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 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.!!
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)
 The Chain Rule applied to compute
Dan Jurafsky

joint probability of words in sentence

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?

• No! Too many possible sentences!

• We’ll never see enough data for estimating these
Dan Jurafsky

Markov Assumption

• Simplifying assumption:
Andrei Markov

• Or maybe
Dan Jurafsky

Markov Assumption

• In other words, we approximate each

component in the product
Dan Jurafsky

Simplest case: Unigram model

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:

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 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
Modeling
Introduction to N-grams
 Language
Modeling
Estimating N-gram
Probabilities
Dan Jurafsky

Estimating bigram probabilities

• The Maximum Likelihood Estimate
Dan Jurafsky

An example

<s> I am Sam </s>

<s> Sam I am </s>
<s> I do not like green eggs and ham </s>
Dan Jurafsky

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
Dan Jurafsky

Raw bigram counts

• Out of 9222 sentences
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)
= .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
• (also adding is faster than multiplying)
Dan Jurafsky

Language Modeling Toolkits

• SRILM
• http://www.speech.sri.com/projects/srilm/
• KenLM
• https://kheafield.com/code/kenlm/
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

• http://ngrams.googlelabs.com/
 Language
Modeling
Estimating N-gram
Probabilities
Languag Evaluation and Perplexity
e
Modelin
g
How to evaluate N-gram models

"Extrinsic (in-vivo) Evaluation"

To compare models A and B
1. Put each model in a real task
• Machine Translation, speech recognition, etc.
2. Run the task, get a score for A and for B
• How many words translated correctly
• How many words transcribed correctly
3. Compare accuracy for A and B
Intrinsic (in-vitro) evaluation
Extrinsic evaluation not always possible
• Expensive, time-consuming
• Doesn't always generalize to other applications
Intrinsic evaluation: perplexity
• Directly measures language model performance at
predicting words.
• Doesn't necessarily correspond with real application
performance
• But gives us a single general metric for language models
• Useful for large language models (LLMs) as well as n-grams
Training sets and test sets
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; different from training set.
◦ Intuition: we want to measure generalization to unseen data
◦ An evaluation metric (like perplexity) tells us how well
our model does on the test set.
Choosing training and test sets

• If we're building an LM for a specific task

• The test set should reflect the task language we
want to use the model for
• If we're building a general-purpose model
• We'll need lots of different kinds of training data
• We don't want the training set or the test set to
be just from one domain or author or language.
Training on the test set
We can’t allow test sentences into the training set
• Or else the LM will assign that sentence an artificially
high probability when we see it in the test set
• And hence assign the whole test set a falsely high
probability.
• Making the LM look better than it really is
This is called “Training on the test set”
Bad science!
33
Dev sets
• If we test on the test set many times we might
implicitly tune to its characteristics
• Noticing which changes make the model better.
• So we run on the test set only once, or a few times
• That means we need a third dataset:
• A development test set or, devset.
• We test our LM on the devset until the very end
• And then test our LM on the test set once
Intuition of perplexity as evaluation metric: How
good is our language model?

Intuition: A good LM prefers "real" sentences

• Assign higher probability to “real” or “frequently
observed” sentences
• Assigns lower probability to “word salad” or
“rarely observed” sentences?
 Intuition of perplexity 2:
Predicting upcoming words
time 0.9
The Shannon Game: How well can we predict
dream 0.03
the next word?
• Once upon a ____ midnight 0.02
• That is a picture of a ____ …
• For breakfast I ate my usual ____ and 1e-100

Claude Shannon
Unigrams are terrible at this game (Why?)

A good LM is one that assigns a higher probability

to the next word that actually occurs
Picture credit: Historiska bildsamlingen
https://creativecommons.org/licenses/by/2.0/
Intuition of perplexity 3: The best language model is
one that best predicts the entire unseen test set

• We said: a good LM is one that assigns a higher

probability to the next word that actually occurs.
• Let's generalize to all the words!
• The best LM assigns high probability to the entire test
set.
• When comparing two LMs, A and B
• We compute PA(test set) and PB(test set)
• The better LM will give a higher probability to (=be less
surprised by) the test set than the other LM.
Intuition of perplexity 4: Use perplexity
instead of raw probability
• Probability depends on size of test set
• Probability gets smaller the longer the text
• Better: a metric that is per-word, normalized by length
• Perplexity is the inverse probability of the test set,
normalized by the number of words
Intuition of perplexity 5: the inverse
Perplexity is the inverse probability of the test set, normalized
by the number of words

(The inverse comes from the original definition of perplexity

from cross-entropy rate in information theory)
Probability range is [0,1], perplexity range is [1,∞]
Minimizing perplexity is the same as maximizing probability
Intuition of perplexity 6: N-grams

Chain rule:

Bigrams:
 Intuition of perplexity 7:
Weighted average branching factor
Perplexity is also the weighted average branching factor of a language.
Branching factor: number of possible next words that can follow any word
Example: Deterministic language L = {red,blue, green}
Branching factor = 3 (any word can be followed by red, blue, green)
Now assume LM A where each word follows any other word with equal probability ⅓
Given a test set T = "red red red red blue"
PerplexityA(T) = PA(red red red red blue)-1/5 = ) )⅓((
1/5- 5
)⅓( = 3=
1-

