Natural language procssing

Perpectivising NLP: Areas of AI and
their inter-dependencies
--:POS Terminology:--
Search
Machine Planning
Learning
Vision
Knowledge
Logic Representation
Expert
NLP Robotics Systems

QSA Triangle
Query
Search Analystics

Areas being investigated
 Business Intelligence on the Internet
Platform
 Opinion Mining
 Reputation Management
 Sentiment Analysis (some observations
at the end)
NLP is thought to play a key role

Books etc.
 Main Text(s):
 Natural Language Understanding: James Allan
 Speech and NLP: Jurafsky and Martin
 Foundations of Statistical NLP: Manning and Schutze
 Other References:
 NLP a Paninian Perspective: Bharati, Cahitanya and Sangal
 Statistical NLP: Charniak
 Journals
 Computational Linguistics, Natural Language Engineering, AI, AI
Magazine, IEEE SMC
 Conferences
 ACL, EACL, COLING, MT Summit, EMNLP, IJCNLP, HLT,
ICON, SIGIR, WWW, ICML, ECML

Allied Disciplines
Philosophy Semantics, Meaning of “meaning”, Logic
(syllogism)
Linguistics Study of Syntax, Lexicon, Lexical Semantics etc.
Probability and Statistics Corpus Linguistics, Testing of Hypotheses,
System Evaluation
Cognitive Science Computational Models of Language Processing,
Language Acquisition
Psychology Behavioristic insights into Language Processing,
Psychological Models
Brain Science Language Processing Areas in Brain
Physics Information Theory, Entropy, Random Fields
Computer Sc. & Engg. Systems for NLP

Topics to be covered
 Shallow Processing
 Part of Speech Tagging and Chunking using HMM, MEMM, CRF, and
Rule Based Systems
 EM Algorithm
 Language Modeling
 N-grams
 Probabilistic CFGs
 Basic Linguistics
 Morphemes and Morphological Processing
 Parse Trees and Syntactic Processing: Constituent Parsing and Dependency
Parsing
 Deep Parsing
 Classical Approaches: Top-Down, Bottom-UP and Hybrid Methods
 Chart Parsing, Earley Parsing
 Statistical Approach: Probabilistic Parsing, Tree Bank Corpora

Topics to be covered (contd.)
 Knowledge Representation and NLP
 Predicate Calculus, Semantic Net, Frames, Conceptual
Dependency, Universal Networking Language (UNL)
 Lexical Semantics
 Lexicons, Lexical Networks and Ontology
 Word Sense Disambiguation
 Applications
 Machine Translation
 IR
 Summarization
 Question Answering

Grading
 Based on
 Midsem
 Endsem
 Assignments
 Seminar
 Project (possibly)
Except the first two everything else in groups
of 4. Weightages will be revealed soon.

What is NLP
 Branch of AI
 2 Goals
 Science Goal: Understand the language
processing behaviour
 Engineering Goal: Build systems that
analyse and generate language; reduce
the man machine gap

The famous Turing Test: Language
Based Interaction
Machine
Test conductor
Human
Can the test conductor find out which is the machine and which
the human

Inspired Eliza
 http://www.manifestation.com/neurotoys
/eliza.php3

Inspired Eliza (another sample
interaction)
 A Sample of Interaction:

“What is it” question: NLP is
concerned with Grounding
Ground the language into
perceptual, motor and
cognitive capacities.

Grounding faces 3 challenges
 Ambiguity.
 Co-reference resolution
(anaphora is a kind of it).
 Elipsis.

Co-reference Resolution
Sequence of commands to the
robot:
Place the wrench on the
table.
Then paint it.
What does it refer to?

Elipsis
Sequence of command to the Robot:
Move the table to the corner.
Also the chair.
Second command needs completing by
using the first part of the previous
command.

Two Views of NLP and the
Associated Challenges
1. Classical View
2. Statistical/Machine Learning
View

Stages of processing (traditional view)
 Phonetics and phonology
 Morphology
 Lexical Analysis
 Syntactic Analysis
 Semantic Analysis
 Pragmatics
 Discourse

Phonetics
 Processing of speech
 Challenges
 Homophones: bank (finance) vs. bank (river
bank)
 Near Homophones: maatraa vs. maatra (hin)
 Word Boundary
 aajaayenge (aa jaayenge (will come) or aaj aayenge (will come
today)
 I got [ua]plate
 Phrase boundary
 mtech1 students are especially exhorted to attend as such seminars
are integral to one's post-graduate education
 Disfluency: ah, um, ahem etc.

