The document outlines a seminar agenda on word embeddings, focusing on Word2Vec and its comparison with models like GloVe. It details the architecture of neural networks for language modeling, including the continuous bag-of-words (CBOW) and skip-gram methods, as well as techniques for representing words through co-occurrence probabilities. Additionally, it references several resources for further learning in natural language processing.

1. Hyunyoung Lee
Seminar for NLP labs
Word2Vec

2. Agenda
1. Word Embedding
- Vectorization of Image and Text
Word2Vec
2. Word2Vec
- One-hot vector and Co-occurrence matrix for word vector
3. Fundamental
- Basic component of word embedding in a neural net
4. Word Vector in a neural net
5. Word2Vec, CBOW and skip-gram
- Comparing Image processing with Word Vector about vector presentation.
6. Glove

3. - Image Vector representation
1. Word Embedding Word2Vec
- RGB Values of every pixel like Height * Width * RGB as a value in a row
So it is easy to make the image a vector in some space, i.e. RGB space.

4. - What is the Word Embedding ?
1. Word Embedding Word2Vec
- In NLP tasks, Before a neural net,
Word vector is represented by Word frequency like TF-IDF and so on.
In a neural net, There are multiple tries for word vector representation :
- Language modeling and Word embedding modeling

5. One-hot representation
Dim = |V| (v is the size of vocabulary)
- motel
- hotel
If you search for [Seattle motel] key word, we want the search engine to match web page containing
“Seattle hotel”
Similarity(motal, hotel) = 0
motel
hotel = 0
If we do inner product with the above vectors, we can not find out similarity between words
2. Word2Vec Word2Vec
T

6. Co-occurrence matrix
Let’s see window based co-occurrence matrix
- Example Corpus :
- I like deep learning.
- I like NLP.
- I enjoy flying.
Total vocabulary size(|V|) = 8
Vector(“I”) = [0, 2, 1, 0, 0, 0, 0, 0]
Vector(“like”) = [2, 0, 0, 1, 0, 1, 0 , 0] …
2. Word2Vec Word2Vec

7. Co-occurrence with SVD
With SVD(Singular Value Decomposition)
- this calculation is so expensive and not efficient. For example, for M * N matrix is O(mn )
- SVD based methods don’t scale well for big matrices, and it is too hard to incorporate new words
or documents
2. Word2Vec Word2Vec
2

8. 3. Fundamental Word2Vec
output layer’s values is regarded as :
- score
- probability
Backpropagation makes those value maximum or
minimum
Feedforward Neural Network(Basic Neural Network)

9. - Embedding Layer(Inner product)
3. Fundamental Word2Vec
- Intermediate Layer(s)
- Softmax Layer
- One or more layer that produce an intermediate representation of the input
For Example, Hidden layer with tanh, sigmoid activation function or RNN(LSTM, GRU) which is
state-of-the-art neural language models.
- The final layer to compute the probability distribution over words in total vocabulary.

10. Language model and Word embedding model with a neural net
- The main purpose of language model is to compute the probability of a sentence or sequence of
words and the probability of an upcoming word
The probability of a sequence of m words {W1, … , Wm} is denoted as P(W1, … , Wm)
P(W1, … , Wm) is conditioned on a widow of n previous words : P(Wt | Wt-1 , … , Wt-n+1)
i.e. The probability of a sentence or sequence of words :
The probability of an upcoming word :
- So, a model that computes either of those probability above is called a language model(LM)
- The Chain Rule applied to compute joint probability of words in sentence
Markov Assumption :
for example, P(“it water is so transparent”) =
P(its) * P(water | its) * P(is | its water) * P(so | its water is) * P(transparent | its water is so)
By Markov Assumption, the probability of the above sentence :
OR
4. Word Vector in a neural net Word2Vec

11. Language model and Word embedding model with a neural net
- How to estimate these probability
In N-gram based language model -
For example, bigram -
trigram -
4. Word Vector in a neural net Word2Vec

