Chinese NER using Bert

BERT for Chinese NER.

update：其他一些可以参考,包括Biaffine、GlobalPointer等:examples

dataset list

1. cner: datasets/cner
2. CLUENER: https://github.com/CLUEbenchmark/CLUENER

model list

1. BERT+Softmax
2. BERT+CRF
3. BERT+Span

requirement

1. 1.1.0 =< PyTorch < 1.5.0
2. cuda=9.0
3. python3.6+

input format

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

run the code

1. Modify the configuration information in run_ner_xxx.py or run_ner_xxx.sh .
2. sh scripts/run_ner_xxx.sh

note: file structure of the model

├── prev_trained_model
|  └── bert_base
|  |  └── pytorch_model.bin
|  |  └── config.json
|  |  └── vocab.txt
|  |  └── ......

CLUENER result

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

ALBERT for CLUENER

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	google	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

Cner result

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)