What is the relationship between Naïve Bayes and expectation-maximization (EM) under latent variables?

Naïve Bayes can be extended using EM when class labels are partially or fully latent.

When labels are missing, EM alternates between estimating class probabilities (E-step) and updating likelihood parameters (M-step). This results in semi-supervised or unsupervised Naïve Bayes variants. EM leverages soft assignments to iteratively refine parameter estimates until convergence.