Ecommerce · European Union

The challenge

The group's automated moderation was tuned for German and English, and Polish, Czech, Ukrainian and Hungarian content passed through with a clearly higher error rate in both directions. The group had no reliable measure of the problem, because its sampling was also German-language.

What we did

We built a managed service combining human moderation and annotation in six languages from Katowice and remotely. Measurement first: we sampled existing decisions to establish the real per-language error rate, then retrained classifiers on the underserved languages. Moderator wellbeing was designed explicitly.

Results

• False blocks in Polish and Ukrainian 73 percent lower within three quarters
• Prohibited content passing moderation clearly lower in every language except German
• Appeals in the underserved languages more than halved
• The group gained a per-language quality measure it had never had

Why it worked

Measuring first, in every language, is what made everything after it work.