ADVA N C ED

NLP

CHAPTER 01: LANGUAGE

MODELLING
 1 2 3 4
Probabilistic LM? Markov hypothesis N-gram Models Evaluation and complexity

COURSE OUTLINE

ADVA N CED N LP Mrs. MEZZI 2

 PROBABILISTIC LANGUAGE MODELS

◾ O bjective: assign a probability to a sentence.

o Automatic translation:
 P(high winds tonight) > P(large winds tonite)
Why?

o Spellchecking
 P(fifteen minutes from) > P(fiften minuetes from)

o Speech recognition
 P(I saw a van) > P(eyes awe of an)
o + Automatic summarization, optical character recognition, document classification, question-answering,
etc., etc.!!

ADVA N CED N LP Mrs. MEZZI 3

 PROBABILISTIC LANGUAGE MODELS

◾Objective: determine the of a phrase or word combination by

computing the following probability: P(W) = P(w1,w2,w3,w4,w5…wn)

◾Related tasks: probability of the next P(w5|w1,w2,w3,w4)

◾A model that makes it possible to compute both scenarios :
P(W) ou P(wn|w1,w2…wn-1) is called a Language Model

ADVA N CED N LP Mrs. MEZZI 4

 PROBABILISTIC LANGUAGE MODELS

◾ Examples:
o Let’s meet instead in…
o The company's stock lost two...
o If you are able to visit …

ADVA N CED N LP Mrs. MEZZI 5

 PROBABILISTIC LANGUAGE MODELS
information provided by "Natural Language Corpus Data: Beautiful Data" and derived from the "google web
trillion word corpus"

ADVA N CED N LP Mrs. MEZZI 6

 PROBABILISTIC LANGUAGE MODELS
information provided by "Natural Language Corpus Data: Beautiful Data" and derived from the "google web
trillion word corpus"

ADVA N CED N LP Mrs. MEZZI 7

 C ALCUATION OF THE PROBABILITY P(W)

◾ How to calculate joint probabilities?

◾P(this, water, is, so, clear, that)

◾ Intuition: Taking conditional probabilities into account

ADVA N CED N LP Mrs. MEZZI 8

 REC ALL: C O N D ITIONAL PROBABILITIES

P(B|A) = P(A,B)/P(A)or P(A,B) = P(A)P(B|A)

◾Many variables:
P(A,B,C,D) = P(A)P(B|A)P(C|A,B)P(D|A,B,C)
◾General Rule:
P(x1,x2,x3,…,xn) = P(x1)P(x2|x1)P(x3|x1,x2)…P(xn|x1,…,xn-1)
ADVA N CED N LP Mrs. MEZZI 9
 LANGUAGE-BASED CONDITIONAL PROBABILITIES

P(w1 w2 … wn )   P(wi | w1 w2 … wi 1)

i

P(“This water is so clear”) =

P(this) × P(water|this) × P(is|this water)
× P(so|this water is) × P(clear|this water is so)

ADVA N CED N LP Mrs. MEZZI 10

 HO W TO ESTIMATE THESE PROBABILITIES

◾ Can we just divide ?

�(the|this water is so clear that) =

�����(this water is so cl��� �ℎ�� �ℎ�)
�����(this water is so cl��� �ℎ��)

◾Wrong! Too many possible phrases! (infinite sentences!!)

◾A large corpus is required to estimate this probability.

ADVA N CED N LP Mrs. MEZZI 11

 MARKOV HYPOTHESIS

◾Simplified hypothesis: Andrei Markov

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

◾Or
P(the | its water is so transparent that)  P(the | transparent that)

ADVA N CED N LP Mrs. MEZZI 12

 MARKOV HYPOTHESIS

P(w1 w2 … wn )  P(w
i i
| wi k … wi 1)
◾ In other words, we approximate each component of the product

P(wi | w1 w2 … wi 1)  P(wi | wi  k … wi 1)

ADVA N CED N LP Mrs. MEZZI 13

 MARKOV HYPOTHESIS

o Calculating the probability of a word is limited to knowing the probability

of the words that precede it.
o Markov's hypothesis allows us to focus only on a limited history.
N-gram models:
o Unigram: P(the) no context
o Bigram: P(the |that)
o Trigram: P(the | transparent that)

ADVA N CED N LP Mrs. MEZZI 14

 N-GRAM MODELS

◾ We can thus extend to trigram models, 4-grams, 5-grams models

Small N -grams  information loss
At times, it can be an insufficient language model, because there are long-
distance dependencies:

“The computer I used yesterday for the Advanced NLP class session,
crashed.”
ADVA N CED N LP Mrs. MEZZI 15
 N-GRAM MODELS
o If we use large N-grams  we obtain a model with a high complexity
 We then need a larger corpus
 Representation of N-grams :VN where V is the vocabulary size and N the number of grams.
o Issue with absent N-grams in the training corpus
 P(<s>he likes swimming </s>) = P(he|<s>)P(likes |he)P(swimming |likes)P(</s> |swimming)
= 32 0 1= 0
43 11
 The idea is to create a probability for absent N-grams by taking a small percentage of the
probabilities of the current N -grams.

