The diffusion model is crazy again! This time the occupied area is

Diffusion model was first used in the field of image generation , Then it is extended to other continuous domains , For example, voice. 、 video 、 Point cloud data , lately Google Published for text to image generation GLIDE Model , Let the diffusion model change from AI The fire inside the circle goes out of the circle .

however , Although the diffusion model is extremely hot , However, the performance of diffusion model for discrete variables has been poor , Especially in language , Figure and other structures .

lately , The natural language processing research group of Stanford University in Using diffusion model to solve naturallanguageprocessing problems New progress has been made in .

say concretely , stay Controllable natural language generation On mission , Researchers Using the continuous diffusion model , Pluggable manipulation of the pre trained language generation model , Can achieve or even surpass in many tasks Fine-Tuning The effect of , Greatly surpassing the previous work .

This work is very novel and solid in terms of methods and experiments , In just one week Twitter I got a thousand praises on , stay Github On the harvest 140 individual stars.

below , Let's take a look at the application of this diffusion model in the language field , It may also inspire your field ~

Paper title :
Diffusion-LM Improves Controllable Text Generation
Thesis link :
https://arxiv.org/pdf/2205.14217.pdf