Hybrid neural model for Hausa text auto completion: addressing data scarcity for a low-resource language

The Hausa language, a major linguistic vehicle for over 70 million people in West Africa, remains critically underserved by contemporary Natural Language Processing (NLP) technologies, exacerbating the digital divide. This paper presents the design and evaluation of a dedicated neural text autocomplete system to address this gap. Confronting the fundamental challenge of data scarcity, we developed a hybrid corpus of 50,000 sentences, merging authentic Hausa text with algorithmically generated sentences created via a rule-based generator. For prediction, we implemented a stacked Long Short-Term Memory (LSTM) neural network, enhanced with a large language model (LLM) fallback mechanism for robustness. The system achieved a Top-1 accuracy of 92.4% and a Top-5 accuracy of 98.6% on a held-out test set, significantly outperforming traditional trigram (Top-3: 82.5%) and simple RNN (Top-3: 89.1%) baselines. A user study with 30 Hausa speakers confirmed its practical utility, demonstrating a 35% average increase in typing speed and a 91% acceptance rate for the primary suggestion. This work provides a reproducible framework for developing NLP tools in low-resource settings, introduces a novel hybrid dataset for Hausa which we publicly release, and delivers empirical evidence of tangible benefits for users. Our findings offer a scalable blueprint for enhancing digital inclusivity for underserved language communities.