Natural Language Processing

with Deep Learning

CS224N/Ling284

Christopher Manning
Lecture 2: Word Vectors,Word Senses, and
Classifier Review
 Lecture Plan
Lecture 2: Word Vectors and Word Senses
1. Finish looking at word vectors and word2vec (10 mins)
2. Optimization basics (8 mins)
3. Can we capture this essence more effectively by counting? (12m)
4. The GloVe model of word vectors (10 min)
5. Evaluating word vectors (12 mins)
6. Word senses (6 mins)
7. Review of classification and how neural nets differ (10 mins)
8. Course advice (2 mins)

Goal:
2 be able to read word embeddings papers by the end of class
1. Review: Main idea of word2vec
• Start with random word vectors
• Iterate through each word in the whole corpus
• Try to predict surrounding words using word vectors
𝑃 𝑤5:9 | 𝑤5 𝑃 𝑤569 | 𝑤5

𝑃 𝑤5:7 | 𝑤5 𝑃 𝑤567 | 𝑤5

… problems turning into banking crises as …

,- )
• 𝑃 𝑜𝑐 =
&'((*+ . • Update vectors so you
,- )
∑1∈3 &'((*1 . can predict better
• This algorithm learns word vectors that capture word
3
similarity and meaningful directions in a wordspace
Word2vec parameters and computations
••••• ••••• • •
• • • •• • • • •• • •
• • • •• • • • •• • •
• • • •• • • • •• • •
• • • •• • • • •• • •
• • • •• • • • •• • •
U V 𝑈. 𝑣>? softmax(𝑈. 𝑣>? )
outside center dot product probabilities

! Same predictions at each position

We want a model that gives a reasonably

high probability estimate to all words
that occur in the context (fairly often)
4
Word2vec maximizes objective function by
putting similar words nearby in space

5
2. Optimization: Gradient Descent
• We have a cost function 𝐽 𝜃 we want to minimize
• Gradient Descent is an algorithm to minimize 𝐽 𝜃
• Idea: for current value of 𝜃, calculate gradient of 𝐽 𝜃 , then take
small step in the direction of negative gradient. Repeat.

Note: Our
objectives
may not
be convex
like this

6
Gradient Descent
• Update equation (in matrix notation):

𝛼 = step size or learning rate

• Update equation (for a single parameter):

• Algorithm:

7
Stochastic Gradient Descent
• Problem: 𝐽 𝜃 is a function of all windows in the corpus
(potentially billions!)
• So is very expensive to compute
• You would wait a very long time before making a single update!

• Very bad idea for pretty much all neural nets!

• Solution: Stochastic gradient descent (SGD)
• Repeatedly sample windows, and update after each one
• Algorithm:

8
Stochastic gradients with word vectors!
• Iteratively take gradients at each such window for SGD
• But in each window, we only have at most 2m + 1 words,
so is very sparse!

9
Stochastic gradients with word vectors!
• We might only update the word vectors that actually appear!

• Solution: either you need sparse matrix update operations to

only update certain rows of full embedding matrices U and V,
or you need to keep around a hash for word vectors

[ ]
d

|V|

• If you have millions of word vectors and do distributed

computing, it is important to not have to send gigantic
updates around!

10
1b. Word2vec: More details
Why two vectors? à Easier optimization. Average both at the end
• But can do algorithm with just one vector per word
Two model variants:
1. Skip-grams (SG)
Predict context (“outside”) words (position independent) given center
word
2. Continuous Bag of Words (CBOW)
Predict center word from (bag of) context words
We presented: Skip-gram model

Additional efficiency in training:

1. Negative sampling
So far: Focus on naïve softmax (simpler, but expensive, training method)
11
The skip-gram model with negative sampling (HW2)

• The normalization factor is too computationally expensive.

,- )
&'((*+ .
•𝑃 𝑜𝑐 = ,- )
∑1∈3 &'((*1 .

• Hence, in standard word2vec and HW2 you implement the skip-

gram model with negative sampling

• Main idea: train binary logistic regressions for a true pair (center
word and word in its context window) versus several noise pairs
(the center word paired with a random word)

12
The skip-gram model with negative sampling (HW2)

• From paper: “Distributed Representations of Words and Phrases

and their Compositionality” (Mikolov et al. 2013)
• Overall objective function (they maximize):

• The sigmoid function:

(we’ll become good friends soon)
• So we maximize the probability
of two words co-occurring in first log
à
13
The skip-gram model with negative sampling (HW2)

• Notation more similar to class and HW2:

• We take k negative samples (using word probabilities)

• Maximize probability that real outside word appears,
minimize prob. that random words appear around center word

• P(w)=U(w)3/4/Z,
the unigram distribution U(w) raised to the 3/4 power
(We provide this function in the starter code).
• The power makes less frequent words be sampled more often
14
3. Why not capture co-occurrence counts directly?

