Notes from a PM's desk.
- 翻訳PMのデスクからのメモ。
- 翻译项目经理的工作笔记。
- 번역 PM의 업무 노트.
- Notizen vom Schreibtisch eines Projektmanagers.
- Заметки менеджера переводческих проектов.
- Tarjima loyihasi menejerining qaydlari.
- ملاحظات من مكتب مدير مشاريع الترجمة.
Translator performance KPIs: how to build a scorecard
A translator scorecard should help a PM choose work and give useful feedback. Start with the text that was reviewed, the errors found and the deadline agreed. Keep enough context to explain each score.
Delivery time alone does not tell you how good a translation is. An LQA score alone does not explain a late file. Read them together, and compare work with similar content and review rules.
Keep the context with each score
Record the language pair, content type and service. Translation, editing and machine translation post-editing are different jobs. Also record the reviewed word count, reviewer and version of the client glossary.
If a source file arrived late or a deadline changed, keep that detail. Otherwise a report can blame a linguist for a delay outside their control.
For quality scores, use the same error definitions and severity rules. Reviewers should check a shared sample and discuss where they disagree before their scores are used to compare linguists.
Five translator performance measures
1. Edit distance from machine translation
Edit distance measures changes between an MT draft and the final text. State how you count them: tools may count character or word edits in different ways.
Use it to choose segments for review. A low edit distance can mean that the draft was good, or that it was barely checked. A high one can reflect poor MT output, a style rewrite or needed corrections. The number alone cannot tell you which happened.
Compare similar language pairs and content. Do not set a minimum edit percentage as proof that someone worked hard enough.
2. Fuzzy matches accepted unchanged
Divide fuzzy-match segments confirmed without edits by all fuzzy-match segments confirmed. Multiply by 100 to get a percentage. Keep exact matches separate, and compare the same fuzzy bands.
A high rate is a reason to review a sample. It is not proof of poor work. A fuzzy match can need no change when only a tag or an irrelevant source detail differs. Check the text before judging the linguist.
This measure needs data about which match was offered and what was finally confirmed. Do not infer it from the overall TM match rate.
3. Review outcomes
If your review process records responses, track how many proposed corrections are accepted, disputed or withdrawn. Keep the reason for a dispute with the segment.
An acceptance rate is the number of accepted corrections divided by the corrections that received a response. Report open items separately. Use the same definition each month.
A linguist who accepts every change is not automatically better than one who disputes a style preference. Review a few examples and look for repeated, confirmed errors.
4. Terminology and consistency errors
Count errors where a term breaks the agreed glossary or repeated text changes without a reason. Divide by the reviewed word count to compare samples of different sizes.
Check context first. The same source word may need different target terms in different parts of a document. A glossary alert asks for a decision; it does not prove an error.
KanCAT shows glossary hits in the CAT editor and alerts when a preferred term is missed. Reviewers can log issues on the segment. See the CAT editor and quality review features.
5. LQA error rate by category
Linguistic quality assurance, or LQA, records errors in reviewed text. Choose categories that suit the work and define severity before scoring. The MQM Council explains how error types and severity weights form a quality score.
For a simple raw error rate, divide confirmed errors by reviewed words, then multiply by 1,000. For example, four terminology errors in a 2,000-word sample give two errors per 1,000 words. This is an invented example, not a passing threshold.
Keep the sample size visible. One error in a short sample can change the rate sharply. A raw count also treats a minor error and a critical one alike. If you use weighted penalties, publish the weights and compare scores only under the same rules.
KanCAT logs MQM issues on segments, with critical, major, minor and neutral severity levels. Full MQM scoring starts on Starter; the freelancer Free plan has pass or fail scoring. The categories in KanCAT are a selection for the product, not the full MQM framework.
Add delivery data without calling it quality
Track on-time jobs against the deadline each linguist accepted. Divide on-time deliveries by jobs due in the period, and show the job count beside the rate.
Record deadline changes and holds. Compare patterns over several jobs before drawing a conclusion. If two linguists have the same quality score but different delivery histories, that can affect whom you offer an urgent job.
Use the scorecard in a conversation
Choose work that fits the linguist's strengths. Compare similar reviewed samples before assigning high-risk content. A large word count alone does not prove experience in a subject.
Use the error category to make feedback specific. Repeated glossary errors call for a check of the glossary, brief and review process. Discuss the examples with the linguist before deciding what needs to change.
Keep agreed fees with the assignment. A score can inform a rate discussion, but it does not change an existing agreement. Our vendor PO guide explains how to record the terms.
Collect only the data you need
Tell linguists what you track and how you use it. Limit access to people who need the report. Give a linguist a way to question an error or supply missing context. Check the privacy and retention rules that apply to your business.
KanCAT has linguist reports and a leaderboard with quality scores and delivery data. The measures described above are a scorecard method; KanCAT does not claim to calculate them all. Read the reporting features and plan limits before choosing a tool.
Start with reviewed word count and LQA errors by category. Test your definitions on a shared sample, then add measures when you have data you can explain.