Moderation & safety
Multilingual moderation
Content moderation pipeline that scores and routes contributions in multiple languages, natively, not via English translation, to preserve nuance, context and local-law alignment.
Multilingual moderation is content moderation that handles each language natively, French French, German German, Italian Italian, Brazilian Portuguese Brazilian Portuguese, instead of routing everything through English translation. For a cross-border European publisher (Mediahuis, Ringier, Funke, Bertelsmann), it is the only model that works.
Why translation-based moderation fails
Generic English-trained models with translation routing miss :
- Linguistic specificity : insults, sarcasm, irony all degrade through translation.
- Cultural reference : a political reference that is normal in French is hate-coded in German.
- Coded language : evolving euphemisms are language-bound; translation strips them.
- Latency : translation + scoring + retranslation doubles or triples decision time.
Languages Logora supports natively
French · German · Italian · Spanish · Portuguese (BR) · English · Polish (since April 2026). Other languages require ~2-3 weeks of dataset preparation and 10k-50k labelled comments for the moderation model, done on-demand for groups operating outside the default seven.
How Logora handles language detection
Auto-detection at submission, with explicit override for outlets running multilingual diasporas (Krone Austria runs DE-Austria-flavoured + DE-Bavarian + IT for the Südtirol diaspora). Each language has its own thresholds, its own banned-terms list, its own reclamation rules.
See AI moderation for the broader pipeline.