Natural Language Processing

with Deep Learning

CS224N/Ling284

Christopher Manning
Lecture 1: Introduction and Word Vectors
Lecture Plan
Lecture 1: Introduction and Word Vectors
1. The course (10 mins)
2. Human language and word meaning (15 mins)
3. Word2vec introduction (15 mins)
4. Word2vec objective function gradients (25 mins)
5. Optimization basics (5 mins)
6. Looking at word vectors (10 mins or less)

Key learning today: The (really surprising!) result that word meaning can be represented
rather well by a (high-dimensional) vector of real numbers

2
Course logistics in brief
• Instructor: Christopher Manning
• Head TA: Anna Goldie
• Coordinator: Amelie Byun
• TAs: Many wonderful people! See website
• Time: Tu/Th 3:15–4:45 Pacific time, Zoom U. (à video)

• We’ve put a lot of other important information on the class webpage. Please read it!
• http://cs224n.stanford.edu/
a.k.a., http://www.stanford.edu/class/cs224n/
• TAs, syllabus, help sessions/office hours, Ed (for all course questions/discussion)
• Office hours start Thursday evening!
• Python/numpy and then PyTorch tutorials: First two Fridays 1:30–2:30 Pacific time on Zoom U.
• Slide PDFs uploaded before each lecture
3
4
What do we hope to teach? (A.k.a. “learning goals”)
1. The foundations of the effective modern methods for deep learning applied to NLP
• Basics first, then key methods used in NLP: Word vectors, feed-forward networks,
recurrent networks, attention, encoder-decoder models, transformers, etc.

2. A big picture understanding of human languages and the difficulties in understanding

and producing them via computers

3. An understanding of and ability to build systems (in PyTorch) for some of the major
problems in NLP:
• Word meaning, dependency parsing, machine translation, question answering

5
Course work and grading policy
• 5 x 1-week Assignments: 6% + 4 x 12%: 54%
• HW1 is released today! Due next Tuesday! At 3:15 p.m.
• Submitted to Gradescope in Canvas (i.e., using @stanford.edu email for your Gradescope account)
• Final Default or Custom Course Project (1–3 people): 43%
• Project proposal: 5%, milestone: 5%, poster or web summary: 3%, report: 30%
• Participation: 3%
• Guest lecture reactions, Ed, course evals, karma – see website!
• Late day policy
• 6 free late days; afterwards, 1% off course grade per day late
• Assignments not accepted more than 3 days late per assignment unless given permission in advance
• Collaboration policy: Please read the website and the Honor Code!
Understand allowed collaboration and how to document it: Don’t take code off the
web; acknowledge working with other students; write your own assignment solutions
6
High-Level Plan for Assignments (to be completed individually!)
• Ass1 is hopefully an easy on ramp – a Jupyter/IPython Notebook
• Ass2 is pure Python (numpy) but expects you to do (multivariate) calculus so you really
understand the basics
• Ass3 introduces PyTorch, building a feed-forward network for dependency parsing
• Ass4 and Ass5 use PyTorch on a GPU (Microsoft Azure)
• Libraries like PyTorch and Tensorflow are now the standard tools of DL
• For Final Project, more details presented later, but you either:
• Do the default project, which is a question answering system
• Open-ended but an easier start; a good choice for many
• Propose a custom final project, which we approve
• You will receive feedback from a mentor (TA/prof/postdoc/PhD)
• Can work in teams of 1–3; can use any language/packages
7
Lecture Plan
1. The course (10 mins)
2. Human language and word meaning (15 mins)
3. Word2vec introduction (15 mins)
4. Word2vec objective function gradients (25 mins)
5. Optimization basics (5 mins)
6. Looking at word vectors (10 mins or less)

8
https://xkcd.com/1576/ Randall Munroe CC BY NC 2.5
Trained on text data, neural machine translation is quite good!

