Great video. One thing to add: instead of always discarding the context vectors at the end, another strategy (mentioned in other videos/articles) is to concatenate or add the two vectors for a word.
Great video. You mentioned that we discard the context vectors and take the main vectors as our final word embeddings. I just want to add that based on the literature, you can also add the main and context vectors, or concatenate them.
The best and most concise explanation of Word2Vec that I've seen so far. I probably need to go back and review gradient descent again, because updating the weights is still confusing.
this is by far the best explanation of word2vec on youtube over any university lecture but one question. where and how are you getting your initial vector values in this example for V(like) and V(data)? can you also clarify the components of the vector? you have two columns/elements for each vector in the main vector and context vector.
Exactly what I was wondering. Where do the initialisations of these vectors come from? I'm guessing they're initialised randomly, with the algorithm evolving the distribution into something that relates to the training data.
Another fantastic video! The main/context embeddings kind of confused me but the ending really cleared it up. Curious to know if there is a deliberate way of choosing # of dimensions or if its simply trial and error. On a side note, will you be participating in the "66 days of data" challenge this July?
Great Word2Vec video. Nice touch with blue for Main and red for Context. Working an example with numbers is also very good. Can we please have a follow up video (from ritvik "from the future" :) ) to do the UPDATE part of the loop? We know there are multiple ways to do this UPDATE, but just expound on the "simple" method of moving the vectors as you illustrated in 2D. Thanks.
why is this word2vec so different from "conventional" word2vec where you use a neural network and bag of words to calculate the weights matrix? ok, my bad, just realized you talked about neural network update method towards the end.
And how do we distinguish between 2 same words in this case "data" if we have larger text where data is next to science but it is also close to other word or words for example data processing, data mining, data validation, data.... data..... everything, is it the case that word data is "closer" to word science in general than it is to word processing or there is some other mechanism for this ?
Simply since the distances are invariant under orthogonal transformations of the space "embeddings" (by the way: Not in the mathematical sense) are not unique. But they also depend on the prupose of the model.
Hello again! After a couple of days of studying another materials, I realized that I don't understand where the formulas for the score and the error come from, I don't see them in the books I read (and I haven't seen this approach with labels yet either). Are there any papers or books where I can find them and the prove that they work?
How is the initial size of the vectors determined? As you mentioned that traditionally there may be vectors with 50 dimensions and your example has used 2 to ease understanding.
Can someone will explain where those 2 numeric values (vector representation) for like word, -.2 and -1 comes from ? Word2vec is used to create numeric representation of words in vector form or create vector in such a way that word which are close in documents are also close in vectors representation as well ?
Just FYI, I just made and uploaded a RU-vid video where I used bits of your video here as a use case to illustrate a graphical ontology language that I am calling UniML, Universal Modeling L (not to be confused with UML, Unified modeling Language). While things like neural nets map transformation or functions (f: X -> Y) UniML attempts to model more than that and model the thing itself, including any such transformations as an ontology model but not using just symbols such as does ontology languages such as OWL or RDF but graphically so that graphically depicts the under lying data structures. Here is a link to my video ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-bioz226CcWY.html
