Translation memory

A translation memory, often shortened to TM, is a database of text you have already translated, stored as pairs: a source segment and its translation. A segment is usually a sentence, sometimes a paragraph or a heading. When the same segment comes up again, the memory offers the stored translation, so nobody translates it twice and it reads the same everywhere it appears.

Exact and fuzzy matches#

A translation tool compares each new segment with the memory and reports what it finds.

An exact match is a segment identical to one already stored, character for character. Wikipedia's article on translation memory notes that these are also called 100% matches.

A fuzzy match is a stored segment that is close but not the same. In an invented example, "Free shipping on orders over 50 euros" comes up against a stored "Free shipping on orders over 40 euros". Tools score the similarity as a percentage, and a translator edits the stored translation to fit.

Memories move between tools in TMX, the Translation Memory eXchange format, so a company that changes agency or tool can take its memory along.

How a translation memory differs from a glossary#

The two are often confused because both store translations for reuse. They work at different sizes and are maintained differently.

A glossary holds single terms and the one way each must be translated: product names, plan names, terms that must never vary. Someone decides every entry by hand, and the glossary is consulted while translating new text.

A translation memory holds whole segments and fills itself as you translate. Nobody curates it entry by entry. It saves the work of retranslating, and it also keeps any mistake that was approved once.

A site needs a glossary to make new translations consistent, and benefits from a memory to avoid paying twice for the same sentence. The article on how to build a translation glossary covers the first.

How LangRelay reuses translations#

LangRelay works with what is on your rendered pages, and its reuse is exact. Identical source text of the same kind shares one translation across a Site: a footer line on five hundred pages is translated and reviewed once, then shown on every page. When machine translation runs on a new page, it reuses every translation that already exists for that text, including those a person wrote or edited, and never overwrites one.

The exception is explicit. If you want new machine output for text that already has a translation, you select it in the Translation grid and choose Machine retranslate, and the result goes through review like any other translation.

LangRelay's design also covers a sentence that changes: its earlier translation can be carried over to the new sentence as a recovered translation, marked as recovered and unpublishable until a person accepts it, so a near match never goes live on its own. Recovered translations are on the roadmap; today a changed sentence is machine-translated afresh and goes live like any other translation: right away by default, unless the free checks hold it.

LangRelay does not score fuzzy matches or exchange TMX files. A translation team that relies on a memory in a translation management system can keep using it for documents and apps, and LangRelay handles the website.

Reuse and the glossary sit with LangRelay's other quality controls on the translation quality page, and post-editing website copy covers the review that follows. For the texts a theme or plugin prints, see string translation.