Combination of Deep Recurrent Neural Networks and Conditional Random Fields for Extracting Adverse Drug Reactions from User Reviews

Table 1

Results of the proposed models and baseline methods.

Method	Exact			Partial
	P	R	F	P	R	F
Baseline CRF	0.6254	0.5972	0.6110	0.8145	0.7539	0.7521
Feature-rich CRF	0.6726	0.6532	0.6628	0.8303	0.7646	0.7622
1-layer LSTM	0.5798	0.6587	0.6167	0.8121	0.8065	0.7809
2-layer LSTM	0.6362	0.7044	0.6686	0.8090	0.8495	0.8005
3-layer LSTM	0.6588	0.7022	0.6798	0.8247	0.8323	0.7997
4-layer LSTM	0.6689	0.7093	0.6885	0.8255	0.8280	0.8000
1-layer GRU	0.5862	0.6772	0.6284	0.7995	0.8368	0.7900
2-layer GRU	0.6384	0.7093	0.6720	0.8165	0.8338	0.8002
3-layer GRU	0.6675	0.7191	0.6923	0.8151	0.8373	0.8009
4-layer GRU	0.6565	0.7262	0.6896	0.8006	0.8665	0.8033
2-layer LSTM + CRF	0.6947	0.6973	0.6960	0.8191	0.8161	0.7872
2-layer LSTM + CNN + CRF	0.6809	0.7039	0.6922	0.8083	0.8488	0.7978
3-layer LSTM + CNN + CRF	0.6868	0.7066	0.6965	0.8270	0.8488	0.8115
3-layer GRU + CNN + CRF	0.7048	0.7082	0.7065	0.8219	0.8311	0.7978