Morphology
 Study of meaningful component of word
 Nouns: Plural (boy-boys); Gender marking (czar-czarina)
 Verbs: Tense (stretch-stretched); Aspect (e.g. perfective sit-had
sat); Modality (e.g. request khaanaa khaaiie)
 First crucial first step in NLP
 Languages rich in morphology: e.g., Dravidian, Hungarian,
Turkish
 Languages poor in morphology: Chinese, English
 Languages with rich morphology have the advantage of easier
processing at higher stages of processing
 A task of interest to computer science: Finite State Machines for
Word Morphology

Lexical Analysis
 Essentially refers to dictionary access and
obtaining the properties of the word
e.g. dog
noun (lexical property)
take-’s’-in-plural (morph property)
animate (semantic property)
4-legged (-do-)
carnivore (-do)
Challenge: Lexical or word sense
disambiguation

Lexical Disambiguation
First step: part of Speech Disambiguation
 Dog as a noun (animal)
 Dog as a verb (to pursue)
Sense Disambiguation
 Dog (as animal)
 Dog (as a very detestable person)
Needs word relationships in a context
 The chair emphasised the need for adult education
Very common in day to day communications
Satellite Channel Ad: Watch what you want, when you
want (two senses of watch)
e.g., Ground breaking ceremony/research

Technological developments bring in new
terms, additional meanings/nuances for
existing terms
 Justify as in justify the right margin (word
processing context)
 Xeroxed: a new verb
 Digital Trace: a new expression

Syntax Processing Stage
Structure Detection
SS
NNPP
VVPP
VV NNPP
II
lliikkee mmaannggooeess

Parsing Strategy
 Driven by grammar
 S-> NP VP
 NP-> N | PRON
 VP-> V NP | V PP
 N-> Mangoes
 PRON-> I
 V-> like

Challenges in Syntactic
Processing: Structural Ambiguity
 Scope
1.The old men and women were taken to safe locations
(old men and women) vs. ((old men) and women)
2. No smoking areas will allow Hookas inside
 Preposition Phrase Attachment
 I saw the boy with a telescope
(who has the telescope?)
 I saw the mountain with a telescope
(world knowledge: mountain cannot be an instrument of
seeing)
 I saw the boy with the pony-tail
(world knowledge: pony-tail cannot be an instrument of
seeing)
Very ubiquitous: newspaper headline “20 years later, BMC
pays father 20 lakhs for causing son’s death”

Structural Ambiguity…
 Overheard
 I did not know my PDA had a phone for 3
months
 An actual sentence in the newspaper
 The camera man shot the man with the
gun when he was near Tendulkar

Headache for parsing: Garden
Path sentences
A garden path sentence is a grammatically
correct sentence that starts in such a way
that a reader's most likely interpretation will
be incorrect
Consider
 The horse raced past the garden (sentence
complete)
 The old man (phrase complete)
 Twin Bomb Strike in Baghdad (news paper
heading: complete)

Headache for Parsing
 Garden Pathing
 The horse raced past the garden fell
 The old man the boat
 Twin Bomb Strike in Baghdad kill 25
(Times of India 5/9/07)

Semantic Analysis
 the study of the meaning of words
 John gave a book to Mary
 Give action: Agent: John, Object: Book,
Recipient: Mary
 Challenge: ambiguity in semantic role labeling
 (Eng) Visiting aunts can be a nuisance
 (Hin) aapko mujhe mithaai khilaanii padegii
(ambiguous in Marathi and Bengali too; not in
Dravidian languages)

Pragmatics
 Very hard problem
 Model user intention
 Tourist (in a hurry, checking out of the hotel,
motioning to the service boy): Boy, go upstairs and
see if my sandals are under the divan. Do not be
late. I just have 15 minutes to catch the train.
 Boy (running upstairs and coming back panting):
yes sir, they are there.
 World knowledge
 WHY INDIA NEEDS A SECOND OCTOBER (ToI,
2/10/07)

