BERT NER

Use google BERT to do CoNLL-2003 NER !

Train model using Python and TensorFlow 2.0

ALBERT-TF2.0

BERT-SQuAD

BERT-NER-Pytorch

Requirements

python3
pip3 install -r requirements.txt

Download Pretrained Models from Tensorflow offical models

bert-base-cased unzip into bert-base-cased
bert-large-cased unzip into bert-large-cased

code for pre-trained bert from tensorflow-offical-models

Run

Single GPU

python run_ner.py --data_dir=data/ --bert_model=bert-base-cased --output_dir=out_base --max_seq_length=128 --do_train --num_train_epochs 3 --do_eval --eval_on dev

Multi GPU

python run_ner.py --data_dir=data/ --bert_model=bert-large-cased --output_dir=out_large --max_seq_length=128 --do_train --num_train_epochs 3 --multi_gpu --gpus 0,1,2,3 --do_eval --eval_on test

Result

BERT-BASE

Validation Data

             precision    recall  f1-score   support

        PER     0.9677    0.9756    0.9716      1842
        LOC     0.9671    0.9592    0.9631      1837
       MISC     0.8872    0.9132    0.9001       922
        ORG     0.9191    0.9314    0.9252      1341

avg / total     0.9440    0.9509    0.9474      5942

Test Data

             precision    recall  f1-score   support

        ORG     0.8773    0.9037    0.8903      1661
        PER     0.9646    0.9592    0.9619      1617
       MISC     0.7691    0.8305    0.7986       702
        LOC     0.9333    0.9305    0.9319      1668

avg / total     0.9053    0.9184    0.9117      5648

Pretrained model download from here

BERT-LARGE

Validation Data

             precision    recall  f1-score   support

        ORG     0.9290    0.9374    0.9332      1341
       MISC     0.8967    0.9230    0.9097       922
        PER     0.9713    0.9734    0.9723      1842
        LOC     0.9748    0.9701    0.9724      1837

avg / total     0.9513    0.9564    0.9538      5942

Test Data

             precision    recall  f1-score   support

        LOC     0.9256    0.9329    0.9292      1668
       MISC     0.7891    0.8419    0.8146       702
        PER     0.9647    0.9623    0.9635      1617
        ORG     0.8903    0.9133    0.9016      1661

avg / total     0.9094    0.9242    0.9167      5648

Pretrained model download from here

Inference

from bert import Ner

model = Ner("out_base/")

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

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"
        }
    ]
}