Introduction

This repository contains the code for BERT-fused NMT, which is introduced in the ICLR2020 paper Incorporating BERT into Neural Machine Translation.

If you find this work helpful in your research, please cite as:

@inproceedings{
Zhu2020Incorporating,
title={Incorporating BERT into Neural Machine Translation},
author={Jinhua Zhu and Yingce Xia and Lijun Wu and Di He and Tao Qin and Wengang Zhou and Houqiang Li and Tieyan Liu},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=Hyl7ygStwB}
}

NOTE: We have updated our code to enable you use more powerful pretrained models contained in huggingface/transformers. With bert-base-german-dbmdz-uncased, we get a new result $37.34$ on IWSLT'14 de->en task.

Requirements and Installation

• PyTorch version == 1.0.0/1.1.0
• Python version >= 3.5

Installing from source

To install fairseq from source and develop locally:

git clone https://github.com/bert-nmt/bert-nmt
cd bertnmt
pip install --editable .

Getting Started

Data Preprocessing

First, you should run Fairseq prepaer-xxx.sh to get tokenized&bped files like:

train.en train.de valid.en valid.de test.en test.de

Then you can use makedataforbert.sh to get input file for BERT model (please note that the path is correct). You can get

train.en train.de valid.en valid.de test.en test.de train.bert.en valid.bert.en test.bert.en

Then preprocess data like Fairseq:

python preprocess.py --source-lang src_lng --target-lang tgt_lng \
  --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \
  --destdir destdir  --joined-dictionary --bert-model-name bert-base-uncased

Note: For more language pairs used in our paper, please refer to another repo.

Train a vanilla NMT model using Fairseq

Using data above and standard Fairseq repository, you can get a pretrained NMT model.

Note: The update_freq in iwslt en->zh translation is set to 2, and other hyper-parameters are the same as de<->en

Train a BERT-fused NMT model

The important options we add:

        parser.add_argument('--bert-model-name', default='bert-base-uncased', type=str)
        parser.add_argument('--warmup-from-nmt', action='store_true', )
        parser.add_argument('--warmup-nmt-file', default='checkpoint_nmt.pt', )
        parser.add_argument('--encoder-bert-dropout', action='store_true',)
        parser.add_argument('--encoder-bert-dropout-ratio', default=0.25, type=float)

1. --bert-model-name specify the BERT model name, provided in file.
2. --warmup-from-nmt indicate you will also use a pretrained NMT model to train your BERT-fused NMT model. If you this option, we suggest you use --reset-lr-scheduler, too.
3. --warmup-nmt-file specify the NMT model name (in your $savedir).
4. --encoder-bert-dropout indicate you will use drop-net trick.
5. --encoder-bert-dropout-ratio specify the ratio ($\in [0, 0.5]$) used in drop-net. This is a training script example:

#!/usr/bin/env bash
nvidia-smi

cd /yourpath/bertnmt
python3 -c "import torch; print(torch.__version__)"

src=en
tgt=de
bedropout=0.5
ARCH=transformer_s2_iwslt_de_en
DATAPATH=/yourdatapath
SAVEDIR=checkpoints/iwed_${src}_${tgt}_${bedropout}
mkdir -p $SAVEDIR
if [ ! -f $SAVEDIR/checkpoint_nmt.pt ]
then
    cp /your_pretrained_nmt_model $SAVEDIR/checkpoint_nmt.pt
fi
if [ ! -f "$SAVEDIR/checkpoint_last.pt" ]
then
warmup="--warmup-from-nmt --reset-lr-scheduler"
else
warmup=""
fi

python train.py $DATAPATH \
-a $ARCH --optimizer adam --lr 0.0005 -s $src -t $tgt --label-smoothing 0.1 \
--dropout 0.3 --max-tokens 4000 --min-lr '1e-09' --lr-scheduler inverse_sqrt --weight-decay 0.0001 \
--criterion label_smoothed_cross_entropy --max-update 150000 --warmup-updates 4000 --warmup-init-lr '1e-07' \
--adam-betas '(0.9,0.98)' --save-dir $SAVEDIR --share-all-embeddings $warmup \
--encoder-bert-dropout --encoder-bert-dropout-ratio $bedropout | tee -a $SAVEDIR/training.log

Generate

Using the generate.py to test model is the same as the Fairseq, but you should add --bert-model-name to indicate your BERT model name.

Using the interactive.py to test model is a little different from the Fairseq. You should follow this procedure:

sed -r 's/(@@ )|(@@ ?$)//g' $bpefile > $bpefile.debpe
$MOSE/scripts/tokenizer/detokenizer.perl -l $src < $bpefile.debpe > $bpefile.debpe.detok
paste -d "\n" $bpefile $bpefile.debpe.detok > $bpefile.in
cat $bpefile.in | python interactive.py  -s $src -t $tgt \
--buffer-size 1024 --batch-size 128 --beam 5 --remove-bpe  > output.log