Table 8.

Comparison with previous systems of CDR extraction

Method		System	P (%)	R (%)	F (%)
without KBs	Feature-based	Gu et al. [6]	62.00	55.10	58.30
	Neural network-based	Nguyen et al. [13]	57.00	68.60	62.30
		Le et al. [14]	58.02	76.20	65.88
		Verga et al. [15]	55.60	70.80	62.10
with KBs	Feature-based	Pons et al. [9]	73.10	67.60	70.20
		Peng et al. [10]	68.15	66.04	67.08
		♠Peng et al. [10]	71.07	72.61	71.83
	Neural network-based	Li et al. [16]	59.97	81.49	69.09
		Zhou et al. [17]	60.51	80.48	69.08
		Ours	69.65	72.98	71.28
		♠ Ours	72.12	68.67	70.35

The descriptions and analysis for Table 8 could be found in subsection “Comparison with previous works”. The marker ^♠ indicates that the system uses additional weakly labeled data for training. The highest F1-score of each subgroup is highlighted in bold