. Author manuscript; available in PMC: 2020 Jul 13.

Published in final edited form as: J Biomed Inform. 2019 Feb 10;91:103123. doi: 10.1016/j.jbi.2019.103123

Table 6:

Results obtained with NN models, with 10-fold cross validation and model averaging. The scores that indicate an improvement over SVM models in Table 4 are underlined. word, pos, and dep indicate word, part-of-speech and dependency-enhanced word embeddings, respectively. feat indicates additional high-level features (posgram + sent + struct + rule).

Experiment	Overall		Per Category
Experiment	Accu.	MacroF₁	Cat	Pr.	Rec.	F₁
CNN(word)	0.858	0.635	POS	0.830	0.579	0.680
			NEG	0.780	0.192	0.305
			NEU	0.863	0.983	0.919
CNN(pos+dep)	0.850	0.632	POS	0.818	0.559	0.661
			NEG	0.674	0.212	0.318
			NEU	0.861	0.975	0.914
CNN(word)+feat	0.887	0.719	POS	0.919	0.655	0.764
			NEG	0.957	0.310	0.464
			NEU	0.882	0.994	0.935
CNN(pos+dep)+feat	0.896	0.757	POS	0.914	0.689	0.783
			NEG	0.898	0.404	0.552
			NEU	0.896	0.990	0.940
BiLSTM(word)	0.857	0.668	POS	0.785	0.651	0.711
			NEG	0.551	0.291	0.375
			NEU	0.882	0.958	0.918
BiLSTM(pos+dep)	0.837	0.644	POS	0.747	0.610	0.670
			NEG	0.456	0.303	0.356
			NEU	0.876	0.938	0.906
BiLSTM(word)+feat	0.883	0.731	POS	0.845	0.695	0.761
			NEG	0.827	0.378	0.509
			NEU	0.892	0.973	0.931
BiLSTM(pos+dep)+feat	0.874	0.732	POS	0.798	0.708	0.747
			NEG	0.706	0.426	0.523
			NEU	0.900	0.952	0.925