Text Snippets to Corroborate Medical Relations: An Unsupervised Approach using a Knowledge Graph and Embeddings

Area of Science:

• Medical Informatics
• Natural Language Processing
• Knowledge Representation

Background:

• Medical knowledge graphs require continuous updates from literature to maintain search relevance.
• Dynamically verifying knowledge graph triples against medical text is crucial for trustworthy decision support.

Purpose of the Study:

• To develop and evaluate an unsupervised approach for mapping medical knowledge graph triples to corresponding text snippets.
• To assess the precision and recall of mapping semantic relations from a medical knowledge graph to literature.

Main Methods:

• Utilized a medical knowledge graph as the source of triples.
• Employed phrase embeddings and cosine similarity for unsupervised mapping.
• Incorporated key concept boosting to enhance candidate text snippet relevance.

Main Results:

• Achieved a precision of 61.4% and a recall of 86.3% in mapping triples to text snippets.
• Demonstrated the effectiveness of the unsupervised approach in identifying relevant medical literature.

Conclusions:

• The developed method accurately maps semantic relations from knowledge graphs to text.
• This approach lays the foundation for a future application to retrieve medical relations and supporting evidence from text.