But often, the N-gram models constitute an interesting solution

ADVA N CED N LP Mrs. MEZZI 16

 LANGUAGE MODELLING

N-GRAM PROBABILITY ESTIMATION

 DATA PREPROCESSING

o How to convert text to sequences �1� ?

 This depends on the application
o The following criteria need to be determined :
 How to delimit sequences (sentence, paragraphs, or documents)?
 How to delimit words ?
 How to normalize words?
 What words are in the model's vocabulary?

ADVA N CED N LP Mrs. MEZZI 18

 DATA PREPROCESSING

o Sentences are typically used to delimit data for sequences because this is the form of
sequence that is handled the most frequently (e.g. automatic translation)

o Reserved words, such as <s> for the beginning and </s> for end of a sentence in
English, are frequently used to indicate sentence boundaries.

<s> The weather is nice today</s>

ADVA N CED N LP Mrs. MEZZI 19

 DATA PREPROCESSING

o For words:A preliminary processing step involves converting a document into

a list of words that appear in it (tokens).
This step is called « Tokenization »
 Each lexical unit (token) corresponds then to a word.
o In English, French and some similar languages, the steps can be as follows:

 Dividing the sentences using spaces and punctuation.

 Considering whether or not the punctuation is a part of the sequence.

ADVA N CED N LP Mrs. MEZZI 20

 DATA PREPROCESSING

o It is best to disregard the punctuation most of the time.

 The punctuation is omitted when it cannot be captured, as in speech recognition,
for instance.
o In other cases, not even the spaces are sufficient to separate the words.

ADVA N CED N LP Mrs. MEZZI 21

 DATA PREPROCESSING

o The problem can be partially handled by treating each punctuation

mark as a separate character.
o However, there will always be exceptions:
o Ph.D.
o google.com
o 555,500.50
o 555 500,50

ADVA N CED N LP Mrs. MEZZI 22

 DATA PREPROCESSING

o We occasionally work with clitics and must return them to their native form:
o j’aime = je and aime
o he’s = he and is
o At times, we would like a certain token to be regarded as a collocation, which
is a grouping of several words:
o New York
o rock ‘n’ roll

ADVA N CED N LP Mrs. MEZZI 23

 DATA PREPROCESSING

o The selection of tokenization rules varies depending on the

application.
o Additionally, certain languages lack separators: the case of
Japanese and Chinese.

ADVA N CED N LP Mrs. MEZZI 24

 DATA PREPROCESSING

o How to normalize words?

o Do we take capitalization into account?
 no in speech recognition yes in automatic translation
o Should we limit to a lemma ?
 yes in document classification  no in many other NLP tasks.
o Do we need to use any other conversions??
 Numbers: <number>, Date: <date>…etc.
o Each of these decisions is based on the current case study.
o It is recommended to conduct multiple experiments and make improvements in reaction to errors.

ADVA N CED N LP Mrs. MEZZI 25

 BI-GRAMS PROBABILITY ESTIMATION

◾ Maximum Likelihood Estimation MLE allows to compute this probability based on the
sequences ferequencies:

P(wi | w i1 )  count(wi1,wi )

)  c(wi1,wi )
P(w i | w i1
count(w i1 ) c(w i1)

o count(wi-1 , wi) : the bigram’s (wi-1 , wi) frequence

o count (wi-1): the unigram’s frequnce wi-1

ADVA N CED N LP Mrs. MEZZI 26

 EXAMPLE

<s> I am Sam </s>

c(w i1,w i )
P(wi | w i1 )  <s> Sam I am </s>
c(wi1)
<s> I do not like green eggs and ham </s>

ADVA N CED N LP Mrs. MEZZI 27

 EXAMPLE 2: BERKELEY RESTAURANT PROJECT

 https://github.com/wooters/berp-trans
 A database of questions about a Berkeley restaurant (9332 sentences)
o can you tell me about any good cantonese restaurants close by
o mid priced thai food is what i’m looking for
o tell me about chez panisse
o can you give me a listing of the kinds of food that are available
o i’m looking for a good place to eat breakfast
o when is caffe venezia open during the day
ADVA N CED N LP Mrs. MEZZI 28
 BIGRAMS’ FREQUENCIES
◾ Counting bigrams

◾ Tokenization information for Unigrams :

ADVA N CED N LP Mrs. MEZZI 29

 BIGRAMS’ PROBABILITIES

◾ Bigram probability results:

ADVA N CED N LP Mrs. MEZZI 30

 BIGRAMS’ ESTIMATION

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

ADVA N CED N LP Mrs. MEZZI 31

 WHAT KIND OF KNOWLEDGE?

