523245 (3) [Avatar] Offline
#1
I believe this should be Sacramento instead of San Francisco

Sacramento - California + Colorado = Denver

Capital of California = Sacramento
Capital of Colorado = Denver
hobs (58) [Avatar] Offline
#2
You're right! I bet you do well on analogy questions.

Fortunately word vectors are fuzzy, so they can handle our fuzzy thinking when we composed this analogy question. Word vector logical expressions like this will work even if you don't know your state capitals. The machine is just finding the "closest" answer to your logical expression, not the most logical or correct way of thinking about the expression. And it definitely can't guess what the questioner was thinking (did they want the largest city or the capital?) Actually, word vector analogy questions are more like a standardized test that is scored based on popularity, rather than correctness. Word vectors are created based on the statistically most popular way that words are used (in the case of Word2vec, by Google News authors).

This is a problem I have with multiple choice standardized tests in general. We're teaching our children to be able to guess what some test writer is thinking, rather than actually thinking analytically. You were thinking analytically, critically, and that helped you see a correct answer that I wasn't thinking of. It's a good thing I don't write standardized tests.
523245 (3) [Avatar] Offline
#3
Great, thanks for clarifying.

Franco

PS: Your awesome book is printed on my desk. Best NLP book for practitioners by a very long stretch (I read them all).
hobs (58) [Avatar] Offline
#4
Your P.S. made my day!