Diffusion-LM Improves Controllable Text Generation

By Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, Tatsunori B. Hashimoto. 2022

1,430 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Natural Language Processing, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Generative Models

Citations in the corpus

Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements (2023, background): “Also, diffusion models are found effective in controlled text generation, emerging as the next SOTA generation model (Li et al., 2022b).”
PEER: A Collaborative Language Model (2022, background): “Various approaches teach models to iteratively improve texts in an unsupervised fashion (e.g., Shen et al., 2020; Donahue et al., 2020; Li et al., 2022) and explore more efficient ways of representing edits (Mallinson et al.”
InstructPix2Pix: Learning to Follow Image Editing Instructions (2022, background): “Diffusion-based generative models Recent advances in diffusion models [59] have enabled state-of-the-art image synthesis [10, 18, 19, 53, 55, 60] as well as generative models of other modalities such as video [21, 58], audio [30], text [35] and network parameters [45].”
InstructTTS: Modelling Expressive TTS in Discrete Latent Space With Natural Language Style Prompt (2023, background): “Diffusion models have been demonstrated as a strong generative model for image generation [33], [34], [35], [36], text generation [37], [38] and audio generation [23], [39], [40], [41].”
Efficient Diffusion Training via Min-SNR Weighting Strategy (2023, background): “…diffusion models have demonstrated superior performance across a range of generation tasks in various modalities, including text-to-image generation [40, 43, 41, 17], image manipulation [26, 34, 4, 56], video synthesis [18, 47, 22], text generation [28, 16, 59], 3D avatar synthesis [39, 53], etc.”
DIFUSCO: Graph-based Diffusion Solvers for Combinatorial Optimization (2023, background): “Another line of work studies diffusion models for discrete data by applying continuous diffusion models with Gaussian noise on the embedding space of discrete data [30, 69, 24], the {− 1 .”
Gradient-based Constrained Sampling from Language Models (2022, background): “Similar ideas have also been applied to train text generation models in concurrent work with promising results for incorporating controls (Li et al., 2022).”
Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion (2024, neutral): “Diffusion has been widely used in sequence modeling.”
DiffProsody: Diffusion-Based Latent Prosody Generation for Expressive Speech Synthesis With Prosody Conditional Adversarial Training (2023, methodology): “Fig.7showstheobjectiveevaluationresultsforDDPMand DDGAN trained over different time steps: [4, 8, 16, 32, 64, 100].”
Simplified and Generalized Masked Diffusion for Discrete Data (2024, neutral): “This has motivated the development of continuous space diffusion models where the discrete data are embedded in the Euclidean space [19, 20, 21, 22, 23] or the simplex [24, 25, 26, 27, 28].”
NADiffuSE: Noise-aware Diffusion-based Model for Speech Enhancement (2023, methodology): “Recently, DDPM-based SE methods [21]–[29] have also been explored for the promising result in speech generation.”
RODIN: A Generative Model for Sculpting 3D Digital Avatars Using Diffusion (2022, background): “While diffusion models have been successfully applied to different modalities [11, 23, 26, 32], its generative capability is much less explored in 3D generation, with only a few attempts on modeling 3D primitives [36, 74, 76].”
Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow (2022, background): “, 33, 40, 60], and text [41, 88], and more tasks such as image editing [97, 50].”
Analog Bits: Generating Discrete Data using Diffusion Models with Self-Conditioning (2022, background): “Another line of discrete diffusion models is based on the embedding of discrete data (Li et al., 2022).”
Simple and Effective Masked Diffusion Language Models (2024, neutral): “Given the state transition matrix Q t , prior π , and the latent variables z s and z t , where s t , let A.1.1 q ( z t | z s ) Thus, the marginals in (3) correspond to the following forward process: The above equation indicates that during each diffusion step from s → t , a fraction (1 − α t | s )…”
All are Worth Words: A ViT Backbone for Diffusion Models (2022, methodology): “In addition to images, transformers in diffusion models are also employed to encode texts [44,47,49,51], decode texts [7, 28, 36, 43] and generate CLIP embeddings [47].”
A Survey on Generative Diffusion Models (2022, background): “Consequently, diffusion models have garnered signiﬁcant attention in various domains, including computer vision [12], [13], [14], natural language processing [15], [16], time series [17], [18], audio processing [19], [20], graph generation [21], [22], and bioinformatics [23], [24].”
VLP: A Survey on Vision-language Pre-training (2022, background): “Is such a structure the optimal structure for VLP? We note that the recently popular diffusion model[150] for im-”
TabDDPM: Modelling Tabular Data with Diffusion Models (2022, methodology): “Beyond computer vision, the DDPM framework is also investigated in other fields, such as NLP (Austin et al., 2021; Li et al., 2022), waveform signal processing (Kong et al., 2020; Chen et al., 2020b), molecular graphs (Jing et al., 2022; Hooge-boom et al., 2022), time series (Tashiro et al., 2021),…”