Table 1.

In-corpus (IC) performance, measured by F₁-score, of the baseline (BL) BiLSTM-CRF compared to a BiLSTM-CRF with variational dropout (VD)

		BL		VD
Entity	Corpus	Average	σ	Average	σ
Chemicals	BC4CHEMD	88.46	0.61	88.71	0.76
	BC5CDR	92.82	0.80	93.08	0.82
	CRAFT	84.98	1.98	85.22	1.37
Disease	BC5CDR	84.49	0.33	85.10	0.56
	NCBI-disease	87.01	1.17	87.60	1.50
	Variome	85.75	2.83	85.69	3.81
Species	CRAFT	96.28	2.21	96.38	2.26
	Linnaeus	89.44	3.91	89.66	7.47
	S800	72.75	2.42	77.39	4.17
Genes/proteins	BC2GM	81.48	0.48	83.10 **	0.50
	CRAFT	84.46	6.08	86.09	5.19
	JNLPBA	80.92	2.50	81.95	2.62

Note: In the BL model, dropout is applied only to the character-enhanced word embeddings. In the VD model, dropout is additionally applied to the input, recurrent and output connections of all LSTM layers. IC performance is derived from 5-fold cross-validation, using exact matching criteria. Statistical significance is measured through a two-tailed t-test. Bold, best scores, σ, standard deviation.

^**Significantly different than the BL (P ≤ 0.01).