TAE

ML
DL

Rema
MBert
Bidirectional 1ST
EncodersDecoders
TransformerModels
Advance Tent preproe
Word Embedding Work

ppLmodls
Run f
LSTM RNN GRU
ML Used Cases
Tent preprocessing Bettertechniques
31 Word toVectors
Tent preprossing Converting words into
2 Velfors
Tent pre
processing TF IDF lemm f1is
 converting sentence into words

with spacy Tent blob

Tensorflow Pytorch

Hugging face
library
Tenenbaum It is the process of breaking up tent in
component pieces tokens

2
Sheng process of reducing words tobase
word God may not have
meaning
any
Fast but the word
removes
meaning of
3
hematization

Acetate
stemming SpamDetection Sentiment

Commetization Chat bot Test Summarization

language Translation

Step 2
Treating words to vectors
 One hot
Encoding
words
Bag of engrams
Tf IDF
ram
word 2Vec
es
Iemedogy chat
Corpus Combined Sentences
Paragraph
cons
Documents T
Sentence

of Vocabulary If I harelokwintoonds in
my
words corpus It means I have 10k
vocab

E
FEI.IE
jLEEns.I ID
N
Tokenization Stemming Btw
Eoding
No one famnetiganon Word2Vec
loweringot
eneral cans stop Words Let's IDF

D I am bad
03 I am
girl
D C1 0007 Da 1007
01007 010
200103 COO D
COOOD

anlage Simple to implement intuitive

Dionage 1 Sparse Matin

2 Out of Vocab when
you have deiveasi
3 Not fined sizeof input
sentences
Variables

4 Semantic between words is

meaning
not Captured
5
Extra words the test data
cominginto
cannot be handled

Bagofwords

th get Etta th sent Information

I Venu
Ift
good boy
Da Venn bad boy
3
Trigrams D K good bad
Bignans Venu stood a
boy

Venu is is good goodboy

f
Venn
Ea f Fy fr to Titan
good boy bad Venngood good
boy boybad

In
Venn BMbad person

Uefa
is a
Htt
a bad a bad person

If combination
of n
grams
is 6,3
We have to go from Uni to trigrams

IEI.siIaa
it.i same
there is a
problem with Vector representation

The food is good

tod god is the not
1 I 1 I 0
The food is not good 1 I i n I
When we do the cosine similarity in this case

the cosine will be less value

similarity very
which two sentences are same
means ten
But in am not
reality thy
This issue is fined by Tf IDF

Ite
Add Intuitive Wordimportance is
Captured

Droege
Sparcity Out of Vocab

It
Word Embedding
Word 2Vec
Tebow Clontrinous Bow

Skipgram
Word Embeddings Technique which converts
m un

words into Vectors

Iwordfmhdhd
Count n
Frequency Deep LearningTrain Models
Bow Ta IDF D word 2Ve
ONE HOT CROW
Encoding
Skip grams
Word 2Ve
man

Problem with previous technique Bow Tf IDF

Semantic is not Captured
meaning
Sparse Matan AngeDimensions
All these problems solved Word'Ve
are
by
WFÉÉr word
every we create Vectors

Vectors are in limited Dimensions

Sparsely is Reduced
Semantic Meaning is maintained
 word is features
Every represented by
The vectors which are
watching from the
result will be represented

King 96 0.95 man

0.94 0.92
Queen1 0 96
0.95 women
L O 94 0.967

mail.FI
matched
by the result
i wins

How we
that is Cosima
get by Similarity

Eating
Distance a
f Cosine
similarity
need to find
a aim we

2 Dist I Costas
I O 7071 0 29
 More the value towards O More Similar

EI I Dist I Cosa
a
I Cos 907

If

Excavate continous Bow

Igg word we training forthis Sent

I
K Data Science
opd.MS
I
j

If we m

lets window
odd sothat left9
say size sywataka
Windom
size is helpful
window size
fotheringham
Bigger windowsign Bettermodel
IndependentVan
shyperparamy

Venugopal studying Data 75 Venn 1 000000

goal is Datascience sway o I

isstudyingsu.cn e7echDak9s00100
 It will take this ifp and based
Aimofwordfect
if will to generate actors
try
The vectors should capture semantic meaning
This data is to Ann
now
given a pp
we have 9 Mp each word is represent by7Vee

Soffman layer
085.020.47
8 gfj9
iii

Embedding layer

Fit
 From input words each word will be passed to

each Ifplayernodes
As wehave 5 window size we will have
hidden layer with 5 nodes

Off is I word Each word is represented

by I vectors

Window size can extend but we will be only

having I hidden Geyer

when the forward propagation happens they

I will be compared when the backward prop
happens the weight get updated
1 node one
9 How will we the necks of
the training happens
after
The node which is representing a word will
be 5 vectors window size
given by hiddenlayer

Talitha in the cage connecting

41,117,1
from 5 nodes to fp node I will be the 5 vectors

We can extend the window size to any no

Eggen Ilp word Bow

Pd studying ok O O

gopal is DataScience 4
studyingGopal 6 1 00 0 00

fgiqgadamiso.io
Bee turtle
Sam Classifier Duet
117 Tent prepossessing Tokenization Stopwords
stemming Cammitization NYK
Tent 5 Vectors BOW TFIF
word Vee
Average
Word'Ve
Worf
17 Skip 47 Bow
gram
 Word'Ve

