Transformer Model Based on Cross-Language Migration for Low-Resource Hausa News Classification

Hausa, a Chadic language spoken by over 80 million people in West Africa, remains critically under-resourced in natural language processing (NLP) due to the scarcity of annotated datasets and limited technical infrastructure. This study tackles the challenge of news classification for low-resource Hausa by innovatively combining cross-lingual transfer learning with Transformer models. We introduce a hierarchical fine-tuning strategy for XLM-RoBERTa, selectively updating its upper layers to preserve cross-lingual knowledge while adapting to Hausa-specific syntax and semantics. To address data scarcity, we curated a labeled Hausa news corpus comprising 2,917 articles across five categories (Nigeria, Africa, World, Health, and Politics) and enhanced it through back-translation and synthetic data generation using mT5. Experimental results demonstrate that our approach achieves state-of-the-art performance, with 84.1% accuracy and 83.4% macro-F1, outperforming baseline models such as TF-IDF+SVM (62.5%) and fully fine-tuned mBERT (71.6). Key contributions include a novel adaptation framework and an available Hausa news dataset. The findings underscore the effectiveness of partial model adaptation and data augmentation for low-resource languages. Future research should prioritize unsupervised pre-training and community-driven resource sharing to further democratize NLP for underrepresented languages.