Questions between teams
Translation review, clarified
These answers distinguish a useful automated suspicion from an assessed translation finding. For the complete method, start with the terminology trail.
Can AI determine whether a translation is correct?
AI can identify possible issues and compare supplied evidence, but its explanation is not independent verification. A qualified reviewer must assess meaning and context. Fluency, confidence and agreement between models do not establish correctness.
Is every glossary difference an error?
No. Inflection, character voice, descriptive references and approved exceptions can justify variation. First confirm that the same concept is involved. A glossary entry needs a definition and scope before it can support a reliable consistency judgment.
What context should accompany a string?
Include the string ID, source and target, speaker, audience, nearby dialogue, relevant game state, interface location and build. Mark missing evidence explicitly. A screenshot can help with layout, while a short capture may be needed to establish timing or interaction.
Can back-translation prove accuracy?
No. Back-translation adds another interpretation. Use it to generate questions, then assess the target text directly with a competent reviewer. A target-language adaptation can be accurate even when a literal back-translation looks unlike the source.
What should a small team review first?
Prioritize progression-critical instructions, consequential choices, high-frequency names and important interface actions. Record exclusions and sample some unflagged content. A narrow, well-documented review is more honest than an unsupported claim of complete coverage.
Can we upload confidential game scripts to an AI tool?
Only when the rights holder and applicable agreements permit the processing arrangement. Check retention, access, data location and training settings before sharing material. Removing the game’s name does not necessarily remove proprietary or personal information.
How should severity be assigned?
Assign severity according to evidenced player impact in context, using a calibrated project policy rather than model confidence alone. The same naming error may have different consequences in optional lore and a required quest objective. See the review severity matrix.
Does a lower flag count mean better translation?
Not necessarily. Input scope, check configuration, glossary and model changes can all alter flag counts. Compare assessed findings using consistent methods. State denominators, sampling limits and unresolved flags before drawing conclusions about quality.
Does this website analyze uploaded game files?
No. It is an independent informational website. The contact tool only prepares editable text locally for copying or downloading and does not send messages or run translation analysis. Please do not include confidential scripts or player data in an editorial inquiry.