https://kiswahili.tuko.co.ke/
GPT-3: A first step on the path to foundation models
The SEC said, “Musk, your tweets are a S: I broke the window.
blight. Q: What did I break?
S: I gracefully saved the day.
They really could cost you your job,
Q: What did I gracefully save?
if you don't stop all this tweeting at night.” S: I gave John flowers.
Then Musk cried, “Why? Q: Who did I give flowers to?
The tweets I wrote are not mean, S: I gave her a rose and a guitar.
I don't use all-caps Q: Who did I give a rose and a guitar to?
and I'm sure that my tweets are clean.” How many users have signed up since the start of 2020?
“But your tweets can move markets SELECT count(id) FROM users
WHERE created_at > ‘2020-01-01’
and that's why we're sore.
What is the average number of influencers each user is
You may be a genius and a billionaire, subscribed to?
but it doesn't give you the right to SELECT avg(count) FROM ( SELECT user_id, count(*)
be a bore!” FROM subscribers GROUP BY user_id )
AS avg_subscriptions_per_user
How do we represent the meaning of a word?

Definition: meaning (Webster dictionary)

• the idea that is represented by a word, phrase, etc.
• the idea that a person wants to express by using words, signs, etc.
• the idea that is expressed in a work of writing, art, etc.

Commonest linguistic way of thinking of meaning:

signifier (symbol) ⟺ signified (idea or thing)
= denotational semantics

tree ⟺ {🌳, 🌲, 🌴, …}
13
How do we have usable meaning in a computer?
Previously commonest NLP solution: Use, e.g., WordNet, a thesaurus containing lists of
synonym sets and hypernyms (“is a” relationships)
e.g., synonym sets containing “good”: e.g., hypernyms of “panda”:
from nltk.corpus import wordnet as wn from nltk.corpus import wordnet as wn
poses = { 'n':'noun', 'v':'verb', 's':'adj (s)', 'a':'adj', 'r':'adv'}
for synset in wn.synsets("good"): panda = wn.synset("panda.n.01")
print("{}: {}".format(poses[synset.pos()], hyper = lambda s: s.hypernyms()
", ".join([l.name() for l in synset.lemmas()]))) list(panda.closure(hyper))

noun: good
[Synset('procyonid.n.01'),
noun: good, goodness
Synset('carnivore.n.01'),
noun: good, goodness
noun: commodity, trade_good, good Synset('placental.n.01'),
Synset('mammal.n.01'),
adj: good
Synset('vertebrate.n.01'),
adj (sat): full, good
Synset('chordate.n.01'),
adj: good
Synset('animal.n.01'),
adj (sat): estimable, good, honorable, respectable
adj (sat): beneficial, good Synset('organism.n.01'),
Synset('living_thing.n.01'),
adj (sat): good
Synset('whole.n.02'),
adj (sat): good, just, upright
Synset('object.n.01'),
…
Synset('physical_entity.n.01'),
adverb: well, good
adverb: thoroughly, soundly, good Synset('entity.n.01')]

14
Problems with resources like WordNet
• A useful resource but missing nuance:
• e.g., “proficient” is listed as a synonym for “good”
This is only correct in some contexts
• Also, WordNet list offensive synonyms in some synonym sets without any
coverage of the connotations or appropriateness of words
• Missing new meanings of words:
• e.g., wicked, badass, nifty, wizard, genius, ninja, bombest
• Impossible to keep up-to-date!
• Subjective
• Requires human labor to create and adapt
• Can’t be used to accurately compute word similarity (see following slides)

15
Representing words as discrete symbols
In traditional NLP, we regard words as discrete symbols:
hotel, conference, motel – a localist representation

Means one 1, the rest 0s

Such symbols for words can be represented by one-hot vectors:

motel = [0 0 0 0 0 0 0 0 0 0 1 0 0 0 0]
hotel = [0 0 0 0 0 0 0 1 0 0 0 0 0 0 0]

Vector dimension = number of words in vocabulary (e.g., 500,000+)

16
 Sec. 9.2.2

Problem with words as discrete symbols

Example: in web search, if a user searches for “Seattle motel”, we would like to match
documents containing “Seattle hotel”

But:
motel = [0 0 0 0 0 0 0 0 0 0 1 0 0 0 0]
hotel = [0 0 0 0 0 0 0 1 0 0 0 0 0 0 0]
These two vectors are orthogonal
There is no natural notion of similarity for one-hot vectors!

