Natural Language Processing

Curated by
Dr. Tohida Rehman
Assistant Professor
Department of Information Technology
Jadavpur University
 Contents

1. Language Modeling
2. References
 What is Language Models(LMs)
 A language model (LM) is a machine learning model that predicts the next word(s) in a
sequence.
 Formally, it assigns a probability to each possible next word(s).
 LMs can also assign a probability to an entire sentence.
 To specify a correct probability distribution, the probability of all sentences in a language
must sum to 1.
 Why Is Word Prediction Important in NLP?
 Many NLP applications (translation, chatbots, autocomplete) rely on predicting likely
word sequences.
 Modern large language models (LLMs) are trained primarily via next-word
prediction.
 If a model predicts words well, it implicitly learns grammar, facts, and reasoning.
 Knowledge is impotant as chain rule of probability.
 Simple Example: N-Gram Language Models
 Basic Approach: Predicts next word based on previous *(n-1)* words (e.g., bigram: "the
cat" → "sat").
 Limitation: Fixed context window (no long-range dependencies).

 Relies on the chain rule of probability (factorizing sentence probability).

 "The sky is ___"

Possible predictions: blue, cloudy, falling, limitless

 Correct predictions (blue) require knowledge of:

 Grammar (adjective follows "is").
 World knowledge (skies are typically blue).
 Probabilistic Language Models
 Main goal: assign a probability to a sentence
 Machine Translation:
 P(high winds tonite) > P(large winds tonite)
 Spell Correction
 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 Models
 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 estimate these probabilities
 Recall the definition of conditional probabilities
 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(w1w2 …wn ) = Õ P(wi | w1w2 …wi-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)
 Probabilistic Language Models
 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
 Language is creative and any particular context might have never occurred
before!
 Markov Assumption
 Simplifying assumption:

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)

P(w1w2 …wn ) » Õ P(wi | wi-k …wi-1 )

i

In other words, we approximate each component in the product

 Markov Assumption

P(w1w2 …wn ) » Õ P(wi | wi-k …wi-1 )

i

In other words, we approximate each component in the product

P(wi | w1w2 …wi-1) » P(wi | wi-k …wi-1)

 Simplest case: Unigram model

P(w1w 2 …w n ) » Õ P(w i )
i
Word Count Probability P(wi)
 Example corpus counts (from a tiny corpus):
"the" 100 100/200 = 0.5
 Calculate Sentence Probability
"cat" 50 50/200 = 0.25
 Example Sentence: "the cat sat on the cat" "sat" 30 30/200 = 0.15
 P(the)×P(cat)×P(sat)×P(on)×P(the)×P(cat) "on" 20 20/200 = 0.1
=0.5×0.25×0.15×0.1×0.5×0.25 Total 200

=0.000234375(or 2.34×10−4)
⁈ Repetition penalty: Repeated words ("the", "cat") reduce probability exponentially.
⁈ Sparsity issue: If a word is unseen(like dog) making the entire sentence probability zero.
 Bigram model
P(wi | w1w2 …wi-1) » P(wi | wi-1)
Word Count Probability P(wi)
Calculating Sentence Probability:
"the" 100 100/200 = 0.5
"cat" 50 50/200 = 0.25
"the cat sat on the cat"
"sat" 30 30/200 = 0.15

P(the)×P(cat∣the)×P(sat∣cat)×P(on∣sat)×P(the∣on)×P(cat∣the) "on" 20 20/200 = 0.1

=0.5×0.4×0.5×0.5×0.5×0.4 Total 200 P(w_i|w_{i-1})
=0.01 40/100=0.4
"the cat" 40
"cat sat" 25 25/50=0.5
"sat on" 15 15/30=0.5
"on the" 10 5/50=0.1
 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

 Estimating bigram probabilities
 The Maximum Likelihood Estimate

count(wi-1,wi )
P(wi | w i-1) =
count(w i-1 )

c(wi-1,wi )
P(wi | w i-1 ) =
c(wi-1)
 An example
<s> I am Sam </s>
<s> Sam I am </s>
<s> I do not like green eggs and ham </s> c(wi-1,wi )
P(wi | w i-1 ) =
c(wi-1)
 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:
 An example
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
 Evaluation and Perplexity
 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.
 Evaluation and Perplexity
 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
 Evaluation and Perplexity
 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.
 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
 Intuition of Perplexity
 The Shannon Game:
 How well can we predict the next word? mushrooms 0.1
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
….
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
 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