With a co-occurrence matrix X

• 2 options: windows vs. full document
• Window: Similar to word2vec, use window around
each word à captures both syntactic (POS) and
semantic information
• Word-document co-occurrence matrix will give
general topics (all sports terms will have similar
entries) leading to “Latent Semantic Analysis”

15
Example: Window based co-occurrence matrix
• Window length 1 (more common: 5–10)
• Symmetric (irrelevant whether left or right context)
• Example corpus:
• I like deep learning.
• I like NLP.
• I enjoy flying.

16
Window based co-occurrence matrix
• Example corpus:
• I like deep learning.
• I like NLP.
• I enjoy flying.
counts I like enjoy deep learning NLP flying .
I 0 2 1 0 0 0 0 0
like 2 0 0 1 0 1 0 0
enjoy 1 0 0 0 0 0 1 0
deep 0 1 0 0 1 0 0 0
learning 0 0 0 1 0 0 0 1
NLP 0 1 0 0 0 0 0 1
flying 0 0 1 0 0 0 0 1
. 0 0 0 0 1 1 1 0
17
Problems with simple co-occurrence vectors

Increase in size with vocabulary

Very high dimensional: requires a lot of storage

Subsequent classification models have sparsity issues

à Models are less robust

18
Solution: Low dimensional vectors
• Idea: store “most” of the important information in a fixed, small
number of dimensions: a dense vector

• Usually 25–1000 dimensions, similar to word2vec

• How to reduce the dimensionality?

19
Method: Dimensionality Reduction on X (HW1)
Singular Value Decomposition of co-occurrence matrix X
Factorizes X into UΣVT, where U and V are orthonormal

k
X

Retain only k singular values, in order to generalize.

𝑋J is the best rank k approximation to X , in terms of least squares.
20
Classic linear algebra result. Expensive to compute for large matrices.
Simple SVD word vectors in Python
Corpus:
I like deep learning. I like NLP. I enjoy flying.

21
Simple SVD word vectors in Python
Corpus: I like deep learning. I like NLP. I enjoy flying.
Printing first two columns of U corresponding to the 2 biggest singular values

22
Hacks to X (several used in Rohde et al. 2005)

Scaling the counts in the cells can help a lot

• Problem: function words (the, he, has) are too
frequent à syntax has too much impact. Some fixes:
• min(X,t), with t ≈ 100
• Ignore them all

• Ramped windows that count closer words more

• Use Pearson correlations instead of counts, then set
negative values to 0
• Etc.
23
Interesting syntactic patterns emerge in the vectors
CHOOSING
CHOOSE
CHOSE
CHOSEN

STOLEN
STEAL
STOLE
STEALING

TAKE
SPOKE SPEAK
SPOKEN
SPEAKING TAKEN TAKING
TOOK
THROW
THROWN THREW
THROWING

SHOWN
SHOWED EATEN
EAT
SHOWING ATE
EATING

SHOW

GROWN
GROW
GREW

GROWING

FigureCOALS model from

11: Multidimensional scaling of present, past, progressive, and past participle forms for eight verb families.
An Improved Model of Semantic Similarity Based on Lexical Co-Occurrence
Rohde et al. ms., 2005
24

22
Rohde, Gonnerman, Plaut Modeling Word Meaning Using Lexical Co-Occurren
Interesting semantic patterns emerge in the vectors
DRIVER

JANITOR
DRIVE SWIMMER
STUDENT

CLEAN TEACHER

DOCTOR
BRIDE
SWIM
PRIEST

LEARN TEACH
MARRY

TREAT PRAY

COALS Figure
model13: Multidimensional scaling for nouns and their associated verbs.
from
An Improved Model of Semantic Similarity Based on Lexical Co-Occurrence
Table 10 Rohde et al. ms., 2005
The 25
10 nearest neighbors and their percent correlation similarities for a set of nouns, under the COALS-14K mode
gun point mind monopoly cardboard lipstick leningrad feet
4. Towards GloVe: Count based vs. direct prediction

• LSA, HAL (Lund & Burgess), • Skip-gram/CBOW (Mikolov et al)

