JCBIE: a joint continual learning neural network for biomedical information extraction

BMC Bioinformatics

Table 3 Comparison between typical joint extraction and JCBIE, based on a multi-corpora learning paradigm

Method\Dataset		\(\mathrm {ADE_1}\)	ADE	DDI + ADE	ADE + DDI + CPR	Avg.
ExtendNER	SP	82.56	86.64	88.24	86.35	85.95
	ET	84.79	87.99	89.85	84.45	86.77
	RE	68.70	76.94	77.50	68.20	72.84
L &R	SP	82.56	89.07	90.04	90.58	88.06
	ET	84.79	90.02	93.25	89.64	89.43
	RE	68.70	81.12	79.12	70.01	74.74
Typical joint extraction	SP	82.56	88.77	92.41	89.35	88.27
	ET	84.79	89.13	92.09	88.05	88.52
	RE	68.70	79.53	78.32	70.16	74.18
JCBIE	SP	87.80	89.17	91.98	91.12	90.02
	ET	87.77	89.65	92.07	90.38	89.97
	RE	74.18	80.56	80.09	72.97	76.95

The bold means the best results
The results are measured by micro-F1. NB: Without a subscript specification, ADE is the combination of ADE\({_1}\) and ADE\({_2}\). When only \(ADE_1\) is employed, ExtendNER, L&R, and Typical Joint Extraction are equal, because they do not start to distill at the first step

ISSN: 1471-2105