Skip to content
LionrapidLionRapid Docs

MT engines and routing

Machine translation is configured per project in translation setup: you choose an engine, and you control how much work is reused from translation memories instead of being machine-translated. Around the MT call itself, the platform protects content that must survive verbatim.

Choosing an engine

Section titled “Choosing an engine”
  • Default — the recommended engine, with automatic fallback for reliability. Pick this unless you have a specific reason not to.
  • On-Prem — the in-house engine only.
  • Cloud — the cloud engine only.
  • Google / DeepL — route to that external engine. Choose these when you have an existing preference for a provider’s output in your language pair. Note that the external engines cover fewer languages than the platform itself.
  • Copy Source — fills every target with the source text unchanged. Useful when the “translation” step is really an editing pass, or targets must start as copies.
  • None — skips machine translation entirely; translators start from empty segments (or from TM matches, which still apply).

Language support

Section titled “Language support”

The platform’s language list covers 345 languages plus about 1,100 script and region variants (for example Brazilian versus European Portuguese, or Cyrillic versus Latin Serbian). Engine coverage varies within that list — the external engines in particular support a subset — so the setup screen only offers what applies to your language pair.

Subject domains

Section titled “Subject domains”

Each file in a project carries a subject domain — one of a fixed set of categories such as legal, financial, medical, software, or marketing. The platform suggests a domain automatically when a file is imported; you can confirm or change it per file during translation setup. The domain describes what the file is about, and the same domain labels are used to organize your TMs and glossaries and appear in translator profiles as areas of experience.

What is reused before MT

Section titled “What is reused before MT”

Machine translation is the fallback, not the first resort. With TMs attached to a language pair, two setup controls decide how much never reaches the engine:

  • Reuse exact TM and repeated segments before MT. When enabled, exact attached-TM matches are filled from the memory, and repeated source segments are translated once and propagated instead of being sent repeatedly.
  • Skip MT at TM match ≥. TM matches scoring at or above this percentage are filled from the memory instead of being machine-translated. At 100 — the default — only matches identical up to case and spacing are reused; lower the threshold to reuse fuzzier matches and post-edit them. The dropdown needs a TM attached to the pair.

Segments filled from a TM keep that provenance, so you can always tell machine output from memory reuse.

Independently of TMs, segments that should not be translated at all are copied instead of machine-translated — for example URLs, email addresses, purely numeric content, and text that is already in the target language.

Protected entities during MT

Section titled “Protected entities during MT”

Some things inside a segment must come out of machine translation exactly as they went in. Before a segment is sent to the engine, recognized entities of the following kinds are replaced by opaque placeholder tokens, and after translation the original text is restored verbatim:

  • person names
  • organization names
  • product and brand names
  • codes and identifiers
  • email addresses
  • URLs
  • IBANs
  • phone numbers
  • acronyms

This is why “Jaguar” the brand survives untouched even in a language where “jaguar” the animal would correctly translate.

Two kinds of content are deliberately not frozen:

  • Dates, money, and measurements are left to the engine so they can be localized (for example 1,200 becoming 1 200). A separate automatic check audits that no numbers were lost or changed in value.
  • Place names translate normally, because languages have their own names for places (Germany is “Saksa” in Finnish).

After restoration, an automatic check flags any placeholder that failed to come back cleanly, so a broken token can never ship silently in a delivered translation. The whole mechanism is fail-soft: if protection cannot be applied to a segment, the segment is simply translated without it rather than failing.

How entities look and behave while editing is covered in QA & entities.

Section titled “Related pages”