Discourse
Processing of sequence of sentences
Mother to John:
John go to school. It is open today. Should you
bunk? Father will be very angry.
Ambiguity of open
bunk what?
Why will the father be angry?
Complex chain of reasoning and application of
world knowledge
Ambiguity of father
father as parent
or
father as headmaster

Complexity of Connected Text
John was returning from school
dejected – today was the math test
He couldn’t control the class
Teacher shouldn’t have made him
responsible
After all he is just a janitor

NLP as an ML task
 France beat Brazil by 1 goal to 0 in the
quarter-final of the world cup football
tournament. (English)
 braazil ne phraans ko vishwa kap
phutbal spardhaa ke kwaartaar
phaainal me 1-0 gol ke baraabarii se
haraayaa. (Hindi)

Categories of the Words in the
Sentence
France beat Brazil by 1 goal to 0 in the quarter
final of the world cup football tournament
by
to
in
the
of
Brazil
beat
France
10
goal
quarter final
world cup
Football
tournament
content
words
function
words

Further Classification 1/2
Brazil
beat
France
1
goal
0
quarter final
world cup
football
tournament
Brazil
France
1
goal
0
quarter final
world cup
football
tournament
beat
Brazil
France
1
goal
0
quarter final
world cup
Football
tournament
noun
verb
proper
noun
common
noun

Further Classification 2/2
by
to
In
the
of
determiner preposition
the
by
to
in
of

Why all this?
 information need
 who did what
 to whom
 by what
 when
 where
 in what manner

Semantic roles
patient/theme
beat
France
Brazil
world
cup
football
1 goal to 0
quarter
finals
agent
manner
time
modifier

Semantic Role Labeling: a
classification task
tournament
 Brazil: agent or object?
 Agent: Brazil or France or Quarter Final or
World Cup?
 Given an entity, what role does it play?
 Given a role, it is played by which
entity?

A lower level of classification:
Part of Speech (POS) Tag
Labeling
tournament
 beat: verb of noun (heart beat, e.g.)?
 Final: noun or adjective?

Uncertainty in classification:
Ambiguity
 Visiting aunts can be a nuisance
 Visiting:
 adjective or gerund (POS tag ambiguity)
 Role of aunt:
 agent of visit (aunts are visitors)
 object of visit (aunts are being visited)
 Minimize uncertainty of classification
with cues from the sentence

What cues?
 Position with respect to the verb:
 France to the left of beat and Brazil to the right:
agent-object role marking (English)
 Case marking:
 France ne (Hindi); ne (Marathi): agent role
 Brazil ko (Hindi); laa (Marathi): object role
 Morphology: haraayaa (hindi); haravlaa
(Marathi):
 verb POS tag as indicated by the distinctive
suffixes

Cues are like
attribute-value pairs
prompting machine learning from NL data
 Constituent ML tasks
 Goal: classification or clustering
 Features/attributes (word position, morphology, word
label etc.)
 Values of features
 Training data (corpus: annotated or un-annotated)
 Test data (test corpus)
 Accuracy of decision (precision, recall, F-value, MAP
etc.)
 Test of significance (sample space to generality)

What is the output of an ML-NLP
System (1/2)
 Option 1: A set of rules, e.g.,
 If the word to the left of the verb is a noun and has
animacy feature, then it is the likely agent of the
action denoted by the verb.
 The child broke the toy (child is the agent)
 The window broke (window is not the agent; inanimate)

What is the output of an ML-NLP
System (2/2)
 Option 2: a set of probability values
 P(agent|word is to the left of verb and has
animacy) > P(object|word is to the left of verb and
has animacy)> P(instrument|word is to the left of
verb and has animacy) etc.

