Keep terminology aligned through translation and client feedback

When this use case fits

Use this approach when a project has specialist, changing or client-specific vocabulary and the organization needs a deliberate record of the terms it accepts. It is designed for project managers and localization leads who want AI help with terminology work without treating a generated suggestion as an automatic glossary decision.

The result is a controlled loop: Term Expert proposes terminology, people approve it, Translator uses the approved terms, and later human corrections can become evidence for a new glossary decision. People remain responsible for approving terminology and confirming translated segments.

Build the first controlled glossary

Start with a termbase that has a writable direction for the job’s language pair. You can check the pair’s termbases under Resources… on the project page.

1. In the job’s AI cell, choose Term Expert (Extract) and start it. Its settings (Settings: Term Expert (Extract)) are used by every job in that language pair from the next run. Term Expert finds terminology in the source text and proposes target equivalents; when it finishes, the cell shows how many terms were proposed.
2. Human verification 1 — proposed terminology. An Admin or PM opens the termbase and reviews the proposed terms on the Terms tab, with their source context.
3. Approve the terms that suit the project, one by one or with Approve all Proposed, and leave or remove the rest. Only approved terms guide translation.

Proposed terms do not act as approved terminology. The setting Approve extracted terms automatically writes extracted terms as approved instead; it is a trust choice for the whole language pair that skips the review step, and it is not the approach described here.

Bring later corrections back into the loop

Imagine that a client or linguist changes a delivered translation. Reconcile reads corrections recorded in TilBridge, so a change that exists only in an external file has to come back first: import the corrected bilingual Word or XLIFF file, or make the correction in the TilBridge editor.

1. In the job’s AI cell, choose Term Expert (Reconcile) and start it after those corrections are recorded.
2. Reconcile compares the human corrections with approved terminology. It does not build a new glossary from corrections; where a correction diverges from an approved term, it opens a reviewable challenge.
3. Human verification 2 — terminology change. In the termbase’s Changes to review tab, an Admin or PM reviews the evidence and chooses one outcome: Add as synonym approves the proposal and keeps the existing term; Replace approved term approves the proposal and marks the previous term Forbidden; Dismiss closes the challenge when no glossary change is warranted.

An open challenge has to be decided in Changes to review; its proposal cannot be edited or deleted as a shortcut.

Carry an approved decision back into translation

Once a term is approved, the next Translator run on a job in that language pair uses it: Translator receives the approved terms whose source occurs in the document. Proposed and Forbidden terms are not used. Machine-translation glossaries follow the same rule.

To check existing translations against approved terms, including segments already confirmed, choose Reviewer (Terminology check) in the job’s AI cell and start it. It checks approved terms only. Segments where no approved term applies, or where the approved term is already present, are usually not sent to the model. Locked segments are excluded, and nothing is saved to translation memory.

This is still not an automatic finish line.

1. Human verification 3 — corrected segments. Reviewer corrections become Draft and unreviewed, including changes to previously confirmed exact matches. Open the run’s review report from the job’s AI run history (Open the review report) to see how many segments were checked, corrected, blocked or not checked.
2. A person reviews those changes in context in the editor and confirms the segments that should become accepted work. Translator output likewise stays unconfirmed until a person confirms it.

Other terminology starting points

When a client’s existing translation memory, rather than the new project’s source text, is the best terminology source, use Extract terms… from the translation memory or termbase page. It proposes terms from the whole memory or a random sample into one termbase direction, and those proposals are reviewed the same way.

Limits and operating choices

• Term Expert needs a writable termbase direction for the job’s language pair before it can start.
• It does not guarantee a complete or correct glossary, and it does not replace human terminology review.
• Proposed terms do not guide translation until an Admin or PM approves them, unless automatic approval is turned on for the language pair.
• Re-running extraction does not replace a weak proposal: a source already present in an attached termbase is skipped in any status. An Admin or PM can edit or delete an ordinary proposal; a proposal under an open challenge must be decided in Changes to review.
• Reconcile records divergence for human decision; it does not change the glossary itself.
• Translator uses approved terms as guidance for the whole document; people still confirm the result.
• Reviewer (Terminology check) leaves its corrections as Draft for a person to confirm and does not save translations to translation memory. Without approved terms, it has nothing to check.

Treat this as a repeatable review routine, not a guarantee that terminology or translation is correct. The value comes from combining useful automation with visible human decisions at each point where organizational terminology changes.