Solution:
• Could try to rely on WordNet’s list of synonyms to get similarity?
• But it is well-known to fail badly: incompleteness, etc.
• Instead: learn to encode similarity in the vectors themselves
17
Representing words by their context
• Distributional semantics: A word’s meaning is given
by the words that frequently appear close-by
• “You shall know a word by the company it keeps” (J. R. Firth 1957: 11)
• One of the most successful ideas of modern statistical NLP!
• When a word w appears in a text, its context is the set of words that appear nearby
(within a fixed-size window).
• We use the many contexts of w to build up a representation of w

!"#$%&'(%')*+%,)*-&#,.%(/*)0&'1'"*1')# ,2'31'" 4&1/%/*2/*52--%'%+*1'*6778!

!/291'"*)52)*:0&#-%*'%%+/*0'1;1%+ ,2'31'" &%"0.2)1#'*)#*&%-.24%*)5%*5#+"%-#+"%!
!<'+12*52/*=0/)*"1$%'*1)/ ,2'31'" /9/)%(*2*/5#)*1'*)5%*2&(!

18 These context words will represent banking

Word vectors
We will build a dense vector for each word, chosen so that it is similar to vectors of words
that appear in similar contexts, measuring similarity as the vector dot (scalar) product

0.286 0.413
0.792 0.582
−0.177 −0.007
banking = −0.107 monetary = 0.247
0.109 0.216
−0.542 −0.718
0.349 0.147
0.271 0.051

Note: word vectors are also called (word) embeddings or (neural) word representations
They are a distributed representation
19
Word meaning as a neural word vector – visualization

0.286
0.792
−0.177
−0.107
expect = 0.109
−0.542
0.349
0.271
0.487

20
3. Word2vec: Overview
Word2vec (Mikolov et al. 2013) is a framework for learning word vectors

Idea:
• We have a large corpus (“body”) of text: a long list of words
• Every word in a fixed vocabulary is represented by a vector
• Go through each position t in the text, which has a center word c and context
(“outside”) words o
• Use the similarity of the word vectors for c and o to calculate the probability of o given
c (or vice versa)
• Keep adjusting the word vectors to maximize this probability

21
Word2Vec Overview
Example windows and process for computing 𝑃 𝑤!"# | 𝑤!

𝑃 𝑤!%$ | 𝑤! 𝑃 𝑤!"$ | 𝑤!
𝑃 𝑤!%# | 𝑤! 𝑃 𝑤!"# | 𝑤!

… problems turning into banking crises as …

outside context words center word outside context words

in window of size 2 at position t in window of size 2

22
Word2Vec Overview
Example windows and process for computing 𝑃 𝑤!"# | 𝑤!

𝑃 𝑤!%$ | 𝑤! 𝑃 𝑤!"$ | 𝑤!
𝑃 𝑤!%# | 𝑤! 𝑃 𝑤!"# | 𝑤!

… problems turning into banking crises as …

outside context words center word outside context words

in window of size 2 at position t in window of size 2

23
Word2vec: objective function
For each position 𝑡 = 1, … , 𝑇, predict context words within a window of fixed size m,
given center word 𝑤! . Data likelihood:
&
Likelihood = 𝐿 𝜃 = , , 𝑃 𝑤!"# | 𝑤! ; 𝜃
𝜃 is all variables !$% '()#)(
to be optimized #*+

sometimes called a cost or loss function

The objective function 𝐽 𝜃 is the (average) negative log likelihood:

&
1 1
𝐽 𝜃 = − log 𝐿(𝜃) = − 5 5 log 𝑃 𝑤!"# | 𝑤! ; 𝜃
𝑇 𝑇
!$% '()#)(
#*+
Minimizing objective function ⟺ Maximizing predictive accuracy
24
Word2vec: objective function
• We want to minimize the objective function:
'
1
𝐽 𝜃 =− * * log 𝑃 𝑤!"* | 𝑤! ; 𝜃
𝑇
!&# %()*)(
*+,

• Question: How to calculate 𝑃 𝑤!"# | 𝑤! ; 𝜃 ?

