This repository contains source code for our EMNLP 2020 paper "Reusing a Pretrained Language Model on Languages with Limited Corpora for Unsupervised NMT" (Paper link)
Introduction
This paper presents a method to fine-tune of a pretrained monolingual LM (on a high-resource language) to a low-resource language, to serve as initialization of an Unsupervised NMT (UNMT) encoder-decoder model. To this end, we propose a vocabulary extension method to allow fine-tuning.
Our method, entitled RE-LM, provides very competitive UNMT results in low-resource - high-resource language pairs, outperforming XLM in English-Macedonian (En-Mk) and English-Albanian (En-Sq), yielding more than +8.3 BLEU points for all four translation directions.
This code is built on using the XLM baseline, which is publicly available.
Prerequisites
Dependencies
- Python 3.6.9
- NumPy (tested on version 1.15.4)
- PyTorch (tested on version 1.2.0)
- Apex (for fp16 training)
Install Requirements
Create Environment (Optional): Ideally, you should create a conda environment for the project.
conda create -n relm python=3.6.9
conda activate relm
Install PyTorch 1.2.0 with the desired cuda version to use the GPU:
conda install pytorch==1.2.0 torchvision -c pytorch
Clone the project:
git clone https://github.com/alexandra-chron/relm_unmt.git
cd relm_unmt
Then install the rest of the requirements:
pip install -r ./requirements.txt
It is HIGHLY recommended to use half precision (using Apex) by simply adding --fp16 True --amp 1 flags to each running command. Without it, you might run out of memory.
To train with multiple GPUs use:
export NGPU=8; python -m torch.distributed.launch --nproc_per_node=$NGPU train.py
To train with multiple GPUs and half precision use:
export NGPU=8; python -m torch.distributed.launch --nproc_per_node=$NGPU train.py --fp16 True --amp 1
Download data
We sample 68M English sentences from Newscrawl.
We use Macedonian and Albanian Common Crawl deduplicated monolingual data from the OSCAR corpus.
Our validation and test data is created by sampling from the SETIMES parallel En-Mk, En-Sq corpora.
To allow reproducing our results, we provide the validation and test data in ./data/mk-en and ./data/sq-en directories.
Preprocessing for pretraining
Before pretraining an HMR (high-monolingual-resource) monolingual MLM, make sure you
have downloaded the HMR data and placed it in ./data/$HMR/ directory.
The data should be in the form: {train_raw, valid_raw, test_raw}.$HMR.
After that, run the following (example for En):
./get_data_mlm_pretraining.sh --src en
RE-LM
1. Train a monolingual LM
Train your monolingual masked LM (BERT without the next-sentence prediction task) on the monolingual data:
python train.py \
--exp_name mono_mlm_en_68m \
--dump_path ./dumped \
--data_path ./data/en/ \
--lgs 'en' \
--mlm_steps 'en' \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout '0.1' \
--attention_dropout '0.1' \
--gelu_activation true \
--batch_size 32 \
--bptt 256 \
--optimizer 'adam,lr=0.0001' \
--epoch_size 200000 \
--validation_metrics _valid_en_mlm_ppl \
--stopping_criterion '_valid_en_mlm_ppl,10'
## There are other parameters that are not specified here (see train.py).
Alternatively, you can download the pretrained En MLM (as well as the vocabulary and BPE codes) that we used using this link:
wget https://www.dropbox.com/s/nrhnyyjuncbwx71/pretrained_en_mlm.zip?dl=0
Preprocessing for fine-tuning (and UNMT)
Before fine-tuning the pretrained MLM and running UNMT, make sure you
have downloaded the LMR data and placed it in ./data/$LMR-$HMR/ directory.
The data should be in the form: {train_raw, valid_raw, test_raw}.$LMR.
Then, run the following (example for En, Mk):
./get_data_and_preprocess.sh --src en --tgt mk
In Step 2, the embedding layer (and the output layer) of the MLM model will be increased by the amount of new items added to the existing vocabulary.
In the directory ./data/$LMR-$HMR/, a file named vocab.$LMR-$HMR-ext-by-$NUMBER has been created.
This number indicates by how many items we need to extend the initial vocabulary, and consequently
the embedding and linear layer, to account for the LMR language.
