We're looking into using topic modeling for many aspects of our project in the upcoming months. Specifically, we think it can provide us the following interesting features:

1. Search the document corpus using topic modeling similarity.
2. Find the similarity between Q-A A-P and Q-P.

But besides those things, we are also simply exploring the topic space (bit of a pun).

Vector generation methods

• Gensim methods
  • LSA, available on Gensim, search index
  • LDA, same
  • word2vec, same
• GloVe vectors: We use Wikipedia 300 dimension (GWP300), and Common Crawl 840B-token 300 dimension (GCC840)
• Composes Best Predict Vectors (CBP)

Simple evaluations

These are using the equation format used in scripts/gensim/analog.py

Analogical

• compare(w('king') - w('man') + w('woman'), w('queen'))
  • Should be high (e.g. > 0.6)
• queenish=w('king') - w('man') + w('woman'); compare(queenish, w('queen')) - compare(queenish, w('king'))
  • Should be positive (low is ok)
• compare(w('putin') - w('russia') + w('usa'), w('obama'))
  • Should be high
• potusish = w('putin') - w('russia') + w('usa'); compare(potusish, w('obama')) - compare(potusish, w('putin'))
  • Should be positive

Synonymical

• compare(w('democrat'), w('republican'))
  • Should be high
• compare(w('party'), w('republican')) - compare(w('party'), w('democrat'))
  • Should have low magnitude