Consistency is the product.

Ask three good translators to translate your company name into Arabic and you will get three defensible answers. All correct, all different, and that is the problem. A ministry that appears as three different names across its own publications looks careless in a way no single mistranslation ever could. Consistency is not a nice-to-have in institutional translation; it is most of what "quality" means to the reader.

Two tools carry the consistency load: the glossary (or term base), which fixes the words, and translation memory, which fixes the sentences. They sound similar and get confused constantly, so this lesson takes them one at a time.

The term base: one approved answer per term

A glossary is a list of source terms, each paired with exactly one approved target translation for a language pair. Not a dictionary of options: a ruling. The debate about how to render "Chief Sustainability Officer" in Arabic happens once, someone decides, and the decision is recorded with a category and a note explaining why. From then on the question is closed for everyone, including the machine.

What belongs in one? Five families cover most of it:

• Names. The organization, its subsidiaries, departments, and key people. Transliteration is a choice, not a fact: is it "Al Dana" or "Al-Danah"? Decide once.
• Titles. Job titles and honorifics, which Arabic renders in more than one legitimate way. "His Excellency" has conventions that vary by country and context.
• Product and program terms. Initiative names, product names, brand vocabulary. Often the ruling is "do not translate," which is itself a glossary entry.
• Legal and technical terms of art. "Force majeure," "net working capital," terms where the approved rendering carries legal weight.
• Transliterations. Foreign words your organization spells a particular way in Arabic, or Arabic words it romanizes a particular way in English.

Governance: a glossary is only as good as its gatekeeper

An unowned glossary rots. Duplicate entries appear, contradictory terms creep in, and within a year nobody trusts it, which means nobody uses it. The fix is boring and organizational, not technical: someone owns the term base. Anyone can propose a term (reviewers hit missing terms daily; they are your best source), but a named owner, often the communications lead or a senior reviewer, approves what enters. At Al Dana Holdings, that is one person: reviewers propose, the communications director approves, and the glossary stays small, current, and trusted. Small matters. Five hundred terms people rely on beat five thousand nobody checks.

Translation memory: never translate the same sentence twice

Where the glossary works at the level of terms, translation memory works at the level of segments: sentences, headings, table cells, the units a document breaks into for translation. Every time a reviewer approves a segment, the source and its approved translation are stored as a pair. When a future document contains a matching segment, the stored translation comes back instead of a fresh machine pass.

Matches come in two strengths. An exact match (100 percent) means the new segment is character-for-character identical to one already approved: the stored translation is reused and locked, because re-translating it could only introduce drift. A fuzzy match is close but not identical, say 85 or 92 percent similar: the stored translation is offered as a suggestion with its score, and the reviewer decides whether to take it, adapt it, or ignore it. The score is honest about what it is: a similarity of characters, not a guarantee of meaning. A one-word difference can be trivial ("2025" becomes "2026") or catastrophic ("shall" becomes "shall not"), which is exactly why fuzzy matches suggest rather than lock.

One rule keeps the asset clean: only approved segments enter the memory. Store raw machine output and the memory becomes a machine echoing its own first drafts back at you. Store only what reviewers signed off, and every match carries a human decision inside it.

Portability: your memory is yours

Years of translation work should not be hostage to any tool, including the one you like. The industry settled this with TMX, a standard exchange format for translation memories: any serious platform imports it and exports it. If your organization has agency history in a TMX file, importing it means your new system starts matching against years of past work on day one. And if you ever leave, the memory leaves with you. Ask any vendor two questions before you commit: can I bring my TMX in, and can I take it out?

In TranslateX, both tools live in the workspace and apply automatically: the glossary is enforced on every job, memory matches surface in the review editor with their scores, and TMX moves in and out freely. Which raises the obvious next question: what does a reviewer actually do with all these machine outputs, matches, and suggestions? That is lesson four, the post-editing workflow.