# Style Transfer in Text: Exploration and Evaluation

**Authors:** Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, Rui Yan  
**arXiv:** [1711.06861](https://arxiv.org/abs/1711.06861)  
**Date:** 2017-11-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1711.06861v1/

## Key points

- Authors address the challenge of style transfer in NLP
- Progress in style transfer slower in NLP due to lack of parallel data and effective evaluation metrics
- Proposed approach uses adversarial networks to learn separate content and style representations
- Introduce new evaluation metrics measuring transfer strength and content preservation
- Experiments conducted on paper-news title transfer and positive-negative review transfer tasks
- Results show proposed models outperform traditional auto-encoder approaches
- Research contributes to advancing style transfer capabilities in NLP by leveraging non-parallel data and introducing innovative evaluation metrics

## Summary

In their paper "Style Transfer in Text: Exploration and Evaluation," authors Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan address the challenge of style transfer in natural language processing (NLP). They highlight that while progress in style transfer has been significant in domains like computer vision, it has been slower in NLP due to the lack of parallel data and effective evaluation metrics. To overcome this limitation, the authors propose a novel approach to learn style transfer using non-parallel data. Their proposed models utilize adversarial networks to learn separate content representations and style representations, aiming to enhance the transfer strength of styles while preserving the original content effectively. Additionally, they introduce new evaluation metrics that measure two crucial aspects of style transfer: transfer strength and content preservation. To validate their models and evaluation metrics, the authors conduct experiments on two tasks: paper-news title transfer and positive-negative review transfer. The results demonstrate that their proposed content preservation metric aligns well with human judgments. Furthermore, their models outperform traditional auto-encoder approaches by generating sentences with higher style transfer strength and comparable content preservation scores. Overall, this research significantly contributes to advancing style transfer capabilities in NLP by leveraging non-parallel data and introducing innovative evaluation metrics for assessing both stylistic transformation and content fidelity. These findings pave the way for further advancements in text-based style transfer applications across various domains.