Word Embeddings - Introduction

INTRODUCTION WORD2VEC WORD MOVERS DISTANCE Q&A
WORD Embeddings
A non-exhaustive introduction to Word Embeddings
x
y
z
w
Christian S. Perone
christian.perone@gmail.com

AGENDA
INTRODUCTION
Philosophy of Language
Vector Space Model
Embeddings
Word Embeddings
Language Modeling
WORD2VEC
Introduction
Semantic Relations
Other properties
WORD MOVERS DISTANCE
Rationale
Model
Results
Q&A

WHO AM I
Christian S. Perone
Machine Learning/Software Engineer
Blog
http://blog.christianperone.com
Open-source projects
https://github.com/perone
Twitter @tarantulae

Section I
INTRODUCTION

PHILOSOPHY OF LANGUAGE
(...) the meaning
of a word is its use
in the language.
—Wittgenstein, Ludwig,
Philosophical Investigations – 1953

VECTOR SPACE MODEL
Interpreted in a lato sensu, VSM is a space where text is
represented as a vector of numbers instead of its original
textual representation

VECTOR SPACE MODEL
Many approaches to go from other spaces to a vector space

VECTOR SPACE MODEL
Many approaches to go from other spaces to a vector space
Many advantages when you have vectors with special
properties

VECTOR SPACE MODEL

EMBEDDINGS
From a space with one dimension per word to a continuous vector
space with much lower dimensionality. From one mathematical object
to another, but preserving “structure”.
Source: Our beloved scikit-learn.

WORD EMBEDDINGS
Word Model
Word
Embedding
V(cat) = [ 1.4, -1.3, ... ]
Cat
sat
mat
on
cat = [ 0, 1, 0, ... ]
Sparse
Dense
From a sparse representation (usually one-hot encoding) to a
dense representation
Embeddings created as by-product vs explicit model

LANGUAGE MODELING
P(w1, · · · , wn) =
i
P(wi | w1, · · · , wi−1)
P(”the cat sat on the mat“) > P(”the mat sat on the cat“)
Useful for many different tasks, such as speech recognition,
handwriting recognition, translation, etc.
Naive counting: doesn’t generalize, too many possible
sentences
A word sequence on which the model will be tested is likely to
be different from all the word sequences seen during training.
[Bengio et al, 2003]
Markov assumption / how to approximate it

MARKOV ASSUMPTION AND N-GRAM MODELS
Markov assumption simpliﬁes the model, it tries to approximate the
components of the product.
Unigram: P(w1, · · · , wn) =
i
P(wi)
Bigram: P(wi | w1, · · · , wi−1) ≈ P(wi | wi−1)
Extend to trigram, 4-gram, etc.

WORD EMBEDDINGS
CHARACTERISTICS
Language modeling, low-dimensional and dense but increased complexity.
Examples: neural language models, word2vec, GloVe, etc.
Source: Bengio et al., 2003
Classic neural language model
proposed by Bengio et al. in
2003.
After that, many other important
works by Collobert and
Weston (2008) and then by
Mikolov et al (2013).

Section II
WORD2VEC

WORD2VEC
Unsupervised technique with supervised tasks, takes a text
corpus and produces word embeddings as output. Two different
architectures:
w(t-2)
w(t+1)
w(t-1)
w(t+2)
w(t)
SUM
INPUT PROJECTION OUTPUT
w(t)
INPUT PROJECTION OUTPUT
w(t-2)
w(t-1)
w(t+1)
w(t+2)
CBOW Skip-gram
Figure 2: Graphical representation of the CBOW model and Skip-gram model. In the CBOW model, the distributed
representations of context (or surrounding words) are combined to predict the word in the middle. In the Skip-gram
model, the distributed representation of the input word is used to predict the context.
Source: Exploiting Similarities among Languages for Machine Translation. Mikolov, Thomas et al.
2013.

WORD2VEC
Source: TensorFlow

AMAZING EMBEDDINGS
Semantic relationships are often preserved on vector operations.
Source: TensorFlow

WORD ANALOGIES
Suppose we have the vector w ∈ Rn of any given word such as wking,
then we can do:

WORD ANALOGIES
then we can do:
wking − wman + wwoman ≈ wqueen
This vector operation shows that the closest word vector to the
resulting vector is the vector wqueen.

WORD ANALOGIES
then we can do:
wking − wman + wwoman ≈ wqueen
This vector operation shows that the closest word vector to the
resulting vector is the vector wqueen.
This is an amazing property for the word embeddings, because it
means that they carry important relational information that can
be used to many different tasks.

LANGUAGE STRUCTURE
−0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
one
two
three
four
five
→
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
−0.6
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
uno (one)
dos (two)
tres (three)
cuatro (four)
cinco (five)
−0.3 −0.25 −0.2 −0.15 −0.1 −0.05 0 0.05 0.1 0.15
−0.3
−0.25
−0.2
−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
cat
dog
cow
horse
pig
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
−0.5
−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
0.4
0.5
gato (cat)
perro (dog)
vaca (cow)
caballo (horse)
cerdo (pig)
Figure 1: Distributed word vector representations of numbers and animals in English (left) and Spanish (right). The ﬁve
vectors in each language were projected down to two dimensions using PCA, and then manually rotated to accentuate
their similarity. It can be seen that these concepts have similar geometric arrangements in both spaces, suggesting that
Source: Exploiting Similarities among Languages for Machine Translation. Mikolov, Thomas et al.
2013.

DEEP LEARNING ?
Word2vec isn’t Deep Learning, the model is actually very shallow.