• COALS, Hellinger-PCA (Rohde • NNLM, HLBL, RNN (Bengio et
et al, Lebret & Collobert) al; Collobert & Weston; Huang et al; Mnih
& Hinton)

• Fast training • Scales with corpus size

• Efficient usage of statistics
• Inefficient usage of statistics
• Primarily used to capture word • Generate improved performance
similarity on other tasks
• Disproportionate importance
given to large counts • Can capture complex patterns
beyond word similarity

26
 Encoding meaning in vector differences
[Pennington, Socher, and Manning, EMNLP 2014]

Crucial insight: Ratios of co-occurrence probabilities can encode

meaning components

x = solid x = gas x = water x = random

large small large small

small large large small

large small ~1 ~1
 Encoding meaning in vector differences
[Pennington, Socher, and Manning, EMNLP 2014]

Crucial insight: Ratios of co-occurrence probabilities can encode

meaning components

x = solid x = gas x = water x = fashion

1.9 x 10-4 6.6 x 10-5 3.0 x 10-3 1.7 x 10-5

2.2 x 10-5 7.8 x 10-4 2.2 x 10-3 1.8 x 10-5

8.9 8.5 x 10-2 1.36 0.96

 Encoding meaning in vector differences
Q: How can we capture ratios of co-occurrence probabilities as
linear meaning components in a word vector space?

A: Log-bilinear model:

with vector differences

 Combining the best of both worlds
GloVe [Pennington et al., EMNLP 2014]

• Fast training
• Scalable to huge corpora
• Good performance even with
small corpus and small vectors
GloVe results

Nearest words to
frog:

1. frogs
2. toad
3. litoria litoria leptodactylidae
4. leptodactylidae
5. rana
6. lizard
7. eleutherodactylus

rana eleutherodactylus
31
5. How to evaluate word vectors?
• Related to general evaluation in NLP: Intrinsic vs. extrinsic
• Intrinsic:
• Evaluation on a specific/intermediate subtask
• Fast to compute
• Helps to understand that system
• Not clear if really helpful unless correlation to real task is established
• Extrinsic:
• Evaluation on a real task
• Can take a long time to compute accuracy
• Unclear if the subsystem is the problem or its interaction or other
subsystems
• If replacing exactly one subsystem with another improves accuracy à
Winning!

32
Intrinsic word vector evaluation
• Word Vector Analogies

a:b :: c:?

man:woman :: king:?

• Evaluate word vectors by how well

their cosine distance after addition
captures intuitive semantic and king
syntactic analogy questions
• Discarding the input words from the
search!
woman
• Problem: What if the information is
man
there but not linear?

33
Glove Visualizations

34
Glove Visualizations: Company - CEO

35
Glove Visualizations: Superlatives

36
 2013); skip-gram (SG) and CBOW results are
SG 300 1B 61 61 61 †
from evaluation
Analogy (Mikolov et and
al., 2013a,b); we trained SG
hyperparameters
s of CBOW † 300 1.6B 16.1 52.63 36.1
(19) and CBOW using the word2vec tool . See text
vLBLvectors
up-Glove word 300 1.5B 54.2 64.8 60.0
evaluation
for details and a description of the SVD models.
fre- ivLBL 300 1.5B 65.2 63.0 64.0
when
er of Model
GloVe Dim. Size Sem.
300 1.6B 80.8 Syn. Tot.
61.5 70.3
e to
er is ivLBL
SVD 100
300 1.5B6B 55.9 6.3 50.18.1 53.2
7.3
arge
(17) HPCA
SVD-S 100
300 1.6B6B 36.7 4.2 16.4
46.6 10.8
42.1
gen-
e we GloVe
SVD-L 100 1.6B
300 6B 67.556.6 54.3
63.0 60.3
60.1
SG † 300
CBOW 1B
6B 63.6 61 67.4 61 65.761
SG†
CBOW 300 1.6B6B 73.016.1 66.0
52.6 69.1
36.1
1,
,(19) vLBL
GloVe 300 1.5B6B 77.454.2 67.0
64.8 60.0
71.7
(20) ivLBL 1000
CBOW 300 1.5B6B 57.365.2 68.9
63.0 63.7
64.0
when
GloVe
SG 300 1.6B
1000 6B 80.866.1 61.5
65.1 70.3
65.6
e to
37 SVD
SVD-L 300 42B 6B 38.4 6.3 58.28.1 49.2
7.3
 Analogy evaluation and hyperparameters

