BERT for Chinese NER.
- cner: datasets/cner
- CLUENER: https://github.com/CLUEbenchmark/CLUENER
- BERT+Softmax
- BERT+CRF
- BERT+Span
- 1.1.0 =< PyTorch < 1.5.0
- cuda=9.0
- python3.6+
Input format (prefer BIOS tag scheme), with each character its label for one line. Sentences are splited with a null line.
美 B-LOC
国 I-LOC
的 O
华 B-PER
莱 I-PER
士 I-PER
我 O
跟 O
他 O
- Modify the configuration information in
run_ner_xxx.pyorrun_ner_xxx.sh. sh scripts/run_ner_xxx.sh
note: file structure of the model
├── prev_trained_model
| └── bert_base
| | └── pytorch_model.bin
| | └── config.json
| | └── vocab.txt
| | └── ......
The overall performance of BERT on dev:
| Accuracy (entity) | Recall (entity) | F1 score (entity) | |
|---|---|---|---|
| BERT+Softmax | 0.7897 | 0.8031 | 0.7963 |
| BERT+CRF | 0.7977 | 0.8177 | 0.8076 |
| BERT+Span | 0.8132 | 0.8092 | 0.8112 |
| BERT+Span+adv | 0.8267 | 0.8073 | 0.8169 |
| BERT-small(6 layers)+Span+kd | 0.8241 | 0.7839 | 0.8051 |
| BERT+Span+focal_loss | 0.8121 | 0.8008 | 0.8064 |
| BERT+Span+label_smoothing | 0.8235 | 0.7946 | 0.8088 |
The overall performance of ALBERT on dev:
| model | version | Accuracy(entity) | Recall(entity) | F1(entity) | Train time/epoch |
|---|---|---|---|---|---|
| albert | base_google | 0.8014 | 0.6908 | 0.7420 | 0.75x |
| albert | large_google | 0.8024 | 0.7520 | 0.7763 | 2.1x |
| albert | xlarge_google | 0.8286 | 0.7773 | 0.8021 | 6.7x |
| bert | 0.8118 | 0.8031 | 0.8074 | ----- | |
| albert | base_bright | 0.8068 | 0.7529 | 0.7789 | 0.75x |
| albert | large_bright | 0.8152 | 0.7480 | 0.7802 | 2.2x |
| albert | xlarge_bright | 0.8222 | 0.7692 | 0.7948 | 7.3x |
The overall performance of BERT on dev(test):
| Accuracy (entity) | Recall (entity) | F1 score (entity) | |
|---|---|---|---|
| BERT+Softmax | 0.9586(0.9566) | 0.9644(0.9613) | 0.9615(0.9590) |
| BERT+CRF | 0.9562(0.9539) | 0.9671(0.9644) | 0.9616(0.9591) |
| BERT+Span | 0.9604(0.9620) | 0.9617(0.9632) | 0.9611(0.9626) |
| BERT+Span+focal_loss | 0.9516(0.9569) | 0.9644(0.9681) | 0.9580(0.9625) |
| BERT+Span+label_smoothing | 0.9566(0.9568) | 0.9624(0.9656) | 0.9595(0.9612) |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent