CONCEPT Cited by 1 source
Translation Edit Rate¶
Translation Edit Rate (TER) measures how much a reviewer changes a machine-generated translation before approval. In an AI-assisted localization workflow, it is an operational proxy for draft usefulness: a lower rate often means the reviewer is polishing a usable draft rather than rewriting it. It is not a semantic-quality score and cannot replace qualified human approval. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
How to interpret it¶
Atlassian reports early-production TER from about 10% to just under 50%, depending on language. The spread makes language and locale a first-class reporting dimension. A fleet average can hide a locale where terminology, morphology, or source context remains poor.
| TER observation | Plausible interpretation | Follow-up |
|---|---|---|
| Low and stable | drafts are frequently close to approved form | preserve human approval; inspect for reviewer under-correction |
| High in one locale | weak language-specific context, terminology, or model behavior | audit context and glossary by locale |
| Rising after a product change | source-text or style-policy regression | inspect changed messages and validation misses |
| Falling over time | corrections may be improving the draft system | verify with independent quality review, not TER alone |
Why it is not a release gate¶
A low edit distance can preserve a plausible but wrong translation, insensitive tone, or culturally inappropriate wording. Conversely, a correct translation may need a large stylistic rewrite. Atlassian therefore keeps professional translators as the final sign-off for every string and uses TER as a quality-and-cost signal only. (Source: sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development)
Seen in¶
- sources/2026-08-06-atlassian-scaling-localization-at-atlassian-keeping-translation-at-the-pace-of-ai-era-development — early production range of ~10% to <50%, with reviewer edits feeding future improvements.