Machine translation & quality

Every project chooses how it is machine-translated, and every machine-translated segment is checked afterwards. This chapter covers both halves: the engines and the routing that decides what actually goes to MT, and the quality layers that review what comes back.

The short version

1. You pick an engine (or none) per project during translation setup. The engine options are: Default (the recommended engine, with automatic fallback for reliability), On-Prem (the in-house engine only), Cloud (the cloud engine only), Google or DeepL (route to that external engine), Copy Source (fills targets with the source text unchanged), and None (skips machine translation entirely).
2. Not everything goes to MT. Exact translation-memory matches and repeated segments can be reused up front, high-scoring TM matches can skip MT by threshold, and content that should survive verbatim — names, codes, URLs and the like — is protected around the MT call. See MT engines and routing.
3. Everything that comes back is checked. Automatic checks flag concrete defects (see QA & entities), and a quality-estimation model scores every machine-translated segment so weak ones are flagged for review and strong ones can be fast-tracked. See Quality estimation.

Run MT & review

You do not have to commit to a workflow before seeing what MT produces. Run MT & review in translation setup fills every target right away while the project stays unactivated. When a target finishes, open it with Review to read — and fix — the results in the editor before deciding the workflow and assigning anyone. The project workflow can be changed freely until a seat is confirmed, so review the MT first, then decide.

In this chapter

• MT engines and routing — what each engine option means, subject domains, TM reuse before MT, and how sensitive content is protected during MT.
• Quality estimation — the per-segment quality score, review flags, pre-confirmation, and AI repair suggestions.