DEEP LEARNING ?
However, there is an important relation here, because word
embeddings are usually used to initialize dense LSTM embeddings
for different tasks using deep architectures.

DEEP LEARNING ?
However, there is an important relation here, because word
embeddings are usually used to initialize dense LSTM embeddings
for different tasks using deep architectures.
Also, you can of course train Word2vec models using techniques
developed in Deep Learning context.

DEMO
Demo time for some word analogies in Portuguese.
Model trained by Kyubyong Park.
Trained on Wikipedia (pt) - 1.3GB corpus
w ∈ Rn
where n is 300
Vocabulary size is 50246
Model available at https://github.com/Kyubyong/wordvectors
For comparison, Wikipedia (en) is 13.5GB

Section III

Word2vec deﬁnes a vector for word, but how can we use its
information to compare documents ?

Word2vec deﬁnes a vector for word, but how can we use its
information to compare documents ?
There are many approaches to represent documents, to mention:
BOW, TF-IDF, N-grams, etc. However, they frequently show
near-orthogonality.

Take the two sentences:
“Obama speaks to the media in Illinois”
and
“The President greets the press in Chicago”

Take the two sentences:
“Obama speaks to the media in Illinois”
and
“The President greets the press in Chicago”
While these sentences have no words in common, they convey
nearly the same information, a fact that cannot be represented by
the BOW model (Kusner, Matt J. et al. 2015).

KILIAN@WUSTL.EDU
is, MO 63130
‘Obama’
word2vec embedding
‘President’
‘speaks’
‘Illinois’
‘media’
‘greets’
‘press’
‘Chicago’
document 2document 1
Obama
speaks
to
the
media
in
Illinois
The
President
greets
the
press
in
Chicago
Figure 1. An illustration of the word mover’s distance. All
non-stop words (bold) of both documents are embedded into a
Source: From Word Embeddings To Document Distances. Kusner, Matt J. et al. 2015.

rd Embeddings To Document Distances
3a) for more
ture and the
m model can
f words per
The ability
del to learn
- vec(sushi)
(Einstein) -
) (Mikolov
g is entirely
xt corpus of
ough we use
ghout, other
eston, 2008;
The President greets the press in Chicago.
Obama speaks in Illinois.
1.30
D1
D2
D3
D0
D0 The President greets the press in Chicago.
Obama speaks to the media in Illinois.
The band gave a concert in Japan.
0.49 0.42 0.44
0.200.240.451.07
1.63
+ +=
= + + + 0.28
0.18+
Figure 2. (Top:) The components of the WMD metric between a
query D0 and two sentences D1, D2 (with equal BOW distance).
The arrows represent ﬂow between two words and are labeled
with their distance contribution. (Bottom:) The ﬂow between two
sentences D3 and D0 with different numbers of words. This mis-

From Word Embeddings To Document Distances
1 2 3 4 5 6 7 8
0
10
20
30
40
50
60
70
twitter recipe ohsumed classic reuters amazon
testerror%
43
33
44
33 32 32
29
66
63 61
49 51
44
36
8.0 9.7
62
44 41
35
6.9
5.0
6.7
2.8
33
29
14
8.16.96.3
3.5
59
42
28
14
17
12
9.3
7.4
34
17
22
21
8.4
6.4
4.3
21
4.6
53 53
59
54
48
45
43
51
56 54
58
36
40
31
29
27
20newsbbcsport
k-nearest neighbor error
BOW [Frakes & Baeza-Yates, 1992]
TF-IDF [Jones, 1972]
Okapi BM25 [Robertson & Walker, 1994]
LSI [Deerwester et al., 1990]
LDA [Blei et al., 2003]
mSDA [Chen et al., 2012]
Componential Counting Grid [Perina et al., 2013]
Word Mover's Distance
Figure 3. The kNN test error results on 8 document classiﬁcation data sets, compared to canonical and state-of-the-art baselines methods.
1 2 3 4 5 6 7 8
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
averageerrorw.r.t.BOW
1.6
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0
1.29
1.15
1.0
0.72
0.60 0.55
0.49 0.42
BOW
TF-IDF
Okapi BM25
LSI
LDA
mSDA
CCG
WMD
Figure 4. The kNN test errors of various document metrics aver-
aged over all eight datasets, relative to kNN with BOW.
w, TF(w, D) is its term frequency in document D, |D| is
Table 2. Test error percentage and standard deviation for different
text embeddings. NIPS, AMZ, News are word2vec (w2v) models
trained on different data sets whereas HLBL and Collo were also
obtained with other embedding algorithms.
DOCUMENT k-NEAREST NEIGHBOR RESULTS
DATASET HLBL CW NIPS AMZ NEWS
(W2V) (W2V) (W2V)
BBCSPORT 4.5 8.2 9.5 4.1 5.0
TWITTER 33.3 33.7 29.3 28.1 28.3
RECIPE 47.0 51.6 52.7 47.4 45.1
OHSUMED 52.0 56.2 55.6 50.4 44.5
CLASSIC 5.3 5.5 4.0 3.8 3.0
REUTERS 4.2 4.6 7.1 9.1 3.5
AMAZON 12.3 13.3 13.9 7.8 7.2

Section IV
Q&A

Q&A

Word Embeddings - Introduction

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Word Embeddings - Introduction

Similar to Word Embeddings - Introduction (20)

More from Christian Perone

More from Christian Perone (11)

Recently uploaded

Recently uploaded (20)

Word Embeddings - Introduction