Encouraging results for Knowledge Graph Extraction by LLM Ontology-prompting
A demonstration of the use of an LLM as an Extract-Load-Transform (ELT) engine from unstructured text to a Knowledge Graph by prompting with a non-trivial ontology
In a previous article, it was shown that an LLM can be prompted with unstructured text and then asked (politely of course) to extract a graph corresponding to a specific ontology/schema. This suggests an effective way to ETL/ELT unstructured data into the, preferred, structured form of a graph.
However, in the original article, the ontology used IRIs that were syntactically close to the structured text (has child <-> :child, etc). Also, the schema was confined to simple node-edge-node relationships. Sceptics could say that was not a realistic ontology/schema.
In this follow-up article, it is shown that the ontology can use syntactically unrelated, anonymized, IRIs along with descriptions, and still achieve the same quality of result (:child -> :op2). Additionally, the LLM can deduce reified/materialized node-edge-node relationships when called for by the schema. ( :a :marriedTo :b -> :abMarriage :spouse :a, :b)
This demonstrates the use of an LLM as an Extract-Load-Transform (ELT) engine from unstructured text to a knowledge graph.
