UnsupNTS: Unsupervised Neural Text Simplification

This is the original implementation of the Unsupervised Neural Text Simplification system and their semi-supervised variants mentioned in the ACL 2019 long paper:

Sai Surya, Abhijit Mishra, Anirban Laha, Parag Jain, and Karthik Sankaranarayanan. Unsupervised Neural Text Simplification arXiv preprint arXiv:1810.07931 (2018).

Training

Download tsdata.zip from link and extract

unzip tsdata.zip

tsdata.zip has

• Unsupervised sets of easy and difficult set of sentences judged on readability ease scores.
• Dict2vec embeddings trained on the above unsupervised sets.
• 10k parallel pairs of difficult and simplified variants.
• Test set and references - eight tab seperated references per each test sentence.

Train the models using

bash train.sh

train.sh has

• UNTS system from unsupervised simplification data using the exact same settings described in the paper.
• UNTS-10k system, using additional 10k supervised pairs of mixture of split-rephrase and simplification parallel pairs.
• UNMT system on the unsupervised simplification data.
• ablations on adversarial and separation/classifier losses.

For more details and additional options, run the above scripts with the --help flag. Alternatively, visit the ipynb in google colaboratory to reproduce the results. To access pretrained models visit link. The folder predictions has the generations from the pretrained models.

Note: Pretrained models were trained with pytorch 0.3.1 and may not exactly reproduce the result in pytorch 1.0.1 .

Generation and Evaluation of Simplifications

bash translate.sh

translate.sh is used for

• Generating simplifications of test dataset.
• Computing stand alone metrics such as Flesch readability ease score difference, Tree similarity and Document similarity metrics.
• Computing SARI, BLEU and Word-diff metrics.

Acknowledgements

A large portion of this repo is borrowed from the following repos: https://github.com/artetxem/undreamt and https://github.com/senisioi/NeuralTextSimplification.

If you use our system for academic research, please cite the following paper:

@inproceedings{surya-etal-2019-unsupervised,
    title = "Unsupervised Neural Text Simplification",
    author = "Surya, Sai  and
      Mishra, Abhijit  and
      Laha, Anirban  and
      Jain, Parag  and
      Sankaranarayanan, Karthik",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1198",
    doi = "10.18653/v1/P19-1198",
    pages = "2058--2068"
}