Language

Modeling
Introduction to N-grams
Dan Jurafsky

Probabilistic Language Models

• Today’s goal: assign a probability to a sentence
• Machine Translation:
• P(high winds tonight) > P(large winds tonight)
• 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.!!
Dan Jurafsky

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
Dan Jurafsky

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

Dan Jurafsky

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)
Dan Jurafsky
The Chain Rule applied to compute
joint probability of words in sentence

P(w1w 2 … w n ) =Õ 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)
Dan Jurafsky

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
Dan Jurafsky

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)
Dan Jurafsky

Markov Assumption
• Probability of a word depends only on the previous word is
called a Markov assumption.
• Markov model is the class of probabilistic models that assume
we can predict the probability of some future unit without
looking too far into the past.
• N-Gram is the augmentation of this idea and hence
generalization of unigram, bigram, … n gram becomes.
Dan Jurafsky

Markov Assumption
• For complete word sequence:
 Dan Jurafsky

Simplest case: Unigram model

P(w1w 2 … w n ) » Õ 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

Dan Jurafsky

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

Dan Jurafsky

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(s) which I had just put into the machine room
on the fifth floor is (are) crashing.”

• But we can often get away with N-gram models

 Language
Modeling
Estimating N-gram
Probabilities
Dan Jurafsky

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)
 Dan Jurafsky

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>
Dan Jurafsky

More examples of User Queries:

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
Dan Jurafsky

Raw bigram counts

• Bigram counts for eight of the words (out of V = 1446) in the Berkeley
Restaurant Project corpus of 9332 sentences. Zero counts are in blue.
Dan Jurafsky

Raw bigram probabilities

• Normalize by unigrams:

• Result:
Dan Jurafsky

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)
= .25×.33×.0011×0.5×0.68
=.000031
Dan Jurafsky

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
Dan Jurafsky

Practical Issues
• We do everything in log space
• Avoid underflow (e.g. 0.002 x 0.002 =0.00004)
• (also adding is faster than multiplying)

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

Dan Jurafsky

Language Modeling Toolkits

• SRILM
• http://www.speech.sri.com/projects/srilm/
• SRILM is a toolkit for building and applying statistical LMs,
primarily for use in speech recognition, statistical tagging and
segmentation, and machine translation.
• KenLM
• https://kheafield.com/code/kenlm/
• KenLM estimates, filters, and queries language models.
Dan Jurafsky

Google N-Gram Release, August 2006

…
Dan Jurafsky

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
Dan Jurafsky

Google Book N-grams

• https://books.google.com/ngrams
Language
Modeling
Evaluation and
Perplexity
Dan Jurafsky

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.
Dan Jurafsky

Training on the test set

• We can’t allow test sentences into the training set
• We will assign it an artificially high probability when we set it in
the test set
• “Training on the test set”
• Bad science!
• And violates the honor code

29
Dan Jurafsky

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
Dan Jurafsky

Difficulty of extrinsic 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 (small scale
preliminary study)
• But is helpful to think about.
Dan Jurafsky

Intuition of Perplexity
mushrooms 0.1
• The Shannon Game:
• How well can we predict the next word? pepperoni 0.1
anchovies 0.01
I always order pizza with cheese and ____ ….
The 33rd President of the US was ____ fried rice 0.0001
I saw a ____ ….
and 1e-100

• Unigrams are terrible at this game. (Why?)

• A better model of a text
• is one which assigns a higher probability to the word that actually occurs
Dan Jurafsky

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 1
of words: = N
P(w1w2 ...wN )

Chain rule:

For bigrams:

Minimizing perplexity is the same as maximizing probability

Dan Jurafsky

Perplexity as branching factor

• Let’s suppose a sentence consisting of N random digits
• What is the perplexity of this sentence according to a model that
assign P=1/10 to each digit?
Dan Jurafsky

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
Modeling
Generalization and zeros
Dan Jurafsky

N-Gram Dependence Over Training Corpus

• Probabilities often encode specific facts about a given training
corpus.
• N-grams do a better and better job of modeling the training
corpus as we increase the value of N.
Dan Jurafsky

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
Dan Jurafsky

The Shannon Visualization Method

• Shannon’s game gives us an intuition: The
generated texts from the higher order <s> I
models sure look better. That is, they look I want
more like the text the model was want to
obtained from. to eat
• Choose a random bigram eat Chinese
(<s>, w) with its probability. Chinese food
• Now choose a bigram (w, x) according to food </s>
its probability
I want to eat Chinese food
• And so on until we choose </s>
• Then string the words together
Dan Jurafsky

Approximating Shakespeare

No Coherence

Word to word coherence

Look like Shakespeare

Look like Shakespeare

Dan Jurafsky

Shakespeare as corpus
• N=884,647 tokens, V=29,066
• Shakespeare produced 300,000 bigram types out of V2= 844
million possible bigrams.
• Test result was quite bad and 99.96% of the possible bigrams
were never seen (have zero entries in the table)
• Quadrigrams: What's coming out looks like Shakespeare
because it is Shakespeare
Dan Jurafsky

Dependence of Grammar on Training Set

• WSJ to get an idea of the dependence of a grammar on its training set.

