🤖 AI Summary
This work addresses offensive content detection in Hausa—a low-resource language—by constructing the first manually annotated Hausa offensive terminology dataset. Grounded in user surveys and empirical analysis, the study focuses on high-risk domains such as religion and politics to develop a domain-adapted detection system. Methodologically, it integrates supervised learning (XGBoost and fine-tuned multilingual BERT) with multilingual baselines, including direct translation via Google Translate. Key contributions are threefold: (1) release of the first open-source Hausa offensive dataset; (2) empirical demonstration that cultural context critically impacts detection performance, rendering literal translation ineffective; and (3) proposal of a localized, multi-stakeholder governance framework. Experiments show the proposed models achieve >70% accuracy—significantly outperforming translation-based baselines—and reveal pronounced concentration of offensive content in religious and political discourse.
📝 Abstract
Hausa, a major Chadic language spoken by over 100 million people mostly in West Africa is considered a low-resource language from a computational linguistic perspective. This classification indicates a scarcity of linguistic resources and tools necessary for handling various natural language processing (NLP) tasks, including the detection of offensive content. To address this gap, we conducted two set of studies (1) a user study (n=101) to explore cyberbullying in Hausa and (2) an empirical study that led to the creation of the first dataset of offensive terms in the Hausa language. We developed detection systems trained on this dataset and compared their performance against relevant multilingual models, including Google Translate. Our detection system successfully identified over 70% of offensive, whereas baseline models frequently mistranslated such terms. We attribute this discrepancy to the nuanced nature of the Hausa language and the reliance of baseline models on direct or literal translation due to limited data to build purposive detection systems. These findings highlight the importance of incorporating cultural context and linguistic nuances when developing NLP models for low-resource languages such as Hausa. A post hoc analysis further revealed that offensive language is particularly prevalent in discussions related to religion and politics. To foster a safer online environment, we recommend involving diverse stakeholders with expertise in local contexts and demographics. Their insights will be crucial in developing more accurate detection systems and targeted moderation strategies that align with cultural sensitivities.