Refrained Trainth model

In WordLuce we convert word into fined dimen

My name is Venu Word'Ve

1111111111

is converted intouecker instead we

a 3000
want the whole sentence to a ND vector
The problem can be fined with AngwordZUec
transpose the Angword2Vec have
If we we

input feature Boop G off father

of
I
soon I
Sentence
 Janie
wusebrary
handles NLP tasks
DIF effectively
with most efficient implementation
of Common Algorithms
one
NLP tasks has only
For all spay
implemented method
the most efficient algo
available we do not have option to choose

other
algorithms

NLTI Natural language Toolkit 2001

It providesmany functionalities but less

efficient Implementations

With us SPACY

more
Form
efficient and faster
NLP tasks spacy is
mon

but the user will

not be able to choose specific Algo

only
 Week has variety of Algo's for all common

tasks

Spacy does not include pre created

models for some apps Sentiment Analysis
Itis easier to perform in NL 717
SPAY BASICS

Estoy IEEE
D loading language library
Building Pipeline Object
3 Using Tokens
4 Parts of speech Tagging
5 Token Attributes
Unduesting

EEE.I.EE
 We create an nape from pacy which

takes tent and automatically takes rain tent

and series of operations to tag

performs a
tent data
parse and describe the

TÉÉÉÉFan of breaking up
ram

ant into component pieces Tokens

Tokens Tent
are pieces of original
understand the
Everything that helps us
is derived fromtokens
the tent
meaning of our another
and thier relationship to

Tphf.sn
Igmngchanfanfin sInbetmem
Suffin onding
to a
Eton special case split
string
into tokens or not to spot a string into
aural tokens b S let's
In
 STEMMINI
when searching certain words the tant
returns similar words
if
boat boats

boating

Spacey doesnt include stammer but a more

advanced one Lemnitization

IIigginners
 KREMER more aculeate It offers slight
It is
improvement
owe the Porter Stemmer both in
speed
legit g

Ignmmigniniteyend word remain and

a
to apply
consider a language's full vocab

Morphological to words
analysis

Lemmetization is much more informative then

 simple Stemming
at words too
Lemmetization looks surrounding
to determine a words Pos part of Speech
can be used in Sentiment Analysis
Stemming
because it to understand the
just needs
barn ward
used in chat bot QIA
can be
hemmetization
response ah
programs

like th them byet

rds words of
value in the sentences
does not add much
remove these words
helpm to
Stop words

may produce It
intermediate
Probl withemming
not harm
the word which may any
representation of
meaning
Sentz
ÉÉ 3 Sta is
Ég
a
my
good gird
7 Bogey
sent 3 Girl am good

If lower cases

stemming Lemetization
He Stop words
Sent 1 good boy
sent 2 good girl
sent 3 7
Boy girl good

Now check the frequency of key

words

3rd
fefuds
good we have 2

y 2 7 my.sn
Fogh
girl 2
words iLEI
 have same
weightage
In sent
Dig
we

we
for both words Usually
good g boy
need to give move for important
weighpage
keywords No sematic MEANIN Sparcity
00
To Overcome this we have TF IDF

From skleann feature infraities tant import

Countbeetorizer ordering ofthe words is
lost
TF IDF

gDamentfvmp
I o
su.IT g
Senta 3 Sta is a
good gird
7 Bogey Girl am good
sent 3

Il lower cases

stemming Lemetization

I Stop words we should

give
 Send 1 mo
good boy garage
sent 2 to word which
good guy
a cane
Sent 3 3
Boy guy good

Now check the frequency of key

words

good
3ha

É
boy

gin
2
27rem Ég
Term frequency the

captured by IDF

IDF

logging
keyword

set
As good is present
in

boy log 3 2 sentence it is not given

more or
 g al g Leightage
p

TxIDI

from stlearn feature extraction tent import

Tfidfvacherizer

W
ÉÉ semantic info is not sand

each word is represented vector 32 or

as a
of
More dimensions instead no
of a
single
The semantic words is
info G relation btwn diff
also captured
 EI

Women 3 216 5

Wedding
Word avec Glove
into Vacher Representation
Converting the
words

of 10,000 One hot Representation t

14
man 5000 indent
women 9000
fighestmanywords
8
hard to generalist the
and
É
I model
go sparse siz.is big
to
so
many
o's less I's
 É Raron
words like boy gig it
there are
If words into
to convert the
We
try
based on some features
Vectors
Place e ti
Aga Area
Ewes Gender

a i.at 4
Raga

04 9 9,4 t

fond
I I

BordinmindammensoneMain
Analogy 9
I
Similarly Kingsmen i

armiggitaniiiigit
Of
Bug IL 520001dm
soggy

Y A keras will convert

Ii if

win
many age
is y

É Gmbedding layer
Nether Representation
the sentence into
It converts

provide how many dimensions you

Ires want
feature Representation of a
It is erasing a
word
all sentenies should be in same