HausaHateData: A Manually Annotated Multi-Source Dataset for Hausa Hate Speech Detection with Benchmark Evaluation Using XLM-RoBERTa-Large, AfriBERTa and mBERT

Automated hate speech detection has advanced considerably for high-resource languages, but the availability of manually annotated resources for Hausa remains limited. This study presents HausaHateData, a multi-source corpus of 6,509 unique Hausa-language text instances collected from Twitter, Facebook, the AfriHate corpus, and other publicly available sources. Each instance was assigned a binary label of hate speech or non-hate speech. Trigger-word annotations were added to hate-speech instances where a relevant word or phrase could be identified, and all instances were assigned a positive or negative sentiment label. Two annotators independently labelled the data, producing an observed agreement of 82.99% and a Cohen’s kappa coefficient of approximately 0.65 before disagreements were resolved through discussion. The final corpus contains 2,565 hate-speech instances (39.41%) and 3,944 non-hate-speech instances (60.59%). To establish benchmark results, XLM-RoBERTa-Large, AfriBERTa and multilingual BERT (mBERT) were fine-tuned for binary classification. At the reported final epoch, mBERT achieved the highest accuracy (96.08%) and precision (98.73%), while AfriBERTa achieved the highest recall (94.15%), F1-score (94.89%) and ROC-AUC (0.977). XLM-RoBERTa-Large recorded an accuracy of 94.93% and an F1-score of 93.25%. These results provide initial benchmark evidence for Hausa hate-speech classification and demonstrate the potential value of HausaHateData for further research on low-resource African language processing. The experimental split and confusion-matrix values should be confirmed against the original prediction files before publication.