Ontology extension by online clustering with large language model agents

Area of Science:

• Computational Linguistics
• Knowledge Representation
• Medical Informatics

Background:

• Ontologies are crucial for structured knowledge representation and shared understanding.
• Existing ontology extension methods often require extensive annotated data and domain expertise.
• Healthcare data, particularly patient-generated content, is often unstructured and vast.

Purpose of the Study:

• To propose a novel framework for automatic ontology extension from streaming data in a zero-shot manner.
• To leverage Large Language Models (LLMs) for symptom typing and taxonomy in medical ontologies.
• To enhance medical knowledge representation for breast and bladder cancer survivors.

Main Methods:

• Developed a zero-shot ontology extension framework utilizing online and hierarchical clustering.
• Employed LLMs for Symptom Typing (classification) and Symptom Taxonomy (integration) from patient forum data.
• Implemented a dual-phase model with multiple LLMs for accurate and seamless integration.

Main Results:

• Demonstrated effective integration of new medical knowledge into existing ontologies.
• Achieved real-time, structured categorization and integration of symptoms with minimal human oversight.
• Validated the framework's effectiveness through quantitative analyses and data visualizations.

Conclusions:

• The proposed framework successfully extends medical ontologies from unstructured streaming data.
• LLMs combined with clustering offer an efficient zero-shot approach for knowledge-based systems in healthcare.
• This method advances the development of dynamic and comprehensive medical knowledge graphs.