darr/nerbert

bert_ner

介绍

Use google BERT to do CoNLL-2003 NER

dataset

CoNLL-2003

model folder

About the parameter --bert_model

You can just input model name like: bert-base-cased
It will call pytorch-pretrained-bert to down the bert-base-cased model files.
And this will work, besides you need to wait a little longer time.
So if you network is unstable, you may fail.

At this time, you need to download the pretrained models manually.
Since this project is based on pytorch, so you need the pytorch version of the BERT.

pytorch-pretrained-bert has changed name to pytorch-transformers. here is the github:
pytorch-transformers

Ok, The Subject create the local model folder
a model folder is like this:

bert-base-cased
    python_model.bin
    bert_config.json
    vocab.txt

When you download the model tar.gz file, untar it then you can get python_model.bin and bert_cofnig.json.
For vocab.txt, you need download the file like bert-base-cased-vocab.txt, and change its name to vocab.txt

how to run?

bash run.sh

This command will create the environment that needed by the models.
This project is created on the purposes of easy-on-run.
If you want to know the details about the models, just read code.

It takes about 10 minutes for one epoch, 50 minutes for five epoch. The results report by BERT paper

https://arxiv.org/pdf/1810.04805.pdf

System	Dev F1	Test F1
ELMo+BiLSTM+CRF	95.7	92.2
CVT+Multi	-	92.6
BERT_BASE	96.4	92.4
BERT_LARGE	96.6	92.8

The result this project get: bert-base-cased

             precision    recall  f1-score   support

        LOC     0.9271    0.9305    0.9288      1668
        ORG     0.8754    0.9055    0.8902      1661
        PER     0.9663    0.9573    0.9618      1617
       MISC     0.7803    0.8348    0.8066       702

avg / total     0.9049    0.9189    0.9117      5648

my device has a 1080Ti, if you use bert-large may have a memory problem, and I do not try.

Inference

from bert import Ner

model = Ner("out_!x/")

output = model.predict("Steve went to Paris")

print(output)
'''
    [
        {
            "confidence": 0.9981840252876282,
            "tag": "B-PER",
            "word": "Steve"
        },
        {
            "confidence": 0.9998939037322998,
            "tag": "O",
            "word": "went"
        },
        {
            "confidence": 0.999891996383667,
            "tag": "O",
            "word": "to"
        },
        {
            "confidence": 0.9991968274116516,
            "tag": "B-LOC",
            "word": "Paris"
        }
    ]
'''

Deploy REST-API

BERT NER model deployed as rest api In the run.sh file

python api.py

API will be live at 0.0.0.0:8000 endpoint predict

cURL request

curl -X POST http://0.0.0.0:8000/predict -H 'Content-Type: application/json' -d '{ "text": "Steve went to Paris" }'

Output

{
    "result": [
        {
            "confidence": 0.9981840252876282,
            "tag": "B-PER",
            "word": "Steve"
        },
        {
            "confidence": 0.9998939037322998,
            "tag": "O",
            "word": "went"
        },
        {
            "confidence": 0.999891996383667,
            "tag": "O",
            "word": "to"
        },
        {
            "confidence": 0.9991968274116516,
            "tag": "B-LOC",
            "word": "Paris"
        }
    ]
}