Legal_BERTimbau
Introduction
Legal_BERTimbau Large is a fine-tuned BERT model based on BERTimbau Large.
"BERTimbau Base is a pretrained BERT model for Brazilian Portuguese that achieves state-of-the-art performances on three downstream NLP tasks: Named Entity Recognition, Sentence Textual Similarity and Recognizing Textual Entailment. It is available in two sizes: Base and Large.
For further information or requests, please go to BERTimbau repository."
The performance of Language Models can change drastically when there is a domain shift between training and test data. In order create a Portuguese Language Model adapted to a Legal domain, the original BERTimbau model was submitted to a fine-tuning stage where it was performed 1 "PreTraining" epoch over 30 000 legal Portuguese Legal documents available online.
Available models
| Model | Arch. | #Layers | #Params |
|---|---|---|---|
rufimelo/Legal-BERTimbau-base | BERT-Base | 12 | 110M |
rufimelo/Legal-BERTimbau-large | BERT-Large | 24 | 335M |
Usage
from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("rufimelo/Legal-BERTimbau-base")
model = AutoModelForMaskedLM.from_pretrained("rufimelo/Legal-BERTimbau-base")
Masked language modeling prediction example
from transformers import pipeline
from transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("rufimelo/Legal-BERTimbau-base")
model = AutoModelForMaskedLM.from_pretrained("rufimelo/Legal-BERTimbau-base")
pipe = pipeline('fill-mask', model=model, tokenizer=tokenizer)
pipe('O advogado apresentou [MASK] para o juíz')
# [{'score': 0.5034703612327576,
#'token': 8190,
#'token_str': 'recurso',
#'sequence': 'O advogado apresentou recurso para o juíz'},
#{'score': 0.07347951829433441,
#'token': 21973,
#'token_str': 'petição',
#'sequence': 'O advogado apresentou petição para o juíz'},
#{'score': 0.05165359005331993,
#'token': 4299,
#'token_str': 'resposta',
#'sequence': 'O advogado apresentou resposta para o juíz'},
#{'score': 0.04611917585134506,
#'token': 5265,
#'token_str': 'exposição',
#'sequence': 'O advogado apresentou exposição para o juíz'},
#{'score': 0.04068068787455559,
#'token': 19737, 'token_str':
#'alegações',
#'sequence': 'O advogado apresentou alegações para o juíz'}]
For BERT embeddings
import torch
from transformers import AutoModel
model = AutoModel.from_pretrained('rufimelo/Legal-BERTimbau-base')
input_ids = tokenizer.encode('O advogado apresentou recurso para o juíz', return_tensors='pt')
with torch.no_grad():
outs = model(input_ids)
encoded = outs[0][0, 1:-1]
#tensor([[ 0.0328, -0.4292, -0.6230, ..., -0.3048, -0.5674, 0.0157],
#[-0.3569, 0.3326, 0.7013, ..., -0.7778, 0.2646, 1.1310],
#[ 0.3169, 0.4333, 0.2026, ..., 1.0517, -0.1951, 0.7050],
#...,
#[-0.3648, -0.8137, -0.4764, ..., -0.2725, -0.4879, 0.6264],
#[-0.2264, -0.1821, -0.3011, ..., -0.5428, 0.1429, 0.0509],
#[-1.4617, 0.6281, -0.0625, ..., -1.2774, -0.4491, 0.3131]])
Citation
If you use this work, please cite BERTimbau's work:
@inproceedings{souza2020bertimbau,
author = {F{\'a}bio Souza and
Rodrigo Nogueira and
Roberto Lotufo},
title = {{BERT}imbau: pretrained {BERT} models for {B}razilian {P}ortuguese},
booktitle = {9th Brazilian Conference on Intelligent Systems, {BRACIS}, Rio Grande do Sul, Brazil, October 20-23 (to appear)},
year = {2020}
}