• More data helps • Dimensionality

• Wikipedia is better than • Good dimension is ~300
news text!
80 70
Semantic Syntactic Overall
85
70 65
80

75 60 60
Accuracy [%]

Accuracy [%]
Accuracy [%]
70
50 55
65

60 40 50
Semantic
55 Syntactic
30 45
Overall
50

Gigaword5 + 20 40
Wiki2010 Wiki2014 Gigaword5 Common Crawl 0 100 200 300 400 500 600 2
Wiki2014 Vector Dimension
1B tokens 1.6B tokens 4.3B tokens 6B tokens 42B tokens

(a) Symmetric context

Figure 3: Accuracy on the analogy task for 300-
dimensional vectors trained on different corpora.
38
Figure 2: Accuracy on the analogy task
entries are updated to assimilate new knowledge, trained on the 6 billion token corpus. In
Another intrinsic word vector evaluation
• Word vector distances and their correlation with human judgments
• Example dataset: WordSim353
http://www.cs.technion.ac.il/~gabr/resources/data/wordsim353/

Word 1 Word 2 Human (mean)

tiger cat 7.35
tiger tiger 10
book paper 7.46
computer internet 7.58
plane car 5.77
professor doctor 6.62
stock phone 1.62
stock CD 1.31
stock jaguar 0.92
39
 Table 3: Spearman rank correlation on word simi-
h the
larity tasks. All vectors are 300-dimensional. The
Correlation ⇤ evaluation
CBOW vectors are from the word2vec website
a va-
• andvector
Word differdistances
in thatand
they contain
their phrase
correlation withvectors.
human judgments
with
vec Model Size WS353 MC RG SCWS RW
With SVD 6B 35.3 35.1 42.5 38.3 25.6
and SVD-S 6B 56.5 71.5 71.0 53.6 34.7
n the SVD-L 6B 65.7 72.7 75.1 56.5 37.0
CBOW † 6B 57.2 65.6 68.2 57.0 32.5
Giga-
SG† 6B 62.8 65.2 69.7 58.1 37.2
most
GloVe 6B 65.8 72.7 77.8 53.9 38.1
f 10.
SVD-L 42B 74.0 76.4 74.1 58.3 39.9
w in
GloVe 42B 75.9 83.6 82.9 59.6 47.8
s. CBOW⇤ 100B 68.4 79.6 75.4 59.4 45.5
ated
how• Some ideas from Glove paper have been shown to improve skip-gram (SG)
L model onsum
this larger corpus. The fact that this
model also (e.g. both vectors)
top basic SVD model does not scale well to large cor-
ypi-40 pora lends further evidence to the necessity of the
Extrinsic word vector evaluation
• Extrinsic evaluation of word vectors: All subsequent tasks in this class Seman
Table 4: F1 score on NER task with 50d vectors. 85

Discrete is the baseline without word vectors. We 80

• One exampleuse where good word vectors
publicly-available should
vectors help directly:
for HPCA, HSMN,named75entity

Accuracy [%]
recognition: finding
and CW.a person,
See text organization
for details. or location 70

Model Dev Test ACE MUC7 65

60
Discrete 91.0 85.4 77.4 73.4
55
SVD 90.8 85.7 77.3 73.7 50
SVD-S 91.0 85.5 77.6 74.3
Wiki2010 Wik
SVD-L 90.5 84.8 73.6 71.5 1B tokens 1.6B

HPCA 92.6 88.7 81.7 80.7

Figure 3: Accura
HSMN 90.5 85.7 78.7 74.7
dimensional vecto
CW 92.2 87.4 81.7 80.2
CBOW 93.1 88.2 82.2 81.1 entries are update
GloVe 93.2 88.3 82.9 82.2 whereas Gigawor
outdated and poss
• Next: How toshown for neural
use word vectorsvectors in (Turian
in neural et al., 2010).
net models!
4.4 Model Analysis: Vector Length and 4.6 Model Ana
41
Context Size
6. Word senses and word sense ambiguity
• Most words have lots of meanings!
• Especially common words
• Especially words that have existed for a long time

• Example: pike

• Does one vector capture all these meanings or do we have a

mess?
42
pike
• A sharp point or staff
• A type of elongated fish
• A railroad line or system
• A type of road
• The future (coming down the pike)
• A type of body position (as in diving)
• To kill or pierce with a pike
• To make one’s way (pike along)
• In Australian English, pike means to pull out from doing
something: I reckon he could have climbed that cliff, but he
piked!

