Browse Teach Curate

Diffusion-Enhanced Document-Topic Sampling in Embedded Topic Models

A method that improves the embedded topic model by replacing its logistic-normal sampling of the document-topic distribution with the forward process of a denoising diffusion model. By diffusing document-derived representations toward a normal distribution, it captures a more complex document-topic distribution while preserving the embedded topic model's original variational objective for stable optimization.

Data Text

DiffETM: Diffusion Process Enhanced Embedded Topic Model 1st Wei Shao

The embedded topic model (ETM) is a variational-autoencoder approach to topic modeling that represents words and topics as vectors in a shared embedding space and assumes each document's topic distri…