. 2019 Apr 26;2(2):261–271. doi: 10.1093/jamiaopen/ooz009

Table 4.

Performance of different NER models at different levels of entity nesting

Level	Model	P (%)	R (%)	F (%)
Innermost	CRF	77.19	68.78	72.74^a
	BiLSTM-CRF	73.93	73.38	73.56 ^a
	Layered BiLSTM-CRF	69.79	70.41	70.10
Outermost	CRF	73.63	66.41	69.83^a
	BiLSTM-CRF	75.61	67.35	71.24^a
	Layered BiLSTM-CRF	74.00	74.54	74.27
All	CRF	75.44	67.61	71.31^a
	BiLSTM-CRF	74.71	70.42	72.50^a
	Layered BiLSTM-CRF	77.02	75.45	76.23

Note: For each different level, the best precision (P), recall (R), and F-score (F) amongst the 3 models is shown in bold.

Abbreviations: NER: named entity recognition; CRF: conditional random field.

^{^a}

A significant difference between CRF and (flat) BiLSTM-CRF models at P < .05. Since the layered BiLSTM-CRF takes as input different entities than the baseline models (ie, all entities vs innermost or outermost entities), we did not apply significance testing between layered and flat models.