A Distributional Lens for Multi-Aspect Controllable Text Generation
This addresses a practical challenge in text generation for applications requiring fine-grained control over multiple attributes, though it is incremental as it builds on prior distributional and fusion techniques.
The paper tackles the problem of attribute degeneration in multi-aspect controllable text generation, where existing methods suffer from mutual interference when fusing multiple controllers. The result is a method that outperforms strong baselines on attribute relevance and text quality, achieving state-of-the-art performance on tasks like sentiment, topic, and detoxification control.
Multi-aspect controllable text generation is a more challenging and practical task than single-aspect control. Existing methods achieve complex multi-aspect control by fusing multiple controllers learned from single-aspect, but suffer from attribute degeneration caused by the mutual interference of these controllers. To address this, we provide observations on attribute fusion from a distributional perspective and propose to directly search for the intersection areas of multiple attribute distributions as their combination for generation. Our method first estimates the attribute space with an autoencoder structure. Afterward, we iteratively approach the intersections by jointly minimizing distances to points representing different attributes. Finally, we map them to attribute-relevant sentences with a prefix-tuning-based decoder. Experiments on the three-aspect control task, including sentiment, topic, and detoxification aspects, reveal that our method outperforms several strong baselines on attribute relevance and text quality and achieves the SOTA. Further analysis also supplies some explanatory support for the effectiveness of our approach.