UrbanFoodKG: human-in-the-loop and LLM-assisted knowledge graph engineering for integrated food-system exploration

Abstract

The increasing population of the United States and the decline in available farmland has resulted in a rise in urban agriculture. Urban food systems encompass all aspects of growing, managing, and distributing food grown in urban environments. The wide array of urban food systems data, often collected in heterogeneous surveys, reports, and studies, presents a challenge for those gathering the data to organize and understand it. A lack of understanding prevents effective decision-making, resulting in less-than-desirable urban agricultural outcomes. This report presents the UrbanFoodKG ontology, created as a part of the GRIP Resilient Urban Food Systems project, meant to assist in collating the information gathered and answering what domain-specific questions the experts have. The UrbanFoodKG knowledge graph takes domain-expert knowledge and the survey data they gathered, using the Knowledge Acquisition and Representation Methodology (KNARM) to produce an application based on the UrbanFoodKG ontology that allows users to reason over the urban agricultural data. The application currently supports 37 domain-expert competency questions. UrbanFoodKG provides an initial infrastructure for standardizing and connecting urban agriculture data while enabling future applications in policy analysis, food-system resilience, and natural-language question answering. The computational approaches employed during this process are geared towards understanding the Human-in-the-loop LLM-assisted Knowledge Graph Engineering better used in concert with the KNARM methodology. Although the current evaluation is limited by the available survey data, the framework can be extended as additional datasets, domain knowledge, and policy information become available.