• Shakespeare and the Wall Street Journal are both English, so we

might expect some overlap between our n-grams for the two genres.
• there is clearly no overlap in generated sentences, and little overlap
even in small phrases.
• Statistical models are likely to be pretty useless as predictors if the
training sets and the test sets are as different as Shakespeare and WSJ.
Dan Jurafsky

The Wall Street Journal is not Shakespeare

Dan Jurafsky

The perils (serious danger) 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
Dan Jurafsky

N-Gram Precautions
• Training and Test set
• Should be from same genre e.g. translating legal documents, question-
answering system, etc.
• Should be of same dialect e.g. social media posts, spoken transcripts, etc.
• Genre and dialect is still not sufficient and our models are subject to
problem of sparsity (Zero probability).
 Dan Jurafsky

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

Dan Jurafsky

Zero probability bigrams

• Bigrams with zero probability
• mean that we will assign 0 probability to the test set!
• And we cannot compute perplexity (can’t divide by 0)!
 Language
Modeling
Smoothing: Add-one
(Laplace) smoothing
 Dan Jurafsky

The intuition of smoothing (from Dan Klein)

• When we have sparse statistics:
P(w | denied the)

allegations
3 allegations

outcome
2 reports

reports

attack
…

claims
1 claims

request

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

attack
reports
0.5 request
…

man
claims

request
2 other
7 total
Dan Jurafsky

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!
c(wi- 1, wi )
PMLE (wi | wi- 1 ) =
• MLE estimate: c(wi- 1 )
c(wi- 1, wi ) +1
• Add-1 estimate: PAdd- 1 (wi | wi- 1 ) =
c(wi- 1 ) +V
Dan Jurafsky

Maximum Likelihood Estimates

• The maximum likelihood estimate
• 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.
Dan Jurafsky

Berkeley Restaurant Corpus: Laplace

smoothed bigram counts
Dan Jurafsky

Laplace-smoothed bigrams

=(5+1)/(2533+1446)
=0.0015
Dan Jurafsky

Reconstituted counts

={(5+1)x2533}/
(2533+1446)
=3.8
Dan Jurafsky

Compare with raw bigram counts

C(want to) went from 608 to 238, P(to|want) from .66 to .26!
Discount d= c*/c and d for “chinese food” =.10!!! A 10x reduction
Dan Jurafsky

Add-1 estimation is a blunt (less sharp)

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
Modeling
Interpolation, Backoff,
and Web-Scale LMs
Dan Jurafsky

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

Dan Jurafsky

Linear Interpolation
• Simple interpolation (Counting all different N gram probabilities first and then
respective lambda values are calculated which would be normalized to sigma as follows )

• Lambdas conditional on context: (same but some context will be used)

Dan Jurafsky

Linear Interpolation

• All lambdas values are the interpolated weights of bigrams as below:

Dan Jurafsky

Example Linear Interpolation

• we are given the following corpus:
the green book STOP
my blue book STOP
his green house STOP
book STOP
• Assume we compute a language model based on this corpus using linear interpolation with
λi=1/3∀i∈{1,2,3}, now
.
• What is the value of the parameter P(book | the, green): (Note: please include STOP words in
your unigram model). Now
• P’(book | the, green) = 1/3×P(book | the, green)
+1/3×P(book | green)
+1/3×P(book)
61
Dan Jurafsky

Example Linear Interpolation

=0.571

62
Dan Jurafsky

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:
• Use the EM algorithm, an iterative learning algorithm that converges on
locally optimal λs (Jelinek and Mercer, 1980).
Dan Jurafsky

Unknown words: Open versus closed

vocabulary tasks
• If we know all the words in advance
• 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
Dan Jurafsky

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 (sth towards declining)-based pruning for less important N grams (Stolcke, 1998)
• Efficiency
• Efficient data structures like trees, hashes, etc.
• Build approximate language models using bloom filter Data Structure (Talbot and Osborne
2007, Church et al. 2007 ).
• Store words as indexes/hashes, not strings
• Use Huffman coding to fit large numbers of words into two bytes
• Quantize probabilities (4-8 bits instead of 8-byte float)
Dan Jurafsky

