New ALTO Dataset Tackles Tech-Facilitated Gender-Based Violence in African Languages

Researchers have introduced ALTO, a novel dataset specifically designed to detect tech-facilitated gender-based violence (TFGBV) in Levantine Arabic, African French, and Swahili. This initiative addresses a critical gap, as most existing models and datasets for identifying online harm are predominantly based on data from Global North languages, often failing to recognize the nuanced and culturally specific forms of TFGBV prevalent in other regions.

The project involved a collaborative, community-led approach, working with civil society organizations (CSOs) in Jordan, Francophone West Africa, and Kenya. This ensured that the crowdsourced data for ALTO was culturally and linguistically relevant, capturing the diverse manifestations of TFGBV. The dataset, though relatively small with 3022 instances, was developed using active learning for annotation, proving to be performant in improving detection capabilities.

Benchmarking and fine-tuning experiments revealed that current harmful content detection models perform poorly on TFGBV recall within the ALTO dataset, highlighting their inadequacy for these specific linguistic and cultural contexts. However, the study demonstrated significant improvement with fine-tuning. A notable finding was the consistent difficulty models had in identifying covert harms across all languages, with error rates being particularly high in Swahili, underscoring the complexities of detecting subtle forms of violence. The researchers have released the annotation taxonomy, the dataset itself, and the trained models to foster further development and application.

This work is crucial for African communities, as it directly contributes to making online spaces safer for women and gender minority groups by enabling more effective detection of online abuse in locally relevant languages. By providing a specialized tool and dataset, ALTO empowers local organizations and tech platforms to better combat TFGBV, which is a pervasive issue globally and within Africa.