Skip to main content
. 2019 Jun 20;36(1):280–286. doi: 10.1093/bioinformatics/btz504

Table 3.

In-corpus (IC) performance, measured by F1-score, of the baseline (BL) BiLSTM-CRF compared to a BiLSTM-CRF trained with transfer learning (TL)

BL
TL
Entity Corpus Average σ Average σ
Chemicals BC4CHEMD 88.46 0.61 88.98 0.63
BC5CDR 92.82 0.80 92.20 0.86
CRAFT 84.98 1.98 85.80 1.74
Disease BC5CDR 84.49 0.33 84.41 0.24
NCBI-disease 87.01 1.17 87.66 0.86
Variome 85.75 2.83 86.69 3.03
Species CRAFT 96.28 2.21 96.55 1.72
Linnaeus 89.44 3.91 90.72 4.90
S800 72.75 2.42 74.93 3.27
Genes/proteins BC2GM 81.48 0.48 80.65* 0.57
CRAFT 84.46 6.08 85.50 4.59
JNLPBA 80.92 2.50 81.56 2.73

Note: The TL model was pre-trained on the CALBC-Small-III corpus. IC performance is derived from 5-fold cross-validation, using exact matching criteria. Statistical significance is measured through a two-tailed t-test. Bold, best scores, σ, standard deviation.

*

Significantly different than the BL (P ≤ 0.05).