[Paper Review] A Reparameterized Discrete Diffusion Model for Text Generation

Zheng Lin, Jianbo Yuan+2 more|arXiv (Cornell University)|Feb 11, 2023

Topic Modeling9 citations

TL;DR

This paper introduces Reparameterized Discrete Diffusion Models (RDMs) for text generation, deriving an equivalent reparameterized backward process, simplifying training, and enabling flexible, efficient decoding that outperforms previous discrete and continuous diffusion approaches on multiple benchmarks.

ABSTRACT

This work studies discrete diffusion probabilistic models with applications to natural language generation. We derive an alternative yet equivalent formulation of the sampling from discrete diffusion processes and leverage this insight to develop a family of reparameterized discrete diffusion models. The derived generic framework is highly flexible, offers a fresh perspective of the generation process in discrete diffusion models, and features more effective training and decoding techniques. We conduct extensive experiments to evaluate the text generation capability of our model, demonstrating significant improvements over existing diffusion models.

Motivation & Objective

• Motivate and analyze discrete diffusion models for natural language generation.
• Derive an equivalent reparameterized backward process and a routing-based sampling mechanism.
• Propose the Reparameterized Diffusion Model (RDM) framework with simplified training and flexible decoding.
• Empirically evaluate RDMs on translation and general text generation tasks, showing improvements over prior diffusion models and competitive performance with autoregressive baselines.

Proposed method

• Derive a compact, equivalent backward transition for discrete diffusion that reveals a route-and-denoise mechanism.
• Introduce RDMs by explicitly modeling a latent routing variable vt−1 that controls whether tokens are denoised or reset to noise.
• Formulate training as a reweighted cross-entropy loss that is invariant to routing distributions up to reweighting.
• Develop adaptive routing during sampling to selectively denoise tokens based on model confidence scores.
• Provide algorithms for training (Algorithm 1) and sampling (Algorithm 2) that leverage the joint diffusion with routing ( vt−1, xt−1 ).
• Show that training can be reduced to a simple cross-entropy objective over noisy tokens, amortized over a family of routing processes.

Experimental results

Research questions

• RQ1 Can a reparameterized backward process yield a more flexible and efficient diffusion-based text generator?
• RQ2 Does explicit routing (vt−1) improve training stability and decoding quality for discrete diffusion models?
• RQ3 Can RDMs achieve better text generation quality with fewer iterations compared to existing discrete and continuous diffusion methods?
• RQ4 How do adaptive routing strategies affect generation speed and sample quality in practice?

Key findings

• RDMs provide significant quality gains over vanilla discrete diffusion models across translation and open-ended generation tasks.
• RDMs outperform continuous diffusion baselines while running orders of magnitude faster (e.g., several hundredfold runtime improvements in some setups).
• A simplified training objective reduces to a reweighted cross-entropy loss, invariant to routing probabilities up to reweighting, enabling a broad family of routings to be trained with a shared objective.
• An adaptive routing strategy that denoises only high-confidence tokens yields strong improvements over uniform routing, with gains coupled to improved decoding strategies.
• Empirical results show substantial BLEU improvements over prior discrete diffusion models and competitive performance with autoregressive baselines on translation benchmarks (IWSLT14 DE-EN, WMT14 EN-DE, WMT16 EN-RO).
• RDMs show notably better speed-quality trade-offs than DiffuSeq and other continuous diffusion approaches, often achieving similar or better quality with far fewer iterations.