Research map / Probabilistic and scientific modeling / Variational and generative sampling
Revision in Continuous Space: Unsupervised Text Style Transfer without Adversarial Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use gradient-based continuous-space revision instead of one-pass transfer methods | Text style transfer · Yelp and Amazon sentiment transfer | Accuracy: 94.8 -> 95.7 (Yelp BTS); 82.6 -> 90.0 (Amazon BTS) | Helped |
| Use gradient-based continuous-space revision instead of back-translation | Yelp gender transfer · Yelp gender transfer | Accuracy: 60.3 -> 79.9 | Helped |
| Increase λc to emphasize content preservation (ablation) | Proposed method · Yelp sentiment transfer | Accuracy: 93.4 -> 88.7 | Hurt |
| Increase λc to emphasize content preservation (ablation) | Proposed method · Yelp sentiment transfer | Overlap: 34.8 -> 35.7 | Helped |
| Retrain the sentiment predictor after joint training (ablation) | Proposed method · Yelp sentiment transfer | Accuracy: 88.3 -> 93.1 | Helped |
| Also retrain the content predictor (ablation) | Proposed method with retrained sentiment predictor · Yelp sentiment transfer | Overlap: 39.7 -> 41.6 | Helped |
| Add a content predictor targeted at all original words rather than removing it (ablation) | Proposed method · Yelp sentiment transfer | Overlap: 33.2 -> 36.3 | Helped |
| Target original nouns only in the content predictor rather than all words (ablation) | Proposed method · Yelp sentiment transfer | Noun%: 60.2 -> 68.6 | Helped |
| Add a content predictor rather than using none (ablation) | Proposed method · Yelp sentiment transfer | Accuracy: 91.9 -> 92.8 | Helped |
| Use lower VAE KL loss (ablation) | Proposed method · Yelp sentiment transfer | Accuracy: 84.7 -> 92.6 | Helped |
| Use lower VAE KL loss (ablation) | Proposed method · Yelp sentiment transfer | Overlap: 43.1 -> 31.7 | Hurt |
| Simultaneously control sentiment, keyword presence, and length | Proposed method · Yelp fine-grained attribute control | Accuracy: 93.0 (Length↑); 87.6 (Length↓) | Mixed |
Closest papers by meaning
- A Probabilistic Formulation of Unsupervised Text Style Transfer (ICLR 2020 · similarity 0.90)
- Style Transfer in Text: Exploration and Evaluation (AAAI 2018 · similarity 0.90)
- Unsupervised Text Style Transfer using Language Models as Discriminators (NeurIPS 2018 · similarity 0.88)
- ParaGuide: Guided Diffusion Paraphrasers for Plug-and-Play Textual Style Transfer (AAAI 2024 · similarity 0.86)
- StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements (CVPR 2025 · similarity 0.85)
- Adaptive Prompt Routing for Arbitrary Text Style Transfer with Pre-trained Language Models (AAAI 2024 · similarity 0.84)
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.