Simplification with Non-Aligned Data

This is a source code that supplements a paper:
Learning to Simplify with Data Hopelessly Out of Alignment, published on ArXiv.

Pre-requisites

• Python 3.6
• torch 1.3.1
• torchaudio 0.10.0+cu113
• torchtext 0.5.0
• torchvision 0.11.1+cu113
• sentencepiece 0.1.9 (together with its python library)

Setting it up

Download and untar the following at the project's top directory.

• asset.tar.bz2
• js-resource.tar.bz2
• wasser-resource.tar.bz2

cd nonaligned_simple
tar jxvf asset.tar.bz2
tar jxvf js-resource.tar.bz2
tar jxvf wasser-resource.tar.bz2

Training

cd js
../util/train.sh -d tsd -b 64 -g 1

• wasser-gan

cd wasser
../util/train.sh -d tsd -b 64 -g 1

-d : dataset name
-b : batch size
-g : GPU ID

Generation

Using pre-trained models

As a trained model is provided as part of the package, you can bypass training and go to the generation step directly. Here is what you do.

• js-gan

cd js
../util/generate.sh -d tsd -g 1

The output is found in js/data/tsd/pred.out.

• wasser-gan

cd wasser
../util/generate.sh -d tsd -g 1

-d : dataset name
-g : GPU ID

The result in wasser/data/tsd/pred.out.

Detokenization

generate.sh gives you a result in a sentence-piece format. Running the following will restore it into a normal text.

cd js
../util/decode_spm.sh 

You find the result in js/data/tsd/pred_decoded.txt.

About Data

The training data is based on the sscorpus. We removed source/target pairs whose similarity exceeds 0.65. The test set is the same as one used by Zhang, et al (2017).

References

@misc{https://doi.org/10.48550/arxiv.2204.00741,
doi = {10.48550/ARXIV.2204.00741},
url = {https://arxiv.org/abs/2204.00741},
author = {Nomoto, Tadashi},
keywords = {Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
title = {Learning to Simplify with Data Hopelessly Out of Alignment},
publisher = {arXiv},
year = {2022},
copyright = {Creative Commons Attribution Share Alike 4.0 International}
}

@inproceedings{zhang-lapata-2017-sentence,
title = "Sentence Simplification with Deep Reinforcement Learning",
author = "Zhang, Xingxing  and Lapata, Mirella",
booktitle = "Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing",
month = sep,
year = "2017",
address = "Copenhagen, Denmark",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/D17-1062",
 }