Let's say that we have executed the LDA topic generation process (with Gibbs sampling) once. Now for the next round of LDA topic generation, how to make use of the already existing topic matrix? Does any library support this kind of feature?
Asked
Active
Viewed 45 times
0
-
Nobody answered this, eh? I have a similar question and there doesn't appear to be much out there. – Evan Zamir Jul 08 '16 at 05:09
-
@EvanZamir Recently I have been ending up in strange places in terms of technical problems. How do you think I got my tumbleweed badge from stackoverflow? :P – genonymous Jul 09 '16 at 23:58
-
@EvanZamir BTW I implemented the solution to this problem myself. What I did was a simple thing. During the topic matrix initialization phase at the start, instead of assigning random topics for each word in each document, I assign them the topics obtained from the previous topic matrix. But the words in the new documents might have shifted/added/removed compared to the old documents. So to assign correct previous topic to current word, I have to take a diff between previous document and current document. I used java-diff-utils library for this. It used Myer's diff algorithm. – genonymous Jul 10 '16 at 00:09