Table 4.

Comparison of the performance of cross-corpus evaluation for comparative methods using strict matching

Strict matching		train CDR → test NCBIa			train NCBI → test CDRb
Model		p	r	f	p	r	f
BiLSTM		57.32	37.92	45.64	55.19	30.79	39.52
BiLSTM-CRF		68.34	36.88	47.90	58.30	38.74	46.55
GRAM-CNN		59.74	42.81	49.88	58.48	33.21	42.36
BERT		68.92	53.13	60.00	54.17	61.44	57.57
CLSTM	word level	62.42	48.96	54.87	60.92	38.09	46.87
	character level (3)c	68.12	44.06	53.51	62.74	32.66	42.96
	character level (7)c	65.08	45.63	53.64	60.69	21.75	32.02
	word+char levels (3, 3)d	66.77	43.75	52.86	54.00	44.08	48.54
	word+char levels (5, 5)d	69.36	42.92	53.02	57.63	39.51	46.88

^aTest the disease entities in the NCBI corpus using the model trained on the CDR corpus

^bTest the disease entities in the CDR corpus using the model trained on the NCBI corpus

^cThe number in parentheses represents the window size at the character level.

^dThe numbers in parentheses represent the window sizes at the word and character level, respectively