That's interesting, intuitively I'd think that you need a certain document length for LDA to reliably capture repeated word co-occurrences as signals for underlying topics... It would be quite useful to do a systematic comparison that doesn't have all the flaws of the above mentioned paper!

Also, I've never quite seen the value of stemming for topic modeling, but I'd argue that preprocessing in general is an important part of any method. That's why I wouldn't necessarily copy every preprocessing step when comparing methods, especially when the latter are quite different.

I suspect that a good number of citations stem from the fact that the paper "confirms" a common experience in the CSS community: that embedding-based topic models tend to produce more interpretable topics than LDA _on very short texts_.

Thanks for sharing!