How Custom Glossaries Improve Translation Accuracy

Generic translation fails on the words that matter most to you: names, products and terms of art. A glossary fixes that in two places, and most people only know about one of them.

The speaker says"The Town Hall is moved to Thursday."
  • KoreanWithout a glossary시청이 목요일로 옮겨졌습니다.With your glossary전사 회의가 목요일로 변경되었습니다.
  • SpanishWithout a glossaryEl ayuntamiento se traslada al jueves.With your glossaryLa reunión general se traslada al jueves.
  • JapaneseWithout a glossary市役所は木曜日に移動します。With your glossary全社ミーティングは木曜日に変更になりました。
An illustrative example. Out of context, a "town hall" is a building, and each underlined word means exactly that. In this company it is the monthly all-staff meeting.

Modern translation is very good at ordinary sentences. Ask it to translate "the meeting starts at three" into forty languages and it will be right forty times. Where it stumbles is on the small set of words that belong to you: the name of your product, what you call your all-staff meeting, the acronym your team coined in 2019.

Those words are a tiny share of what gets said, and they carry most of the meaning. A listener can forgive an awkward phrase. They cannot follow a meeting in which the thing being discussed has a different name every time it is mentioned.

Two places a term can go wrong

Live translation is two steps. First the system has to hear the word. Then it has to translate it. Your special terms can fail at either step, and the failures look different.

A hearing failure produces nonsense. Say "QolSync AI" to a speech recogniser that has never met the word and you may well get "coal sync". No translator can rescue that, because the right word never reached it. A translation failure is subtler: the word was heard correctly and then rendered as a plausible phrase that is not the one your people use.

So a glossary has to work at both steps. Ours does.

Your glossaryNames, products, acronyms and their approved translations
  1. Step 1 · HearingTerms are given to speech recognition as hintsSo an unusual word is recognised as itself, not as the nearest common one.
  2. Step 2 · TranslatingTerms found in the sentence go along with itSo the translator uses your word for it, fitted to the grammar of the sentence.
The captionHeard correctly, and named the way your people name it
One glossary, used twice on the way from speech to caption.

Step one: helping it hear

Before a session begins, your terms are handed to speech recognition as context: these are words you are likely to hear. There is only so much room for hints, so they are ranked. Terms marked as case-sensitive go first, because those are usually proper names. High-priority terms come next, then multi-word phrases, which are the easiest to mishear in pieces.

Step two: only the terms in this sentence

The obvious way to use a glossary is to hand the whole list to the translator with every sentence. It is also the wrong way. A long list is slow to read, and the more terms a translator is shown at once, the more likely it is to use one where it does not belong.

Instead, each sentence is scanned for the glossary terms that actually appear in it, and only those travel with it.

A glossary of 1,000 termsTwo of them appear in this sentence.
What travels with the sentence
  • Town Hall전사 회의 · reunión general · 全社ミーティング
  • QolSync AIQolSync AI (keep as is)
The search checks for every term in a single pass over the sentence, so a glossary of ten thousand terms takes about as long to check as a glossary of ten.

That last point is what makes a large glossary practical in a live setting. The time it takes to find your terms depends on how long the sentence is, not on how long your list is. You can keep adding to a glossary for years without slowing down a single caption.

Fitted, not pasted

Swapping one word for another works in a spreadsheet. It does not work in a sentence. In Korean, a noun takes a different ending depending on whether it is doing something or having something done to it. In Spanish, the words around it change to agree with its gender. Paste a term in blindly and the result reads like a ransom note.

"The Town Hall starts at three."전사 회의가 3시에 시작합니다.La reunión general empieza a las tres.全社ミーティングが3時に始まります。
"Please join the Town Hall."전사 회의에 참석해 주세요.Únase a la reunión general.全社ミーティングにご参加ください。
Korean, Spanish and Japanese. The highlighted term is fixed. The small words around it, shown in brown, change with its role in the sentence.

So the translator is told two things: use this term for this word, and adjust the grammar around it so the sentence still reads naturally. Your word is not negotiable. The ending is the translator's job.

One deliberate trade-off: the fast draft captions that appear while someone is still speaking skip the glossary, to stay fast. The finished caption that replaces the draft a moment later is the one that carries your terms. In practice you see a word settle into its proper form as the sentence completes.

Building one in ten minutes

The reason most glossaries never get built is that somebody has to sit down and type them. We tried to remove as much of that as we could.

Let the AI read your document

The fastest route starts from something you already have: the agenda, last quarter's report, the slide deck for the event. Upload a PDF, a Word file, a PowerPoint or plain text, choose your two languages, and the AI reads it for you. It usually takes one to three minutes, and you can close the window while it works.

What it does with your file
  1. Reads the document, scanned pages included
  2. Works out what subject it is about
  3. Finds the names, products, acronyms and terms that should stay untranslated
  4. Proposes a translation for each one
  5. Hands the list to you to review
Review termsEnglish to Korean
Ready to add
  • SelectedTown Hall전사 회의
  • SelectedQolSync AIQolSync AI
  • SelectedOKROKR
Needs a look
  • Not selectedAtlas아틀라스 · 72% sure
  • Not selectedNorth Star북극성 · 64% sure
An illustrative review screen. Terms the AI is confident about are ticked for you. Uncertain ones are left unticked, with how sure it is.

Three things about this step are deliberate. First, it works out the subject of the document before it looks for terms, because the same word means different things in different places. "Sprint" is one thing at a software company and another at an athletics club.

Second, every proposed term is checked against the actual text of your file before you see it. A term that does not appear in the document cannot appear in the list.

Third, you stay in charge. Nothing is saved until you press Add, and uncertain terms are never added unless you tick them yourself. You are shown what the analysis will use from your allowance before it runs, and cancelling at that point costs nothing. The file you uploaded is deleted when the job is finished.

  • Or type a few inFor the handful of words you already know matter, adding them by hand takes a minute.
  • Bring a spreadsheetA plain two-column file of term and translation is enough. If a file arrives in an encoding that would garble the text, it is refused with an explanation, not imported broken.
  • Share it with the teamA team glossary applies to everyone's sessions, so the product is called the same thing whoever is hosting.
Three other ways to build and share a glossary.

What to put in it

Start smaller than you think. Ten well-chosen terms do more than three hundred ordinary ones, because ordinary words are already translated well. The terms worth adding are the ones a stranger to your organisation would not know:

  • People and places. Names of speakers, offices, venues and teams.
  • Things you make. Product and feature names, especially those that should never be translated.
  • Acronyms. The ones you invented, and technical ones such as API or SDK that should stay as they are.
  • Words with an official translation. Titles, roles and terms your community has already agreed on in each language.

Then run one meeting, read the transcript, and add whatever came out wrong. A glossary is finished the same way a dictionary is: never, and usefully so from the first day.