Table 3.

The lenient F1-score of different machine learning models and rule-based postprocessing for the end-to-end task. The boldface represents the best performance on each type of relation

	SVM	SVM + postprocessing	CNN-RNN	CNN-RNN + postprocessing	JOINT	JOINT + postprocessing
Strength → Drug	0.9574	0.9646	0.9637	0.9720	0.9644	0.9644
Dosage → Drug	0.9218	0.9337	0.9231	0.9353	0.9245	0.9245
Duration → Drug	0.7395	0.7735	0.7400	0.7861	0.7366	0.7366
Frequency → Drug	0.9361	0.9522	0.9405	0.9582	0.9425	0.9425
Form → Drug	0.941	0.9510	0.9404	0.9516	0.9363	0.9363
Route → Drug	0.9228	0.9350	0.9299	0.9415	0.9287	0.9287
Reason → Drug	0.5626	0.5756	0.5722	0.5792	0.5637	0.5630
ADE → Drug	0.4734	0.4718	0.4749	0.4755	0.3821	0.3790
Overall	0.8750	0.8853	0.8792	0.8905	0.8775	0.8774

Abbreviations: ADE, adverse drug event; CNN, convolution neural network; RNN, recurrent neural network; SVM, support vector machine.