Reproduce “StyleFlow: Disentangle Latent Representations via Normalizing Flow for Unsupervised Text Style Transfer” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kangchen Zhu, Zhiliang Tian, Ruifeng Luo, Xiaoguang Mao
StyleFlow introduces a model for unsupervised text style transfer that leverages Normalizing Flow and an attention-aware coupling layer to enhance content preservation. The method achieves state-of-the-art results in content preservation on Yelp and IMDb datasets while maintaining competitive style transfer accuracy and fluency.
Key insights
- Normalizing Flow's inherent reversibility enables theoretically lossless encoding and decoding, directly addressing content preservation challenges in text style transfer.
- Employing an attention-aware mechanism within coupling layers allows for more effective token-level disentanglement of style and content, even for words with mixed information.
- Data augmentation by perturbing latent representations in Normalizing Flow's learned manifold can significantly improve model robustness, particularly in data-scarce scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.