legal-bert-ner-base-cased-ptbr

(BERT base) NER model in the legal domain in Portuguese

README under construction

ner-legal-bert-base-cased-ptbr is a NER model (token classification) in the legal domain in Portuguese that was finetuned from the model dominguesm/legal-bert-base-cased-ptbr by using a NER objective.

The model is intended to assist NLP research in the legal field, computer law and legal technology applications. Several legal texts in Portuguese (more information below) were used with the following labels:

• PESSOA
• ORGANIZACAO
• LOCAL
• TEMPO
• LEGISLACAO
• JURISPRUDENCIA

The labels were inspired by the LeNER_br dataset.

Training Dataset

The dataset of ner-legal-bert-base-cased-ptbr include:

• 971932 examples of miscellaneous legal documents (train split)
• 53996 examples of miscellaneous legal documents (valid split)
• 53997 examples of miscellaneous legal documents (test split)

The data used was provided by the BRAZILIAN SUPREME FEDERAL TRIBUNAL, through the terms of use: LREC 2020.

The results of this project do not imply in any way the position of the BRAZILIAN SUPREME FEDERAL TRIBUNAL, all being the sole and exclusive responsibility of the author of the model.

Using the model for inference in production

from transformers import AutoModelForTokenClassification, AutoTokenizer
import torch

# parameters
model_name = "dominguesm/ner-legal-bert-base-cased-ptbr"
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.

from transformers import pipeline

model_name = "dominguesm/ner-legal-bert-base-cased-ptbr"

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

ner(input_text, aggregation_strategy="average")

Training procedure

Hyperparameters

batch, learning rate...

• per_device_batch_size = 64
• gradient_accumulation_steps = 2
• learning_rate = 2e-5
• num_train_epochs = 3
• weight_decay = 0.01
• optimizer = torch.optim.AdamW
• epsilon = 1e-08
• lr_scheduler_type = linear

save model & load best model

• save_total_limit = 3
• logging_steps = 1000
• 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

Training results

Num examples = 971932
Num Epochs = 3
Instantaneous batch size per device = 64
Total train batch size (w. parallel, distributed & accumulation) = 128
Gradient Accumulation steps = 2
Total optimization steps = 22779
Evaluation Infos:
  Num examples = 53996
  Batch size = 128

Validation metrics by Named Entity (Test Dataset)

• Num examples = 53997
• overall_precision: 0.9432396865925381
• overall_recall: 0.9614334116769161
• overall_f1: 0.9522496545298874
• overall_accuracy': 0.9894741602608071

Notes

• For the production of this readme, i used the readme written by Pierre Guillou (available here) as a basis, reproducing some parts entirely.