Hate speech detection

Hate speech detection is the moderation sub-task focused on contributions that target a group with hostility, defined by ethnicity, religion, political belief, sexual orientation, gender, disability. It is more specific than toxicity detection, and the legal stakes are higher because hate speech is criminally regulated in most EU jurisdictions (LCEN in France, NetzDG in Germany).

What makes it hard

Hate speech detection has more false positives than generic toxicity scoring because :

• Quoted hate : a journalist or commenter citing hate to denounce it gets flagged.
• Reclaimed slurs : terminology that is hate-speech in one context, in-group identity in another.
• Coded language : evolving euphemisms (“globalists”, numerical codes) require constant model updates.
• Cross-cultural variance : what reads as a casual political insult in France can be hate speech under German law.

Generic English-trained models miss most of this. Press-grade hate detection needs : multilingual training, regular dataset refresh on press-specific corpora, and a strong human queue for the edge cases.

Legal framework

• France (LCEN) : 24-hour takedown obligation on flagged hate content. Logora’s Article 14 statement-of-reasons covers the LCEN takedown notice.
• Germany (NetzDG) : 24-hour takedown for “manifestly illegal” hate content. Logora’s German moderation pipeline is calibrated to NetzDG thresholds.
• EU (DSA) : harmonised takedown obligations + transparency reports + statements of reasons. The umbrella framework.

How Logora handles hate speech

Logora’s hate-detection layer runs on top of the toxicity score, with category-specific scoring (ethnic, religious, political, sexual orientation, gender, disability). Thresholds calibrated per outlet, per language. Audit logs feed the DSA Article 24 transparency report.

See AI moderation for the broader pipeline.