Table 4.

Performance comparison with other state-of-art methods.

	Models	F-Score of Each DDI Type				Overall
		Advice	Effect	Mechanism	Int	Precision	Recall	F-Score
	UTurku [12]	63.0	60.0	58.2	50.7	73.20	49.90	59.40
Feature-based	FBK irst [49]	69.20	62.80	67.90	54.70	64.60	65.60	65.10
method	Kim [14]	72.50	66.20	69.30	48.30	-	-	67.00
	Raihani [15]	77.40	69.60	73.60	52.40	73.70	68.70	71.10
	CNN [20]	77.72	69.32	70.23	46.37	75.70	64.66	69.75
	SCNN [21]	-	-	-	-	72.50	65.10	68.60
	MCCNN [22]	78.00	68.20	72.20	51.00	75.99	65.25	70.21
Neural	GRU [50]	-	-	-	-	73.67	70.79	72.20
network-based	CNN-GCNs [23]	81.62	71.03	73.83	45.83	73.31	71.81	72.55
method	SVM-LSTM [25]	71.50	72.00	73.80	54.90	75.30	63.70	69.00
	Joint-LSTMs [26]	79.41	67.57	76.32	43.07	73.41	69.66	71.48
	Hierarchical RNNs [27]	80.30	71.80	74.00	54.30	74.10	71.80	72.90
	PM-BLSTM [28]	81.60	71.28	74.42	48.57	75.80	70.38	72.99
Our method	RHCNN	80.54	73.49	78.25	58.90	77.30	73.75	75.48