MASS is a novel pre-training method for sequence to sequence based language generation tasks. It randomly masks a sentence fragment in the encoder, and then predicts it in the decoder.
MASS can be applied on cross-lingual tasks such as neural machine translation (NMT), and monolingual tasks such text summarization. The current codebase supports unsupervised and supervised NMT, text summarization and conversational response generation. We will release our implementation for other sequence to sequence generation tasks in the future.
MASS contains the following codes:
- Unsupervised Neural Machine Translation
- Supervised Neural Machine Translation
- Text Summarization
- Conversational Response Generation
!!!In order to satisfy more users and developers in sequence to sequence tasks, we also implement MASS on fairseq, which will foucs more on the pre-training and fine-tuning in large scale supervised tasks, including neural machine translation, text summarization, grammatical error correction, etc. We release the codes in MASS-fairseq folder.
Currently we implement MASS based on the codebase of XLM. The depencies are as follows:
- Python 3
- NumPy
- PyTorch (version 0.4 and 1.0)
- fastBPE (for BPE codes)
- Moses (for tokenization)
- Apex (for fp16 training)
Unsupervised Neural Machine Translation just uses monolingual data to train the models. During MASS pre-training, the source and target languages are pre-trained in one model, with the corresponding langauge embeddings to differentiate the langauges. During MASS fine-tuning, back-translation is used to train the unsupervised models. We provide pre-trained and fine-tuned models:
| Languages | Pre-trained Model | Fine-tuned Model | BPE codes | Vocabulary |
|---|---|---|---|---|
| EN - FR | MODEL | MODEL | BPE codes | Vocabulary |
| EN - DE | MODEL | MODEL | BPE codes | Vocabulary |
| En - RO | MODEL | MODEL | BPE_codes | Vocabulary |
We are also preparing larger models on more language pairs, and will release them in the future.
We use the same BPE codes and vocabulary with XLM. Here we take English-French as an example.
cd MASS
wget https://dl.fbaipublicfiles.com/XLM/codes_enfr
wget https://dl.fbaipublicfiles.com/XLM/vocab_enfr
./get-data-nmt.sh --src en --tgt fr --reload_codes codes_enfr --reload_vocab vocab_enfr
python train.py \
--exp_name unsupMT_enfr \
--data_path ./data/processed/en-fr/ \
--lgs 'en-fr' \
--mass_steps 'en,fr' \
--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 3000 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 100 \
--eval_bleu true \
--word_mass 0.5 \
--min_len 5 \
During the pre-training prcess, even without any back-translation, you can observe the model can achieve some intial BLEU scores:
epoch -> 4
valid_fr-en_mt_bleu -> 10.55
valid_en-fr_mt_bleu -> 7.81
test_fr-en_mt_bleu -> 11.72
test_en-fr_mt_bleu -> 8.80
After pre-training, we use back-translation to fine-tune the pre-trained model on unsupervised machine translation:
MODEL=mass_enfr_1024.pth
python train.py \
--exp_name unsupMT_enfr \
--data_path ./data/processed/en-fr/ \
--lgs 'en-fr' \
--bt_steps 'en-fr-en,fr-en-fr' \
--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 2000 \
--batch_size 32 \
--bptt 256 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 30 \
--eval_bleu true \
--reload_model "$MODEL,$MODEL" \
During MASS pre-training, the source and target languages are pre-trained in one model, with the corresponding langauge embeddings to differentiate the langauges. During MASS fine-tuning, supervised sentence pairs are directly used to train the NMT models. We provide pre-trained and fine-tuned models:
| Languages | Fine-tuned Model | BPE codes | Vocabulary | BLEU |
|---|---|---|---|---|
| Ro-En | MODEL | BPE codes | Vocabulary | WMT16 Ro-En: 39.1 |
We will release the pre-trained and fine-tuned models for other langauge pairs in the future.
Here is an example to show how to run mass fine-tuning on the WMT16 en-ro dataset.
wget https://dl.fbaipublicfiles.com/XLM/codes_enro
wget https://dl.fbaipublicfiles.com/XLM/vocab_enro
./get-data-bilingual-enro-nmt.sh --src en --tgt fr --reload_codes codes_enro --reload_vocab vocab_enro
Download the mass pre-trained model from the above link. And use the following command to fine tune:
DATA_PATH=./data/processed/en-ro
MODEL=mass_enro_1024.pth
python train.py \
--exp_name unsupMT_enro \
--dump_path ./models/en-ro/ \
--exp_id wmt16_enro_ft \
--data_path $DATA_PATH \
--lgs 'en-ro' \
--bt_steps 'en-ro-en,ro-en-ro' \
--encoder_only false \
--mt_steps 'en-ro,ro-en' \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--tokens_per_batch 2000 \
--batch_size 32 \
--bptt 256 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 50 \
--eval_bleu true \
--reload_model "$MODEL,$MODEL"
To apply MASS on text summarization, we provide an example of how to run MASS pre-training and fine-tuning on the Gigaword dataset.
| Pre-trained Model | BPE codes | Vocabulary |
|---|---|---|
| MODEL | BPE codes | Vocabulary |
Our model is still under training. We will update it if have a better model.
For pre-training, we use the following command:
python train.py \
--exp_name mass_english \
--data_path ./data/processed/en/ \
--lgs 'en' \
--mass_steps 'en' \
--encoder_only false \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.1 \
--attention_dropout 0.1 \
--gelu_activation true \
--tokens_per_batch 3000 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 100 \
--eval_bleu true \
--word_mass 0.5 \
--min_len 5 \
--english_only true
Different from unsupervised NMT tasks, we directly use paired data (article-title) to fine-tune the pre-trained model. The fine-tuning command is:
MODEL=mass_en_1024.pth
python train.py \
--exp_name mass_summarization \
--data_path ./data/processed/summarization/ \
--lgs 'ar-ti' \
--mt_steps 'ar-ti' \
--encoder_only false \
--emb_dim 1024 \
--n_layers 6 \
--n_heads 8 \
--dropout 0.2 \
--attention_dropout 0.2 \
--gelu_activation true \
--tokens_per_batch 3000 \
--optimizer adam_inverse_sqrt,beta1=0.9,beta2=0.98,lr=0.0001 \
--epoch_size 200000 \
--max_epoch 20 \
--eval_bleu true \
--english_only true \
--reload_model "$MODEL,$MODEL"
To be updated soon.
If you find MASS useful in your work, you can cite the paper as below:
@article{song2019mass,
title={Mass: Masked sequence to sequence pre-training for language generation},
author={Song, Kaitao and Tan, Xu and Qin, Tao and Lu, Jianfeng and Liu, Tie-Yan},
journal={arXiv preprint arXiv:1905.02450},
year={2019}
}