12. Language model and Word embedding model with a neural net
- The first deep neural network architecture model For NLP presented by Bengio et al(2003) to predict
P(Wt | Wt-1 , … , Wt-n+1)
- This model is prototype which we now refer to as a word embedding.
There is some issue :
- softmax layer
- computing power
4. Word Vector in a neural net Word2Vec
Classic neural language model (Bengio et al. 2003)

13. Language model and Word embedding model with a neural net
- A little more model than Begino et al is C&W model(2011)
There is some variation :
- changing cost function like the above
4. Word Vector in a neural net Word2Vec
The C&W model without ranking objective(collobert et al. 2011)

14. Language model and Word embedding model with a neural net
- Another way to make word2vec in a neural net
- In NLP, transfer learning is word2vec, BUT Sometimes
we could make word2vec on the specific task using a neural net
4. Word Vector in a neural net Word2Vec

15. Distributional similarity based representations
A lot of value by presenting a word by means of its neighbors
One of the most successful ideas of modern statistical NLP
5. Word2Vec, CBOW, skip-gram Word2Vec
Banking

16. Google’s Word2Vec – CBOW, skip-gram
Goal : simple (shallow) neural network model
Learning from billion words scale corpus
Predict middle word from neighbors with
A fixed size context window
1. Skip-gram
2. CBOW(continuous bag-of-words)
5. Word2Vec, CBOW, skip-gram Word2Vec

17. Skip-gram
Method : Predict neighbor Wt+j given word Wt
Maximizes following average log probability
5. Word2Vec, CBOW, skip-gram Word2Vec
Skip-gram(Mikolov et al. 2013)

18. CBOW
Method : Predict word given bag-of-neighbors
Loss function =
5. Word2Vec, CBOW, skip-gram Word2Vec
CBOW(Mikolov et al. 2013)

19. Skip-gram & CBOW
WV*N (WIN)and W’N*V (WOUT) is embedding layer.
N of these embedding layer is word2vec’s dimension
5. Word2Vec, CBOW, skip-gram Word2Vec

20. Let’s see an example of skip-gram
5. Word2Vec, CBOW, skip-gram Word2Vec

21. Word Analogies with Word2Vec
[king] – [man] + [woman] ≈ [queen]
5. Word2Vec, CBOW, skip-gram Word2Vec

22. Word Analogies with Word2Vec
[king] – [man] + [woman] ≈ [queen]
5. Word2Vec, CBOW, skip-gram Word2Vec

23. Global statistics of co-occurrence probability
6. Glove Word2Vec

24. Global statistics of co-occurrence probability
6. Glove Word2Vec
Glove visualization Company – CEO Superlatives

25. Word2Vec vs Glove
6. Glove Word2Vec

26. Stanford lecture(Online)
CS224n : Natural Language Processing with Deep Learning
- lecture note1 : http://web.stanford.edu/class/cs224n/lecture_notes/cs224n-2017-notes1.pdf
- lecture note2 : http://web.stanford.edu/class/cs224n/lecture_notes/cs224n-2017-notes2.pdf
- lecture note5 : http://web.stanford.edu/class/cs224n/lecture_notes/cs224n-2017-notes5.pdf
- lecture slide2 :http://web.stanford.edu/class/cs224n/lectures/cs224n-2017-lecture2.pdf
CS223n : Convolutional Natural Networks For Visual Recognition
- lecture note : Neural networks Part 1: Setting up the Architecture http://cs231n.github.io/neural-networks-1/
- lecture note : Linear classification : Support Vector Machine, Softmax http://cs231n.github.io/linear-classify/
Sebastian Ruder blog : http://ruder.io/word-embeddings-1/index.html#fn:2
Colah’s blog : http://colah.github.io/posts/2014-07-NLP-RNNs-Representations/
Neural Text Embedding for information Retrieval (WSDM 2017) by MicroSoft
Mikolov, T., Corrado, G., Chen, K., & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. Proceedings
of the International Conference on Learning Representations (ICLR 2013), 1–12
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Distributed Representations of Words and Phrases and their Compositionali
ty. NIPS, 1–9.
Reference Word2Vec
Private Blog
ACM International Conference on Web Search and Data mining
Paper