Extract terms from a previous translation

Two ways to find terms. Create glossary from files gives a quick list of English terms. The terminology extraction wizard reads a translated file, shows what is sent, extracts candidates, lets you filter, edit and approve them, finds translations from the file's own usage, and commits the approved pairs into the termbase as Provisional after a pre-flight check.

1. Two ways to find terms

   bAIbel AV can ask an AI model to find the terms in your documents in two ways. Create glossary from files is quick: a list of English terms for the Project Glossary. The terminology extraction wizard builds termbase entries from a translated file, and you approve every step.

   

2. In Project Glossaries, click Create from files…, tick files and Extract terms

   In the project, open Project Glossaries and click Create from files…. Tick the files, choose the AI model, and click Extract terms. The source text is sent in batches; terms the glossary already has are skipped.

   

3. Open the glossary and fill in translations by hand

   A message gives the count and the cost: 34 terms, for about three cents. Open the glossary: every term is there with no translation, and nothing was reviewed before it was written. Fill in the German by hand (drive belt → Antriebsriemen), and delete what is not a term, like seal.

   

4. Open Terminology: each extraction is a task

   Open Terminology. Each extraction is a task: Create New Extraction starts one, and a saved task opens again where it was left. Liesel saved this one before running it.

   

5. Open the task with ▶ and choose Bilingual

   Open the task with ▶. Bilingual reads a translated file, here the project's own: 395 words in 43 segments. Only its English is sent to find terms; the German is kept for the contexts and for the next pass.

   

6. Check Confidentiality and the batch size

   Confidentiality shows the project's own settings: the wizard masks the text as translation does, and checks every request for protected names before it is sent. This manual is public, so nothing is masked. Candidates already in the parent termbase are marked. The English goes in batches of about 150 words: three requests here.

   

7. Click Next: the prompts

   Click Next. Each side has its prompt; Default is the one shipped with AV. The target side sends each approved term with up to five translated examples, the most different ones. Special Instructions are added to every request for terms: leave them empty unless the client has a point to make.

   

8. Click Next, Start Extraction

   Click Next, then Start Extraction, and confirm. The run takes a few seconds: all 43 segments, 34 candidate terms, and the cost of the whole run.

   

9. Click Next: Review Terms; click one to see its contexts

   Click Next: Review Terms. Each candidate shows how many segments it is in. Those already in the termbase are marked EXISTING and left unticked. Click one to see it in context: drive belt is translated two ways in the file.

   

10. Set Min to 2 and Apply

    Set Min to 2 and click Apply: candidates found in only one segment are unticked. Nothing is ticked by it, and a real term found once can be ticked again by hand.

    

11. Click the pencil to edit a candidate

    The pencil edits a candidate. The AI model proposed blade bolts; the entry should be in the singular, blade bolt. It keeps its contexts.

    

12. Tick the terms to approve

    Untick the everyday words the AI model picked up: engine, mower, wheel. Tick fuel stabiliser again: it is found once, but it is a real term. Ticked means approved, and only the 13 ticked terms go on to the next step.

    

13. Click Next, Start Target Extraction

    Click Next, then Start Target Extraction. Only the approved terms are sent: one request per term, each with up to five of its segments in English and German. The AI model proposes 14 translations for the 13 terms, for under eight cents.

    

14. Approve (✓) or reject (✕) each pair

    Click Next. Each proposal shows how often the file uses it: drive belt is Antriebsriemen in 67% of its segments and Keilriemen in 33%. Reject Keilriemen (✕) and approve the other 13 pairs (✓).

    

15. Click Next: Export

    Click Next: Export. The approved pairs are merged into the parent termbase by default. Choosing Create New Termbase with a name already taken is refused by Run Pre-flight Check, which says why.

    

16. Choose Merge into Existing and Run Pre-flight Check; Commit & Export

    Choose Merge into Existing and Thistlecombe terms, and tick the MultiTerm XML copy. The check counts what Commit will do: 13 entries, 26 new terms marked Provisional. Existing terms are left as they are. Commit & Export is enabled only after a check passes.

    

17. In the termbase editor, filter on Provisional

    In the termbase editor, filter on Provisional: the 13 new entries are there, waiting to be agreed, beside one that was Provisional before. Each new entry says it was added by the term extraction wizard, and its note names the task.

    

18. Existing entries are left alone

    Search for blade: the company's blade → Messer is still Approved, as it was. The new blade adapter and blade bolt sit beside it, Provisional.

    

19. The AI model proposes; you approve

    The AI model proposes; you approve, edit and commit.