Language Introduction to N-grams

Modeling
 Probabilistic Language Models

Today’s goal: assign a probability to a sentence

◦ Machine Translation:
◦ P(high winds tonite) > P(large winds tonite)
◦ Spell Correction
Why? ◦ The office is about fifteen minuets from my house
◦ P(about fifteen minutes from) > P(about fifteen minuets from)
◦ Speech Recognition
◦ P(I saw a van) >> P(eyes awe of an)
◦ + Summarization, question-answering, etc., etc.!!
 Probabilistic Language Modeling

Goal: compute the probability of a sentence or sequence of

words:
P(W) = P(w1,w2,w3,w4,w5…wn)
Related task: probability of an upcoming word:
P(w5|w1,w2,w3,w4)
A model that computes either of these:
P(W) or P(wn|w1,w2…wn-1) is called a language model.
Better: the grammar But language model or LM is standard
How to compute P(W)
How to compute this joint probability:

◦ P(its, water, is, so, transparent, that)

Intuition: let’s rely on the Chain Rule of Probability

 Reminder: The Chain Rule

Recall the definition of conditional probabilities

p(B|A) = P(A,B)/P(A) Rewriting: P(A,B) = P(A)P(B|A)

More variables:
P(A,B,C,D) = P(A)P(B|A)P(C|A,B)P(D|A,B,C)
The Chain Rule in General
P(x1,x2,x3,…,xn) = P(x1)P(x2|x1)P(x3|x1,x2)…P(xn|x1,…,xn-1)
The Chain Rule applied to compute joint
probability of words in sentence

P(w1w 2 …w n ) = ∏ P(w i | w1w 2 …w i−1 )

i

P(“its water is so transparent”) =

P(its) × P(water|its) × P(is|its water)
× P(so|its water is) × P(transparent|its water is so)
How to estimate these probabilities
Could we just count and divide?
P(the | its water is so transparent that) =
Count(its water is so transparent that the)
Count(its water is so transparent that)

No! Too many possible sentences!

We’ll never see enough data for estimating these
 Markov Assumption

Simplifying assumption:
Andrei Markov

P(the | its water is so transparent that) ≈ P(the | that)

Or maybe
P(the | its water is so transparent that) ≈ P(the | transparent that)
 Markov Assumption

P(w1w 2 …w n ) ≈ ∏ P(w i | w i−k …w i−1 )

i
In other words, we approximate each
component in the product

P(w i | w1w 2 …w i−1 ) ≈ P(w i | w i−k …w i−1 )

 Simplest case: Unigram model

P(w1w 2 …w n ) ≈ ∏ P(w i )
i
Some automatically generated sentences from a unigram model

fifth, an, of, futures, the, an, incorporated, a,

a, the, inflation, most, dollars, quarter, in, is,
mass
€
thrift, did, eighty, said, hard, 'm, july, bullish

that, or, limited, the

 Bigram model

Condition on the previous word:

P(w i | w1w 2 …w i−1 ) ≈ P(w i | w i−1 )

texaco, rose, one, in, this, issue, is, pursuing, growth, in,
a, boiler, house, said, mr., gurria, mexico, 's, motion,
control, proposal, without, permission, from, five, hundred,
fifty, five, yen

outside, new, car, parking, lot, of, the, agreement, reached

this, would, be, a, record, november

N-gram models

We can extend to trigrams, 4-grams, 5-grams

In general this is an insufficient model of language
◦ because language has long-distance dependencies:
“The computer which I had just put into the machine room
on the fifth floor crashed.”

But we can often get away with N-gram models

Language Introduction to N-grams
Modeling
Language Estimating N-gram
Probabilities
Modeling
 Estimating bigram probabilities
The Maximum Likelihood Estimate

count(w i−1,w i )
P(w i | w i−1 ) =
count(w i−1 )

c(w i−1,w i )
P(w i | w i−1 ) =
c(w i−1 )
€
 An example
<s> I am Sam </s>
c(w i−1,w i )
P(w i | w i−1 ) = <s> Sam I am </s>
c(w i−1 )
<s> I do not like green eggs and ham </s>
 More examples:
Berkeley Restaurant Project sentences

can you tell me about any good cantonese restaurants close by

mid priced thai food is what i’m looking for
tell me about chez panisse
can you give me a listing of the kinds of food that are available
i’m looking for a good place to eat breakfast
when is caffe venezia open during the day
 Raw bigram counts
Out of 9222 sentences
Raw bigram probabilities
Normalize by unigrams:

