Diffusion-LM Improves Controllable Text Generation

Xiang Li · John Thickstun · Ishaan Gulrajani · Percy Liang · Tatsunori Hashimoto

Hall J #606

Keywords: [ infilling ] [ controllable text generation ] [ controlled generation ] [ language model ] [ Diffusion model ]

[ Abstract ]
[ Paper [ OpenReview
Wed 30 Nov 9 a.m. PST — 11 a.m. PST


Controlling the behavior of language models (LMs) without re-training is a major open problem in natural language generation. While recent works have demonstrated successes on controlling simple sentence attributes (e.g., sentiment), there has been little progress on complex, fine-grained controls (e.g., syntactic structure). To address this challenge, we develop a new non-autoregressive language model based on continuous diffusions that we call Diffusion-LM. Building upon the recent successes of diffusion models in continuous domains, Diffusion-LM iteratively denoises a sequence of Gaussian vectors into word vectors, yielding a sequence of intermediate latent variables. The continuous, hierarchical nature of these intermediate variables enables a simple gradient-based algorithm to perform complex, controllable generation tasks. We demonstrate successful control of Diffusion-LM for six challenging fine-grained control tasks, significantly outperforming prior work.

Chat is not available.