How is this different from
classical NLP
 The burden is on the data as opposed
to the human.
corpus
Linguist
Text data
Computer
rules
rules/probabilities
Classical NLP
Statistical NLP

Classification appears as
sequence labeling

A set of Sequence Labeling
Tasks: smaller to larger units
 Words:
 Part of Speech tagging
 Named Entity tagging
 Sense marking
 Phrases: Chunking
 Sentences: Parsing
 Paragraphs: Co-reference annotating

Example of word labeling: POS Tagging
<s>
Come January, and the IIT campus is abuzz with
new and returning students.
</s>
<s>
Come_VB January_NNP ,_, and_CC the_DT
IIT_NNP campus_NN is_VBZ abuzz_JJ with_IN
new_JJ and_CC returning_VBG students_NNS
._.
</s>

Example of word labeling: Named
Entity Tagging
<month_name>
January
</month_name>
<org_name>
IIT
</org_name>

Example of word labeling: Sense
Marking
Word Synset WN-synset-no
come {arrive, get, come} 01947900
.
.
.
abuzz {abuzz, buzzing, droning}
01859419

Example of phrase labeling:
Chunking
the IIT campus
Come July, and is
abuzz with new and returning students
.

Example of Sentence labeling:
Parsing
[S1[S[S[VP[VBCome][NP[NNPJuly]]]]
[,,]
[CC and]
[S [NP [DT the] [JJ IIT] [NN campus]]
[VP [AUX is]
[ADJP [JJ abuzz]
[PP[IN with]
[NP[ADJP [JJ new] [CC and] [ VBG returning]]
[NNS students]]]]]]
[..]]]

Modeling Through the Noisy
Channel
5 problems in NLP

5 Classical Problems in NLP: being
tackled now by statistical approaches
 Part of Speech Tagging
 Statistical Spell Checking
 Automatic Speech Recognition
 Probabilistic Parsing
 Statistical Machine Translation

Problem-1: PoS tagging
Input:
1. sentences (string of words to be
tagged)
2. tagset
Output: single best tag for each word

PoS tagging: Example
Sentence:
The national committee remarked on a number of
other issues.
Tagged output:
The/DET national/ADJ committee/NOU
remarked/VRB on/PRP a/DET number/NOU of/PRP
other/ADJ issues/NOU.

Stochastic Models (Contd..)
t* argmax P(t | w)
t
=
P w t
( , )
( )
P ( t | w ) = P ( t ) P ( w | t )
=
( )
P w
P w
P(w,t)
P(t | w)
Best tag t*,
Bayes Rule gives,
Joint Distribution
Conditional Distribution

Problem 2: Probabilistic Spell Checker
w Noisy Channel
t
(wn, wn-1, … , w1) (tm, tm-1, … , t1)
Given t, find the most probable w : Find that ŵ for which P(w|t) is
maximum, where t, w and ŵ are strings:
ý = P w t
argmax ( | )
w
Correct word Wrongly spelt word
Guess at the
correct word
ŵ

Spell checker: apply Bayes
Rule
ŵý = p w p t w
 Why apply Bayes rule?
argmax ( ( ) . ( | ))
 Finding p(w|t) vs. p(t|w) ?
w
 p(w|t) or p(t|w) have to be computed by counting
c(w,t) or c(t,w) and then normalizing them
 Assumptions :
 t is obtained from w by a single error.
 The words consist of only alphabets

Spell checker: Confusion
Matrix (1/3)
Confusion Matrix: 26x26
 Data structure to store c(a,b)
 Different matrices for insertion, deletion,
substitution and transposition
 Substitution
 The number of instances in which a is wrongly
substituted by b in the training corpus
(denoted sub(x,y) )

Confusion Matrix (2/3)
 Insertion
 The number of times a letter y is inserted after
x wrongly( denoted ins(x,y) )
 Transposition
 The number of times xy is wrongly transposed
to yx ( denoted trans(x,y) )
 Deletion
 The number of times y is deleted wrongly after
x ( denoted del(x,y) )

Confusion Matrix (3/3)
 If x and y are alphabets,
 sub(x,y) = # times y is written for x (substitution)
 ins(x,y) = # times x is written as xy
 del(x,y) = # times xy is written as x
 trans(x,y) = # times xy is written as yx

Spell checking: Example
 Correct document has ws
 Wrong document has ts
 P(maple|aple)=
# (maple was wanted instead of aple) / # (aple)
 P(apple|aple) and P(applet|aple) calculated
similarly
 Leads to problems due to data sparsity.
 Hence, use Bayes rule.

