I'm very interested in ways we can expand LDA (and other topic models) to retain more of the meaning of the documents, especially for small feedback (like reviews), such that a human could explore the results and find impactful, actionable data.

Using N-grams increases the dimensionality way too much.

I've thought about word-vectors as an option, but am unsure how similar terms could be grouped internally in LDA.

This kind of subject is probably worth a PHD thesis, honestly.