Disclaimer: This is not an official Google product.

Overview

This directory contains code for generating the data and model described in "SummAE: Zero-Shot Abstractive Text Summarization using Length-Agnostic Auto-Encoders".

For questions or issues, contact [email protected].

Install dependencies (run.sh)

This package depends on Tensorflow and google_research/rouge. See the included run.sh for how to install and run a unit test inside of a virtualenv.

Generate rocstories data.

Request (free) and download raw data for ROCStories corpora into a directory pointed to by the environment variable ROCSTORIES_RAW. This directory should contain the following files:

• "ROCStories_winter2017 - ROCStories_winter2017.csv"
• "ROCStories__spring2016 - ROCStories_spring2016.csv"

Then run the data processing script:

export ROCSTORIES_RAW=absolutepathto/raw_rocstories
export ROCSTORIES_DATA=absolutepathto/processed_rocstoriesdata

Inside google_research/google_research directory:

bash summae/generate_data.sh $ROCSTORIES_RAW summae/testdata $ROCSTORIES_DATA

Verify data

python -m summae.verify_data --data_dir=$ROCSTORIES_DATA

Running the model code

Train from scratch for a few steps using run_locally.sh

export HYPERS=`pwd`/testdata/hypers.json
bash summae/run_locally.sh train /tmp/testmodel

Decode latest model checkpoint

bash summae/run_locally.sh decode /tmp/testmodel 0

Run decode of best model

mkdir /tmp/best
cp -r summae/testdata/best /tmp/
bash summae/run_locally.sh decode /tmp/best 358000

Decodes output is saved to /tmp/best/decodes/.

Citation

If you use this code in research please cite:

@article{liu2019summae,
  title={SummAE: Zero-Shot Abstractive Text Summarization using Length-Agnostic
  Auto-Encoders},
  author={Liu,  Peter J. and Chung, Yu-An and Ren, Jie},
  journal={arXiv preprint arXiv:1910.00998},
  url={http://arxiv.org/abs/1910.00998},
  year={2019}
}