tESA: a distributional measure for calculating semantic relatedness

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Text Analysis

Background:

• Semantic relatedness quantifies concept links, often approximated by word-based methods.
• Biomedical informatics relies on semantic relatedness for text and knowledge processing.
• Current methods often depend on specialized structured resources, limiting adaptability.

Purpose of the Study:

• To present tESA, an extension of the Explicit Semantic Relatedness (ESA) method.
• To address the challenge of utilizing unstructured life sciences domain knowledge.
• To improve semantic relatedness calculations for biomedical texts.

Main Methods:

• tESA utilizes two separate sets of vectors from different document sections (e.g., titles).
• This contrasts with the original ESA method's single vector space.
• The study evaluates tESA and domain-adapted ESA on standard biomedical semantic relatedness benchmarks.

Main Results:

• tESA achieves results comparable to or better than current state-of-the-art methods.
• The proposed method demonstrates strong applicability within the Life Sciences domain.
• Performance was evaluated against established benchmarks for biomedical semantic relatedness.

Conclusions:

• Combining semantics from different document sections enhances distributional semantic relatedness measures.
• Extending ESA with title vectors improves performance on large reference datasets.
• The study discusses the impact of the extension on distributional representation size.