I am having a little friendly debate with my coworker on how to properly/optimally do topic modeling. I am just using the regular traditional nmf/lda approach and he decided to do it using "skip grams" in the sense that he just took every single bigram word combination in every document within N words (he did it within 5) and labels it as a topic, sorts it for those that appear most commonly and filters out non sensical topics.

Can you please tell me if there is any validity to this approach over LDA/NMF?

To me it seems like this is just equivalent to running regexes within 5 words on every possible bigram and not modeling, so there would be no reason to invent LDA/NMF if this "skiagram" approach can possibly to better. But I could be wrong.


1 Answer 1


From what I understand, it sounds like your coworker is only considering a bigram as a topic? Wouldn't this result in a number of topics equal to the number of bigrams he/she finds? This is less than ideal in a situation where you would only want a small number of topics.

LDA represents every document in a corpus as a mixture of topics and each word in every document is mapped to a corresponding topic (For a better explanation, see here).

Topics in a topic model are generally represented a list of unigrams. There are many algorithms that can be used that will construct topic phrases based upon unigram lists outputted by topics (Here is one from 2007 for example).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.