• Answer: We will use two vectors per word w:
• 𝑣- when w is a center word
• 𝑢- when w is a context word
• Then for a center word c and a context word o:

exp(𝑢-& 𝑣. )
𝑃 𝑜𝑐 = &𝑣 )
∑/∈1 exp(𝑢/ .
25
Word2Vec with Vectors
• Example windows and process for computing 𝑃 𝑤!"* | 𝑤!
• 𝑃 𝑢./0123(4 | 𝑣56!0 short for P 𝑝𝑟𝑜𝑏𝑙𝑒𝑚𝑠 | 𝑖𝑛𝑡𝑜 ; 𝑢./0123(4 , 𝑣56!0 , 𝜃

All words vectors 𝜃

appear in denominator

𝑃 𝑢-*(!./0+ | 𝑣%#'( 𝑃 𝑢)*%+%+ |𝑣%#'(

𝑃 𝑢',#%#& | 𝑣%#'( 𝑃 𝑢!"#$%#& |𝑣%#'(

… problems turning into banking crises as …

outside context words center word outside context words

in window of size 2 at position t in window of size 2

26
Word2vec: prediction function
② Exponentiation makes anything positive
① Dot product compares similarity of o and c.
𝑢1 𝑣 = 𝑢. 𝑣 = ∑#%23 𝑢% 𝑣%
exp(𝑢-& 𝑣. ) Larger dot product = larger probability
𝑃 𝑜𝑐 = &𝑣 )
∑/∈1 exp(𝑢/ .
③ Normalize over entire vocabulary
to give probability distribution

• This is an example of the softmax function ℝ6 → (0,1)6 Open

region
exp(𝑥7 )
softmax 𝑥7 = 6 = 𝑝7
∑#$% exp(𝑥# )
• The softmax function maps arbitrary values 𝑥7 to a probability distribution 𝑝7
• “max” because amplifies probability of largest 𝑥5
But sort of a weird name
• “soft” because still assigns some probability to smaller 𝑥5 because it returns a distribution!
• Frequently used in Deep Learning

27
To train the model: Optimize value of parameters to minimize loss
To train a model, we gradually adjust parameters to minimize a loss

• Recall: 𝜃 represents all the

model parameters, in one
long vector
• In our case, with
d-dimensional vectors and
V-many words, we have à
• Remember: every word has
two vectors

• We optimize these parameters by walking down the gradient (see right figure)
• We compute all vector gradients!
28
 4.

29
30
31
32
5. 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 direction
of negative gradient. Repeat.

Note: Our
objectives
may not
be convex
like this L

But life turns

out to be
okay J

33
Gradient Descent
• Update equation (in matrix notation):

𝛼 = step size or learning rate

• Update equation (for single parameter):

• Algorithm:

34
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:

35
Lecture Plan
1. The course (10 mins)
2. Human language and word meaning (15 mins)
3. Word2vec introduction (15 mins)
4. Word2vec objective function gradients (25 mins)
5. Optimization basics (5 mins)
6. Looking at word vectors (10 mins or less)
• See Jupyter Notebook

36
37
4. Word2vec derivations of gradient
• Zoom U. Whiteboard – see video or revised slides
• The basic Lego piece: The chain rule
• Useful basic fact:

• If in doubt: write it out with indices

38
Chain Rule
• Chain rule! If y = f(u) and u = g(x), i.e., y = f(g(x)), then:

• Simple example:

𝑑𝑦
= 20(𝑥 ; + 7);. 3𝑥 <
𝑑𝑥
39
Interactive Whiteboard Session!

Let’s derive gradient for center word together

For one example window and one example outside word:

exp(𝑢@ A 𝑣B )
log 𝑝 𝑜 𝑐 = log F
∑CDE exp(𝑢C A 𝑣B )

You then also need the gradient for context words (it’s similar;
left for homework). That’s all of the parameters 𝜃 here.
40
Calculating all gradients!
• We went through the gradient for each center vector v in a window
• We also need gradients for outside vectors u
• Derive at home!
• Generally, in each window we will compute updates for all parameters that are being
used in that window. For example:

𝑃 𝑢',*#%#& |𝑣!"#$%#& 𝑃 𝑢"+ |𝑣!"#$%#&

𝑃 𝑢#(+, | 𝑣&'()#(* 𝑃 𝑢!"#$%$ |𝑣&'()#(*

… problems turning into banking crises as …

outside context words center word outside context words

in window of size 2 at position t in window of size 2
41
Word2vec: More details
Why two vectors? à Easier optimization. Average both at the end.

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
This lecture so far: Skip-gram model

Additional efficiency in training:

1. Negative sampling
So far: Focus on naïve softmax (simpler but more expensive training method)

42