Problem 3: Probabilistic Speech
Recognition
 Problem Definition : Given a sequence of
speech signals, identify the words.
 2 steps :
 Segmentation (Word Boundary Detection)
 Identify the word
 Isolated Word Recognition :
 Identify W given SS (speech signal)
^
W = P W SS
argmax ( | )
W

Speech recognition: Identifying the word
^
W P W SS
argmax ( | )
W
argmax ( ) ( | )
W
P W P SS W
=
=
 P(SS|W) = likelihood called “phonological
model “  intuitively more tractable!
 P(W) = prior probability called “language
model” ( ) # W appears in the corpus
# words in the corpus
P W =

Pronunciation Dictionary
Word Pronunciation Automaton
s4
ae
1.0
0.73
1.0 1.0 1.0 1.0
t o m t
o
aa
end
s1 s2 s3
s5
s6 s7
1.0
0.27
Tomato
 P(SS|W) is maintained in this way.
 P(t o m ae t o |Word is “tomato”) = Product of arc probabilities

Problem 4: Statistical Machine
Translation
Source language Noisy Channel
sentences
Target language
sentences
 What sentence in the target language will
maximise the probability
P(target sentence|source sentence)

Statistical MT: Parallel Texts
 Parallel texts
 Instruction manuals
 Hong Kong
legislation
 Macao legislation
 Canadian
parliament
Hansards
 United nation
reports
 Official journal of
the European
Communities
 Trilingual
documents in
Indian states
 Observation:
Every time I see banco, the
translation is bank or bench
… if it is banco de, then it
always becomes bank and
never bench
Courtsey: a presentation by K. Knight

SMT: formalism
 Source language: F
 Target language: E
 Source language sentence: f
 Target language sentence: e
 Source language word: wf
 Target language word: we

SMT Model
 To translate f:
 Assume that all sentences in E are
translations of f with some probability!
 Choose the translation with the highest
probability
e^ argmax ( p(e | f ))
e
=

SMT: Apply Bayes Rule
e^ argmax( p(e).p( f | e))
e
=
P(e) is called the language model and
stands for fluency
and
P(f|e} is called the translation model and
stands for faithfulness

Reason for Applying Bayes Rule
 The way P(f|e) and P(e|f) are usually
calculated
 Word translation based
 Word order
 Collocations (For example, strong tea)
 Example:
 f: It is raining
 Candidates for e (in Hindi):
 bAriSa Ho raHI HE (rain happening is)
 Ho bAriSa raHI HE (is rain happening)
 bAriSa Ho raHA HE (rain happening_masculine is)

Post-1
 POST----5 TITLE: "Wants to invest in IPO? Think again" | <br /><br
/>Hereâ€™s a sobering thought for those who believe in investing in IPOs.
Listing gains â€” the return on the IPO scrip at the close of listing day
over the allotment price â€” have been falling substantially in the past two
years. Average listing gains have fallen from 38% in 2005 to as low as 2% in the first half of
2007.Of the 159 book-built initial public offerings (IPOs) in India between 2000 and 2007,
two-thirds saw listing gains. However, these gains have eroded sharply in recent
years.Experts say this trend can be attributed to the aggressive pricing strategy that
investment bankers adopt before an IPO. â€&oelig;While the drop in average
listing gains is not a good sign, it could be due to the fact that IPO issue managers are
getting aggressive with pricing of the issues,â€ says Anand Rathi, chief
economist, Sujan Hajra.While the listing gain was 38% in 2005 over 34 issues, it fell to 30%
in 2006 over 61 issues and to 2% in 2007 till mid-April over 34 issues. The overall listing
gain for 159 issues listed since 2000 has been 23%, according to an analysis by Anand
Rathi Securities.Aggressive pricing means the scrip has often been priced at the high end
of the pricing range, which would restrict the upward movement of the stock, leading to
reduced listing gains for the investor. It also tends to suggest investors should not
indiscriminately pump in money into IPOs.But some market experts point out that India
fares better than other countries. â€&oelig;Internationally, there have been
periods of negative returns and low positive returns in India should not be considered a bad
thing.

