(BERT large) NER model in the legal domain in Portuguese (LeNER-Br)

ner-bert-large-portuguese-cased-lenerbr is a NER model (token classification) in the legal domain in Portuguese that was finetuned on 20/12/2021 in Google Colab from the model pierreguillou/bert-large-cased-pt-lenerbr on the dataset LeNER_br by using a NER objective.

Due to the small size of the finetuning dataset, the model overfitted before to reach the end of training. Here are the overall final metrics on the validation dataset (note: see the paragraph "Validation metrics by Named Entity" to get detailed metrics):

f1: 0.9082022949426265
precision: 0.8975220495590088
recall: 0.9191397849462366
accuracy: 0.9808310603867311
loss: 0.1228889599442482

Check as well the base version of this model with a f1 of 0.893.

Note: the model pierreguillou/bert-large-cased-pt-lenerbr is a language model that was created through the finetuning of the model BERTimbau large on the dataset LeNER-Br language modeling by using a MASK objective. This first specialization of the language model before finetuning on the NER task allows to get a better NER model.

Blog post

NLP | Modelos e Web App para Reconhecimento de Entidade Nomeada (NER) no domínio jurídico brasileiro (29/12/2021)

Widget & App

You can test this model into the widget of this page.

Use as well the NER App that allows comparing the 2 BERT models (base and large) fitted in the NER task with the legal LeNER-Br dataset.

Using the model for inference in production

# install pytorch: check https://pytorch.org/
# !pip install transformers 
from transformers import AutoModelForTokenClassification, AutoTokenizer
import torch

# parameters
model_name = "pierreguillou/ner-bert-large-cased-pt-lenerbr"
model = AutoModelForTokenClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

input_text = "Acrescento que não há de se falar em violação do artigo 114, § 3º, da Constituição Federal, posto que referido dispositivo revela-se impertinente, tratando da possibilidade de ajuizamento de dissídio coletivo pelo Ministério Público do Trabalho nos casos de greve em atividade essencial."

# tokenization
inputs = tokenizer(input_text, max_length=512, truncation=True, return_tensors="pt")
tokens = inputs.tokens()

# get predictions
outputs = model(**inputs).logits
predictions = torch.argmax(outputs, dim=2)

# print predictions
for token, prediction in zip(tokens, predictions[0].numpy()):
    print((token, model.config.id2label[prediction]))

You can use pipeline, too. However, it seems to have an issue regarding to the max_length of the input sequence.

!pip install transformers
import transformers
from transformers import pipeline

model_name = "pierreguillou/ner-bert-large-cased-pt-lenerbr"

ner = pipeline(
    "ner",
    model=model_name
) 

ner(input_text)

Training procedure

Notebook

The notebook of finetuning (HuggingFace_Notebook_token_classification_NER_LeNER_Br.ipynb) is in github.

Hyperparameters

batch, learning rate...

per_device_batch_size = 2
gradient_accumulation_steps = 2
learning_rate = 2e-5
num_train_epochs = 10
weight_decay = 0.01
optimizer = AdamW
betas = (0.9,0.999)
epsilon = 1e-08
lr_scheduler_type = linear
seed = 42

save model & load best model

save_total_limit = 7
logging_steps = 500
eval_steps = logging_steps
evaluation_strategy = 'steps'
logging_strategy = 'steps'
save_strategy = 'steps'
save_steps = logging_steps
load_best_model_at_end = True
fp16 = True

get best model through a metric

metric_for_best_model = 'eval_f1'
greater_is_better = True

Training results

Num examples = 7828
Num Epochs = 20
Instantaneous batch size per device = 2
Total train batch size (w. parallel, distributed & accumulation) = 4
Gradient Accumulation steps = 2
Total optimization steps = 39140

Step   Training Loss  Validation Loss  Precision  Recall    F1        Accuracy
500    0.250000       0.140582         0.760833   0.770323  0.765548  0.963125
1000   0.076200       0.117882         0.829082   0.817849  0.823428  0.966569
1500   0.082400       0.150047         0.679610   0.914624  0.779795  0.957213
2000   0.047500       0.133443         0.817678   0.857419  0.837077  0.969190
2500   0.034200       0.230139         0.895672   0.845591  0.869912  0.964070
3000   0.033800       0.108022         0.859225   0.887312  0.873043  0.973700
3500   0.030100       0.113467         0.855747   0.885376  0.870310  0.975879
4000   0.029900       0.118619         0.850207   0.884946  0.867229  0.974477
4500   0.022500       0.124327         0.841048   0.890968  0.865288  0.975041
5000   0.020200       0.129294         0.801538   0.918925  0.856227  0.968077
5500   0.019700       0.128344         0.814222   0.908602  0.858827  0.969250
6000   0.024600       0.182563         0.908087   0.866882  0.887006  0.968565
6500   0.012600       0.159217         0.829883   0.913763  0.869806  0.969357
7000   0.020600       0.183726         0.854557   0.893333  0.873515  0.966447
7500   0.014400       0.141395         0.777716   0.905161  0.836613  0.966828
8000   0.013400       0.139378         0.873042   0.899140  0.885899  0.975772
8500   0.014700       0.142521         0.864152   0.901505  0.882433  0.976366

9000   0.010900       0.122889         0.897522   0.919140  0.908202  0.980831

9500   0.013500       0.143407         0.816580   0.906667  0.859268  0.973395
10000  0.010400       0.144946         0.835608   0.908387  0.870479  0.974629
10500  0.007800       0.143086         0.847587   0.910108  0.877735  0.975985
11000  0.008200       0.156379         0.873778   0.884301  0.879008  0.976321
11500  0.008200       0.133356         0.901193   0.910108  0.905628  0.980328
12000  0.006900       0.133476         0.892202   0.920215  0.905992  0.980572
12500  0.006900       0.129991         0.890159   0.904516  0.897280  0.978683

Validation metrics by Named Entity

{'JURISPRUDENCIA': {'f1': 0.8135593220338984,
  'number': 657,
  'precision': 0.865979381443299,
  'recall': 0.7671232876712328},
 'LEGISLACAO': {'f1': 0.8888888888888888,
  'number': 571,
  'precision': 0.8952042628774423,
  'recall': 0.882661996497373},
 'LOCAL': {'f1': 0.850467289719626,
  'number': 194,
  'precision': 0.7777777777777778,
  'recall': 0.9381443298969072},
 'ORGANIZACAO': {'f1': 0.8740635033892258,
  'number': 1340,
  'precision': 0.8373205741626795,
  'recall': 0.914179104477612},
 'PESSOA': {'f1': 0.9836677554829678,
  'number': 1072,
  'precision': 0.9841269841269841,
  'recall': 0.9832089552238806},
 'TEMPO': {'f1': 0.9669669669669669,
  'number': 816,
  'precision': 0.9481743227326266,
  'recall': 0.9865196078431373},
 'overall_accuracy': 0.9808310603867311,
 'overall_f1': 0.9082022949426265,
 'overall_precision': 0.8975220495590088,
 'overall_recall': 0.9191397849462366}

andreycolombo
/

sophia-sentitive-filter