Table 2 Doc-level performance comparison over our proposed model without and with knowledge on the CDR dataset
From: Biomedical relation extraction via knowledge-enhanced reading comprehension
KBs | Model | P (%) | R (%) | F1 (%) |
|---|---|---|---|---|
Without KBs | ||||
Traditional ML | ME [8] | 62.00 | 55.10 | 58.30 |
Kernel-based SVM [24] | 53.20 | 69.70 | 60.30 | |
NN-based ML | ||||
Relation classification | CNN+SDP [6] | 58.02 | 76.20 | 65.88 |
LSTM+CNN [25] | 56.20 | 68.00 | 61.50 | |
BRAN(Transformer) [3] | 55.60 | 70.80 | 62.10 | |
CNN+CNNchar [11] | 57.00 | 68.60 | 62.30 | |
GCNN [2] | 52.80 | 66.00 | 58.60 | |
Sequence labeling | Bio-Seq(LSTM+CRF) [23] | 60.00 | 67.50 | 63.50 |
Reading comprehension | RC (Ours) | 65.83 | 66.32 | 66.07 |
With KBs | ||||
Traditional ML | ||||
SVM+Rules(+CTD)[26] | 68.15 | 66.04 | 67.08 | |
SVM(+CTD+SIDER+MEDI+MeSH) [9] | 65.80 | 68.57 | 67.16 | |
Kernel-based models(+CTD) [10] | 60.84 | 76.36 | 67.72 | |
SVM(+Euretos KB) [27] | 73.10 | 67.60 | 70.20 | |
NN-based ML | ||||
Relation classification | CAN(+CTD) [7] | 60.51 | 80.48 | 69.08 |
LSTM+CNN(+CTD) [4] | 63.60 | 76.80 | 69.60 | |
RPCNN(+CTD+SIDER+MEDI+MeSH fea) [5] | 65.24 | 77.21 | 70.77 | |
KCN(+CTD) [1] | 69.65 | 72.98 | 71.28 | |
Reading comprehension | KRC(+DCh-Miner) (Ours) | 65.33 | 67.17 | 66.23 |
KRC(+CTD) (Ours) | 71.93 | 70.45 | 71.18 | |