Text Analytics Forum

Future Knowledge Graphs Will Be a Collaboration Between LLMs & Humans

Jans Aasman is very much interested in how to build structured, reliable knowledge graphs from the ocean of unstructured text that is out there in the world. His thesis is that you need a collaboration between LLMs like GPT3/4 and very smart people that know a) the structure they want in their knowledge graph, b) how to do the prompting to get structured data out of LLMs, and c) most importantly, how to validate the output of LLMs. As an exercise, Aasman recently built a wine knowledge graph using GPT3/4. The LLM created the ontology and taxonomy, and it was used to fill the knowledge graph. Because you never can trust the output of GPT4 entirely, a validation approach was used that also employs a combination of GPT4 and web search. We were surprised at how well it worked. In the process, tools were created to make it easier to build knowledge graphs. Aasman shows demos in which SPARQL and some new magic predicates were used to orchestrate the building and validation of new knowledge.

Text Analytics Forum 2023

Speaker:
Jans Aasman, CEO, Franz Inc.