Combining Symbolic and Generative AI to Explore Knowledge Base and Control Cabbage Pests in West-Africa

Presentation online: youtube.com International audience Suggesting agroecological solutions based on local resources for crop protection to farmers in the South is a challenge. To this end, the Knomana initiative gathered knowledge on plant-based solutions derived from local flora to replace synthetic pesticides and antimicrobial products. It relies on data collected in scientific literature which describes protection systems, mainly composed of a plant which controls a pest on a crop, and includes additional information such as the location, the used plant part, and the plant extract type. To facilitate dataset exploration and produce recommendations, knowledge is formally represented and then analyzed using a symbolic AI method (Formal Concept Analysis, i.e. FCA) and its relational extension (RCA). FCA and RCA enable the extraction of implication rules (the Duquenne-Guigues Basis of Implications) that reveal relevant dependencies and generalizations in the data. As a symbolic method, FCA and RCA produce robust and reliable results, but that are not directly intelligible by domain experts. This is where Generative AI (LLM) could help, by presenting the results in plain language that is expected to be understandable and provide an adequate level of detail for domain experts.As a use case, we focus on the protection of Brassicaceae in West Africa, a major crop of interest affected by many pests causing significant losses. In addition to data from Knomana, the dataset also indicates, for each crop and plant, whether it is used for human care, consumed, cultivated, or spontaneous.After extracting the rules using RCA, rules are reformulated using three LLMs (ChatGPT, Claude, and Gemini) and submitted to domain experts (entomologists). They evaluate to what extent this reformulation is complete, correct (i.e., free of misinterpretation), non-redundant, adds information coming from LLM general knowledge which is relevant, and provides useful insights. These results are the first step in developing AI-based recommendation systems for the protection of agricultural crops in the South.