 Common Smoothing Techniques
 Laplace (Add-1) Smoothing
 Adds 1 to all counts.
 Problem: Over-smoothes, distorts probabilities.
 Good-Turing Smoothing
 Adjusts counts based on frequency of frequencies.
 Problem: Doesn’t handle high n-grams well.
 Interpolation & Backoff
 Mixes higher and lower-order n-grams.
 Problem: May not generalize well.
 Smoothing: Add-one (Laplace) smoothing
Solution of Generalization and zeros
 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 ) =
 (Maximum Likelihood estimate) c(wi-1 )
c(wi-1, wi ) +1
 Add-1 estimate: PAdd-1 (wi | wi-1 ) =
c(wi-1 ) +V
 Laplace-smoothed bigrams

where V is the vocabulary size

 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.

Reminder: Add-1 (Laplace) Smoothing, Add-k

c(wi-1, wi ) +1
PAdd-1 (wi | wi-1 ) =
c(wi-1 ) +V

c(wi-1, wi ) + k
PAdd-k (wi | wi-1 ) =
c(wi-1 ) + kV

1
c(wi-1, wi ) + m( )
PAdd-k (wi | wi-1 ) = V
c(wi-1 ) + m
 Unigram prior smoothing

1
c(wi-1, wi ) + m( )
PAdd-k (wi | wi-1 ) = V
c(wi-1 ) + m

c(wi-1, wi ) + mP(wi )
PUnigramPrior (wi | wi-1 ) =
c(wi-1 ) + m
 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

 Linear Interpolation

 Simple interpolation

 Lambdas conditional on context:

 How to set the lambdas?
 Use a held-out corpus

Training Data Held-Out Test

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(l1...lk )) = å log PM ( l1... lk ) (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
 Create an unknown word token <UNK>
 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)

38
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-1
)
if count(wi-k+1 ) > 0
i

S(wi | wi-k+1 ) = í count(wi-k+1 )

i-1

ï i-1
ïî 0.4S(w i | w i-k+2 ) otherwise

count(wi )
S(wi ) =
N
39
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
 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?)

c(w Î history)
PCACHE (w | history) = l P(wi | wi-2 wi-1 ) + (1- l )
| history |
 Advanced smoothing algorithms
 Intuition used by many smoothing algorithms
 Good-Turing

 Kneser-Ney

 Witten-Bell

 Use the count of things we’ve seen once

 to help estimate the count of things we’ve never seen
 Why Kneser-Ney Smoothing?
 Problem with Traditional Smoothing:
 Handles unseen n-grams but ignores context diversity.
 Example:
 "San Francisco" vs. "Francisco" (rare alone but frequent after "San").
 Traditional methods overestimate "Francisco" in new contexts.
Kneser-Ney Smoothing – Key Idea
 Instead of raw counts, it considers how many different contexts a word appears in.
 Advantages of Kneser-Ney:
 Handles Rare Words Better:
 Considers context diversity.
 Works well for unseen n-grams.
 Outperforms other smoothing techniques in practice.
 Notation: Nc = Frequency of frequency c
 Nc = the count of things we’ve seen c times
 Sam I am I am Sam I do not eat
 I 3
 sam 2
 am 2
 do 1
 not 1
N1 = 3
 eat 1
N2 = 2

N3 = 1
 Good-Turing smoothing intuition
 You are fishing (a scenario from Josh Goodman), and caught:
 10 carp, 3 perch, 2 whitefish, 1 trout, 1 salmon, 1 eel = 18 fish
 How likely is it that next species is trout?
 1/18
 How likely is it that next species is new (i.e. catfish or bass)
 Let’s use our estimate of things-we-saw-once to estimate the new things.
 3/18 (because N1=3)
 Assuming so, how likely is it that next species is trout?
 Must be less than 1/18
 How to estimate?
 Good Turing calculations

