bertcner
Code for paper Chinese clinical named entity recognition with variant neural structures based on BERT methods
Paper url: https://www.sciencedirect.com/science/article/pii/S1532046420300502
We pre-trained BERT model to improve the performance of Chinese CNER. Different layers such as Long Short-Term Memory (LSTM) and Conditional Random Field (CRF) were used to extract the text features and decode the predicted tags respectively. And we also proposed a new strategy to incorporate dictionary features into the model. Radical features of Chinese characters were also used to improve the model performance.
Model structure
Usage
Pre-trained models
For replication, we uploaded two models in Baidu Netdisk.
Link: https://pan.baidu.com/s/1obzG6OSbu77duhusWg2xmQ
Code: k53q
Examples
To replicate the result of CCKS-2018 dataset
python main.py \
--data_dir=data/ccks_2018 \
--bert_model=model/ \
--output_dir=./output \
--terminology_dicts_path="{'medicine':'data/ccks_2018/drug_dict.txt','surgery':'data/ccks_2018/surgery_dict.txt'}" \
--radical_dict_path data/radical_dict.txt \
--constant=0 \
--add_radical_or_not=True \
--radical_one_hot=False \
--radical_emb_dim=20 \
--max_seq_length=480 \
--do_train=True \
--do_eval=True \
--train_batch_size=6 \
--eval_batch_size=4 \
--hidden_dim=64 \
--learning_rate=5e-5 \
--num_train_epochs=5 \
--gpu_id=3 \
Results
CCKS-2018 dataset
| Method | P | R | F1 |
|---|---|---|---|
| FT-BERT+BiLSTM+CRF | 88.57 | 89.02 | 88.80 |
| +dictionary | 88.58 | 89.17 | 88.87 |
| +radical(one-hot encoding) | 88.51 | 89.39 | 88.95 |
| +radical(random embedding) | 89.24 | 89.11 | 89.17 |
| +dictionary +radical | 89.42 | 89.22 | 89.32 |
| ensemble | 89.59 | 89.54 | 89.56 |
| Team Name | Method | F1 |
|---|---|---|
| Yang and Huang (2018) | CRF(feature-rich + rule) | 89.26 |
| heiheihahei | LSTM-CRF(ensemble) | 88.92 |
| Luo et al.(2018) | LSTM-CRF(ensemble) | 88.63 |
| dous12 | - | 88.37 |
| chengachengcheng | - | 88.30 |
| NUBT-IBDL | - | 87.62 |
| Our | FT-BERT+BiLSTM +CRF+Dictionary(ensemble) | 89.56 |
CCKS-2017 dataset
| Method | P | R | F1 |
|---|---|---|---|
| FT-BERT+BiLSTM+CRF | 91.64 | 90.98 | 91.31 |
| +dictionary | 91.49 | 90.97 | 91.23 |
| +radical(one-hot encoding) | 91.83 | 90.80 | 91.35 |
| +radical(random embedding) | 92.07 | 90.77 | 91.42 |
| +dictionary+radical | 91.76 | 90.88 | 91.32 |
| ensemble | 92.06 | 91.15 | 91.60 |
| Team Name | Method | F1 |
|---|---|---|
| Qiu et al. (2018b) | RD-CNN-CRF | 91.32 |
| Wang et al. (2019) | BiLSTM-CRF+Dictionary | 91.24 |
| Hu et al. (2017) | BiLSTM-FEA(ensemble) | 91.03 |
| Zhang et al. (2018) | BiLSTM-CRF(mt+att+ms) | 90.52 |
| Xia and Wang (2017) | BiLSTM-CRF(ensemble) | 89.88 |
| Ouyang et al. (2017) | BiRNN-CRF | 88.85 |
| Li et al. (2017) | BiLSTM-CRF(specialized +lexicons) | 87.95 |
| Our | FT-BERT+BiLSTM +CRF+Dictionary(ensemble) | 91.60 |
