bAIbelAV
/av/term-extraction03 / 11 · INTAKE

Term extraction

Build a termbase from any previous translation — your own, your client's or other reference material — without leaving the tool you translate in.

Point it at your files. It proposes the terms; you approve the ones you want. Then it finds how each one was translated, says how sure it is, and puts the pairs you accept into your termbase. No second program to buy, learn or feed.

In numbers

0candidates that go forward unapproved
2passes: the terms, then their translations
6scripts split by character, not by space

It proposes the terms. You approve them. Then it goes to work.

Run the extraction and you get candidates: terms the AI model found in this document, not terms it has added to anything. You work through them. Only the ones you approve go forward into the target-side pass — the approved list is what the second stage reads, so an unapproved candidate has nowhere to go.

CANDIDATES THE AI MODEL FOUND · YOU TICK THE ONES YOU WANT force majeure New 7 the parties 41 limitation of liability New 5 hereinafter 12 notice period New 6 business day Existing 3 Showing 6 of 6 (Selected: 3) the number: segments it was found in THE SECOND PASS READS THIS LIST, AND ONLY THIS Target Term Extraction force majeure limitation of liability notice period how each was translated, found in the segments it occurs in AN UNTICKED CANDIDATE HAS NOWHERE TO GO

Then it finds how each term was translated, and says how sure it is.

For every term you approved, it reads the segments the term occurs in and proposes the translation your documents actually use. The confidence beside it is measured, not guessed: the share of that term's segments whose translation contains the proposal.

Into your termbase, marked as new.

Commit, and the pairs you accepted go into your termbase as new, provisional entries, each saying it was added by the extraction and naming the task. Nothing already in the termbase is touched. And the whole termbase exports to MultiTerm XML, every language, whenever you want it elsewhere.

Stop halfway. Come back on Thursday.

Candidates are kept with the task. Close the project, reopen it, and the run picks up where it was rather than paying to find the same terms twice. You see what each run cost while it happens.

Stand-ins here too.

Term extraction reads your document, so whatever it sends to an AI model goes through the same confidentiality handling as translation — your project's settings, or a choice made for this task alone — and every request is read back before it goes and refused if a protected name is still in it. If you protect it, the candidates show the stand-ins the AI model saw.

Confidentiality

It splits words correctly in scripts that do not use spaces.

Six scripts — Chinese, Japanese, Korean, Thai, Lao and Khmer — go through character-based tokenisation rather than a whitespace split. If you have ever watched a term tool return one "term" per paragraph of Chinese, that is the problem this avoids.

A term list, when there is no translation yet.

Run it on a source document alone and it ends with a term list: the terms you approve, each with one sentence it appears in, saved as a CSV file that opens in Excel. Termbase entries need a translation, so those come from a translated file.