Scientific knowledge graph and ontology generation using open large language models

Abstract

Knowledge graphs (KGs) are powerful tools for structured information modeling, increasingly recognized for their potential to enhance the factuality and reasoning capabilities of Large Language Models (LLMs). However, in scientific domains, KG representation is often constrained by the absence of ontologies capable of modeling complex hierarchies and relationships inherent in the data. Moreover, the manual curation of KGs and ontologies from scientific literature remains a time-intensive task typically performed by domain experts. This work proposes a novel method leveraging LLMs for zero-shot, end-to-end ontology, and KG generation from scientific literature; implemented exclusively using open-source LLMs. We evaluate our approach by assessing its ability to reconstruct an existing KG and ontology of chemical elements and functional groups. Furthermore, we apply the method to the emerging field of Single Atom Catalysts (SACs), where information is scarce and unstructured. Our results demonstrate the effectiveness of our approach in automatically generating structured knowledge representations from complex scientific literature in areas where manual curation is challenging or time-consuming. The generated ontologies and KGs provide a foundation for improved information retrieval and reasoning in specialized fields, opening new avenues for LLM-assisted scientific research and knowledge management.