Research map / Image generation / Editing and style transfer
ParaGuide: Guided Diffusion Paraphrasers for Plug-and-Play Textual Style Transfer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces ParaGuide, a paraphrase-conditioned word-embedding diffusion model that performs plug-and-play textual style transfer by guiding denoising with gradients from off-the-shelf attribute classifiers and style embedders, evaluated on the Enron Email Corpus.
Problem
Textual style transfer usually needs labeled data for a fixed set of styles or large LLMs; existing unsupervised methods must be retrained for each new target style, and they do not allow flexible guidance from off-the-shelf classifiers or authorship embeddings.
Approach
Train a paraphrase-conditioned diffusion model over noised word embeddings (initialized from SSD-LM RoBERTa-large) to reconstruct text from an AR paraphrase, then at inference steer token logits with gradients from any differentiable RoBERTa-compatible guidance model (attribute classifiers or style embedders) using a sinusoidal guidance schedule.
Key results
- On formality transfer, ParaGuide (lambda=2e2) achieves the highest Joint score (0.48) and beats all baselines on both external accuracy and similarity among automatic metrics.
- In human formality evaluation, ParaGuide (lambda=2e2) reaches 0.65 accuracy, 0.58 similarity and 0.69 fluency, with Joint 0.33 versus 0.20 for M&M (Hamming), with significance at p=0.05.
- On low-resource authorship transfer with the Style embedder, ParaGuide (lambda=1.5e3) achieves Joint 0.434 versus 0.344 for fine-tuned STRAP and 0.338 for ChatGPT-3.5, while ChatGPT-3.5 leads on UAR Joint (0.280).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ParaGuide rather than the strongest listed baseline for formality transfer | Enron formality transfer · Enron formality transfer | Joint: 0.31 (0.54, 0.08) -> 0.48 (0.38, 0.59) | Helped |
| Use ParaGuide rather than M&M (Disc) for formality transfer | Enron formality transfer · Enron formality transfer | External Accuracy: 0.55 (0.44, 0.65) -> 0.83 (0.68, 0.99) | Helped |
| Use ParaGuide rather than M&M (Hamming) for human-rated formality transfer | Enron formality transfer · Enron formality transfer | Human Joint: 0.20 (0.03, 0.36) -> 0.33 (0.23, 0.43) | Helped |
| Use ParaGuide rather than STRAP fine-tuned for sentiment transfer | Enron sentiment transfer · Enron sentiment transfer | Joint: 0.18 (0.24, 0.12) -> 0.29 (0.33, 0.25) | Helped |
| Use ParaGuide rather than M&M (Disc) for sentiment transfer | Enron sentiment transfer · Enron sentiment transfer | External Accuracy: 0.5 (0.32, 0.67) -> 0.8 (0.86, 0.74) | Helped |
| Use ParaGuide rather than ChatGPT-3.5 in the Style embedding space | Enron authorship transfer · Enron authorship transfer | Style Confusion: 0.54 -> 0.74 | Helped |
| Increase authorship guidance strength from λ=2e2 to λ=2.5e3 (ablation) | ParaGuide · Enron authorship transfer | Style Confusion: 0.50 -> 0.74 | Helped |
| Increase authorship guidance strength from λ=2e2 to λ=2.5e3 (ablation) | ParaGuide · Enron authorship transfer | Similarity: 0.52 -> 0.42 | Hurt |
| Increase authorship guidance strength from λ=2e2 to λ=2.5e3 (ablation) | ParaGuide · Enron authorship transfer | Fluency: 0.78 -> 0.64 | Hurt |
| Increase classifier guidance strength from λ=2e2 to λ=1e4 (ablation) | ParaGuide · Enron formality transfer | External Accuracy: 0.76 (0.58, 0.95) -> 0.83 (0.68, 0.99) | Helped |
| Increase classifier guidance strength from λ=2e2 to λ=1e4 (ablation) | ParaGuide · Enron formality transfer | Similarity: 0.52 (0.5, 0.53) -> 0.40 (0.37, 0.44) | Hurt |
| Increase classifier guidance strength from λ=2e2 to λ=1e4 (ablation) | ParaGuide · Enron sentiment transfer | External Accuracy: 0.56 (0.61, 0.51) -> 0.8 (0.86, 0.74) | Helped |
Surprising findings
- ParaGuide scores lower than baselines on automatic fluency metrics, yet human annotators rated its fluency higher than the top baselines.
- Fine-tuned STRAP produces highly repetitive text on the limited Enron data and performs far worse in human evaluation than its automatic scores suggested.
Limitations
- Automatic fluency and similarity scores are lower than baselines, creating a trade-off controlled by lambda.
- Authorship transfer on UAR lags ChatGPT-3.5 in Joint score.
- Human annotator agreement on formality is low (Krippendorff alpha 0.13).
Experimental setup
Datasets: Enron Email Corpus Reddit Million User Dataset (MUD) XFORMAL Twitter sentiment data (cardiffnlp) CoLA
Benchmarks: Enron Email Corpus formality transfer Enron Email Corpus sentiment transfer Enron Email Corpus authorship transfer
Metrics: Internal accuracy External accuracy Confusion Mutual Implication Score similarity Fluency (CoLA classifier) Joint (geometric mean) human accuracy, similarity, fluency
Baselines: STRAP Mix and Match (Disc) Mix and Match (Hamming) ChatGPT-3.5 BERT LING PARA
Builds on: SSD-LM Mix and Match STRAP Diffusion-LM classifier guidance
Closest papers by meaning
- Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer (CVPR 2024 · similarity 0.86)
- PaRa: Personalizing Text-to-Image Diffusion via Parameter Rank Reduction (ICLR 2025 · similarity 0.85)
- OmniStyle: Filtering High Quality Style Transfer Data at Scale (CVPR 2025 · similarity 0.85)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.85)
- Adaptive Prompt Routing for Arbitrary Text Style Transfer with Pre-trained Language Models (AAAI 2024 · similarity 0.85)
- Controlling Language and Diffusion Models by Transporting Activations (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.