rationale for finding topic words on original (non-reduced) vectors

[rbroc]

Hi there! Quick clarification question. To find topic words, it seems like you compute similarities between topic vectors and word vectors using the full-dimensionality vectors (i.e., not what you get after dimensionality reduction, which is, on the other hand, what the clustering models get as input). Just curious whether there a rationale for that, and whether you have experimented with using low-dimensional vectors?