How to abstract bigram topics instead of unigrams using Latent Dirichlet Allocation (LDA) in python- gensim?

Question

LDA Original Output

• Uni-grams

  • topic1 -scuba,water,vapor,diving

  • topic2 -dioxide,plants,green,carbon

Required Output

• Bi-gram topics

  • topic1 -scuba diving,water vapor

  • topic2 -green plants,carbon dioxide

Any idea?

Solution 1

You can use word2vec to get most similar terms from the top n topics abstracted using LDA.

LDA Output

Create a dictionary of bi-grams using topics abstracted (for ex:-san_francisco)

check http://www.markhneedham.com/blog/2015/02/12/pythongensim-creating-bigrams-over-how-i-met-your-mother-transcripts/

Then, do word2vec to get most similar words (uni-grams,bi-grams etc)

Word and Cosine distance

los_angeles (0.666175)
golden_gate (0.571522)
oakland (0.557521)

check https://code.google.com/p/word2vec/ (From words to phrases and beyond)

Solution 2

Given I have a dict called docs, containing lists of words from documents, I can turn it into an array of words + bigrams (or also trigrams etc.) using nltk.util.ngrams or your own function like this:

from nltk.util import ngrams

for doc in docs:
    docs[doc] = docs[doc] + ["_".join(w) for w in ngrams(docs[doc], 2)]

Then you pass the values of this dict to the LDA model as a corpus. Bigrams joined by underscores are thus treated as single tokens.