New dataset improves detection of gender based violence in African and Levantine languages

Researchers developed ALTO, a dataset of 3022 instances of tech-facilitated gender-based violence to improve detection in Levantine Arabic, African French, and Swahili.

The project involved collaboration with civil society organizations in Jordan, Kenya, and Francophone West Africa. The team used a community-led approach to gather culturally and linguistically relevant examples of online harm. This effort addresses a gap where existing models primarily rely on data from Global North languages, often failing to recognize local nuances of gendered violence.

Testing showed that current harmful content detection models have poor recall for these specific types of violence. The researchers found that these models frequently miss covert harms across all three languages, with error rates being notably higher in Swahili. The team released their annotation taxonomy, the ALTO dataset, and trained models to help improve safety tools in these regions.