undirected-generation-dev

This repo contains the source code of the models described in the following paper

• "Learning and Analyzing Generation Order for Undirected Sequence Models" in Findings of EMNLP, 2021. (paper).

The basic code structure was adapted from the NYU dl4mt-seqgen. We also use the pybleu from fairseq to calculate BLEU scores during the reinforcement learning.

0. Preparation

0.1 Dependencies

• PyTorch 1.4.0/1.6.0/1.8.0

0.2 Data

The WMT'14 De-En data and the pretrained De-En MLM model are provided in the dl4mt-seqgen.

• Download WMT'14 De-En valid/test data.
• Then organize the data in data/ and make sure it follows such a structure:

------ data
--------- de-en
------------ train.de-en.de.pth
------------ train.de-en.en.pth
------------ valid.de-en.de.pth
------------ valid.de-en.en.pth
------------ test.de-en.de.pth
------------ test.de-en.en.pth

• Download pretrained models.
• Then organize the pretrained masked language models in models/ make sure it follows such a structure:

------ models
--------- best-valid_en-de_mt_bleu.pth
--------- best-valid_de-en_mt_bleu.pth

2. Training the order policy network with reinforcement learning

Train a policy network to predict the generation order for a pretrained De-En masked language model:

./train_scripts/train_order_rl_deen.sh

• By defaults, the model checkpoints will be saved in models/learned_order_deen_uniform_4gpu/00_maxlen30_minlen5_bsz32.
• By using this script, we are only training the model on De-En sentence pairs where both the German and English sentences with a maximum length of 30 and a minimum length of 5. You can change the training parameters max_len and min_len to change the length limits.

3. Decode the undirected generation model with learned orders

• Set the MODEL_CKPT parameter to the corresponding path found under models/00_maxlen30_minlen5_bsz32. For example:

export MODEL_CKPT=wj8oc8kab4/checkpoint_epoch30+iter96875.pth

• Evaluate the model on the SCAN MCD1 splits by running:

export MODEL_CKPT=...
./eval_scripts/generate-order-deen.sh $MODEL_CKPT

4. Decode the undirected generation model with heuristic orders

• Left2Right

./eval_scripts/generate-deen.sh left_right_greedy_1iter

• Least2Most

./eval_scripts/generate-deen.sh least_most_greedy_1iter

• EasyFirst

./eval_scripts/generate-deen.sh easy_first_greedy_1iter

• Uniform

./eval_scripts/generate-deen.sh uniform_greedy_1iter

Citation

@inproceedings{jiang-bansal-2021-learning-analyzing,
    title = "Learning and Analyzing Generation Order for Undirected Sequence Models",
    author = "Jiang, Yichen  and
      Bansal, Mohit",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-emnlp.298",
    pages = "3513--3523",
}