CoRTEx: contrastive learning for representing terms via explanations with applications on constructing biomedical...

Area of Science:

• Biomedical Informatics
• Artificial Intelligence
• Natural Language Processing

Background:

• Biomedical Knowledge Graphs are essential for research.
• Term clustering, identifying synonyms, is key to building these graphs.
• Existing contrastive learning models struggle with difficult terms and lack generalizability beyond the Unified Medical Language System (UMLS).

Purpose of the Study:

• To enhance term representation and improve term clustering for biomedical knowledge graphs.
• To leverage world knowledge from large language models (LLMs) to overcome limitations of previous models.
• To introduce a novel method, Contrastive Learning for Representing Terms via Explanations (CoRTEx).

Main Methods:

• Generated explanations for UMLS terms using ChatGPT.
• Employed contrastive learning with simultaneous term and explanation embeddings.
• Introduced hard negative samples and a ChatGPT-assisted BIRCH algorithm for efficient clustering.

Main Results:

• CoRTEx achieved the highest F1 score on established clustering and hard negative test sets.
• Clustered over 35 million terms from the Biomedical Informatics Ontology System (BIOS) into over 22 million clusters.
• Demonstrated efficacy in handling challenging samples through case studies.

Conclusions:

• CoRTEx shows superior accuracy and robustness compared to benchmark models.
• Aligning terms with their explanations enhances clustering performance.
• The method is suitable for large-scale biomedical ontology term clustering.