STAS: A Python repository from xssstory

This is an implementation of the STAS (Sentence-level Transformer based Attentive Summarization) model described in Unsupervised Extractive Summarization by Pre-training Hierarchical Transformers

Installation

You need to install python3 and following libararies

pip install pytorch==1.2
pip install pyrouge==0.1.3
pip install pytorch-transformers==1.1.0
python setup.py build
python setup.py develop

# For rouge-1.5.5.pl
sudo apt-get update
sudo apt-get install expat
sudo apt-get install libexpat-dev -y

sudo cpan install XML::Parser
sudo cpan install XML::Parser::PerlSAX
sudo cpan install XML::DOM

We also provide the Dockerfile we used to train and evaluate the model.

Trained models

You can download our released models from here, the files are organized as follows:

.
├── README.md
└── released_model
    ├── cnndm_model
        ├── checkpoint85.pt
        └── ensemble_result
            ├── pacsum
                ├── 61.test.txt
                └── 61.valid.txt
            └── stas
                ├── 13.test.txt
                └── 13.valid.txt
    └── nyt_model
        ├── checkpoint65.pt
        └── ensemble_result
            ├── pacsum
                └── ...
            └── stas
                └── ...

We provide the sentence scores given by STAT and PASUM in the ensemble_result, you can combine the scores following Evaluation 3.

data preprocess

You should split your data into train/validation/test subsets and get 6 files like train.article, train.summary, valid.article, valid.summary, test.article and test.summary, and make sure that each line has one article/summary, the sentence in the article/summary is splited by "<S_SEP>". (we only use summaries for evaluation and test). Here is an example:

Apple 's first generation iPad launched on 3 April 2010 <S_SEP> In its five years on the market , 225 million devices have been sold <S_SEP> But larger smartphones and smart watches may herald its end <S_SEP> Sales for the iPad dropped 18 per cent in the final quarter of 2014

Then run the get-data-bpe.sh (modify the file path in the script accroding to you situation) and you will get a file folder for training and evaluating our model.

Training

We provide the scripts for training on the CNN/DM and NYT datasets, We trained our models with 4 Nvidia Tesla V100GPUs and employed gradient accumulation technique.

bash train_cnndm.sh # For cnndm
bash train_nyt.sh # For nyt

Evaluation

We also provide the steps to evaluate the models.

run the scripts to score the sentences

bash extract_cnndm.sh # for cnndm
bash extract_nyt.sh # for nyt

computing the ROUGE scores

python sum_eval_pipe.py -raw_test=data/cnndm/test -raw_valid=data/cnndm/validation -model_dir=released_model/cnndm_model/85/ # for cnndm
python sum_eval_pipe.py -raw_test=data/nyt/test -raw_valid=data/nyt/valid  -model_dir=released_model/nyt_model/65/ #for nyt

combine the scores given by STAS and PACSUM

python ensemble.py
python evaluate_ensemble.py
# for nyt
python ensemble.py --raw-valid=data/nyt/valid.article --raw-test=data/nyt/test.article --stas-dir=released_model/nyt_model/ensemble_result/stas/ --pacsum-dir=released_model/nyt_model/ensemble_result/pacsum/ --outdir=released_model/nyt_model/ensemble_result/ensenble/ --rerank=False

The generated summaries and ROGUE socres will be stored in the released_model/cnndm_model/ensemble_result/ensemble/test and released_model/cnndm_model/ensemble_result/ensemble/valid .

Citation

@inproceedings{xu-etal-2020-unsupervised,
    title = "Unsupervised Extractive Summarization by Pre-training Hierarchical Transformers",
    author = "Xu, Shusheng  and
      Zhang, Xingxing  and
      Wu, Yi  and
      Wei, Furu  and
      Zhou, Ming",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.findings-emnlp.161",
    pages = "1784--1795",
}