Is it possible for word2vec to produce similar embedding vectors for two words which never share any common words in the sentences that the words are found in?

Specifically, imagine I have the words A and B. Next imagine that I have words(A) and words(B), which represent the set of all words which show up in the same sentences as A and B, respectively. If the intersection of words(A) and words(B) is the null set (meaning the two words never have common words), then would it be possible for word2vec to put the embedding vectors for A and B in similar regions of the vectorspace?

  • $\begingroup$ Yes, as long as word(A) and words(B) have similar embeddings. $\endgroup$ – Emre Aug 26 '18 at 1:29

As you send a bag of words into the CBOW as an input, and it works based on the n-grams, as in your case words A and B do not share any co-occurrences with each other, it means their not presented in any shared n-grams, and their vectors should not be near each other at all.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.