Replacing non-biomedical concepts improves embedding of biomedical concepts.

Area of Science:

• Natural Language Processing
• Bioinformatics
• Machine Learning

Background:

• Embeddings represent words in vector spaces, crucial for machine learning.
• Biomedical embeddings traditionally standardize concept synonyms.
• The impact of non-biomedical synonym replacement on embedding quality is understudied.

Purpose of the Study:

• To investigate the effect of replacing non-biomedical synonyms on embedding quality.
• To reduce noise in embeddings by standardizing non-biomedical terms.
• To enhance the coherence of biomedical concept representations.

Main Methods:

• Leveraged WordNet to identify and replace non-biomedical synonyms.
• Applied a combined approach to 1,055 biomedical concept sets.
• Used Word2Vec on 10 million PubMed abstracts to generate embeddings.

Main Results:

• Non-biomedical synonym replacement reduced mean pairwise distance by 8%.
• This indicates improved intra-cluster coherence and higher embedding quality.
• The complementary approach enhances the integrity of biomedical concept representations.

Conclusions:

• Replacing non-biomedical synonyms is a valuable complementary technique for improving biomedical embeddings.
• This method offers a practical way to reduce noise and enhance embedding quality.
• Further research will explore applicability to other embedding techniques and downstream tasks.