ADVA N CED N LP Mrs. MEZZI 32

 ISSUE

o We take all necessary steps in logarithmic space to prevent

underflow.
o It is also faster to add (sum) than to multiply (multiplication).

log(p1  p2  p3  p4 )  log p1  log p2  log p3  log p4

ADVA N CED N LP Mrs. MEZZI 33
 GOOGLE N-GRAMS

o https://books.google.com/ngrams
o Models pre-processed from books
o free download:
https://storage.googleapis.com/books/ngrams/books/datasetsv3.html

ADVA N CED N LP Mrs. MEZZI 34

 ESTIMATION OF BIGRAMS

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

ADVA N CED N LP Mrs. MEZZI 35

 LANGUAGE MODELLING

EVALUATION A N D COMPLEXITY
 PERFORMANCE EVALUATION
◾ Does our linguistic model prefer good sentences to bad ones?
◾ Assign a higher probability to "real" or "frequently observed" sentences than to
"ungrammatical" or "rarely observed" sentences? than "ungrammatical" or "rarely
observed" sentences?
◾  Our the model's parameters are trained using a learning base.
◾  We evaluate the model's performance on data that we did not observe.
◾ A test database is an unnoticed, entirely unused set of data that is distinct from
our training set.
◾ An evaluation measure shows us how well our model performs on the test set.

ADVA N CED N LP Mrs. MEZZI 37

 TEST-BASED LEARNING

In terms of ethics:
o Test sentences are not allowed in learning.
o We will assign them an artificially high probability when we define them in the test
set .
o Test-based learning is a practice that violates the code of honour and distorts
results.

ADVA N CED N LP Mrs. MEZZI 38

 IN-VIVO EVALUATION

◾ The best rating for models A and B

◾ Put each template into a specific task
◾ Spelling corrector, speech recognition, Machine translation systems

◾ Run the task and get the relevance of A and B

◾ How many errors correctly fixed?
◾ How many words are translated correctly?

◾ Compare the relevance of Model A to Model B.

ADVA N CED N LP Mrs. MEZZI 39

 ISSUES
◾ In-vivo evaluation is:
◾ Long (time-consuming); can take days or weeks...
Donc
◾ To perform an internal (in-vitro) evaluation
◾ Measure: perplexity
◾ Approximations are ineffective.
◾ Unless the test collection is similar to the learning collection.
◾Typically useful for benchmarking experiments.
ADVA N CED N LP Mrs. MEZZI 40
 INTUITION CONCERNING PERPLEXITY

mushrooms 0.1
◾ The Shannon Game:
pepperoni 0.1
◾ In which manner, will we be able to predict the next word?

I always order pizza with cheese and anchovies 0.01

….
The 33rd President of the US was
fried rice 0.0001
I saw a
….
◾ Unigrams are not a good solution for this game (why?).
and 1e-100
◾ A more suitable model for the text
◾ Is a model that assigns a high probability to words that actually appear.

ADVA N CED N LP Mrs. MEZZI 41

 PERPLEXITY

The best linguistic model is one that best predicts a set of invisible tests 
Give the highest P (sentence). 1
PP(W )  P(w1w2...wN )
N

1
 N
The perplexity is the inverse probability of the total P(w1w2...wN )

number of tests, normalized by the number of words:

Conditional perplexities "chain rule":

For bigrams:

ADVA N CED N LP Mrs. MEZZI 42

 PERPLEXITY

o The perplexity allows us to calculate the factor of medium

ponderation in order to predict the next word.
 Average number of hesitations (equivocal predictions) during a
prediction.
o Perplexity occurs typically between 40 and 400.

Reducing perplexity (confusion) equates to increasing probability.

ADVA N CED N LP Mrs. MEZZI 43

 LOW PERPLEXITY = SUPERIOR MODEL

◾The following are the outcomes of an N-gram model:

◾Learning on 38 million words, testing on 1.5 million
words (Wall Street Journal):
N-gram
Unigram Bigram Trigram
Order
Perplexity 962 170 109
ADVA N CED N LP Mrs. MEZZI 44
 LOW PERPLEXITY = SUPERIOR MODEL

o The trigram model is the best (it has the lowest perplexity).
o To interpret these results, we conclude that the uni-gram model must make
numerous assumptions.
o Perplexity has the advantage of being a measure that is not dependent on a specific
application.
o However, it is difficult to predict whether a gain in perplexity would translate into a
real-world gain (for example, the number of translation errors).
o The goal remains to assess the system's performance using the extrinsic (in-vivo)
language evaluation model.

ADVA N CED N LP Mrs. MEZZI 45

 Additional links

o Probabilistic models in NLP

ADVA N CED N LP Mrs. MEZZI 46

 THANK YOU!
Keep in touch: [email protected]

ADVA N CED N LP Mrs. MEZZI 47