/pegasus

Primary LanguagePythonApache License 2.0Apache-2.0

PEGASUS library

Pre-training with Extracted Gap-sentences for Abstractive SUmmarization Sequence-to-sequence models, or PEGASUS, uses self-supervised objective Gap Sentences Generation (GSG) to train a transformer encoder-decoder model. The paper can be found on arXiv. ICML 2020 accepted.

If you use this code or these models, please cite the following paper:

@misc{zhang2019pegasus,
    title={PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization},
    author={Jingqing Zhang and Yao Zhao and Mohammad Saleh and Peter J. Liu},
    year={2019},
    eprint={1912.08777},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Results update

We train a dynamic gap sentence generation model on both C4 and HugeNews, and dynamicly sample important sentences. The updated the results are reported in this table.

dataset C4 HugeNews Mixed & Dynamic
xsum 45.20/22.06/36.99 47.21/24.56/39.25 47.60/24.83/39.64
cnn_dailymail 43.90/21.20/40.76 44.17/21.47/41.11 44.16/21.56/41.30
newsroom 45.07/33.39/41.28 45.15/33.51/41.33 45.98/34.20/42.18
multi_news 46.74/17.95/24.26 47.52/18.72/24.91 47.65/18.75/24.95
gigaword 38.75/19.96/36.14 39.12/19.86/36.24 39.65/20.47/36.76
wikihow 43.07/19.70/34.79 41.35/18.51/33.42 46.39/22.12/38.41 *
reddit_tifu 26.54/8.94/21.64 26.63/9.01/21.60 27.99/9.81/22.94
big_patent 53.63/33.16/42.25 53.41/32.89/42.07 52.29/33.08/41.66 *
arxiv 44.70/17.27/25.80 44.67/17.18/25.73 44.21/16.95/25.67
pubmed 45.49/19.90/27.69 45.09/19.56/27.42 45.97/20.15/28.25
aeslc 37.69/21.85/36.84 37.40/21.22/36.45 37.68/21.25/36.51
billsum 57.20/39.56/45.80 57.31/40.19/45.82 59.67/41.58/47.59

The "Mixed & Dynamic" model has the following changes:

  • trained on both C4 and HugeNews (dataset mixture is weighted by their number of examples).
  • trained for 1.5M instead of 500k (we observe slower convergence on pretraining perplexity).
  • the model dynamicly choose 15%-45% important sentences to generate
  • importance sentences are sampled instead of using a fixed strategy (This is done by adding a 20% noise to importance scores.)
  • the sentencepiece tokenizer is updated to be able to encode newline character.

(*) the numbers of wikihow and big_patent datasets are not comparable because of change in tokenization and data:

  • wikihow dataset contains newline characters which is useful for paragraph segmentation, the C4 and HugeNews model's sentencepiece tokenizer doesn't encode newline and loose this information.
  • we update the BigPatent dataset to preserve casing, some format cleanings are also changed, please refer to change in TFDS.

Setup

create an instance on google cloud with GPU (optional)

Please create a project first and create an instance

gcloud compute instances create \
  ${VM_NAME} \
  --zone=${ZONE} \
  --machine-type=n1-highmem-8 \
  --accelerator type=nvidia-tesla-v100,count=1 \
  --boot-disk-size=500GB \
  --image-project=ml-images \
  --image-family=tf-1-15 \
  --maintenance-policy TERMINATE --restart-on-failure

install library and dependencies

Clone library on github and install requirements.

git clone https://github.com/google-research/pegasus
cd pegasus
export PYTHONPATH=.
pip3 install -r requirements.txt

Follow the instruction and install gsutil.

Download vocab, pretrained and fine-tuned checkpoints of the "Mixed & Dynamic" model .

mkdir ckpt
gsutil cp -r gs://pegasus_ckpt/ ckpt/

Finetuning on downstream datasets

on existing dataset

Finetune on an existing dataset aeslc.

python3 pegasus/bin/train.py --params=aeslc_transformer \
--param_overrides=vocab_filename=ckpt/pegasus_ckpt/c4.unigram.newline.10pct.96000.model \
--train_init_checkpoint=ckpt/pegasus_ckpt/model.ckpt-1500000 \
--model_dir=ckpt/pegasus_ckpt/aeslc

Evaluate on the finetuned dataset.

python3 pegasus/bin/evaluate.py --params=aeslc_transformer \
--param_overrides=vocab_filename=ckpt/pegasus_ckpt/c4.unigram.newline.10pct.96000.model,batch_size=1,beam_size=5,beam_alpha=0.6 \
--model_dir=ckpt/pegasus_ckpt/aeslc

Note that the above example is using a single GPU so the batch_size is much smaller than the results reported in the paper.

add new finetuning dataset

Two types of dataset format are supported: TensorFlow Datasets (TFDS) or TFRecords.

This tutorial shows how to add a new dataset in TFDS. (The fine-tuning dataset is expected to be supervised, please provide supervised_keys in dataset info).

Tfrecords format requires each record to be a tf example of {"inputs":tf.string, "targets":tf.string}.

For example, if you registered a TFDS dataset called new_tfds_dataset for training and evaluation, and have some files in tfrecord format called new_dataset_files.tfrecord* for test, they can be registered in /pegasus/params/public_params.py.

@registry.register("new_params")
def my_param(param_overrides):
  return public_params.transformer_params(
      {
          "train_pattern": "tfds:new_tfds_dataset,train",
          "dev_pattern": "tfds:new_tfds_dataset,validation",
          "test_pattern": "tfrecord:new_dataset_files.tfrecord*",
          "max_input_len": 512,
          "max_output_len": 128,
          "train_steps": 10000,
          "learning_rate": 0.0001,
          "batch_size": 8,
      }, param_overrides)

Evaluation metrics.

Evaluation results can be found in mode_dir. Summarization metrics are automatically calculated for each evaluation point.

  • ROUGE is the main metric for summarization quality.

  • BLEU is an alternative quality metric for language generation.

  • Extractive Fragments Coverage & Density are metrics that measures the abstractiveness of the summary.

  • Repetition Rates measures generation repetition failure modes.

  • Length statistics measures the length distribution of decodes comparing to gold summary.

Several types of output files can be found in model_dir

  • text_metrics-*.txt: above metrics in text format. Each row contains metric name, 95% lower bound value, mean value, 95% upper bound value.
  • inputs-.txt, targets-.txt, predictions-*.txt: raw text files of model inputs/outputs.

Pre-training

Pretraining (on C4 or any other corpus) requires a customly built tensorflow that includes ops for on-the-fly parsing that processes raw text document into model inputs and targets ids. Please refer to pegasus/ops/pretrain_parsing_ops.cc and pegasus/data/parsers.py for details.