Translation memory and glossary

The translation memory (TM) and the glossary are what keep a multilingual site consistent over time. Both belong to the organization and are shared by all its sites and adapters: a term on a django CMS page and the same term in a product model are translated the same way.

The translation memory remembers approved work

Every approved translation goes into the memory: approvals your editors report, including their corrections, existing translations imported from your site, and TMX files you import. The memory stores one translation per source text and language pair, and the latest approval wins. So when an editor corrects a translation, the correction is what gets reused.

Exact matches are free and still reviewed

Before a job sends a segment to an engine, Ponyglot looks it up in the memory. An exact match is the same source text after normalization, in the same format and language pair. It’s the same comparison as the fingerprint.

An exact match isn’t sent to any engine, so it costs nothing: TM hits don’t count towards your characters. It also works when an engine is unavailable.

An exact match is still delivered as a draft and reviewed. The same English sentence can need a different translation in a different context, and approving without an editor would break the promise that nothing is published automatically.

Repeats within a job are sent once

If the same text occurs several times in one job, for example a “Read more” link on fifty pages, Ponyglot sends it to the engine once and reuses the result for the other segments. The copies are free as well.

Fuzzy matches are only references

A fuzzy match is a similar, but not identical, source text: the same sentence with a different number, for example. Applying its translation automatically would be wrong often enough to be dangerous. Instead, Ponyglot gives fuzzy matches to the LLM as reference translations, so it can follow your established wording. They are never applied as they are, and DeepL doesn’t use them.

The glossary enforces terms

The glossary is the explicit counterpart: rules you write down instead of examples the memory collects. Each term has a translation per language, or is marked do not translate for brand and product names.

Ponyglot applies the glossary in two ways. It tells the engines: DeepL gets a glossary for the language pair, the LLM gets the terms that occur in the text. Then it checks the result: a changed do not translate term is an error that holds the translation back; a missing target term is a warning for the editor. If DeepL’s glossary can’t be prepared, Ponyglot doesn’t translate the language’s segments at all rather than translating them without the glossary. Consistency is the point.

How they work together

For each segment, Ponyglot first checks the memory. Only segments without an exact match go to an engine, together with the glossary terms, the brand voice and, for the LLM, fuzzy matches. The result is checked, delivered as a draft, and, once approved, flows back into the memory. Over time, more segments come from the memory and fewer characters are billed.