Mona
Mona
You can certainly try! Maybe run ETM first to see how that performs as it runs much faster than D-ETM and the code is easier to understand. If that works...
Other languages should be fine if you use pre-trained embeddings in that language or train Word2Vec yourself.
D-ETM was based on ETM but added the temporal evolution of topics as a feature. ETM: https://github.com/adjidieng/ETM
However, there are also models designed specifically for short texts which might work better
Personally I don't really know of anything similar to D-ETM that works better on short texts. A different way to look at (e.g., popularity of) topics over time is doing...
Do you know why in the paper it says they divide by 45 to compute topic coherence?
Might be too obvious, but could it just be because of the random permutation with no seed? Apart from that, I've observed a lot of things I had to change...
You might want to have a look at the paper on Dynamic Mixture Models by Wei et al. (2007) and similar papers on topic modeling for streaming data.
The following paper might be helpful: https://arxiv.org/abs/2002.07514
Is it simply hard-coded cosine distance?