PidginAirAlert: Domain NLP for Community Soot and Health Alerts in Nigerian Pidgin

Communities in Port Harcourt and the wider Niger Delta often report black soot, gas-flare smoke, and related health effects in Nigerian Pidgin is the lingua franca of WhatsApp groups, radio call-ins, and neighbourhood alerts, yet most air-quality systems assume English text and calibrated sensor dashboards. We present PidginAirAlert, a low-resource pipeline that maps short Pidgin messages to an actionable six-way alert taxonomy and a four-level severity score. The design prioritizes informed consent, CPU-only inference, and an optional speech path that cascades automatic speech recognition into the same classifier. We contribute (i) a domain taxonomy grounded in local soot phenomenology, (ii) a participatory WhatsApp/voice-note protocol with a real-only evaluation rule, (iii) a reproducible classical NLP baseline with explicit compute, split, and latency reporting, and (iv) pilot experiments showing category accuracy 0.944 and severity accuracy 0.852 at ≈1.2 ms/message and ≈118 KB model size. We discuss why English-lexified Pidgin is still a low-resource problem, how language triage complements sparse monitors, and how the same schema can transfer to Yoruba, Igbo, and Hausa in future work