Replacing non-biomedical concepts improves embedding of biomedical concepts

Area of Science:

• Computational Biology
• Natural Language Processing
• Bioinformatics

Background:

• Concept embeddings represent biomedical concepts as vectors, capturing semantic relationships.
• The quality of these embeddings is crucial for downstream applications in biomedical research.
• Synonymy can introduce noise and reduce the effectiveness of concept embeddings.

Purpose of the Study:

• To test if replacing non-biomedical synonyms improves biomedical concept embeddings.
• To evaluate the impact of synonym normalization on semantic similarity in vector space.

Main Methods:

• Developed an approach using WordNet to identify and replace synonym sets.
• Replaced synonyms with the most representative term within each synonym set.
• Applied the Word2Vec algorithm to generate concept embeddings before and after synonym replacement.

Main Results:

• Tested on 1055 concept sets, demonstrating an average 8% reduction in mean intra-cluster distance.
• The results suggest improved homophily (clustering) of related concepts in the vector space.
• Synonym replacement positively impacted the quality of biomedical concept embeddings.

Conclusions:

• Non-biomedical synonym replacement is a viable strategy to enhance biomedical concept embeddings.
• The proposed method, implemented in a Python package (wn2vec), offers practical utility.
• This pilot study indicates a promising direction for improving vector-based semantic representations in bioinformatics.