43
Improving Word Representations Via Global Context
And Multiple Word Prototypes (Huang et al. 2012)
• Idea: Cluster word windows around words, retrain with each
word assigned to multiple different clusters bank1, bank2, etc

44
Linear Algebraic Structure of Word Senses, with
Applications to Polysemy (Arora, …, Ma, …, TACL 2018)
• Different senses of a word reside in a linear superposition (weighted
sum) in standard word embeddings like word2vec
• 𝑣pike = 𝛼7 𝑣pikeP + 𝛼9 𝑣pikeR + 𝛼S 𝑣pikeT
UP
• Where 𝛼7 = , etc., for frequency f
UP 6UR 6UT
• Surprising result:
• Because of ideas from sparse coding you can actually separate out
the senses (providing they are relatively common)

45
7. Classification review and notation
• Generally we have a training dataset consisting of samples

{xi,yi}Ni=1

• xi are inputs, e.g. words (indices or vectors!), sentences,

documents, etc.
• Dimension d

• yi are labels (one of C classes) we try to predict, for example:

• classes: sentiment, named entities, buy/sell decision
• other words
• later: multi-word sequences

46
Classification intuition
• Training data: {xi,yi}Ni=1

• Simple illustration case:

• Fixed 2D word vectors to classify
• Using softmax/logistic regression
• Linear decision boundary Visualizations with ConvNetJS by Karpathy!
http://cs.stanford.edu/people/karpathy/convnetjs/demo/classify2d.html

• Traditional ML/Stats approach: assume xi are fixed,

train (i.e., set) softmax/logistic regression weights 𝑊 ∈ ℝX×Z
to determine a decision boundary (hyperplane) as in the picture

• Method: For each fixed x, predict:

47
Details of the softmax classifier

We can tease apart the prediction function into three steps:

1. Take the yth row of W and multiply that row with x:

Compute all fc for c = 1, …, C

2. Apply softmax function to get normalized probability:

= softmax(𝑓\ )

3. Choose the y with maximum probability

48
Training with softmax and cross-entropy loss

• For each training example (x,y), our objective is to maximize the

probability of the correct class y

• Or we can minimize the negative log probability of that class:

49
Background: What is “cross entropy” loss/error?
• Concept of “cross entropy” is from information theory
• Let the true probability distribution be p
• Let our computed model probability be q
• The cross entropy is:

• Assuming a ground truth (or true or gold or target) probability

distribution that is 1 at the right class and 0 everywhere else:
p = [0,…,0,1,0,…0] then:
• Because of one-hot p, the only term left is the negative log
probability of the true class
50
Classification over a full dataset
• Cross entropy loss function over
full dataset {xi,yi}Ni=1

• Instead of

We will write f in matrix notation:

51
Traditional ML optimization
• For general machine learning 𝜃 usually
only consists of columns of W:

• So we only update the decision

boundary via Visualizations with ConvNetJS by Karpathy

52
Neural Network Classifiers
• Softmax (≈ logistic regression) alone not very powerful
• Softmax gives only linear decision boundaries
This can be quite limiting
• à Unhelpful when a
problem is complex

• Wouldn’t it be cool to
get these correct?

53
Neural Nets for the Win!
• Neural networks can learn much more complex
functions and nonlinear decision boundaries!
• In the original space

54
Classification difference with word vectors
• Commonly in NLP deep learning:
• We learn both W and word vectors x
• We learn both conventional parameters and representations
• The word vectors re-represent one-hot vectors—move them
around in an intermediate layer vector space—for easy
classification with a (linear) softmax classifier via layer x = Le
Very large number of
parameters!

55
 8. The course
A note on your experience 😀

“Terrible class” “Best class at Stanford”

“Don’t take it” “Changed my life”
“Too much work” “Obvious that instructors care”
“Learned a ton”
“Hard but worth it”

• This is a hard, advanced, graduate level class

• I and all the TAs really care about your success in this class
• Give Feedback. Take responsibility for holes in your knowledge
• Come to office hours/help sessions
56
Office Hours / Help sessions
• Come to office hours/help sessions!
• Come to discuss final project ideas as well as the assignments
• Try to come early, often and off-cycle
• Help sessions: daily, at various times, see calendar
• Attending in person: Just show up! Our friendly course staff
will be on hand to assist you
• SCPD/remote access: Use queuestatus
• Chris’s office hours:
• Mon 4–6pm. Come along next Monday?

57