Post-2
 POST----7TITLE: "[IIM-Jobs] ***** Bank: International Projects Group -
Manager"| <br />Please send your CV & cover letter to
anup.abraham@*****bank.com ***** Bank, through its International Banking
Group (IBG), is expanding beyond the Indian market with an intent to become a
significant player in the global marketplace. The exciting growth in the overseas
markets is driven not only by India linked opportunities, but also by opportunities
of impact that we see as a local player in these overseas markets and / or as a
bank with global footprint. IBG comprises of Retail banking, Corporate banking
& Treasury in 17 overseas markets we are present in. Technology is seen
as key part of the business strategy, and critical to business innovation &
capability scale up. The International Projects Group in IBG takes ownership of
defining & delivering business critical IT projects, and directly impact
business growth. Role: Manager Â– International Projects Group
Purpose of the role: Define IT initiatives and manage IT projects to achieve
business goals. The project domain will be retail, corporate & treasury. The
incumbent will work with teams across functions (including internal technology
teams & IT vendors for development/implementation) and locations to
deliver significant & measurable impact to the business. Location: Mumbai
(Short travel to overseas locations may be needed) Key Deliverables:
Conceptualize IT initiatives, define business requirements

Sentiment Classification
 Positive, negative, neutral – 3 class
 Sports, economics, literature - multi class
 Create a representation for the document
 Classify the representation
The most popular way of representing a
document is feature vector (indicator
sequence).

Established Techniques
 Naïve Bayes Classifier (NBC)
 Support Vector Machines (SVM)
 Neural Networks
 K nearest neighbor classifier
 Latent Semantic Indexing
 Decision Tree ID3
 Concept based indexing

Successful Approaches
The following are successful approaches
as reported in literature.
 NBC – simple to understand and
implement
 SVM – complex, requires foundations of
perceptions

Mathematical Setting
We have training set
A: Positive Sentiment Docs
B: Negative Sentiment Docs
Let the class of positive and negative
documents be C+ and C- , respectively.
Given a new document D label it positive if
P(C+|D) > P(C-|D)
Indicator/feature
vectors to be formed

Priori Probability
Docu
ment
Vector Classi
ficatio
n
D1 V1 +
D2 V2 -
D3 V3 +
.. .. ..
D4000 V4000 -
Let T = Total no of documents
And let |+| = M
So,|-| = T-M
P(D being
positive)=M/T
Priori probability is calculated without
considering any features of the new
document.

Apply Bayes Theorem
Steps followed for the NBC algorithm:
 Calculate Prior Probability of the classes. P(C+ ) and P(C-)
 Calculate feature probabilities of new document. P(D| C+ ) and
P(D| C-)
 Probability of a document D belonging to a class C can be
calculated by Baye’s Theorem as follows:
P(C|D) = P(C) * P(D|C)
P(D)
• Document belongs to C+ , if
P(C+ ) * P(D|C+) > P(C- ) * P(D|C-)

Calculating P(D|C+)
P(D|C+) is the probability of class C+ given D. This is calculated as
follows:
 Identify a set of features/indicators to evaluate a document and
generate a feature vector (VD). VD = <x1 , x2 , x3 … xn >
 Hence, P(D|C+) = P(VD|C+)
= P( <x1 , x2 , x3 … xn > | C+)
= |<x1,x2,x3…..xn>, C+ |
| C+ |
 Based on the assumption that all features are Independently
Identically Distributed (IID)
= P( <x1 , x2 , x3 … xn > | C+ )
= P(x1 |C+) * P(x2 |C+) * P(x3 |C+) *…. P(xn |C+)
=Π i=1 n P(xi |C+)
P(xi |C+) can now be calculated as |xi |

Baseline Accuracy
 Just on Tokens as features, 80%
accuracy
 20% probability of a document being
misclassified
 On large sets this is significant

To improve accuracy…
 Clean corpora
 POS tag
 Concentrate on critical POS tags (e.g.
adjective)
 Remove ‘objective’ sentences ('of'
ones)
 Do aggregation
Use minimal to sophisticated NLP

Natural language procssing

More Related Content

What's hot (20)

Similar to Natural language procssing (20)

Recently uploaded (20)

Natural language procssing