Better synonyms for enriching biomedical search.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Biomedical literature searches often fail due to differing terminology for the same concepts.
• Distributional word embeddings and cosine similarity are commonly used to find related words.

Purpose of the Study:

• To assess and compare word embedding methods for detecting synonyms in biomedical text.
• To address the limitations of cosine similarity scores across different embedding models.

Main Methods:

• Created the BioSearchSyn set, a manually annotated synonym corpus.
• Evaluated word2vec, fastText, and GloVe embedding methods.
• Utilized pool adjacent violators (PAV) isotonic regression to convert cosine similarity to synonym probability.

Main Results:

• Identified the best performing embedding methods for biomedical synonym detection using the BioSearchSyn set.
• Demonstrated that cosine similarity scores have method-dependent interpretations.
• Enabled uniform interpretation of synonymy scores across different methods by converting them to probabilities.

Conclusions:

• Introduced the BioSearchSyn corpus to identify optimal embedding methods for biomedical synonymy.
• Developed PubTermVariants2.0, an automatically extracted synonym set, enhancing PubMed searches since 2019.