Structured Prompt Interrogation and Recursive Extraction of Semantics (SPIRES): a method for populating knowledge...

Area of Science:

• Bioinformatics
• Computational Biology
• Knowledge Representation

Background:

• Manual curation of knowledge bases and ontologies is labor-intensive.
• Existing AI/NLP methods require substantial training data and struggle with complex schemas.

Purpose of the Study:

• To introduce Structured Prompt Interrogation and Recursive Extraction of Semantics (SPIRES), an AI-driven approach for knowledge extraction.
• To leverage Large Language Models (LLMs) for zero-shot learning in populating knowledge bases with complex, nested schemas.

Main Methods:

• SPIRES utilizes LLMs for zero-shot learning and flexible prompt-based querying.
• It recursively interrogates LLMs with user-defined schemas to extract information and match specified formats.
• The system integrates existing ontologies and vocabularies to ground extracted entities with unique identifiers.

Main Results:

• SPIRES demonstrates successful application across diverse domains, including recipes, cellular pathways, and disease treatments.
• Achieves accuracy comparable to mid-range Relation Extraction methods.
• Significantly enhances an LLM's ability to ground entities with unique identifiers.

Conclusions:

• SPIRES offers a flexible, customizable method for knowledge base assembly, requiring no new training data for novel tasks.
• It effectively leverages LLM capabilities to support and accelerate manual knowledge curation.
• The approach facilitates validation against external databases and ontologies.