Smoothing for Web-scale N-grams

• “Stupid backoff” (Brants et al. 2007)
• No discounting, just use relative frequencies
ì i
ïï count(wi- k+1 ) i
i- 1 i- 1
if count(wi- k+1 ) > 0
S(wi | wi- k+1 ) =í count(wi- k+1 )
ï i- 1
ï
î 0.4S(w i | wi- k+2 ) otherwise

count(wi )
S(wi ) =
66 N
Dan Jurafsky

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

67
Dan Jurafsky

Advanced Language Modeling

• Discriminative models:
• A discriminative language model (DLM) is typically a linear or log-linear model
consisting of a weight vector associated with a feature vector representation of
a sentence.
• choose n-gram weights to improve a task, not to fit the training set
• Parsing-based models: like CFG, PCFG, DCG, HPSG, etc.
• Caching Models
• Recently used words are more likely to appear
• These perform very poorly for speech recognition.
c(w Î history)
PCACHE (w | history) =l P(wi | wi- 2 wi- 1 ) + (1- l)
| history |
• Noisy, new instances may not be in the history, speech recognition
performs better in large amount of historic data.
 Language
Modeling

Advanced:
Kneser-Ney Smoothing
 Absolute discounting: just subtract a
Dan Jurafsky

little from each count

• Suppose we wanted to subtract a little from a Bigram count Bigram count in
in training heldout set
count of 4 to save probability mass for the
0 .0000270
zeros
1 0.448
• How much to subtract ? 2 1.25
3 2.24
• Church and Gale (1991)’s clever idea 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
8 7.21
• see the actual count in the held-out set!
9 8.26
Dan Jurafsky

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 ) = + l (wi- 1 )P(wi )
c(wi- 1 )
unigram

• (Maybe keeping a couple extra values of d for counts 1 and 2).

• How to calculate Lambda? is given next.
71
Dan Jurafsky

Kneser-Ney Smoothing
• Better estimate for probabilities of lower-order unigrams!
• I can’t see without my reading___________? Francisco
• “Francisco” is more common than “glasses” glasses
• … but “Francisco” always follows “San”
• 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?
• Explanation next.
PCONTINUATION (w) µ {wi- 1 : c(wi- 1, w) > 0}
Dan Jurafsky

Kneser-Ney Smoothing
• Number of unique words that come before w
PCONTINUATION (w) µ {wi- 1 : c(wi- 1, w) > 0}
• Total unique bigrams

{(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}
• A frequent word (Francisco) occurring in only one context (San) will have a
low continuation probability.
Dan Jurafsky

Kneser-Ney Smoothing

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

d
l (wi- 1 ) = {w : c(wi- 1, w) > 0}
c(wi- 1 )
The number of unique word that can follow wi-1
the normalized discount = # of times we applied normalized discount
74
Dan Jurafsky

Kneser-Ney Smoothing: Recursive

formulation
i
i- 1 max(c KN (wi- n+1 ) - d, 0) i- 1 i- 1
PKN (wi | wi- n+1 ) = i- 1
+ l (w )P
i- n+1 KN (wi | wi- n+2 )
cKN (wi- n+1 )

75
Dan Jurafsky

Kneser-Ney Smoothing

76
Dan Jurafsky

Kneser-Ney Smoothing

77
Dan Jurafsky

Home Work #03

a) Write a program to compute unsmoothed unigrams and bigrams.
b) Run your N-gram program on two different small corpora of your choice
(you might use email text or newsgroups). Now compare the statistics of the
two corpora. What are the differences in the most common unigrams
between the two? How about interesting differences in bigrams?
c) Add an option to your program to generate random sentences.
d) Add an option to your program to compute the perplexity of a test set.
e) Use Kneser Ney smoothing on the data of corpora built in (b) and then
predict sentences according to trigram model.

78 Note: Use Python to do your work

Dan Jurafsky

Articles to get ideas to work on

• Chen, S. F. and Goodman, J. (1999). An empirical study of smoothing techniques for language modeling.
Computer Speech and Language, 13, 359–394. (Comparison of different LMs)
• Goodman, J. (2006). A bit of progress in language mod- eling: Extended version. Tech. rep. MSR-TR-2001-72,
Machine Learning and Applied Statistics Group, Microsoft Research, Redmond, WA.
• Stolcke, A. (2002). SRILM – an extensible language mod- eling toolkit. In ICSLP-02, Denver, CO. (SRILM)
• Heafield, K., Pouzyrevsky, I., Clark, J. H., and Koehn, P. (2013). Scalable modified Kneser-Ney language model
es- timation.. In ACL 2013, pp. 690–696. (KenLM)
• Bengio, Y., Schwenk, H., Sene ́cal, J.-S., Morin, F., and Gau- vain, J.-L. (2006). Neural probabilistic language
models. In Innovations in Machine Learning, pp. 137–186. Springer.
• Schwenk, H. (2007). Continuous space language models. Computer Speech & Language, 21(3), 492–518.
• Mikolov, T. (2012). Statistical language models based on neural networks. Ph.D. thesis, Ph. D. thesis, Brno
Univer- sity of Technology. (Neural Nets)
• Liu, X., Gales, M. J. F., and Woodland, P. C. (2013). Use of contexts in language model interpolation and
adaptation. Computer Speech & Language, 27(1), 301–321. (LMs Adaptation)
79