CSpace: a concept embedding space for biomedical applications.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Transformer models have advanced natural language analysis but are computationally intensive.
• General-purpose models possess complex architectures with billions of parameters, often exceeding requirements.
• There is a need for efficient and specialized models in biomedical text analysis.

Purpose of the Study:

• To introduce CSpace, a novel, concise word embedding for biomedical concepts.
• To evaluate CSpace's performance against existing methods in semantic similarity tasks.
• To demonstrate CSpace's utility for semantic search and biomedical knowledge discovery.

Main Methods:

• Developed CSpace, a concise word embedding model for biomedical concepts.
• Evaluated CSpace on out-of-vocabulary ratio and semantic textual similarity tasks.
• Compared CSpace's sentence similarity performance with transformer-based models.

Main Results:

• CSpace achieved superior performance in out-of-vocabulary ratio and semantic textual similarity.
• CSpace demonstrated comparable performance to transformer models in sentence similarity tasks.
• Ontological identifiers (MeSH, NCBI) integrated into CSpace facilitate efficient relatedness measurement.

Conclusions:

• CSpace offers an efficient and effective approach to biomedical concept representation.
• CSpace can serve as a foundation for enhanced semantic search in biomedical literature.
• CSpace's ability to measure relatedness may uncover novel associations between diseases, genes, and conditions.