kifish / NER-demo

Pytorch-BERT-CRF-NER;Chinese-Named-Entity-Recognition

NER-demo

Method

1.HMM
https://github.com/kifish/NER-demo/tree/hmm
2.CRF
https://github.com/kifish/NER-demo/tree/crf
3.BiLSTM-viterbi
https://github.com/kifish/NER-demo/tree/BiLSTM-viterbi
4.BiLSTM-CRF
https://github.com/kifish/NER-demo/tree/BiLSTM-crf
5.BiLSTM-CNN-CRF
https://github.com/kifish/NER-demo/tree/BiLSTM-cnn-crf
Update:
6.BERT-Softmax
https://github.com/kifish/NER-demo/tree/bert
7.BERT-CRF
https://github.com/kifish/NER-demo/tree/bert

See more

http://nlpprogress.com/english/named_entity_recognition.html

Experiment

Environment

python3.6+ (according to the branch)
pip install -r requirements.txt (according to the branch)
(use pip install git+https://www.github.com/keras-team/keras-contrib.git to install keras-contrib)

Result

	            precision	   recall        f1-score	   support
BERT                   0.9458      0.9090          0.9263             6195
BERT-CRF               0.9338      0.8901          0.9097             6195
BiLSTM-CRF	       0.8616      0.7138	   0.7806	      6181
BiLSTM-CNN-CRF	       0.8406	   0.7185	   0.7686	      6181
CRF	               0.8420	   0.6279	   0.7170	      6181
BiLSTM-viterbi	       0.8512	   0.5700	   0.6809	      6181
HMM	               0.4911	   0.4341	   0.4479	      6181

注: bert做数据预处理的实现和之前的模型不一样, 有一些出入, 导致support有差异, 待对齐。

Analysis

Todo