N1 (c +1)N c+1
P (things with zero frequency) =
*
GT c* =
N Nc
 Unseen (bass or catfish)
 c = 0:
• Seen once (trout)
•c=1
 MLE p = 0/18 = 0
• MLE p = 1/18

 P*GT (unseen) = N1/N = 3/18

• C*(trout) = 2 * N2/N1
= 2 * 1/3
= 2/3
• P*GT(trout) = 2/3 / 18 = 1/27
46

Ney et al.’s Good Turing Intuition

H. Ney, U. Essen, and R. Kneser, 1995. On the estimation of 'small' probabilities by leaving-one-out.
IEEE Trans. PAMI. 17:12,1202-1212

Held-out words:
 Training Held out
Ney et al. Good Turing Intuition
(slide from Dan Klein)
N1 N0
 Intuition from leave-one-out validation
 Take each of the c training words out in turn
 c training sets of size c–1, held-out of size 1
N2 N1
 What fraction of held-out words are unseen in training?
 N1/c
 What fraction of held-out words are seen k times in N3 N2
training?

....
....
 (k+1)Nk+1/c
 So in the future we expect (k+1)Nk+1/c of the words to be
those with training count k
 There are Nk words with training count k
 Each should occur with probability: N3511 N3510
(k +1)N k+1
 (k+1)Nk+1/c/Nk k* =
Nk N4417 N4416
 Good-Turing complications
(slide from Dan Klein)

 Problem: what about “the”? (say c=4417)

 For small k, Nk > Nk+1
N1
 For large k, too jumpy, zeros wreck estimates N2 N
3

 Simple Good-Turing [Gale and Sampson]:

replace empirical Nk with a best-fit power law
once counts get unreliable
N1
N2
 Resulting Good-Turing numbers
 Numbers from Church and Gale (1991) Count Good Turing c*
c
 22 million words of AP Newswire
0 .0000270
(c +1)N c+1 1 0.446
c* = 2 1.26
Nc
3 2.24
4 3.24
5 4.22
6 5.19
7 6.21
8 7.24
9 8.25
Language Modeling
Advanced:
Kneser-Ney Smoothing
 Resulting Good-Turing numbers

 Numbers from Church and Gale (1991) Count Good Turing c*

c
 22 million words of AP Newswire
0 .0000270
(c +1)N c+1 1 0.446
c* = 2 1.26
Nc
3 2.24
4 3.24
 It sure looks like c* = (c - .75) 5 4.22
6 5.19
7 6.21
8 7.24
9 8.25
52

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(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)?
 Kneser-Ney Smoothing I
 Better estimate for probabilities of lower-order unigrams!
 Shannon game: I can’t see without my reading___________?
Francisco
glasses
 “Francisco” is more common than “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?
 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)µ {wi-1 : c(wi-1, w) > 0}

 Kneser-Ney Smoothing II
 How many times does w appear as a novel continuation:

PCONTINUATION (w)µ {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) =
å {w' i-1 : c(w'i-1, w') > 0}
w'

 A frequent word (Francisco) occurring in only one context (San) will have a low
continuation probability
 Kneser-Ney Smoothing IV

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 word types that can follow wi-1
the normalized discount = # of word types we discounted
= # of times we applied normalized discount
 Kneser-Ney Smoothing: Recursive formulation

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

ìï count(·) for the highest order

cKN (·) = í
ïî continuationcount(·) for lower order

Continuation count = Number of unique single word contexts for 

 HW
 What is the primary limitation of Maximum Likelihood Estimation (MLE) in n-gram

models that Good-Turing smoothing addresses?

 Explain the intuition behind Good-Turing smoothing. How does it adjust the counts of
unseen or rare n-grams?

 Explain how Kneser-Ney smoothing would assign a probability to a word in the real world
example.
 Reference Books
1. Daniel Jurafsky and James H. Martin. 2020. Speech and Language Processing.
2. 3rd Edition Christopher D. Manning and Hinrich Schütze. 1999. Foundations of Statistical
Natural Language Processing. MIT Press.
3. Sowmya Vajjala, Bodhisattwa Majumder, Anuj Gupta, Harshit Surana. 2020. Practical
Natural Language Processing. O'Reilly.
4. NPTEL NLP course.
5. https://www.google.co.in/
6. Coursera course - Natural Language Processing