Chinese NER using Bert

BERT for Chinese NER.

model list

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

requirement

1. pytorch=1.1.0
2. cuda=9.0

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
谈	O
笑	O
风	O
生	O 

run the code

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

note: file structure of the model

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

result

Tne 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)	train_max_length=128 eval_max_length=512 epoch=4 lr=3e-5 batch_size=24
BERT+CRF	0.9562(0.9539)	0.9671(0.9644)	0.9616(0.9591)	train_max_length=128 eval_max_length=512 epoch=10 lr=3e-5 batch_size=24
BERT+Span	0.9604(0.9620)	0.9617(0.9632)	0.9611(0.9626)	train_max_length=128 eval_max_length=512 epoch=4 lr=3e-5 batch_size=24

The entity performance performance of BERT on test:

	CONT	ORG	LOC	EDU	NAME	PRO	RACE	TITLE
BERT+Softmax
Accuracy	1.0000	0.9446	1.0000	0.9911	1.0000	0.8919	1.0000	0.9545
Recall	1.0000	0.9566	1.0000	0.9911	1.0000	1.0000	1.0000	0.9508
F1 Score	1.0000	0.9506	1.0000	0.9911	1.0000	0.9429	1.0000	0.9526
BERT+CRF
Accuracy	1.0000	0.9446	1.0000	0.9823	1.0000	0.9687	1.0000	0.9591
Recall	1.0000	0.9566	1.0000	0.9911	1.0000	0.9697	1.0000	0.9534
F1 Score	1.0000	0.9506	1.0000	0.9867	1.0000	0.9697	1.0000	0.9552
BERT+Span
Accuracy	1.0000	0.9378	1.0000	0.9911	1.0000	0.9429	1.0000	0.9685
Recall	1.0000	0.9548	1.0000	0.9911	1.0000	1.0000	1.0000	0.9560
F1 Score	1.0000	0.9462	1.0000	0.9911	1.0000	0.9706	1.0000	0.9622