Toxic Speech Detection in Social Media

Challenge

Standard content moderation tools failed to detect conversational or context-specific toxic speech on Persian-language Twitter.

Solution

We built a custom toxic speech classification model fine-tuned on culturally-specific slang, metaphor, and social context. We also created a flagging interface for human moderators.

Outcome

Enabled 3x more accurate detection of nuanced threats and helped develop targeted reporting and intervention workflows.

Why It Matters

Toxic speech doesn’t look the same in every language — and neither should the AI that detects it.