I will finish this
week by showing you how you can use fast
k-nearest neighbor to search for pieces
of text related to a query in a large
collection of documents. You simply create vectors for both and find the
nearest neighbors. To get ready to perform
document search, first, think about how to
represent documents as vectors instead of
just words as vectors. Let's say you have this documents composed of three words. I
love learning. How can you represent this entire
documents as a vector? Well, you can find the word vectors for
each individual word. I love learning. Then just add them together, so the sum of
all
these words vectors becomes a documents vector, where the same dimension
as the word vectors, in this case, three-dimensions. You can then apply
document search by using k-nearest neighbors.
Let's go this up. Create a mini dictionary
for word embeddings. Here's the list of what's
contained in the document. You're going to
initialize the documents embedding as an array of zeros. Now for each word in a
document, you'll get the word
vector if the word exists in the dictionary else zero. You add these all up and
return
the documents embedding. Please try it out. You learned in this
video an example of a very general method that text can be embedded into
vector space so that nearest neighbors refer to
text with similar meaning. Well, you will learn more
advanced ways to embed text. This basic structure
will reappear again and again as it's used
throughout modern NLP.
