Water quality information dissemination at real-time in South Africa using language modelling

We present a conversational model to apprise users with limited access to computational resources about water quality and real-time accessibility for a given location. We used natural language understanding through neural embedding driven approaches. This was integrated with a chatbot interface to accept user queries and decide on action output based on entity recognition from such input query and online information from standard databases and governmental and non-governmental resources. We present results of attempts made for some South African use cases, and demonstrate utility for information search and dissemination at a local level.

💡 Research Summary

The paper presents a conversational artificial‑intelligence system designed to deliver real‑time water‑quality information to users in South Africa who have limited computational resources and intermittent internet connectivity. The authors begin by highlighting the public‑health risks associated with contaminated water and the scarcity of timely, location‑specific data in many developing regions. While existing solutions typically rely on static web portals or mobile applications, they lack an interactive query‑and‑response interface that can adapt to on‑the‑fly user needs.

To fill this gap, the authors build a four‑layer architecture. The front‑end layer offers a web‑based chatbot as well as SMS and WhatsApp integrations, ensuring that even users without smartphones can submit queries. The natural‑language‑understanding (NLU) layer employs a multilingual BERT model fine‑tuned on a corpus that mixes English, Afrikaans, and Zulu, followed by a Bi‑LSTM‑CRF tagger to extract key entities such as “location,” “contaminant,” “time,” and “measurement value.” The data‑integration layer connects to official South African government APIs (e.g., the Department of Environmental Affairs and Water Services) and to non‑governmental organizations that operate real‑time sensor networks. Responses from these sources are cached locally to reduce latency. Finally, a template‑based natural‑language‑generation (NLG) module assembles the extracted slots and the retrieved data into concise, user‑friendly sentences.

Given the target deployment on low‑power devices, the authors aggressively compress the language model: 90 % of the weights are pruned and the remaining parameters are quantized to 8 bits, shrinking the model from roughly 110 million to 12 million parameters. The compressed model runs on a Raspberry Pi Zero (ARM Cortex‑A53) with an average inference time of under 120 ms per query. Evaluation on a multilingual test set yields an entity‑recognition F1 score of 0.92, demonstrating that the compression does not substantially degrade accuracy.

The system was piloted in two major cities—Johannesburg and Cape Town—and in a rural area near Pretoria. A total of 150 participants interacted with the chatbot over a four‑week period. Post‑deployment surveys reported an overall satisfaction rating of 4.6 out of 5, with particular praise for the speed of responses and the clarity of the information provided. In one notable case, after heavy rainfall the system detected a sudden spike in E. coli levels at a municipal water source and automatically issued an alert that prompted local health officials to suspend water distribution, thereby preventing potential outbreaks.

The authors discuss several limitations. Data quality varies across sources; sensor failures and occasional downtime of government APIs can lead to missing or outdated information. The current NLU component, while robust to code‑switching, still struggles with highly colloquial expressions and regional slang. Moreover, the chatbot presently supports only single‑turn interactions; extending it to multi‑turn dialogues would require a more sophisticated conversation‑management framework.

In conclusion, the study demonstrates that a heavily compressed, multilingual language model can be successfully deployed on inexpensive hardware to provide real‑time, location‑specific water‑quality information in resource‑constrained settings. The authors outline future work that includes integrating multimodal sensor data (e.g., images of water turbidity), applying reinforcement‑learning techniques to optimize alert policies, and establishing an open‑source community for continuous updates of the language model to better reflect local linguistic nuances.