Roundtrip Translation Enables Text Style Transfer Without Parallel Training Data

Researchers have developed a practical approach to text style transfer—rewriting sentences to match a target tone or register while keeping the underlying meaning intact—without requiring the paired training examples that typically supervise such tasks.

The Core Challenge#

Most style transfer work assumes access to parallel corpora: matched pairs showing the same content in both source and target styles. Such datasets exist for only a handful of style domains, making supervised learning infeasible for new applications. Prior attempts have relied on lightweight "normalizers" that run only at inference time, limiting the sophistication of downstream style adaptation.

A Scalable Normalizer#

The key insight is that neural machine translation (MT) systems, trained on hundreds of millions of general-domain sentence pairs, already function as natural style strippers. When text is translated through a pivot language and back again—a process called roundtrip translation—it retains content while regressing toward generic phrasing, effectively removing stylistic signal without task-specific training.

By repurposing roundtrip translation as a data-generation tool rather than an inference-time patch, researchers can create pseudo-parallel corpora. A monolingual corpus in the target style is roundtrip-translated to produce neutral versions, yielding pairs suitable for fine-tuning.

The RT-SFT Method#

The resulting method, RT-SFT, uses LoRA fine-tuning on an instruction-tuned large language model as the stylizer. The same MT-based normalizer applied during training is also applied to test inputs, ensuring the model operates in-distribution. Across four style domains, the approach substantially outperforms state-of-the-art baselines, including few-shot in-context learning with frontier models.

The work also explores retrieval augmentation for specialized domains requiring strict adherence to terminology and naming conventions.