Ontology enrichment using a large language model: Applying lexical, semantic, and knowledge network-based similarity...

Area of Science:

• Biomedical Informatics
• Knowledge Representation
• Ontology Engineering

Background:

• Ontologies are crucial for domain knowledge representation and require comprehensive domain views for utility.
• Ontology enrichment is necessary to incorporate new concepts and advancements in the field.
• The Social Determinants of Health (SDoH) domain requires continuous updates to its knowledge representation.

Purpose of the Study:

• To develop an automatic pipeline for ontology enrichment using a seed ontology, a Large Language Model (LLM), and a text corpus.
• To apply and demonstrate the effectiveness of this pipeline for extending the SDoH Ontology (SOHOv1).
• To establish a generalizable methodology for ontology enrichment in other domains.

Main Methods:

• Retrieved PubMed abstracts using existing SOHOv1 concepts as search terms.
• Employed GPT-4-1201 to extract semantic triples from abstracts, followed by filtering using lexical, semantic, and knowledge network-based approaches.
• Compared the granularity of extracted triples with SemMedDB and validated results using human experts and ontology tools.

Main Results:

• Expanded SOHOv1 (173 concepts, 585 axioms) to SOHOv2 (572 concepts, 1,542 axioms), significantly increasing concept and axiom counts.
• The LLM-based method identified a greater number of concepts compared to those extracted from SemMedDB.
• Demonstrated the feasibility and effectiveness of the automated enrichment pipeline for the SDoH domain.

Conclusions:

• Successfully extracted semantic triples from PubMed abstracts using GPT-4-1201 via prompt chaining.
• Demonstrated the superiority of GPT-4-1201 extracted triples over SemMedDB for SDoH ontology enrichment.
• Utilized advanced search techniques for concept identification and confirmed concept quality through expert evaluation, validating the generalizability of the methodology.