Case studies
Multilingual annotation and content moderation for a European retail group
For a retail group whose moderation was tuned for German and English, measuring first and retraining classifiers cut errors in the underserved languages.
The challenge
The marketplace carried content in six languages and its automated moderation was tuned for German and English. Turkish, Arabic and Russian passed with materially higher error rates in both directions. The group's own quality sampling was also English-first, so it could not measure the problem.
What Corpshore did
We built a managed service combining human moderation and annotation across all six languages. The first deliverable was measurement: we sampled and labelled existing decisions per language to establish the real error rate. Then we ran moderation at volume and annotation to retrain the classifiers on the underserved languages, gated by quality.
Results
- False suppression in Turkish and Arabic reduced 71 percent within three quarters
- Prohibited content surviving review fell substantially across all non-German languages
- Seller appeals in the underserved languages dropped by more than half
- The group gained per-language quality measurement it had never had
Why it worked
The group knew it had a moderation problem and could not see its shape, because its measurement had the same blind spot as its models. Measuring first, in every language, is what made everything after it work.
Let us build your Türkiye team
Tell us your function, your scale and your language needs. We will come back to you within six hours.