Skip to main content

Word embeddings

In this post, we are going to talk about word embedding (or word vector), which is how we represent words in NLP. Word embedding is used in many higher-level applications such as sentiment analysis, Q&A, etc. Let's have a look at the most currently widely used models.

One-hot vector

is a vector of size V, with V is the vocabulary size. It has value 1 in one position (represents the value of this word "appears") and 0 in all other positions.
[0, 0, ... 1, .., 0]
This is usually used as the input of a word2vec model. It is just operating as a lookup table.
So this one-hot encoding treats words as independent units. In fact, we want to find the "similarity" between words for many other higher-level tasks such as document classification, Q&A, etc.
The idea is: To capture the meaning of a word, we look at the words that frequently appear close-by this word. Let's have a look at some state-of-the-art architectures that give us the results of word vectors.

The Skip-gram model


http://mccormickml.com/2016/04/19/word2vec-tutorial-the-skip-gram-model/

The intuition behind this model is to predict the "context" words given the "center" word.
Given a center word, we want to tell what the probability of another word appearing in its context is. In order to do that, we can use an architecture as illustrated in the picture.
  • Input: One-hot vector, with the size = the size of the whole vocabulary
  • Training examples: word pairs (words that appear together in a window with a certain size)
  • Output: Vector representation of each word, or more specific in the testing phase, given an input word, the output is a probability vector (how likely each word to be the context word of the given input word is).
Now let's have a look at the architecture:
  • The hidden layer: Let's say we want to learn a word vector with 500 features (e.g., 500 dimensions), our vocabulary size is 10K. Then the hidden layer is going to be represented by a matrix of size (10K, 500). 
  • The output layer (softmax classifier): The output layer is a softmax regression classifier, which gives output between 0 and 1. So each given word, it will output a vector size 500 to represent this word. 
A more details about this model can be found here.

The CBOW (Continuous Bag of Words) model

Its architecture is very similar to the skip-gram model, except for the fact that, in skip-gram, we predict context based on center words, while CBOW predicts a "center" word given a bag of "context" words.

GloVe model

This model combines traditional count-based model with direct prediction models as we have been above (Skip-gram and CBOW). This is one of the most popular model that is currently used nowadays.
The model is pretty fast training, scalable to huge corpora and has good performance even with small corpus, and small vectors.
Labels:

Comments

Post a Comment

Popular posts from this blog

Quick text files merging, data preparation

It's very often that in natural language processing, you will have to re-format your data to take as inputs to different systems. In this case, these simple linux commands will help you do it much quicker without having to write a script. 1. Merging two files to one file with two column Input f1 looks like this: 1 2 3 4 Input f2 looks like this: a b c d Output f3 will look like this: 1  a 2  b 3  c 4  d Command: paste f1 f2 > f3  The delimiter by default is a tab. You can also define it (for example, separated by a comma) as follows: paste -d ',' f1 f2 > f3 2.  Create a line number to each line of a text file Assume that you want to create an index to each line in a text file, i.e. inserting a line number and then a tab before the content of each line: Input f1: a b c d Output f2: 1  a 2  b 3  c 4  d Command: nl f1 > f2 3. Joining two files with a common field Input f1: 1   aaa...
1 comment
Read more

Skip-gram model and negative sampling

In the previous post , we have seen the 3 word2vec models: skip-gram, CBOW and GloVe. Now let's have a look at negative sampling and what it is used to make training skip-gram faster. The idea is originated from this paper: " Distributed Representations of Words and Phrases and their Compositionality ” (Mikolov et al. 2013) In the previous example , we have seen that if we have a vocabulary of size 10K, and we want to train word vectors of size 300. Then the number of parameters we have to estimate in each layer is 10Kx300. This number is big and makes training prone to over-fitting and gives too much focus on words that appear often, and less focus on rare words. Subsampling of frequent words So the idea of subsampling is that: we try to maximize the probability that "real outside word" appears, and minimize the probability that "random words" appear around center word. Real outside words are words that characterize the meaning of the center word, wh...
Post a Comment
Read more