Result:
 Bigram estimates of sentence probabilities

P(<s> I want english food </s>) =

P(I|<s>)
× P(want|I)
× P(english|want)
× P(food|english)
× P(</s>|food)
= .000031
What kinds of knowledge?
P(english|want) = .0011
P(chinese|want) = .0065
P(to|want) = .66
P(eat | to) = .28
P(food | to) = 0
P(want | spend) = 0
P (i | <s>) = .25
 Practical Issues

We do everything in log space

◦ Avoid underflow
◦ (also adding is faster than multiplying)

log( p1 × p2 × p3 × p4 ) = log p1 + log p2 + log p3 + log p4

Language Modeling Toolkits

SRILM
◦ http://www.speech.sri.com/projects/srilm/
KenLM
◦ https://kheafield.com/code/kenlm/
 Google N-Gram Release, August 2006

…
 Google N-Gram Release
serve as the incoming 92
serve as the incubator 99
serve as the independent 794
serve as the index 223
serve as the indication 72
serve as the indicator 120
serve as the indicators 45
serve as the indispensable 111
serve as the indispensible 40
serve as the individual 234

http://googleresearch.blogspot.com/2006/08/all-our-n-gram-are-belong-to-you.html
Google Book N-grams
http://ngrams.googlelabs.com/
Language Estimating N-gram
Probabilities
Modeling
Language Evaluation and Perplexity
Modeling
 Evaluation: How good is our model?
Does our language model prefer good sentences to bad ones?
◦ Assign higher probability to “real” or “frequently observed” sentences
◦ Than “ungrammatical” or “rarely observed” sentences?

We train parameters of our model on a training set.

We test the model’s performance on data we haven’t seen.
◦ A test set is an unseen dataset that is different from our training set,
totally unused.
◦ An evaluation metric tells us how well our model does on the test set.
Extrinsic evaluation of N-gram models
Best evaluation for comparing models A and B
◦ Put each model in a task
◦ spelling corrector, speech recognizer, MT system
◦ Run the task, get an accuracy for A and for B
◦ How many misspelled words corrected properly
◦ How many words translated correctly
◦ Compare accuracy for A and B
Difficulty of extrinsic (in-vivo) evaluation of
N-gram models
Extrinsic evaluation
◦ Time-consuming; can take days or weeks
So
◦ Sometimes use intrinsic evaluation: perplexity
◦ Bad approximation
◦ unless the test data looks just like the training data
◦ So generally only useful in pilot experiments
◦ But is helpful to think about.
 Intuition of Perplexity
The Shannon Game: mushrooms 0.1
◦ How well can we predict the next word? pepperoni 0.1
I always order pizza with cheese and ____ anchovies 0.01
The 33rd President of the US was ____ ….
I saw a ____ fried rice 0.0001

◦ Unigrams are terrible at this game. (Why?) ….

and 1e-100
A better model of a text
◦ is one which assigns a higher probability to the word that
actually occurs
 Perplexity
The best language model is one that best predicts an unseen test set
• Gives the highest P(sentence) −
1
PP(W ) = P(w1w2 ...wN ) N
Perplexity is the inverse probability of
the test set, normalized by the number
of words: = N
1
P(w1w2 ...wN )

Chain rule:

For bigrams:

Minimizing perplexity is the same as maximizing probability

The Shannon Game intuition for perplexity
From Josh Goodman
Perplexity is weighted equivalent branching factor
How hard is the task of recognizing digits ‘0,1,2,3,4,5,6,7,8,9’
◦ Perplexity 10

How hard is recognizing (30,000) names at Microsoft.

◦ Perplexity = 30,000

Let's imagine a call-routing phone system gets 120K calls and has to recognize
◦ "Operator" (let's say this occurs 1 in 4 calls)
◦ "Sales" (1in 4)
◦ "Technical Support" (1 in 4)
◦ 30,000 different names (each name occurring 1 time in the 120K calls)
◦ What is the perplexity? Next slide
The Shannon Game intuition for perplexity
Josh Goodman: imagine a call-routing phone system gets 120K calls and has
to recognize
◦ "Operator" (let's say this occurs 1 in 4 calls)
◦ "Sales" (1in 4)
◦ "Technical Support" (1 in 4)
◦ 30,000 different names (each name occurring 1 time in the 120K calls)
We get the perplexity of this sequence of length 120Kby first multiplying 120K
probabilities (90K of which are 1/4 and 30K of which are 1/120K), nd then
taking the inverse 120,000th root:
Perp = (¼ * ¼ * ¼* ¼ * ¼ * …. * 1/120K * 1/120K * ….)^(-1/120K)
But this can be arithmetically simplified to just N = 4: the operator (1/4), the
sales (1/4), the tech support (1/4), and the 30,000 names (1/120,000):
Perplexity= ((¼ * ¼ * ¼ * 1/120K)^(-1/4) = 52.6
Perplexity as branching factor
Let’s suppose a sentence consisting of random
digits
What is the perplexity of this sentence according to
a model that assign P=1/10 to each digit?
Lower perplexity = better model

