Hate Speech and Disinformation Detection Model in the Political Discourse of Hausa-Language

The digital revolution in West Africa has positioned Hausa-language social media as a dominant arena for political discourse; however, coordinated campaigns of identity-based hate speech and strategic disinformation have increasingly weaponized this space, threatening social cohesion and electoral integrity. Despite the urgency, a critical research gap exists because state-of-the-art automated detection systems remain underdeveloped for low-resource languages like Hausa, which fails to capture its linguistic nuances, such as proverbial expressions and code-mixing. To address this, the study construct Hausa HarmBench, a novel expert-annotated dataset of political text, and conduct a rigorous evaluation by fine-tuning multilingual transformer models (mBERT, XLM-R, AfroXLMR) for classification. HausaHarmBench, a dataset for detecting hate speech and disinformation in Hausa political discourse, built from 1.2 million texts collected during the 2022–2023 Nigerian elections from Twitter (X), Facebook, and Hausa news platforms (BBC Hausa, DW Hausa, VOA Hausa). After preprocessing, 20,000 posts were annotated, with 15,000 samples forming the final dataset. Our key findings reveal a performance disparity: while models achieve strong results on explicit hate speech (F1 = 0.84), detecting nuanced disinformation remains challenging (F1 = 0.76), with systematic failures in detecting culturally grounded insults and context-dependent narratives. This study underscores that effective moderation requires tools built on linguistically and culturally representative data, providing both a practical resource for platforms and advancing a theoretical framework for equitable content moderation in the Global South, emphasizing the need to integrate computational methods with sociolinguistic context. The study recommends multimodal detection, culturally aware models using Hausa contextual knowledge, real-time monitoring tools, comparative cross-language studies, and human-AI moderation frameworks for effective content moderation.