Planning and Iterative Refinement in Pre-trained Transformers for Long Text Generation

pair-emnlp2020

Official repository for the paper:

Xinyu Hua and Lu Wang: PAIR: Planning and Iterative Refinement in Pre-trained Transformers for Long Text Generation

If you find our work useful, please cite:

@inproceedings{hua-wang-2020-pair,
    title = "PAIR: Planning and Iterative Refinement in Pre-trained Transformersfor Long Text Generation",
    author = "Hua, Xinyu  and
      Wang, Lu",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
}

Requirements

Python 3.7
PyTorch 1.4.0
PyTorchLightning 0.9.0
transformers 3.3.0
numpy
tqdm
pycorenlp (for preprocessing nytimes data)
nltk (for preprocessing nytimes data)

Data

We release the data sets in the following link(1.2G uncompressed)
Please download and uncompress the file, and put under ./data directory.
For opinion and news domains, the The New York Times Annotated Corpus is licensed by LDC.
We therefore only provide the ids for train/dev/test.
Please follow the instructions to generate the dataset.

Text Planning

To train a BERT planner:

cd planning
python train.py \
    --data-path=../data/ \
    --domain=[arggen,opinion,news] \
    --exp-name=demo \
    --save-interval=1 \ # how frequent to save checkpoints 
    --max-epoch=30 \
    --lr=5e-4 \
    --warmup-updates=5000 \
    --train-set=train \
    --valid-set=dev \
    --tensorboard-logdir=tboard/ \
    --predict-keyphrase-offset \
    --max-samples=32 \ # max number of samples per batch
    [--quiet] \ # whether to print intermediate information

The checkpoints will be dumped to checkpoints/planning/[domain]/[exp-name].
Tensorboard will be available under planning/tboard/.

To run inference using a trained model, with greedy decoding:

cd planning
python decode.py \
    --data-path=../data/ \
    --domain=arggen \
    --test-set=test \
    --max-samples=32 \
    --predict-keyphrase-offset \
    --exp-name=demo \
    [--quiet]

The results will be saved to planning/output/.

Iterative Refinement

We provide implementations for four different setups:

Seq2seq: prompt -> tgt
KPSeq2seq: prompt + kp-set -> tgt
PAIR-light: prompt + kp-plan + masks -> tgt
PAIR-full: prompt + kp-plan + template -> tgt

To train a model:

cd refinement
python train.py \
    --domain=[arggen,opinion,news] \
    --setup=[seq2seq,kpseq2seq,pair-light,pair-full] \
    --train-set=train \
    --valid-set=dev \
    --train-batch-size=10 \
    --valid-batch-size=5 \
    --num-train-epochs=20 \
    --ckpt-dir=../checkpoints/[domain]/[setup]/demo \
    --tensorboard-dir=demo \
    [--quiet]

To run iterative refinement:

cd refinement
python generate.py \
    --domain=[arggen,opinion,news] \
    --setup=[seq2seq,kpseq2seq,pair-light,pair-full] \
    --test-set=test \
    --output-name=test_demo \
    --enforce-template-strategy=flexible \
    --do-sampling \
    --sampling-topk=100 \
    --sampling-topp=0.9 \
    --sample-times=3 \
    --ckpt-dir=../checkpoints/[domain]/[setup]/demo