I came across this post from two years ago on the `topic-models` listserv while looking up the difference between LSA and LDA:
> LDA isn’t necessarily Bayesian (at least not the original variational algorithm — it produces point estimates). In my view, the dispute is really about those who want a probabilistic model (LDA) versus those that don’t care (LSA). Unfortunately, the probabilistic model produced by LDA is only of limited use, because it doesn’t model the document length. But it does have clear probabilistic semantics.
> My impression is that neither model is particularly better at handling polysemy and synonymy. They both do ok if you have enough documents to train on. But perhaps other members of the list can enlighten me.
The note is by Thomas G. Dietterich of Oregon State University. Thank you, Tom.