Table 5.

Machine learning-based NER methods

Publication	Dataset	Dataset size	Method	Features	P	R	F
Tang et al. [132]	i2b2 2010	349 train, 477 test	SSVM	Word + context + sentence + section + cTAKES + MetaMap + ConText + Brown clustering	87.38%	84.31%	85.82%
Wu et al. [148]	i2b2 2010	349 train, 477 test	LSTM	Word embedding	85.33%	86.56%	85.94%
Lee et al. [74]	i2b2 2010	170 train, 256 test	BERT model	Pre-trained and fine-tuned BioBERT	-	-	86.46%
Zhou et al. [163]	i2b2 2010	349 train, 477 test	LSTM-CRF	Pre-trained contextualized embeddings + Glove embedding	-	-	87.45%
Zhou et al. [163]	NCBI-disease 2014	593 train, 100 valid, 100 test	LSTM-CRF	Pre-trained contextualized embeddings + Glove embedding	-	-	87.88%
Lee et al. [74]	NCBI-disease 2014	593 train, 100 valid, 100 test	BERT model	Pre-trained and fine-tuned BioBERT	-	-	89.36%
Yang et al. [153]	n2c2 2019 for family history extraction	99 train, 117 test	majority voting of LSTM-CRF models with BERT fine-tuning	Fasttext embedding + pre-trained BERT	79.69%	79.20%	79.44%
Uzuner et al. [139]	SemEval 2014 Task 7A	199 train, 99 valid, 133 test	CRF models	Textual features enhanced with a rule-based system	91.1%	85.6%	88.3%
Vunikili et al. [140]	CANTEMIST-NER sub-task with tumor morphology mentions	501 train, 500 valid, 5232 test	Transfer learning to fine-tune the BERT model	BERT contextual embeddings pre-trained on general domain Spanish text	72.7%	74.1%	73.4%
Deng et al. [33]	crawled TCM patents’ abstract texts annotated with herb names, disease names, symptoms, and therapeutic effects.	1600 copies: 60% train, 20% valid, 20% test characters	BiLSTM-CRF	Pre-trained and fine-tuned character embedding	94.63%	94.47%	94.48%
Zhou et al. [163]	MACCROBAT 2018 case reports	160 train, 20 valid, 20 test	LSTM-CRF	Pre-trained contextualized embeddings + Glove embedding	-	-	65.75%
Lee et al. [74]	MACCROBAT 2018 case reports	160 train, 20 valid, 20 test	BERT model	Pre-trained and fine-tuned BioBERT	-	-	64.38%