[Paper Review] Deep Learning for Text Style Transfer: A Survey

Di Jin, Zhijing Jin+3 more|arXiv (Cornell University)|Nov 1, 2020

Topic ModelingComputer Science244 references17 citations

TL;DR

This survey provides a comprehensive overview of deep learning methods for text style transfer (TST), covering task formulation, datasets, evaluation, and methodologies using both parallel and non-parallel data. It systematically categorizes approaches—disentanglement, prototype editing, and pseudo-parallel data construction—highlighting their strengths and limitations, and discusses ethical implications and future research directions in NLP.

ABSTRACT

Text style transfer is an important task in natural language generation, which aims to control certain attributes in the generated text, such as politeness, emotion, humor, and many others. It has a long history in the field of natural language processing, and recently has re-gained significant attention thanks to the promising performance brought by deep neural models. In this paper, we present a systematic survey of the research on neural text style transfer, spanning over 100 representative articles since the first neural text style transfer work in 2017. We discuss the task formulation, existing datasets and subtasks, evaluation, as well as the rich methodologies in the presence of parallel and non-parallel data. We also provide discussions on a variety of important topics regarding the future development of this task. Our curated paper list is at https://github.com/zhijing-jin/Text_Style_Transfer_Survey

Motivation & Objective

• To provide a unified and systematic review of neural text style transfer (TST) methods since 2017.
• To standardize terminology, task definitions, and evaluation practices across diverse TST research.
• To analyze and compare methodologies for TST under parallel and non-parallel data settings.
• To identify key challenges and future research directions in TST, including ethical concerns.
• To explore the impact of TST on downstream NLP applications such as dialogue systems, summarization, and text debiasing.

Proposed method

• Classifies TST into parallel and non-parallel data settings, with distinct methodological approaches for each.
• Reviews disentanglement-based methods that separate content and style in latent space (e.g., VAEs, adversarial training).
• Examines prototype editing methods that use sentence templates and attribute markers for controlled generation.
• Analyzes pseudo-parallel corpus construction techniques that simulate parallel data for supervised training.
• Discusses the role of Transformer-based models in improving performance across TST tasks.
• Outlines multi-task learning, data augmentation, and inference techniques for enhancing robustness and controllability.

Experimental results

Research questions

• RQ1 How are style and content formally defined and separated in text style transfer?
• RQ2 What are the dominant methodological paradigms in TST, and how do they differ in performance and generalization?
• RQ3 How do evaluation metrics and benchmark datasets vary across TST subtasks and settings?
• RQ4 What are the ethical risks associated with TST, particularly in sentiment and political slant transfer?
• RQ5 How can TST be extended to benefit downstream NLP tasks like persona-based dialogue and text debiasing?

Key findings

• The survey identifies three main paradigms in non-parallel TST: disentanglement, prototype editing, and pseudo-parallel data construction, each with distinct strengths.
• Disentanglement methods achieve strong style control but often struggle with content preservation and fluency.
• Prototype editing enables high-style control with minimal content change but requires careful template design and is less generalizable.
• Pseudo-parallel data construction enables supervised-style training without real parallel data, improving performance on low-resource settings.
• Ethical concerns are prominent in sentiment and political slant transfer, where the technology can be misused for review manipulation or propaganda.
• TST shows strong potential in applications like text debiasing, writing assistance, and persona-consistent dialogue, but requires careful ethical oversight.