A Comparative Evaluation of Afribert Base and Afribert Large for Hausa Hate Speech Detection

Abstract

Hate speech detection is an important task in Natural Language Processing (NLP), particularly with the growth of online communication. Automatic detection remains challenging for low-resource African languages such as Hausa because annotated datasets and language-specific computational resources are limited. This study comparatively evaluated AfriBERT Base and AfriBERT Large (AfriBERTXL) for Hausa hate speech detection using 6,000 annotated Hausa text samples classified into hate speech and non-hate speech. Both configurations were trained for three epochs and evaluated using accuracy, precision, recall, F1-score and Receiver Operating Characteristic Area Under the Curve (ROC-AUC). On the final test set, AfriBERT Base achieved 79.32% accuracy, 77.98% precision, 66.23% recall, 71.63% F1-score and 87.03% ROC-AUC. AfriBERT Large achieved 80.45% accuracy, 79.94% precision, 67.27% recall, 73.06% F1-score and 88.41% ROC-AUC. Thus, AfriBERT Large recorded higher values across all five reported metrics, although the improvements were modest. The findings provide a direct comparison of the two AfriBERT configurations for Hausa hate speech classification. However, the results should be interpreted in light of the dataset size and the methodological details that require fuller reporting, including dataset sampling, annotation procedures and experimental configuration.