This kind of latent variable model is not only highly unstable but also tends to tell you more about the person interpreting the results than the data itself, e.g.: Chang et al., Reading Tea Leaves: How Humans Interpret Topic Models, http://umiacs.umd.edu/~jbg/docs/nips2009-rtl.pdf
While I agree that it can be unstable (inference can get stuck in local maxima), latent variable models like LDA can be used to rigorously evaluate textual categories (e.g. journal articles). We take for granted that the categories we set are "useful", in some sense, so it's interesting to see that quantitatively questioned.
