Skip to content
This repository has been archived by the owner on Aug 21, 2020. It is now read-only.


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Exploring Neural Text Simplification


We present the first attempt at using sequence to sequence neural networks to model text simplification (TS). Unlike the previously proposed automated methods, our neural text simplification (NTS) systems are able to simultaneously perform lexical simplification and content reduction. An extensive human evaluation of the output has shown that NTS systems achieve good grammaticality and meaning preservation of output sentences and higher level of simplification than the state-of-the-art automated TS systems. We train our models on the Wikipedia corpus containing good and good partial alignments.

	  author    = {Sergiu Nisioi and Sanja Štajner and Simone Paolo Ponzetto and Liviu P. Dinu},
	  title     = {Exploring Neural Text Simplification Models},
	  booktitle = {{ACL} {(2)}},
	  publisher = {The Association for Computational Linguistics},
	  year      = {2017}

Simplify Text | Generate Predictions (no GPUs needed)

  1. OpenNMT dependencies
    1. Install Torch
    2. Install additional packages:
    luarocks install tds
  2. Checkout this repository including the submodules:
   git clone --recursive
  1. Download the pre-trained released models NTS and NTS-w2v (NOTE: when using the released pre-trained models, due to recent changes in third party software, the output of our systems might not be identical to the one reported in the paper.)
   python src/ ./models
  1. Run from the scripts dir:
   cd src/scripts
  1. Check the predictions in the results directory:
   cd ../../results_NTS
  1. Run automatic evaluation metrics
    1. Install the python requirements (only nltk is needed)
       pip install -r src/requirements.txt
    1. Run the evaluate script
       python src/ ./data/test.en ./data/references/references.tsv ./predictions/

The Content of this Repository


  • a script to download the pre-trained models. The models are released to be usable on machines with or without GPUs. They can't be used to continue the training session. In case the download script fails, you may use the direct links for NTS and NTS-w2v
  • a script that creates a word2vec model from a local corpus, using gensim
  • a copy of the SARI implementation
  • evaluates BLEU and SARI scores given a source file, a directory of predictions and a reference file in tsv format
  • ./scripts - contains some of our scripts that we used to preprocess the data, output translations, and create the concatenated embeddings
  • ./patch - the patch with some changes that need to be applied, in case you may want to use the latest checkout of OpenNMT. Alternatively, you may use our forked code which comes directly as a submodule.


Contains the OpenNMT config file. To train, please update the config file with the appropriate data on your local system and run

	th train -config $PATH_TO_THIS_DIR/configs/NTS.cfg


Contains predictions from previous systems (Wubben et al., 2012), (Glavas and Stajner, 2015), and (Xu et al., 2016), and the generated predictions of the NTS models reported in the paper:

  • NTS_default_b5_h1 - the default model, beam size 5, hypothesis 1

  • NTS_BLEU_b12_h1 - the BLEU best ranked model, beam size 12, hypothesis 1

  • NTS_SARI_b5_h2 - the SARI best ranked model, beam size 12, hypothesis 1

  • NTS-w2v_default_b5_h1 - the default model, beam size 5, hypothesis 1

  • NTS-w2v_BLEU_b12_h1 - the BLEU best ranked model, beam size 12, hypothesis 1

  • NTS-w2v_SARI_b12_h2 - the SARI best ranked model, beam size 12, hypothesis 2


Contains the training, testing, and reference sentences used to train and evaluate our models.