Reproduce “Multi-Aspect Controllable Text Generation with Disentangled Counterfactual Augmentation” 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
Yi Liu, Xiangyu Liu, Xiangrong Zhu, Wei Hu
MAGIC, a method for multi-aspect controllable text generation from Nanjing University, addresses challenges from imbalanced attribute correlations in training data by using disentangled counterfactual augmentation in the latent space. It achieved an average 7.4% improvement in attribute relevance over the second-best method, PriorControl, in imbalanced scenarios.
Key insights
- Imbalanced attribute correlations in real-world data are a critical, often overlooked, challenge in multi-aspect controllable text generation, leading to models forming 'stereotypes' and performing poorly on underrepresented combinations.
- Applying counterfactual feature generation directly within the latent space, rather than just at the surface text level, can effectively balance attribute correlations and mitigate biases in generative models.
- Disentangling explicit and implicit attribute control factors within the latent representation enables more precise and robust multi-aspect control, even with correlated attributes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.