# Learning to Simplify with Data Hopelessly Out of Alignment

arXiv: https://arxiv.org/abs/2204.00741 · HTML: https://mlsanity.com/paper/2204.00741/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 8 of its 15 references that are in the map.
- First posted on arXiv 2 Apr 2022.

## Builds on (most cited first, 8 of 8)

- [Improved Training of Wasserstein GANs](https://mlsanity.com/paper/1704.00028/) (2017)
- [fairseq: A Fast, Extensible Toolkit for Sequence Modeling](https://mlsanity.com/paper/1904.01038/) (2019)
- [Towards Principled Methods for Training Generative Adversarial Networks](https://mlsanity.com/paper/1701.04862/) (2017)
- [Unsupervised Machine Translation Using Monolingual Corpora Only](https://mlsanity.com/paper/1711.00043/) (2017)
- [Are GANs Created Equal? A Large-Scale Study](https://mlsanity.com/paper/1711.10337/) (2017)
- [Unsupervised Neural Machine Translation](https://mlsanity.com/paper/1710.11041/) (2017)
- [Sentence Simplification with Deep Reinforcement Learning](https://mlsanity.com/paper/1703.10931/) (2017)
- [Unsupervised Neural Text Simplification](https://mlsanity.com/paper/1810.07931/) (2018)

## Abstract (arXiv)

We consider whether it is possible to do text simplification without relying on a "parallel" corpus, one that is made up of sentence-by-sentence alignments of complex and ground truth simple sentences. To this end, we introduce a number of concepts, some new and some not, including what we call Conjoined Twin Networks, Flip-Flop Auto-Encoders (FFA) and Adversarial Networks (GAN). A comparison is made between Jensen-Shannon (JS-GAN) and Wasserstein GAN, to see how they impact performance, with stronger results for the former. An experiment we conducted with a large dataset derived from Wikipedia found the solid superiority of Twin Networks equipped with FFA and JS-GAN, over the current best performing system. Furthermore, we discuss where we stand in a relation to fully supervised methods in the past literature, and highlight with examples qualitative differences that exist among simplified sentences generated by supervision-free systems.