Training 38 million words, test 1.5 million words, WSJ

N-gram Unigram Bigram Trigram

Order
Perplexity 962 170 109
Language Evaluation and Perplexity
Modeling
Language Generalization and zeros
Modeling
 The Shannon Visualization Method

Choose a random bigram

<s> I
(<s>, w) according to its probability I want
Now choose a random bigram want to
(w, x) according to its probability to eat
eat Chinese
And so on until we choose </s> Chinese food
Then string the words together food </s>
I want to eat Chinese food
 bigrams by first generating a random bigram that starts with <s> (according to its
bigram probability), then choosing a random bigram to follow (again, according to

Approximating Shakespeare
its bigram probability), and so on.
To give an intuition for the increasing power of higher-order N-grams, Fig. 4.3
shows random sentences generated from unigram, bigram, trigram, and 4-gram
models trained on Shakespeare’s works.

–To him swallowed confess hear both. Which. Of save on trail for are ay device and
1gram rote life have
–Hill he late speaks; or! a more to leg less first you enter
–Why dost stand forth thy canopy, forsooth; he is this palpable hit the King Henry. Live
2gram king. Follow.
–What means, sir. I confess she? then all sorts, he is trim, captain.
–Fly, and will rid me these news of price. Therefore the sadness of parting, as they say,
3gram ’tis done.
–This shall forbid it should be branded, if renown made it empty.
–King Henry. What! I will go seek the traitor Gloucester. Exeunt some of the watch. A
4gram great banquet serv’d in;
–It cannot be but so.
Figure 4.3 Eight sentences randomly generated from four N-grams computed from Shakespeare’s works. All
characters were mapped to lower-case and punctuation marks were treated as words. Output is hand-corrected
Shakespeare as corpus

N=884,647 tokens, V=29,066

Shakespeare produced 300,000 bigram types
out of V = 844 million possible bigrams.
2

◦ So 99.96% of the possible bigrams were never seen

(have zero entries in the table)
Quadrigrams worse: What's coming out looks
like Shakespeare because it is Shakespeare
 The Wall Street Journal is not Shakespeare
(no offense)
4.3 • G ENERALIZATION AND Z EROS 11

1
gram
Months the my and issue of year foreign new exchange’s september
were recession exchange new endorsed a acquire to six executives
Last December through the way to preserve the Hudson corporation N.
2
gram
B. E. C. Taylor would seem to complete the major central planners one
point five percent of U. S. E. has already old M. X. corporation of living
on information such as more frequently fishing to keep her
They also point to ninety nine point six billion dollars from two hundred
3
gram
four oh six three percent of the rates of interest stores as Mexico and
Brazil on market conditions
Figure 4.4 Three sentences randomly generated from three N-gram models computed from
40 million words of the Wall Street Journal, lower-casing all characters and treating punctua-
Can you guess the training set author of the LM
that generated these random 3-gram sentences?
They also point to ninety nine point six billion dollars
from two hundred four oh six three percent of the rates
of interest stores as Mexico and gram Brazil on market
conditions
This shall forbid it should be branded, if renown made
it empty.
“You are uniformly charming!” cried he, with a smile of
associating and now and then I bowed and they
perceived a chaise and four to wish for.
44
The perils of overfitting
N-grams only work well for word prediction if the
test corpus looks like the training corpus
◦ In real life, it often doesn’t
◦ We need to train robust models that generalize!
◦ One kind of generalization: Zeros!
◦ Things that don’t ever occur in the training set
◦ But occur in the test set
 Zeros
Training set: • Test set
… denied the allegations … denied the offer
… denied the reports … denied the loan
… denied the claims
… denied the request

P(“offer” | denied the) = 0

Zero probability bigrams
Bigrams with zero probability
◦ mean that we will assign 0 probability to the test set!
And hence we cannot compute perplexity (can’t
divide by 0)!
Language Generalization and zeros
Modeling
Language Smoothing: Add-one
(Laplace) smoothing
Modeling
 The intuition of smoothing (from Dan Klein)
When we have sparse statistics:
P(w | denied the)

allegations
3 allegations

outcome
2 reports

reports

attack
1 claims
…

request
claims

man
1 request
7 total
Steal probability mass to generalize better
P(w | denied the)
2.5 allegations

allegations
1.5 reports

allegations

outcome
0.5 claims

reports

attack
0.5 request
…

man
claims

request
2 other
7 total
 Add-one estimation
Also called Laplace smoothing
Pretend we saw each word one more time than we did
Just add one to all the counts!

