Machine translation evaluation and post-editing
Professional translators score your machine translation or LLM translation output, mark each error by type and severity, and post-edit segments into clean references. You find out which engine to use for each language pair and get human data to train or tune it.
- CriticalMistranslation. "Después de" means after. The safety step is reversed.
- MajorTerminology. "Portada" is a book cover. A printer has a "tapa".
- We evaluate
- MT engines, LLM translation and your own trained models
- Methods
- Error annotation, segment scores, side-by-side ranking and post-editing
- Reviewers
- Professional translators, native in the target language
- Output
- Scores per engine and language pair, error data and references
What MT evaluation tells you
Machine translation evaluation is measuring how good a translation system's output is, usually by having professional translators rate it, mark its errors or correct it. Automatic metrics such as BLEU and COMET are fast, but they need human references to compare against and they can miss errors that change the meaning.
Human scores tell you which engine handles German legal text best, or if your new model beats the old one. They also show which content is fine after a light edit and which still needs a full translation.
What we deliver
For engine selection, model training or ongoing quality checks.
Error annotation
Each error marked with a category and severity, using MQM (Multidimensional Quality Metrics) or your own typology.
Segment scores
Direct assessment scores on a 0 to 100 scale, or adequacy and fluency ratings, per segment.
Side-by-side ranking
Output from two or more engines ranked blind for the same source, to pick a winner per language pair.
Post-editing
Light or full post-editing to your guidelines, with edit distance you can use to measure effort.
Reference translations
Human translations of your test set, so automatic metrics compare against a trusted reference.
Domain test sets
Source sentences chosen from your own content type, such as manuals, contracts or support tickets.
How an MT evaluation runs
Reviewers don't know which engine produced which output.
Test set
Source segments from your real content, in each language pair.
Method
Error typology, scoring scale or post-editing level agreed.
Calibration
Translators score a shared sample and align on severity.
Agreement check QAEvaluation
Blind review of each engine's output, segment by segment.
Report
Scores and error profiles per engine, language pair and content type.
Related AI language services
Questions about MT evaluation
What is machine translation evaluation?
Machine translation evaluation measures the quality of a translation system's output. Human evaluation uses professional translators to score segments, mark errors by type and severity, rank engines side by side or post-edit the output.
What is MQM?
MQM, or Multidimensional Quality Metrics, is a framework for classifying translation errors by category, such as accuracy, fluency, terminology and style, and by severity. Error counts and weights give a quality score you can compare across engines and languages.
What is MT post-editing?
Post-editing is when a translator corrects machine translation output to an agreed standard. Light post-editing fixes meaning and serious errors. Full post-editing brings the text to the quality of a human translation.
Can you evaluate LLM translation?
Yes. We evaluate translation from large language models the same way as from MT engines, and can compare the two on your content.
Send sample output
Send source text and output from the engines you are comparing, or book a call. We score a pilot set in each language pair.