Abstract:- The talk will focus on how Neural Networks are applied in the field of NLP for tasks like classification. Building blocks like Word Embeddings, Recurrent NN, LSTM, GRU, Convolutional NN, Sentence Representation and how they are applied to a piece of text in Tensorflow will be covered. These building blocks can be stacked together in various ways to form deeper network architectures. We will discuss one such architecture which is used within GumGum Inc to do Sentiment Analysis on web pages using NN in Tensorflow.
3. GumGum Inc
3
Artificial Intelligence Company
9 year old, 225 Employees
Based in Santa Monica, CA
Offices in London, Sydney, New York,
Chicago
Thousands of Publishers and Advertisers
Billions of Impressions per day
8. Brain’s mapping to Deep Learning
8
USC is the oldest private
research university in
California.
USCFight On!!
0.58 1.60 0.10 -0.85 -0.18 -0.97 -0.14 1.22 1.89 1.91 0.78 0.33 -0.66 -0.19 1.59
Embeddings
9. Deep Learning in NLP...
9
Machine Learning
Bags of
n-grams
Index
Weight
Deep Learning
Word
Sequence
Vector
Index
17. Sentences : Recurrent Neural Network (RNN)
17
I like playing soccer
RNN RNN RNN RNN
y1
y2
y3
y4
x1
x2
x3
x4
h1
h2
h3
h4
WX
WX
WX
WX
Wh
Wh
Wh
Wo
Wo
Wo
Wo
h0
Wh
18. Sentences : RNN : Long Short Term Memory (LSTM)
18
x
LSTM
y1
I
LSTM
y1
x1
h1
WX
Wh
Wo
h0
19. LSTM
Sentences : RNN : Long Short Term Memory (LSTM)
19
xt
tanh
tanh
ht
Ct
Ct-1
ht-1
20. Sentences : RNN : Long Short Term Memory (LSTM)
20
xt
tanh
tanh
ht
Ct
Ct-1
ht-1
forget gate
ft
ft
= ( Wxf
xt
+ Whf
ht-1
)
Whf
Wxf
21. Sentences : RNN : Long Short Term Memory (LSTM)
21
xt
tanh
tanh
ht
Ct
Ct-1
ht-1
input gate
ft
it
= ( Wxi
xt
+ Whi
ht-1
)
Whf
Wxf
Whi
Wxi
it
Ct
~
Whc
Wxc
= tanh ( Wxc
xt
+ Whc
ht-1
)Ct
~
22. Sentences : RNN : Long Short Term Memory (LSTM)
22
xt
tanh
tanh
ht
Ct
Ct-1
ht-1
cell state update
ft
Whf
Wxf
Whi
Wxi
it
Ct
~
Whc
Wxc
Ct
= ft
* Ct-1
+ it
*Ct
~
23. Sentences : RNN : Long Short Term Memory (LSTM)
23
xt
tanh
tanh
ht
Ct
Ct-1
ht-1
output gate
ft
Whf
Wxf
Whi
Wxi
it
Ct
~
Whc
Wxc
ot
= ( Wxo
xt
+ Who
ht-1
)
Who
Wxo
ht
= ot
* tanh( Ct
)
ot