MLE estimate: c(wi−1, wi )

PMLE (wi | wi−1 ) =
c(wi−1 )
Add-1 estimate:
c(wi−1, wi ) +1
PAdd−1 (wi | wi−1 ) =
c(wi−1 ) +V
 Maximum Likelihood Estimates
The maximum likelihood estimate
◦ of some parameter of a model M from a training set T
◦ maximizes the likelihood of the training set T given the model M
Suppose the word “bagel” occurs 400 times in a corpus of a million words
What is the probability that a random word from some other text will be “bagel”?
MLE estimate is 400/1,000,000 = .0004
This may be a bad estimate for some other corpus
◦ But it is the estimate that makes it most likely that “bagel” will occur 400 times in a
million word corpus.
Berkeley Restaurant Corpus: Laplace
smoothed bigram counts
Laplace-smoothed bigrams
Reconstituted counts
Compare with raw bigram counts
Add-1 estimation is a blunt instrument
So add-1 isn’t used for N-grams:
◦ We’ll see better methods
But add-1 is used to smooth other NLP models
◦ For text classification
◦ In domains where the number of zeros isn’t so huge.
Language Smoothing: Add-one
(Laplace) smoothing
Modeling
Language Interpolation, Backoff, and
Web-Scale LMs
Modeling
 Backoff and Interpolation

Sometimes it helps to use less context

◦ Condition on less context for contexts you haven’t learned much about
Backoff:
◦ use trigram if you have good evidence,
◦ otherwise bigram, otherwise unigram
Interpolation:
◦ mix unigram, bigram, trigram

Interpolation works better

 from all the N-gram estimators, weighing and combining the trigram, bigram, and
P̂(wn |wn 2 wn 1 ) = l1 P(wn |w
Linear Interpolation
unigram counts.
In simple linear interpolation, we combine different order N-grams by linearly
1 ) 2 P(wn |
interpolating all the models. Thus, we estimate the trigram probability P(wn |wn 2 wn +l
by mixing together the unigram, bigram, and trigram probabilities, each weighted
by a l : +l3 P(wn )
Simple interpolation
such that the l s sum to 1:
P̂(wn |wn 2 wn 1 ) = l1 P(wn |wn 2 wn 1) X
+l2 P(wn |wn 1 ) li = 1
+l3 P(wn ) i (4.24)
Lambdas
such that the conditional onX context:
l s sum to 1:In a slightly more sophisticated version of linear i
li = 1 (4.25)
computed ini a more sophisticated way, by condition
ifsophisticated
In a slightly more we have version
particularly accurateeach
of linear interpolation, counts foris a particul
l weight
computed in a morecounts of way,
sophisticated the bytrigrams
conditioningbased on thisThisbigram
on the context. way, will be
if we have particularly accurate counts for a particular bigram, we assume that the
make
counts of the trigrams based the l sbigram
on this for those trigrams
will be more higher
trustworthy, so we and
can thus giv
 How to set the lambdas?

Use a held-out corpus

Held-Out Test
Training Data Data Data

Choose λs to maximize the probability of held-out data:

◦ Fix the N-gram probabilities (on the training data)
◦ Then search for λs that give largest probability to held-out set:
log P(w1...wn | M (λ1...λk )) = ∑ log PM ( λ1... λk ) (wi | wi−1 )
i
 Unknown words: Open versus closed
vocabulary tasks
If we know all the words in advanced
◦ Vocabulary V is fixed
◦ Closed vocabulary task
Often we don’t know this
◦ Out Of Vocabulary = OOV words
◦ Open vocabulary task
Instead: create an unknown word token <UNK>
◦ Training of <UNK> probabilities
◦ Create a fixed lexicon L of size V
◦ At text normalization phase, any training word not in L changed to <UNK>
◦ Now we train its probabilities like a normal word
◦ At decoding time
◦ If text input: Use UNK probabilities for any word not in training
 Huge web-scale n-grams
How to deal with, e.g., Google N-gram corpus
Pruning
◦ Only store N-grams with count > threshold.
◦ Remove singletons of higher-order n-grams
◦ Entropy-based pruning
Efficiency
◦ Efficient data structures like tries
◦ Bloom filters: approximate language models
◦ Store words as indexes, not strings
◦ Use Huffman coding to fit large numbers of words into two bytes
◦ Quantize probabilities (4-8 bits instead of 8-byte float)
Smoothing for Web-scale N-grams
“Stupid backoff” (Brants et al. 2007)
No discounting, just use relative frequencies
" i
$$ count(wi−k+1 ) i
i−1 i−1
if count(wi−k+1 ) > 0
S(wi | wi−k+1 ) = # count(wi−k+1 )
$ i−1
$% 0.4S(wi | wi−k+2 ) otherwise

count(wi )
S(wi ) =
N
65
N-gram Smoothing Summary
Add-1 smoothing:
◦ OK for text categorization, not for language modeling
The most commonly used method:
◦ Extended Interpolated Kneser-Ney
For very large N-grams like the Web:
◦ Stupid backoff

