How the translator learns
The AI translator does not learn by itself. It follows the style guide, the glossary and the memory, and those change only when a person approves a change. The learning loop collects what people correct, proposes rules from it, and measures whether translations improve.
1. What is collected #
Each of these is recorded as a correction, under Translation → Translation corrections (editors and admins):
| Kind | Where it comes from |
|---|---|
| Edit | A person changed an AI translation before publishing it. Recorded when the translation is published. |
| Report | Someone pressed 🚩 Report translation in a language and wrote what is wrong. Everyone can report; the page does not change. |
| Copy desk | A rule someone stated at the 💬 copy desk and accepted ("never call a speaker a solution"). |
| Review | A change the review pass made, with its reason. A weak signal. |
| Proofread | A correction to the default language's text. Kept for the record, never learned from. |
Publishing a translation also approves its texts in Translation memory, so they are reused word for word from then on (see Memory, glossary and style guide). Both need the site's own AI translation.
2. Distill #
In Translation status, the quality panel counts the corrections waiting for each language. Distill (editors and admins) reads them and proposes, for that language:
- glossary terms: a term needs the same change at least twice, or a report that asks for it outright. New terms are added to the Glossary unreviewed; for a term that already exists, it says so and leaves the decision to you;
- notes for the style guide, such as "use informal address";
- example translations to learn from.
Nothing changes until a person approves it. The corrections it read are marked as distilled.
3. Approve #
Under Waiting for approval in the quality panel, or in the Glossary and under Translation → Translation proposals:
- Approve a term to mark it reviewed: it steers translations from then on. To drop a suggested term, delete it in the Glossary.
- Approve a note to add it to that language's notes in the style guide, or an example to add it to the style guide's examples. Dismiss deletes the proposal.
Suggestions from people's own AI assistants arrive the same way, marked with who made them.
4. Measure #
An evaluation measures the translator against your golden references: approved Translation memory entries ticked as model translations (aim for 50 to 100 per language; an evaluation needs at least five).
Run evaluation translates those texts again, without the memory, and an AI judge compares the result with your references: how well it keeps the glossary, how close it comes, and how natural it reads. It runs in the background for a few minutes, and costs about as much as translating those texts twice. Nothing on the site changes.
The quality panel shows the last runs side by side, so you can see whether a new glossary term or style guide note made translations better. All runs lists every run under Translation → Translation evaluations.