NaijaHate - AI Hate Speech Detection for Nigerian Social Media
NaijaHate is a World Bank SocialAI initiative developing context-specific AI algorithms for hate speech detection on Nigerian Twitter. Recognizing that AI models trained on Western datasets perform poorly in Global South contexts, the project created a dataset of 35,973 annotated Nigerian tweets and a pretrained language model (NaijaXLM-T) tailored to Nigerian English and Pidgin. Four Nigerian annotators representing Hausa, Yoruba, Igbo, and Fulani ethnic groups labeled content as hateful, offensive, or neutral. The research found traditional HSD models overestimate real-world performance by two-fold, highlighting the need for culturally-aware AI in content moderation.
Theme Areas
Behavior Goal
Reduce online hate speech and harmful content targeting ethnic, religious, and gender communities; enable platforms to moderate content at scale while respecting local cultural contexts
Target Audiences
Methods & Approaches
Implementers & Partners
- World Bank
- University of Oxford
- NYU
- MIT
Donors & Sponsors
- World Bank
Key Takeaways
- 1AI models trained on Western data fail to generalize to Global South contexts
- 2Culturally-aware annotation by local experts significantly improves model accuracy
- 3Hate speech targeting ethnic and religious groups requires nuanced understanding
- 4Reviewing top 1% of flagged tweets could moderate 60% of hateful content
- 5Open-source datasets and models enable replication in other countries