Concept embedding to measure semantic relatedness for biomedical information ontologies

Area of Science:

• Biomedical Informatics
• Computational Biology
• Natural Language Processing

Background:

• Biomedical ontologies like the Unified Medical Language System (UMLS) are crucial for organizing biological concepts.
• Vector representations of concepts from UMLS definitions are commonly used to assess semantic relatedness.
• Existing methods face limitations due to restricted word coverage, impacting model performance.

Purpose of the Study:

• To develop an improved concept-embedding model for measuring semantic relatedness within biomedical ontologies.
• To address the limited word coverage issue of conventional relatedness measures.
• To enhance the performance and applicability of semantic relatedness models.

Main Methods:

• Utilized Wikipedia as an external knowledge base to obtain context texts for concepts not defined in UMLS.
• Derived concept vector representations from these external context texts.
• Defined semantic relatedness as the cosine similarity between the derived concept vectors.

Main Results:

• The proposed concept-embedding model demonstrated significantly higher coverage of biological concepts.
• Validated superior performance of the new method compared to conventional semantic relatedness measures.
• Successfully overcame the limitations associated with restricted word coverage in prior models.

Conclusions:

• The novel concept-embedding approach effectively expands the scope of semantic relatedness analysis in biomedical contexts.
• Leveraging external knowledge sources like Wikipedia enhances the robustness of concept representation.
• This method offers a more comprehensive and accurate way to measure relatedness between biological concepts.