Multi-Aspect Controllable Text Generation with Disentangled Counterfactual Augmentation

Multi-aspect controllable text generation aims to control the generated textsin attributes from multiple aspects (e.g., "positive" from sentiment and"sport" from topic). For ease of obtaining training samples, existing worksneglect attribute correlations formed by the intertwining of differentattributes. Particularly, the stereotype formed by imbalanced attributecorrelations significantly affects multi-aspect control. In this paper, wepropose MAGIC, a new multi-aspect controllable text generation method withdisentangled counterfactual augmentation. We alleviate the issue of imbalancedattribute correlations during training using counterfactual feature vectors inthe attribute latent space by disentanglement. During inference, we enhanceattribute correlations by target-guided counterfactual augmentation to furtherimprove multi-aspect control. Experiments show that MAGIC outperformsstate-of-the-art baselines in both imbalanced and balanced attributecorrelation scenarios. Our source code and data are available athttps://github.com/nju-websoft/MAGIC.