Chinese Named Entity Recognition (中文命名实体识别).

Warning: When the training loss drops but the evaluation metrics are 0, please check whether the suitable learning rate is selected for Bert and Bi-LSTM respectively.

• Fixed a bug that failed to load the pre-trained model.
• Use the representation of the first sub-token after the wordpiece for classification.

• python>=3.6.4
• pytorch==1.5.0
• transformers==2.11.0
• seqeval==0.0.12
• tqdm

• BiLSTM
• BiLSTM + CRF
• BERT
• BERT + CRF

Each line contains a character and its label, separated by "\t" or space. Each sentence is followed by a blank line.

中	B-LOC
国	I-LOC
很	O
大	O

句	O
子	O
结	O
束	O
是	O
空	O
行	O

We use the renminribao2014 (人民日报2014) dataset (https://pan.baidu.com/s/1LDwQjoj7qc-HT9qwhJ3rcA password: 1fa3). Download and unzip it in data/renminribao2014/, then run process_rmrb.py.

python process_rmrb.py 

Modify the labels in main.py according to your dataset:

labels = ['O', 'B-LOC', 'B-ORG', 'B-T', 'I-LOC', 'I-PER', 'B-PER', 'I-ORG', 'I-T']

# run bert
python main.py --model bert

# run bert+crf
python main.py --model bert --crf

# run bilstm
# set the learning rate to 1e-2
python main.py --model bilstm \
        --learning_rate 1e-2 \
        --num_train_epochs 20 \
        --train_batch_size 64 \
        --dev_batch_size 32

# run bilstm+crf
python main.py --model bilstm --crf \
        --learning_rate 1e-2 \
        --num_train_epochs 20 \
        --train_batch_size 64 \
        --dev_batch_size 32

# run bert
python main.py --model bert --mode test

# run bert+crf
python main.py --model bert --crf --mode test

# run bilstm
python main.py --model bilstm --mode test

# run bert+crf
python main.py --model bilstm --crf --mode test

Report on the subset of renminribao2014 due to the limitation of computing resources. The size of subset is as follows:

BiLSTM

BiLSTM + CRF

BERT

BERT + CRF

• pytorch CRF: https://github.com/kmkurn/pytorch-crf