Temporal theme modeling and theme evolution Sankey diagram

[superseanyoung]

I want to model the topic in time, then get the topic in different time periods, and calculate the cosine similarity between these topic vectors, but I don't know how to find a vector representation of a topic.

[MaartenGr]

That is currently not possible since these slices are not saved (they are quite big if you capture them all). That said, you might be able to extract them if you manually adjust some of the code here:

BERTopic/bertopic/_bertopic.py

Line 874 in 84dbf36

c_tf_idf, words = self._c_tf_idf(documents_per_topic, fit=False)

You would have to save and return all the c_tf_idf matrices (note that the topic vectors are not re-calculated here, so they are not used).