Notebook / Terminology

Follow a term across the game

Start with the thing, not the word

Players do not encounter a glossary. They encounter a sword in an inventory, its name in a dialogue choice and an instruction that asks them to equip it. Terminology consistency matters because those appearances must point to the same thing. A superficial word match can miss that relationship, while a rigid checker can flag perfectly reasonable grammar.

Begin by identifying concepts: named items, abilities, factions, locations and interface actions. Give each concept a stable identifier and a short definition. Record approved target-language forms, forbidden alternatives and the contexts where an exception is allowed. Without those boundaries, an AI system is comparing strings against an assumption rather than a usable language policy.

Keep common words separate from protected names. “Charge” may refer to a combat move, an electrical resource or a shop transaction. One preferred translation cannot safely cover all three. A concept record should tell the reviewer which meaning is intended and which evidence establishes it.

Two language editors comparing concept notes at a sunlit desk, with all document text indistinct
Record the identity behind a term before reviewing its appearances.

Collect a traceable comparison set

Export source and target strings with stable IDs, resource paths, language tags and build identifiers. Include descriptions, character names and developer comments when permission allows. Preserve placeholders rather than replacing them with plausible words: the difference between an item variable and ordinary text often determines whether a term can be checked automatically.

Formats such as OASIS XLIFF and TBX can carry useful localization or terminology information, although a format alone does not guarantee complete context. BCP 47 language tags help distinguish language varieties; they do not establish a project’s preferred vocabulary. Keep the source export and glossary revision associated with each analysis run.

Source / quest.vault.014

Equip the Moonshard before entering the vault.

French candidate

Équipez le fragment lunaire avant d’entrer dans le caveau.

Approved concept / item.moonshard

Éclat de lune — a named inventory item used to open the vault.

In this invented example, the candidate may communicate a similar literal meaning. The potential problem is recognition: the player’s inventory displays a different name. Confirm that the quest refers to the same item before calling the translation incorrect.

Let deterministic checks go first

Check exact prohibited terms, missing required names and malformed placeholders with predictable rules. Normalize Unicode deliberately; visually identical characters can have different underlying representations. Unicode Standard Annex #15 describes normalization forms, but normalization should never silently rewrite a release asset without a controlled process.

Then use AI for cases where surface matching is insufficient: an inflected name, a paraphrased ability reference or a possible conflict between distant strings. Ask for the concept identifier, both string IDs, the relevant excerpts and the reason for suspicion. Require an “insufficient context” outcome. Do not ask the model to invent missing glossary entries.

Review variation before correcting it

A reviewer should decide whether the two occurrences describe the same referent, whether the approved form can fit the grammar and whether the variation is intentional. Case endings, gender agreement and articles may change a surface form without changing the concept. Character nicknames may deliberately differ from the inventory label.

Classify the outcome as confirmed inconsistency, valid grammatical variation, approved narrative exception or unresolved context. These labels preserve more information than a binary pass or fail. If the termbase itself is ambiguous, assign a terminology-policy task instead of forcing translators to guess which interpretation will be accepted.

Margin decision

Confirmed inconsistency: the quest and inventory reference the same object, the glossary applies in both contexts, and no narrative exception is documented. Update the quest string and verify the rendered result.

A correction may alter line length or agreement elsewhere in the sentence. Test the complete revised string, not just the replaced token. Search related quest steps and tutorial entries so a local fix does not leave a second naming system behind.

Colleagues tracing an item-name issue between a screen and handwritten notes in a bright workspace
A valid exception deserves a reason, not a permanent blanket suppression.

Maintain the trail after the fix

Record the original flag, reviewer reasoning, approved replacement, affected IDs and verification build. If a term changes intentionally, version the concept record and explain the migration. An exception should have a scope and an owner; otherwise it can become an invisible rule that future reviewers cannot reconstruct.

Evaluate your checker on a small human-reviewed set containing both genuine inconsistencies and acceptable variation. Count false positives as well as confirmed findings. A system that catches every spelling difference but burdens editors with hundreds of harmless inflections may be less useful than a narrower check with a clearly defined scope.

The goal is not perfect uniformity. It is dependable recognition: players can connect the words they read to the objects, actions and places they encounter.

Next: read beyond the string