wbcbugfree
commented
1 week ago Agrontology is now used to better express the relationships between concepts, but the problem then becomes whether to introduce more general relation ontologies, such as Relation Ontology. The advantage of doing this is that Agrontology is limited in its ability to express relationships; it only has 170 properties. But RO can express almost any relationship. However, because Agrontology has few properties, it is easier to find the most suitable ones when labeling text to RDF triples and subsequently using LLMs. RO has more properties; finding the most appropriate one is very time-consuming and conceivably not easy when using LLMs later on. So the current strategy is to use Agrontology wherever possible and to use RO only when inevitable, or to avoid it outright.
Currently, when using LLMs to convert text to RDF statements, only "skos:related" is used to describe the relationship between concepts. Doing so simplifies the task of LLMs but should provide stronger relationships in the final knowledge graph. SKOS provides more explicit relationships, such as "skos:broader" and "skos:narrower". We can always give concepts clearer relationships by post-processing, but it's worth trying to introduce more complex relationships when annotating the data. Therefore, these text-RDF pairs can be fed into LLMs as part of the prompt, thus providing LLMs with instructions for generating more complicated relationships.