But now suppose red was very likely in training set, such that for LM B:
◦ P(red) = .8 p(green) = .1 p(blue) = .1
We would expect the probability to be higher, and hence the perplexity to be smaller:
PerplexityB(T) = PB(red red red red blue)-1/5
= (.8 * .8 * .8 * .8 * .1) -1/5 =.04096 -1/5 = .527-1 = 1.89
Holding test set constant:
Lower perplexity = better language model

Training 38 million words, test 1.5 million words, WSJ

N-gram Unigram Bigram Trigram

Order
Perplexity 962 170 109
Languag Evaluation and Perplexity
e
Modelin
g
Languag Sampling and Generalization
e
Modelin
g
The Shannon (1948) Visualization Method
Sample words from an LM
Claude Shannon
Unigram:
REPRESENTING AND SPEEDILY IS AN GOOD APT OR COME
CAN DIFFERENT NATURAL HERE HE THE A IN CAME THE TO
OF TO EXPERT GRAY COME TO FURNISHES THE LINE
MESSAGE HAD BE THESE.

Bigram:
THE HEAD AND IN FRONTAL ATTACK ON AN ENGLISH WRITER
THAT THE CHARACTER OF THIS POINT IS THEREFORE
ANOTHER METHOD FOR THE LETTERS THAT THE TIME OF WHO
EVER TOLD THE PROBLEM FOR AN UNEXPECTED.
 How Shannon sampled those words in 1948

"Open a book at random and select a letter at random on the

page. This letter is recorded. The book is then opened to another
page and one reads until this letter is encountered. The
succeeding letter is then recorded. Turning to another page this
second letter is searched for and the succeeding letter recorded,
etc."
Sampling a word from a
distribution
 Visualizing Bigrams the Shannon Way

Choose a random bigram (<s>, w)

<s> I
according to its probability p(w|<s>) I want
Now choose a random bigram (w, x) want to
according to its probability p(x|w) to eat
And so on until we choose </s> eat Chinese
Chinese food
Then string the words together
food </s>
I want to eat Chinese food
Note: there are other sampling
methods
Used for neural language models
Many of them avoid generating words from the very
unlikely tail of the distribution
We'll discuss when we get to neural LM decoding:
◦ Temperature sampling
◦ Top-k sampling
◦ Top-p sampling
Approximating Shakespeare
Shakespeare as corpus

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

Shakespeare produced 300,000 bigram types out of
V2= 844 million possible bigrams.
◦ So 99.96% of the possible bigrams were never seen (have
zero entries in the table)
◦ That sparsity is even worse for 4-grams, explaining why
our sampling generated actual Shakespeare.
The Wall Street Journal is not
Shakespeare
Can you guess the author? These 3-gram
sentences are sampled from an LM trained on who?
1) 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
2) This shall forbid it should be branded,
if renown made it empty.
3) “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. 53
Choosing training data
If task-specific, use a training corpus that has a similar
genre to your task.
• If legal or medical, need lots of special-purpose documents
Make sure to cover different kinds of dialects and
speaker/authors.
• Example: African-American Vernacular English (AAVE)
• One of many varieties that can be used by African Americans and others
• Can include the auxiliary verb finna that marks immediate future tense:
• "My phone finna die"
The perils of overfitting
N-grams only work well for word prediction if the
test corpus looks like the training corpus
• But even when we try to pick a good training
corpus, the test set will surprise us!
• 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
… ate lunch … ate lunch
… ate dinner … ate breakfast
… ate a
… ate the
P(“breakfast” | ate) = 0
Zero probability bigrams
Bigrams with zero probability
◦ Will hurt our performance for texts where those words
appear!
◦ And mean that we will assign 0 probability to the test set!
And hence we cannot compute perplexity (can’t
divide by 0)!
Languag Sampling and Generalization
e
Modelin
g
 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
reports
2 reports

attack
…

request
claims
1 claims

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

allegations
allegations
1.5 reports

outcome
0.5 claims

reports

attack
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!

• MLE estimate:

• Add-1 estimate:
Dan Jurafsky

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.
Dan Jurafsky

Berkeley Restaurant Corpus: Laplace

smoothed bigram counts
Dan Jurafsky

Laplace-smoothed bigrams
Dan Jurafsky

Reconstituted counts
Dan Jurafsky

Compare with raw bigram counts

Dan Jurafsky

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
Modeling
Smoothing: Add-one
(Laplace) smoothing
 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

• Lambdas conditional on context:

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:
Dan Jurafsky

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
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-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)
Dan Jurafsky

Smoothing for Web-scale N-grams

• “Stupid backoff” (Brants et al. 2007)
• No discounting, just use relative frequencies

75
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

76
Dan Jurafsky

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

• These perform very poorly for speech recognition (why?)

 Language
Modeling
Interpolation, Backoff,
and Web-Scale LMs