1
$\begingroup$

I have been familiarizing myself with Word2Vec and Doc2Vec. After reading multiple papers including the the ones by T Mikolov (the creator of Doc2Vec), I am not clear on how does the neural network for Doc2Vec looks like.

I get Word2Vec: a neural network with 3 layers (1 hidden). The input layer has all words in the vocab. The hidden layer is the same size as that of embedding, while the output layer, again the same size as vocab. This is, of course, the CBOW framework.

How does Doc2Vec change things?

$\endgroup$
1
$\begingroup$

Nothing just the starting randomly initialized paragraph vector is concatnated/ averaged with the randomly initialized word vector. Everything else pretty much works the same.

$\endgroup$
  • $\begingroup$ But then, how do we get the vectors for a doc D_i $\endgroup$ – sandyp Jul 12 '18 at 19:41
  • $\begingroup$ You have a vector for doc D_i , which is concatenated with the vectors of the words, w_i in document so you get a , then you train it on the task of predicting the target word given the center word, as a result both the vectors get trained by gradient descent. $\endgroup$ – Himanshu Rai Jul 12 '18 at 19:54
  • $\begingroup$ Not able to visualize it. Would a pic help? Thanks $\endgroup$ – sandyp Jul 12 '18 at 22:21

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.