Surveillance system in Africa using artificial intelligence methods

Monitoring crises in agriculture and health is crucial for early response and implementing control measures. While traditional monitoring systems are vital, there is increasing reliance on informal and local data sources, such as online news and radio, for daily insights. Advances in deep learning and large language models (LLMs) have significantly improved the processing of such textual data. However, these models predominantly support widely spoken languages like English and French, and their performance is limited by the availability of training data. This is especially problematic for the world's many under-resourced languages, which often lack sufficient written or digital resources, resulting in diminished performance and overfitting in language models. Audio data faces even greater challenges, despite the widespread use of radio for information dissemination. In this context, the SurvAAI project addresses the following question: how to unlock the performance of monitoring systems by integrating local and under-resourced languages? The SurvAAI project aims to improve monitoring systems by integrating under-resourced languages, focusing on Africa's rich linguistic diversity. Africa has over 2,000 languages, with 542 of them underrepresented in terms of available data. SurvAAI will enhance resources and NLP processes for three West African languages—Wolof, Fulfuldé, and Ewondo—chosen for their local relevance, data scarcity, and linguistic expertise within the project. By addressing these challenges, SurvAAI seeks to improve early warning systems and make its approach transferable to other low-resource languages. This project aims to address a gap not filled by tech giants like OpenAI or Microsoft, whose large, generic models are expensive and inadequate for agriculture and health monitoring. These models struggle with domain-specific tasks, particularly detecting emerging trends like new diseases that aren't present in historical data. SurvAAI seeks to create specialized models that are robust in new spatio-temporal contexts by overcoming data limitations and minimizing overfitting. SurvAAI focuses on three key challenges: 1. Media-Driven Data Drought (C1): This involves collecting multimodal resources (text, audio, video) for low-resource languages (Wolof, Fulfuldé, Ewondo) and building annotated datasets to evaluate model performance and generalizability. 2. New Learning Breakthroughs (C2): The project will develop methods that combine existing data and models with innovative approaches to optimize performance, interpretability, and adaptability. Strategies like model fine-tuning, data augmentation, and transfer learning will be used to improve models for low-resource languages. Techniques from automatic speech processing will be employed to extract relevant audio features, linking them with textual data. 3. Real-World Impact Efficiency (C3): The project will integrate its developments into two monitoring platforms—epidemiological surveillance and food security. Key NLP tasks include event detection, text classification, and information extraction. These tools will help identify and track crises by processing large volumes of textual data, detecting relevant events, and extracting essential information for decision-making. Ultimately, SurvAAI aims to create efficient, portable models tailored to low-resource languages, enhancing the effectiveness of crisis monitoring systems.