66
 Advanced Language Modeling

Discriminative models:
◦ choose n-gram weights to improve a task, not to fit the training set
Parsing-based models
Caching Models
◦ Recently used words are more likely to appear
c(w ∈ history)
PCACHE (w | history) = λ P(wi | wi−2 wi−1 ) + (1− λ )
| history |
◦ These turned out to perform very poorly for speech recognition
(why?)
Language Interpolation, Backoff, and
Web-Scale LMs
Modeling
 Advanced:
Language Kneser-Ney Smoothing
Modeling
 Absolute discounting: just subtract a little
from each count
Suppose we wanted to subtract a little from a Bigram count Bigram count in
count of 4 to save probability mass for the zeros in training heldout set
0 .0000270
How much to subtract ? 1 0.448
2 1.25
Church and Gale (1991)’s clever idea 3 2.24
4 3.23
Divide up 22 million words of AP Newswire
5 4.21
◦ Training and held-out set 6 5.23
◦ for each bigram in the training set 7 6.21
◦ see the actual count in the held-out set! 8 7.21
It sure looks like c* = (c - .75) 9 8.26
 Absolute Discounting Interpolation

Save ourselves some time and just subtract 0.75 (or some d)!
discounted bigram Interpolation weight

c(wi−1, wi ) − d
PAbsoluteDiscounting (wi | wi−1 ) = + λ (wi−1 )P(w)
c(wi−1 )
unigram
◦ (Maybe keeping a couple extra values of d for counts 1 and 2)
But should we really just use the regular unigram P(w)?

71
 Kneser-Ney Smoothing I
Better estimate for probabilities of lower-order unigrams!
Kong
glasses
◦ Shannon game: I can’t see without my reading___________?
◦ “Kong” turns out to be more common than “glasses”
◦ … but “Kong” always follows “Hong”
The unigram is useful exactly when we haven’t seen this bigram!
Instead of P(w): “How likely is w”
Pcontinuation(w): “How likely is w to appear as a novel continuation?
◦ For each word, count the number of bigram types it completes
◦ Every bigram type was a novel continuation the first time it was seen

PCONTINUATION (w) ∝ {wi−1 : c(wi−1, w) > 0}

 Kneser-Ney Smoothing II
How many times does w appear as a novel continuation:
PCONTINUATION (w) ∝ {wi−1 : c(wi−1, w) > 0}

Normalized by the total number of word bigram types

{(w j−1, w j ) : c(w j−1, w j ) > 0}

{wi−1 : c(wi−1, w) > 0}

PCONTINUATION (w) =
{(w j−1, w j ) : c(w j−1, w j ) > 0}
 Kneser-Ney Smoothing III
Alternative metaphor: The number of # of word types seen to precede w

| {wi−1 : c(wi−1, w) > 0} |

normalized by the # of words preceding all words:

{wi−1 : c(wi−1, w) > 0}

PCONTINUATION (w) =
∑ {w' i−1 : c(w'i−1, w') > 0}
w'
A frequent word (Kong) occurring in only one context (Hong) will have a low
continuation probability
 Kneser-Ney Smoothing IV

max(c(wi−1, wi ) − d, 0)
PKN (wi | wi−1 ) = + λ (wi−1 )PCONTINUATION (wi )
c(wi−1 )
λ is a normalizing constant; the probability mass we’ve discounted

d
λ (wi−1 ) = {w : c(wi−1, w) > 0}
c(wi−1 )
The number of word types that can follow wi-1
the normalized discount = # of word types we discounted
= # of times we applied normalized discount
75
 Kneser-Ney Smoothing: Recursive formulation

i
i−1 max(cKN (wi−n+1 ) − d, 0) i−1 i−1
PKN (wi | wi−n+1 ) = i−1
+ λ (wi−n+1 )PKN (wi | wi−n+2 )
cKN (wi−n+1 )

!# count(•) for the highest order

cKN (•) = "
#$ continuationcount(•) for lower order

Continuation count = Number of unique single word contexts for Ÿ

76
 Advanced:
Language Kneser-Ney Smoothing
Modeling