Table of contents
BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese
Two BARTpho versions BARTpho-syllable and BARTpho-word are the first public large-scale monolingual sequence-to-sequence models pre-trained for Vietnamese. BARTpho uses the "large" architecture and pre-training scheme of the sequence-to-sequence denoising model BART, thus especially suitable for generative NLP tasks. Experiments on a downstream task of Vietnamese text summarization show that in both automatic and human evaluations, BARTpho outperforms the strong baseline mBART and improves the state-of-the-art.
The general architecture and experimental results of BARTpho can be found in our paper:
@article{bartpho,
title = {{BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese}},
author = {Nguyen Luong Tran and Duong Minh Le and Dat Quoc Nguyen},
journal = {arXiv preprint},
volume = {arXiv:2109.09701},
year = {2021}
}
Please CITE our paper when BARTpho is used to help produce published results or incorporated into other software.
Using BARTpho in fairseq
Installation
There is an issue w.r.t. the encode function in the BART hub_interface, as discussed in this pull request https://github.com/pytorch/fairseq/pull/3905. While waiting for this pull request's approval, please install fairseq as follows:
git clone https://github.com/datquocnguyen/fairseq.git
cd fairseq
pip install --editable ./
Pre-trained models
| Model | #params | Download | Input text |
|---|---|---|---|
| BARTpho-syllable | 396M | fairseq-bartpho-syllable.zip | Syllable level |
| BARTpho-word | 420M | fairseq-bartpho-word.zip | Word level |
unzip fairseq-bartpho-syllable.zipunzip fairseq-bartpho-word.zip
Example usage
from fairseq.models.bart import BARTModel
#Load BARTpho-syllable model:
model_folder_path = '/PATH-TO-FOLDER/fairseq-bartpho-syllable/'
spm_model_path = '/PATH-TO-FOLDER/fairseq-bartpho-syllable/sentence.bpe.model'
bartpho_syllable = BARTModel.from_pretrained(model_folder_path, checkpoint_file='model.pt', bpe='sentencepiece', sentencepiece_model=spm_model_path).eval()
#Input syllable-level/raw text:
sentence = 'Chúng tôi là những nghiên cứu viên.'
#Apply SentencePiece to the input text
tokenIDs = bartpho_syllable.encode(sentence, add_if_not_exist=False)
#Extract features from BARTpho-syllable
last_layer_features = bartpho_syllable.extract_features(tokenIDs)
##Load BARTpho-word model:
model_folder_path = '/PATH-TO-FOLDER/fairseq-bartpho-word/'
bpe_codes_path = '/PATH-TO-FOLDER/fairseq-bartpho-word/bpe.codes'
bartpho_word = BARTModel.from_pretrained(model_folder_path, checkpoint_file='model.pt', bpe='fastbpe', bpe_codes=bpe_codes_path).eval()
#Input word-level text:
sentence = 'Chúng_tôi là những nghiên_cứu_viên .'
#Apply BPE to the input text
tokenIDs = bartpho_word.encode(sentence, add_if_not_exist=False)
#Extract features from BARTpho-word
last_layer_features = bartpho_word.extract_features(tokenIDs)
Using BARTpho in transformers
Installation
- Installation with pip (v4.12+):
pip install transformers - Installing from source:
git clone https://github.com/huggingface/transformers.git
cd transformers
pip install -e .
Pre-trained models
| Model | #params | Input text |
|---|---|---|
vinai/bartpho-syllable |
396M | Syllable level |
vinai/bartpho-word |
420M | Word level |
Example usage
import torch
from transformers import AutoModel, AutoTokenizer
#BARTpho-syllable
syllable_tokenizer = AutoTokenizer.from_pretrained("vinai/bartpho-syllable", use_fast=False)
bartpho_syllable = AutoModel.from_pretrained("vinai/bartpho-syllable")
TXT = 'Chúng tôi là những nghiên cứu viên.'
input_ids = syllable_tokenizer(TXT, return_tensors='pt')['input_ids']
features = bartpho_syllable(input_ids)
from transformers import MBartForConditionalGeneration
bartpho_syllable = MBartForConditionalGeneration.from_pretrained("vinai/bartpho-syllable")
TXT = 'Chúng tôi là <mask> nghiên cứu viên.'
input_ids = syllable_tokenizer(TXT, return_tensors='pt')['input_ids']
logits = bartpho_syllable(input_ids).logits
masked_index = (input_ids[0] == syllable_tokenizer.mask_token_id).nonzero().item()
probs = logits[0, masked_index].softmax(dim=0)
values, predictions = probs.topk(5)
print(syllable_tokenizer.decode(predictions).split())
#BARTpho-word
word_tokenizer = AutoTokenizer.from_pretrained("vinai/bartpho-word", use_fast=False)
bartpho_word = AutoModel.from_pretrained("vinai/bartpho-word")
TXT = 'Chúng_tôi là những nghiên_cứu_viên .'
input_ids = word_tokenizer(TXT, return_tensors='pt')['input_ids']
features = bartpho_word(input_ids)
bartpho_word = MBartForConditionalGeneration.from_pretrained("vinai/bartpho-word")
TXT = 'Chúng_tôi là những <mask> .'
input_ids = word_tokenizer(TXT, return_tensors='pt')['input_ids']
logits = bartpho_word(input_ids).logits
masked_index = (input_ids[0] == word_tokenizer.mask_token_id).nonzero().item()
probs = logits[0, masked_index].softmax(dim=0)
values, predictions = probs.topk(5)
print(word_tokenizer.decode(predictions).split())
- Following mBART, BARTpho uses the "large" architecture of BART with an additional layer-normalization layer on top of both the encoder and decoder. Thus, when converted to be used with
transformers, BARTpho can be called via mBART-based classes.
Notes
- Before fine-tuning BARTpho on a downstream task, users should perform Vietnamese tone normalization on the downstream task's data as this pre-process was also applied to the pre-training corpus. A Python script for Vietnamese tone normalization is available at HERE.
- For
BARTpho-word, users should use VnCoreNLP to segment input raw texts as it was used to perform both Vietnamese tone normalization and word segmentation on the pre-training corpus.
License
MIT License
Copyright (c) 2021 VinAI Research
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
27 Oct 30, 2022
41 Dec 15, 2022
17 Nov 22, 2022
126 Jan 02, 2023
8 Dec 25, 2022
3 Jul 24, 2022
19 Dec 13, 2022
19 Oct 28, 2022
62 Oct 02, 2022
409 Jan 06, 2023
1 Feb 14, 2022
4 Feb 18, 2022
2.2k Dec 27, 2022
37 Dec 14, 2022
34 Dec 18, 2022
3 Jun 20, 2022
46 Dec 15, 2022
167 Jan 03, 2023
586 Dec 24, 2022
7 Aug 25, 2022