Significance of Topics for LDA

[Ry C]

Hi all, 

In gensim LDA  the default print_topics returns the top 20 significant topics and the top 10 significant terms associate with them. 

How does gensim rank these topics as significant and is there a paper that it refers to for this topic?

For experiment, I summed up the weights of each inferred topic from the corpus, the ranking based on the aggregated weights is different from the print_topics output. So I am very curious. 

Thank you for your time.

[H Y]

Hi Ry C, I'm having the same question. Have you got any ideas yet? Cheers.

[kor...@asu.edu]

If you look at a pyLDAvis output (sample attached), there are two links to papers in the bottom right of the output, referencing relevancy and saliency. If you change lamba in the output you can see that both the relevance and salience of the topic terms changes. 

 

Not sure if Gensim uses this but the source code calcs look close to relevance

links to papers on saliency and relevance:

1. saliency(term w) = frequency(w) * [sum_t p(t | w) * log(p(t | w)/p(t))] for topics t; see Chuang et. al (2012)

2. relevance(term w | topic t) = λ * p(w | t) + (1 - λ) * p(w | t)/p(w); see Sievert & Shirley (2014)  

[H Y]

Hi there,

Thank you very much! Yeah the salience & relevance calculation looks like a reasonable way to rank topics. However I couldn't find the gensim implementation in the code...worth investigating more.  

Best regards,

Hang

--
You received this message because you are subscribed to a topic in the Google Groups "Gensim" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/gensim/M8hNAsWu-jQ/unsubscribe.
To unsubscribe from this group and all its topics, send an email to gensim+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/gensim/d32f3fb6-6bae-4178-83b6-68e04405e960n%40googlegroups.com.