Table 1.

Description of word embedding sources used.

Name	Author and source	Training source	Unit	Context window	Preprocess (reduce case, remove stop words, term types)	Embedding technology (gensim, FastText, GloVea, BERTb, ELMO, etc)	Returned unit (1-3 ngrams)	Embedding size	Vocab size
BioNLPc Lab PubMed + PMCd W2V	Pyysalo et al 2013 [14,33]	PubMed/ PMC articles	Token	5	Mixed case, no stop words, skip-grams	word2Vec	1 ngram	200	~4 billion tokens
BioNLP LabWiki + PubMed + PMC W2V	Pyysalo et al 2013 [14,33]	Wikipedia, PubMed/ PMC articles	Token	5	Mixed case, no stop words, skip-grams	word2Vec	1 ngram	200	~5.4 billion tokens
BioASQ	Tsatsaronis et al 2015 [13,34]	PubMed abstracts	Token	5	Lowercase, no stop words, continuous bag of words	word2Vec	1 ngram	200	~1.7 billion tokens
Clinical Embeddings W2V300	Flamholz et al 2019 [16,35]	PubMed/ PMC/ MIMIC IIIe	Token	7	Lowercase, include stop words, skip-grams	word2Vec	1-3 ngrams	300	~300k tokens
BioWordVec Extrinsic	Zhang et al 2019 [15,36]	PubMed + MeSHf	Character	5	lowercase, include stop words	FastText	1-3 ngrams	200	~2.3 billion tokens
BioWordVec Intrinsic	Zhang et al 2019 [15,36]	PubMed + MeSH	Character	20	Lowercase, include stop words	FastText	1-3 ngrams	200	~2.3 million tokens
Standard GloVe Embeddings	Pennington et al 2014 [11]	Common Crawl	Token	10	Mixed case	GloVe	1 ngram	300	~2.1 billion tokens

^aGloVe: global vectors for word representation.

^bBERT: Bidirectional Encoder Representations from Transformers.

^cBioNLP: biomedical natural language processing.

^dPMC: PubMed Central.

^eMIMIC III: Medical Information Mart for Intensive Care III.

^fMeSH: Medical Subject Headings.