A post-edit analytics table
| Metric | Main question |
|---|---|
| Post-edit rate | How often did reviewers change a translation? |
| Avg. edit distance | How much did changed text move? |
| No edits | How much was accepted unchanged? |
| Light edits | How much needed small cleanup? |
| Medium edits | How much needed meaningful repair? |
| Heavy edits | Where did output require substantial rewriting? |
| New text | Where was no prior target available? |
| Retranslation | Where was the prior target effectively replaced? |
Use the metrics together.
Build a fair comparison design
If you want to compare two translators, two MT profiles, or two review workflows, control the obvious confounders.
Try to keep similar:
- language pair;
- content type;
- source difficulty;
- review brief;
- reviewer pool;
- sample size.
Otherwise the numbers can look precise while comparing different jobs.
A fair comparison is not perfect experimental science. It is simply careful enough that the result supports a practical decision.
Compare the same source through two routes
For engine evaluation, the strongest simple design is paired comparison.
Take the same source sample.