Recognizing Continuous and Discontinuous Adverse Drug Reaction Mentions from Social Media Using LSTM-CRF

Table 3

Performances of different methods on continuous and discontinuous ADR mentions, respectively (using strict criterion).

Method	Representation	Continuous ADR mention			Discontinuous ADR mention
		CRF	BIOHD	0.6882	0.5950	0.6382	0.3093	0.2495	0.2762
Multilabel	0.6896		0.5993	0.6413	0.5217	0.2405	0.3293
LSTM-CRF (baseline)	BIOHD	0.6720	0.7002	0.6858	0.3579	0.3563	0.3571
	Multilabel	0.6703	0.7062	0.6877	0.4825	0.3533	0.4079
LSTM-CRF + subject	BIOHD	0.6800	0.7092	0.6943	0.3573	0.3613	0.3593
	Multilabel	0.6882	0.7109	0.6994	0.4400	0.3263	0.3747
LSTM-CRF + knowledge	BIOHD	0.6698	0.7165	0.6924	0.3417	0.3802	0.3600
	Multilabel	0.6753	0.7062	0.6904	0.5096	0.3553	0.4187
LSTM-CRF + all	BIOHD	0.6813	0.7173	0.6988	0.3406	0.3513	0.3459
	Multilabel	0.6718	0.7154	0.6929	0.4841	0.3623	0.4144