Multiple-level biomedical event trigger recognition with transfer learning

BMC Bioinformatics

Table 6 Detailed results achieved by the proposed MLTrigNer Model, Basic Model A and TL Model C on Data_MLEE

Trigger type	Basic Model A			TL Model C			MLTrigNer Model
	P	R	F1	P	R	F1	P	R	F1
Cell proliferation	85.37	81.40	83.33	83.33	81.40	82.35	81.40	81.40	81.40
Development	66.37	76.53	71.09	74.51	77.55	76.00	78.35	77.55	77.95
Blood vessel develop	97.33	94.19	95.74	98.64	93.87	96.20	98.99	94.84	96.87
Growth	96.00	85.71	90.57	88.89	85.71	87.27	92.45	87.50	89.91
Death	73.68	75.68	74.67	66.67	81.08	73.17	66.67	81.08	73.17
Breakdown	82.35	63.64	71.79	73.68	63.64	68.29	87.50	63.64	73.68
Remodeling	71.43	50.00	58.82	75.00	30.00	42.86	66.67	40.00	50.00
Synthesis	50.00	25.00	33.33	33.33	25.00	28.57	20.00	25.00	22.22
Gene expression	91.67	83.33	87.30	85.51	89.39	87.41	89.05	92.42	90.71
Transcription	0.0	0.0	0.0	50.00	16.67	25.00	100.0	16.67	28.57
Protein Catabolism	0.0	0.0	0.0	0.0	0.0	0.0	33.33	20.00	25.00
Phosphorylation	75.00	100.0	85.71	100.0	100.0	100.0	100.0	100.0	100.0
Dephosphorylation	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
Localization	78.83	81.20	80.00	77.14	81.20	79.12	82.17	79.70	80.92
Binding	86.96	70.18	77.67	83.02	77.19	80.00	79.25	73.68	76.36
Regulation	59.80	58.93	59.37	65.13	61.65	63.34	65.17	63.29	64.22
Positive regulation	80.88	81.90	81.39	81.11	82.91	82.00	81.96	82.22	82.09
Negative regulation	84.73	65.71	74.02	77.18	75.61	76.39	80.72	73.47	76.92
Planned process	78.69	48.98	60.38	66.86	57.65	61.92	71.07	57.65	63.66
TOTAL	81.63	74.26	77.77	79.69	77.62	78.64	81.76	77.71	79.68

The Basic Model A is trained only on the training and development sets of Data_MLEE without transfer learning. The TL Model C and the MLTrigNer model are jointly trained on the source dataset Data_EPI11 and the training and development sets of the target dataset Data_MLEE using different transfer learning approaches, respectively. The three models are tested on the test set of Data_MLEE. In the results of MLTrigNer Model, the improved F1 values are marked in bold

ISSN: 1471-2105