ABioNER: A BERT-Based Model for Arabic Biomedical Named-Entity Recognition

Table 1

An overview of methods used for biomedical named-entity recognition.

Paper	Language	NE	Method	Results
[8]	English	Gene protein	SVM	Best balanced F1-score = 0.79
[9]	English	Chemical mentions	Hybrid (CRF + dictionary)	F1-score = 68.1
[10]	English	Problem treatment test protein DNA RNA cell type cell line	Unsupervised learning	Overall performance (exact micro-F) Pittsburgh dataset: 53.1; GENIA dataset: 39.5
[11]	English	Disease	Hybrid (stacked ensemble + fuzzy matching)	F1-score = 89.12%
[12]	English	Disease	Multiple label convolutional neural networks	F1-score NCBI corpus: 85.17%; CDR corpus: 87.83%
[13]	English	Document-level chemical NER	Hybrid (attention-based BiLSTM-CRF)	F1-score HEMDNER corpus: 91.14%; CDR corpus: 92.57%
[14]	English	Genes diseases protein DNA RNA cell type cell line	n-Gram character and word embeddings via convolutional neural network	F1-score: NCBI dataset: 87.26%; Biocreative II dataset: 87.26%; JNLPBA dataset: 72.57%
[15]	English	Chemicals	Transfer learning	F1-score:
		Diseases		88.21
		Species		82.09
		Gene		87.01
		Protein		83.09
[16]	English	Diseases	Bidirectional encoder representations from transformers	Best F1-score; 89.71; 75.31
		Species
[17]	Spanish/Swedish	Spanish: disease/drug Swedish: body part/disorder/finding	Bidirectional long short-term memory network	Avg. F1-score: Spanish: 75.25; Swedish: 76.04
[2]	Arabic	Disease diagnosis symptoms treatment methods	Bayesian belief network (BBN)	Avg. F1-score: 71.05%