Knowledge Graph Construction from LLMs

Knowledge Graph Construction from LLMs

Speaker: Édouard Albert Roulhac

Topics: KG, LLM

Abstract

Large Language Models (LLMs) store a lot of knowledge in their weights, but this knowledge is unstructured and challenging to access systematically. This work explores how LLMs can be used to construct structured knowledge databases without external resources. Unlike the conventional knowledge graph (KG) completion tasks that rely on a pre-existing KG, the LM-KBC challenge only allows to use the LLM’s internal knowledge. We explore different strategies, first focusing on building representations for relations and then enabling LLMs to materialize their knowledge on entities before responding.