This repository contains the code for our EMNLP 2020 paper:
Blank Language Models
Tianxiao Shen*, Victor Quach*, Regina Barzilay, and Tommi Jaakkola (*: Equal contribution)

Given partially specified text with one or more blanks, BLM will fill in the blanks with a variable number of tokens consistent with the context, making it ideal for text editing and rewriting.

Input: They also have ___ which ___ .
Output: They also have ice cream which is really good .

We have an online demo built using streamlit, available here

Or try locally by running:

streamlit run app.py

Our code is based on the PyTorch Lightning framework.

It has been tested in PyTorch 1.6.0, PyTorch Lightning 1.0.7

Download the processed Yelp and Yahoo datasets by running:

bash download_data.sh

To train a BLM on Yelp negative sentences:

python train.py --train data/yelp/train.0 --valid data/yelp/valid.0 --root_dir checkpoints/yelp/neg/blm/ \
--vocab_size 10000 --max_len 20 --model_type blm --share_emb_prj_weight

Yelp positive sentences:

python train.py --train data/yelp/train.1 --valid data/yelp/valid.1 --root_dir checkpoints/yelp/pos/blm/ \
--vocab_size 10000 --max_len 20 --model_type blm --share_emb_prj_weight

Yahoo documents:

python train.py --train data/yahoo/train.txt --valid data/yahoo/valid.txt --root_dir checkpoints/yahoo/blm/ \
--vocab_size 20000 --max_len 205 --model_type blm --share_emb_prj_weight

Run python train.py -h to see all training options.

You can use Tensorboard to monitor the training progress.

After training, we can evaluate the model's perplexity by Monte Carlo estimate, and use the model to generate text from scratch or fill in the blanks in the input.

For all of the following, replace epoch\=???.ckpt with the checkpoint saved in training.

• The following command evaluates for Yelp negative sentences:

python test.py --checkpoint checkpoints/yelp/neg/blm/lightning_logs/version_0/checkpoints/epoch\=???.ckpt \
--eval data/yelp/test.0 --n_mc 10

• The following command samples from the model trained on Yelp negative sentences:

python test.py --checkpoint checkpoints/yelp/neg/blm/lightning_logs/version_0/checkpoints/epoch\=???.ckpt \
--sample 1000 --decode sample --output sample.txt

• The following command uses the model trained on Yelp negative sentences to fill in blanked positive sentences to achieve sentiment transfer:

python test.py --checkpoint checkpoints/yelp/neg/blm/lightning_logs/version_0/checkpoints/epoch\=???.ckpt \
--fill data/yelp/blank/test.1.blank --output test.1.tsf

To output the whole generation trajectory, turn on the --write_mid option.

The output file will be stored in outputs/ within the checkpoint directory.

We use the Transformer implementation from https://github.com/jadore801120/attention-is-all-you-need-pytorch

If you use our work, please cite:

@inproceedings{shen2020blank,
    title = "{Blank Language Models}",
    author = "Shen, Tianxiao  and
      Quach, Victor  and
      Barzilay, Regina  and
      Jaakkola, Tommi",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics"
}