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(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)
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

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 )
 Dan  Jurafsky

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

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
Introduction  to  N-­‐grams
 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:  
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)

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/
• 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
 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

30
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  (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.
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  ____ ….
• 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
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  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
Evaluation  and  
Perplexity
 Language  
Modeling
Generalization  and  
zeros
Dan  Jurafsky

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

The  wall  street  journal  is  not  shakespeare  

(no  offense) 4.3 • GENERALIZATION AND ZEROS 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-
Dan  Jurafsky

Can  you  guess  the  author  of  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.  

43
Dan  Jurafsky

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
 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  hence  we  cannot  compute  perplexity  (can’t  divide  by  0)!
 Language  
Modeling
Generalization  and  
zeros
 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
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
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
• 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

 from
Dan   Jurafskyall
the N-gram estimators, weighing and combining the trigram, bigram, and
unigram counts. P̂(wn |wn 2 wn 1 ) = l1 P(wn |w
In simple linear interpolation, we combine different order N-grams by linearly
Linear  Interpolation 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 conditional  
that the l s sum oXn  context:
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
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:

log P(w1...wn | M (λ1...λk )) = ∑ log PM ( λ1... λk ) (wi | wi−1 )

i
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  
" 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 ) =
64 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

65
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
c(w ∈ history)
PCACHE (w | history) = λ P(wi | wi−2 wi−1 ) + (1− λ )
| history |
• These  perform  very  poorly  for  speech  recognition  (why?)
 Language  
Modeling
Interpolation,  Backoff,  
and  Web-­‐Scale  LMs
 Language
Modeling

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

little  from  each  count

• Suppose  we  wanted  to  subtract  a  little   Bigram  count   Bigram  count  in  
in  training heldout set
from  a  count  of  4  to  save  probability  
0 .0000270
mass  for  the  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   5 4.21
Newswire 6 5.23
• Training  and  held-­‐out  set 7 6.21
• for  each  bigram  in  the  training  set 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 ) = + 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)?
70
Dan  Jurafsky

Kneser-­Ney  Smoothing  I
• Better  estimate  for  probabilities  of  lower-­‐order  unigrams!
Francisco
glasses
• Shannon  game:    I  can’t  see  without  my  reading___________?
• “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) ∝ {wi−1 : c(wi−1, w) > 0}
Dan  Jurafsky

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}
Dan  Jurafsky

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  (Francisco)  occurring  in  only  one  context  (San)  will  have  a  
low  continuation  probability
Dan  Jurafsky

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
74 = # of times we applied normalized discount
 Dan  Jurafsky

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

75
 Language
Modeling

Advanced:
Kneser-Ney Smoothing