Reading "Kernel Topic Models" by Hennig et al. confirms my belief that Bayesians model nicely and then loose it all in the approximation.