Reformulating Unsupervised Style Transfer as Paraphrase Generation
Reformulating unsupervised style transfer as paraphrase generation using pretrained GPT-2, achieving 2-3x improvements over SOTA on automatic metrics.
Key Findings
Methodology
This paper introduces a framework that treats style transfer as controlled paraphrase generation. It involves generating diverse pseudo-parallel data via a paraphrasing model, training style-specific inverse paraphrasers, and performing style transfer by combining these components. The core relies on fine-tuning GPT-2 models to produce diverse paraphrases, which normalize input sentences by removing style cues, enabling the inverse models to learn style-specific features without semantic distortion. The approach avoids reinforcement learning or variational inference, simplifying training while maintaining high performance.
Key Results
- On formality transfer and Shakespeare author imitation datasets, STRAP surpasses state-of-the-art models (UNMT, DLSM) by 2-3 times on automatic metrics like J(·), and achieves 4-5x improvements in human evaluations, demonstrating superior semantic preservation and style accuracy.
- When tested on a newly collected large-scale dataset of 15 million sentences spanning 11 styles, STRAP maintains robust performance, validating its scalability and versatility.
- Ablation studies confirm that diversity in paraphrasing and pretraining are critical; removing these components results in significant performance drops, emphasizing the importance of each step.
Significance
This work advances the field of style transfer by providing a simple yet effective unsupervised method that significantly outperforms previous approaches. It addresses the core challenge of semantic preservation while enabling style flexibility, making it highly relevant for real-world applications like content personalization, text editing, and data augmentation. The large-scale dataset and evaluation framework set new standards for future research, promoting more realistic and diverse style transfer systems.
Technical Contribution
The key technical innovation lies in framing style transfer as a controlled paraphrase task, leveraging pretrained GPT-2 models for diverse paraphrase generation and style-specific inverse models. The approach simplifies training, eliminates reliance on complex reinforcement learning, and introduces new evaluation metrics at the sentence level. The large-scale multi-style dataset further enriches the research landscape, enabling more comprehensive benchmarking.
Novelty
This is the first work to systematically formulate unsupervised style transfer as a controlled paraphrase generation problem using large pretrained language models. It departs from prior attribute transfer or attribute manipulation methods by focusing on semantic consistency through normalization, offering a more principled and scalable solution.
Limitations
- The method may struggle with highly stylized or low-resource styles, where pseudo-paraphrase diversity is limited, leading to potential semantic drift or style ambiguity.
- Computational costs are high due to large pretrained models and filtering strategies, which may hinder real-time deployment.
- Current focus is on sentence-level transfer; extending to paragraph or document level remains a challenge.
Future Work
Future directions include integrating multimodal signals (images, audio) for richer style definitions, optimizing model efficiency for deployment, and exploring multi-sentence or multi-paragraph style transfer to handle longer texts. Additionally, refining evaluation metrics and expanding style datasets will further advance the field.
AI Executive Summary
Text style transfer has become a vital task in natural language processing, enabling the transformation of text from one stylistic form to another while preserving its core meaning. Traditional approaches rely heavily on parallel datasets or complex training strategies, limiting their applicability in real-world scenarios. This paper introduces a novel framework that redefines style transfer as a controlled paraphrase generation problem, leveraging the power of pretrained GPT-2 models. The method involves generating diverse pseudo-parallel data by paraphrasing sentences from different styles, training style-specific inverse models to reconstruct original sentences, and performing style transfer by combining these components. This approach is remarkably simple yet highly effective, avoiding the complexities of reinforcement learning or variational inference.
Experimental results demonstrate that STRAP outperforms existing state-of-the-art models by a significant margin on benchmark datasets for formality transfer and Shakespeare author imitation. The automatic evaluation metrics (J-score) show improvements of 2-3 times over previous methods, while human assessments confirm a 4-5 fold enhancement in style accuracy and semantic fidelity. Furthermore, the authors curated a large-scale, diverse dataset of 15 million sentences spanning 11 styles, providing a robust testbed for future research.
Ablation studies highlight the importance of paraphrase diversity and pretrained language models, with results indicating performance drops when these components are removed. The proposed evaluation framework, emphasizing sentence-level metrics, addresses the shortcomings of prior aggregate scores and reveals the true capabilities of the models. Overall, this work pushes the boundaries of unsupervised style transfer, offering a scalable, effective, and versatile solution with broad implications for NLP applications and future research directions.
Deep Dive
Plain Language Accessible to non-experts
想象你在做菜，想换一种味道，但又不想改变菜的基本味道。传统方法可能需要反复试验，调不同的调料。而这篇论文像是发明了一台智能厨师，它能根据你喜欢的味道，自动调整调料比例，保证菜的核心味道不变，同时让风味变得不同。这台“厨师”用一种特别的食谱（模型），可以学习不同风味的特点，然后用“调味”技巧（风格迁移）把菜变成你想要的味道。你只需告诉它你喜欢的风格，它就能帮你把菜变得更有趣、更漂亮。这个方法既简单又高效，不需要你反复试验，就能得到满意的结果。它还能处理各种不同的菜式（风格），让厨房变得更智能、更有趣。
ELI14 Explained like you're 14
想象你有一台神奇的变装机，可以把你的衣服变成不同的风格，比如运动风、正式风，但又不改变你原本的内容。你只要把衣服放进去，它就会用一种特别的魔法，把衣服变成你想要的风格。这台“变装机”用的是一种叫GPT-2的超级聪明模型，它学会了各种风格的衣服样式。你告诉它你想要运动风，它就会帮你把衣服变得像运动员一样酷炫，又不会改变你原本的内容。这样，你每天都可以换不同的风格，既方便又有趣。就像玩换装游戏一样，但用的是文字和智能模型，超级酷！
Abstract
Modern NLP defines the task of style transfer as modifying the style of a given sentence without appreciably changing its semantics, which implies that the outputs of style transfer systems should be paraphrases of their inputs. However, many existing systems purportedly designed for style transfer inherently warp the input's meaning through attribute transfer, which changes semantic properties such as sentiment. In this paper, we reformulate unsupervised style transfer as a paraphrase generation problem, and present a simple methodology based on fine-tuning pretrained language models on automatically generated paraphrase data. Despite its simplicity, our method significantly outperforms state-of-the-art style transfer systems on both human and automatic evaluations. We also survey 23 style transfer papers and discover that existing automatic metrics can be easily gamed and propose fixed variants. Finally, we pivot to a more real-world style transfer setting by collecting a large dataset of 15M sentences in 11 diverse styles, which we use for an in-depth analysis of our system.