You will need to give this value to the --increase_vocab_by argument so that you successfully run fine-tuning (step 2).
2. Fine-tune it on both the LMR and HMR languages
python train.py \
--exp_name finetune_en_mlm_mk \
--dump_path ./dumped/ \
--reload_model 'mono_mlm_en_68m.pth' \
--data_path ./data/mk-en/ \
--lgs 'en-mk' \
--mlm_steps 'mk,en' \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--batch_size 32 \
--bptt 256 \
--optimizer adam,lr=0.0001 \
--epoch_size 50000 \
--validation_metrics _valid_mk_mlm_ppl \
--stopping_criterion _valid_mk_mlm_ppl,10 \
--increase_vocab_for_lang en \
--increase_vocab_from_lang mk \
--increase_vocab_by NUMBER #(see ./data/mk-en/vocab.mk-en-ext-by-$NUMBER)
3. Train a UNMT model (encoder and decoder initialized with RE-LM)
python train.py \
--exp_name unsupMT_ft_mk \
--dump_path ./dumped/ \
--reload_model 'finetune_en_mlm_mk.pth,finetune_en_mlm_mk.pth' \
--data_path './data/mk-en' \
--lgs en-mk \
--ae_steps en,mk \
--bt_steps en-mk-en,mk-en-mk \
--word_shuffle 3 \
--word_dropout 0.1 \
--word_blank 0.1 \
--lambda_ae 0:1,100000:0.1,300000:0 \
--encoder_only False \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--tokens_per_batch 1000 \
--batch_size 32 \
--bptt 256 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 50000 \
--eval_bleu true \
--stopping_criterion valid_mk-en_mt_bleu,10 \
--validation_metrics valid_mk-en_mt_bleu \
--increase_vocab_for_lang en \
--increase_vocab_from_lang mk
RE-LM + adapters
1. This step is the same as with RE-LM.
2. Fine-tune part of the model on the target language only using adapters
python train.py \
--exp_name finetune_en_mlm_mk_adapters \
--dump_path ./dumped/ \
--reload_model 'mono_mlm_en_68m.pth' \
--data_path ./data/mk-en \
--lgs 'en-mk' \
--clm_steps '' \
--mlm_steps 'mk' \
--mlm_eval_steps 'en' \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--use_adapters True \
--adapter_size 256 \
--batch_size 32 \
--bptt 256 \
--optimizer adam,lr=0.0001 \
--epoch_size 50000 \
--validation_metrics _valid_mk_mlm_ppl \
--stopping_criterion _valid_mk_mlm_ppl,10 \
--increase_vocab_for_lang en \
--increase_vocab_from_lang mk \
--increase_vocab_by NUMBER #(see ./data/mk-en/vocab.mk-en-ext-by-$NUMBER)
3. Train a UNMT model (encoder and decoder initialized with RE-LM + adapters)
python train.py \
--exp_name unsupMT_ft_mk \
--dump_path ./dumped/ \
--reload_model 'finetune_en_mlm_mk_adapters.pth,finetune_en_mlm_mk_adapters.pth' \
--data_path './data/mk-en' \
--lgs en-mk \
--ae_steps en,mk \
--bt_steps en-mk-en,mk-en-mk \
--word_shuffle 3 \
--word_dropout 0.1 \
--word_blank 0.1 \
--lambda_ae 0:1,100000:0.1,300000:0 \
--encoder_only False \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--use_adapters True \
--adapter_size 256 \
--tokens_per_batch 1000 \
--batch_size 32 \
--bptt 256 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 50000 \
--eval_bleu true \
--stopping_criterion valid_mk-en_mt_bleu,10 \
--validation_metrics valid_mk-en_mt_bleu \
--increase_vocab_for_lang en \
--increase_vocab_from_lang mk
For the XLM baseline, follow the instructions in XLM github page
If you use our work, please cite our paper:
Reference
@inproceedings{chronopoulou-etal-2020-reusing,
title = "{R}eusing a {P}retrained {L}anguage {M}odel on {L}anguages with {L}imited {C}orpora for {U}nsupervised {NMT}",
author = "Chronopoulou, Alexandra and
Stojanovski, Dario and
Fraser, Alexander",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.214",
pages = "2703--2711",
}