A new corpus tests abuse detection in three languages
Detectors built on Northern data miss online abuse of women in Levantine Arabic, African French and Swahili.
Picture made for ABUREAU.
Researchers have released a small dataset of online abuse aimed at women and gender minorities. The paper says most harmful-content datasets and models are built on languages from the Global North.
The team worked with civil society groups in Jordan, Francophone West Africa and Kenya to gather examples. The result is 3,022 annotated instances in Levantine Arabic, African French and Swahili, labelled with active learning.
Existing detectors scored poorly on it. The authors report that models often miss covert harms in all three languages, with higher error rates for Swahili, and that fine-tuning on the new data improves recall. The taxonomy, data and trained models are released under a CC BY 4.0 licence.
For people who own or use data, the point is simple. A few thousand carefully gathered local examples can show a gap that a much larger imported set hides. Data from places and languages that are thin in public collections is the data that tests whether a model works for everyone, and that gives it a clear use.
