Sentiment…Human Intelligence

advertisement
Sentiment Analysis
Sivaji Bandyopadhyay
Jadavpur University
Kolkata, India
Sentiment…Human Intelligence
Overview
Sentiment Analysis is a multifaceted problem

Sentiment Knowledge Acquisition

Sentiment / Subjectivity Detection

Sentiment Polarity Detection

Sentiment Structurization

Sentiment Summary
Sentiment Knowledge Acquisition
Involving Human Intelligence


Prior Polarity Sentiment Lexicon
Automatic Computational Processes
 WordNet
 Dictionary Based
 Antonym
 Involving Human Intelligence
 Dr Sentiment
 Cross Lingual Projection of Sentiment Lexicons
Sentiment Analysis
IITH is a very good institution.
I love Hyderabad, the city is famous for its Biriyani, Pearl and old
Mughal architecture!
Summer in Hyderabad is too scorching.
Sentiment Analysis
Sentiment Detection
Sentiment Classification
Andrea Esuli and Fabrizio Sebastiani. SentiWordNet: A publicly
available lexical resource for opinion mining. In Proceedings of
Language Resources and Evaluation (LREC), 2006.
4
What is SentiWordNet?

Prior Polarity Lexicon
POS
Offset
Positivity
Negativity
Synset
Adjective
1006361
0.875
0.0
happy
Noun
4466580
0.375
0.0
friendliness
Adverb
214589
0.625
0.125
sharply
Verb
2471993
0.0
0.125
shame
5
Prior Polarity Lexicon
Sentiment Bearing Words:
love, hate, good, favorite
Challenges for Polarity Identification:
Context Information (Pang et al., 2002)
Domain Pragmatic Knowledge (Aue and Gamon, 2005)
Time Dimension (Read, 2005)
Language/Culture Properties (Wiebe and Mihalcea, 2006)
6
Continue….(2)
Prior Polarity Lexicon
Context Information
I prefer Limuzin as it is longer than Mercedes.
Avoid longer baggage during excursion in Amazon.
Language/Culture Properties
सहे रा (Sahera: A marriage-wear of India)
দুর্গ াপু জ া (Durgapujo: A festival of Bengal)
Domain Pragmatic Knowledge
Sensex go high.
Price go high.
Time Dimension
During 90’s mobile phone users generally reported in various
online reviews about their color-phones but in recent times
color-phone is not just enough. People are fascinated and
influenced by touch screen and various software(s) installation
facilities on these new generation gadgets.
7
Continue….(3)
Prior Polarity Lexicon
Suppose total occurrence of a word “long” in a domain corpus is n.
The positive and negative occurrence of that word are Sp and Sn
respectively.
Therefore in a developed sentiment lexicon the assigned positivity
and negativity score of that word will be as follows:
Positivity
Negativity
:
:
Sp /n
Sn /n
These associative positive and negative scores are called prior
polarity.
8
Source Lexicon Acquisition
Available Resources for English
 SentiWordNet (Esuli et. al., 2006)

SentiWordNet is an automatically constructed lexical
resource for English that assigns a positivity score and a
negativity score to each WordNet synset.
 WordNet Affect List (Strapparava et al., 2004)

WordNet synsets tagged with six basic emotions:
anger, disgust, fear, joy, sadness, surprise.
 Taboada’s Adjective List (Voll et al., 2006)

An automatically constructed adjective list with
positivity and negativity polarity assignment.
 Subjectivity Word List (Wilson et. al., 2005)

The entries in the subjectivity word list have been
manually labeled with part of speech (POS) tags as well
as either strong or weak subjective tag depending on
the reliability of the subjective nature of the entry.
9
Continue….(1)
Source Language Acquisition
 Chosen Source Lexicon Resources
 SentiWordNet
 SentiWordNet is most widely used in several applications
such as sentiment analysis, opinion mining and emotion
analysis.
 Subjectivity Word List (Wilson et. al., 2005)
 Subjectivity Word List is most trustable as the opinion mining
system OpinionFinder that uses the subjectivity word list
has reported highest score for opinion/sentiment
subjectivity (Wiebe and Riloff, 2006) (Das and
Bandyopadhyay, 2010)
10
Continue….(2)
Source Language Acquisition
 Noise-Reduction
 A merged sentiment lexicon has been developed from both
the resources by removing the duplicates.
 It has been observed that 64% of the single word entries
are common in the Subjectivity Word List and SentiWordNet.
 The new merged sentiment lexicon consists of 14,135
numbers of tokens.
 Several filtering techniques have been applied to generate
the new list.
11
Continue….(3)
Source Language Acquisition
SentiWordNet
Unambiguous
Words
Discarded
Ambiguous
Words
Subjectivity Word List
Single
Multi
Single
Multi
115424
79091
5866
990
20789
30000
4745
963
Threshold
Orientation
Strength
Subjectivity
Strength
POS
86944
30000
2652
928
12
Target Language Generation
Generation Strategies
Bilingual Dictionary Based Approach
WordNet Based Approach
Antonym Generation
Corpus Based Approach
Dr Sentiment (A Gaming Approach)
13
Continue….(1)
Target Language Generation
Bilingual Dictionary Based Approach
 A word-level translation technique adopted.
 Robust and reliable synsets (approx 9966) are created by
native speakers as well as linguistics experts of the specific
languages as a part of English to Indian Languages Machine
Translation Systems (EILMT).
 Various language specific dictionaries acquired.
14
Continue….(2)
Target Language Generation
Bilingual Dictionary Based Approach
Hindi (90,872)
SHABDKOSH (http://www.shabdkosh.com/)
Shabdanjali
(http://www.shabdkosh.com/content/category/download
s/)
Bengali (102119)
Samsad Bengali-English Dictionary
(http://dsal.uchicago.edu/dictionaries/biswas_bengali/)
Telugu (112310)
Charles Philip Brown English-Telugu Dictionary
(http://dsal.uchicago.edu/dictionaries/brown/)
Aksharamala English-Telugu Dictionary
(https://groups.google.com/group/aksharamala)
English-Telugu Dictionary
(http://ltrc.iiit.ac.in/onlineServices/Dictionaries/Dict_Fr
ame.html)
15
Continue….(3)
Target Language Generation
Bilingual Dictionary Based Approach
Hindi
 Translation process has resulted 22,708 Hindi entries
Bengali
 Translation process has resulted 34,117 Bengali
entries
Telugu
 Translation process has resulted 30,889 Telugu entries
 Almost 88% Telugu SentiWordNet generated by this
process
16
Continue….(4)
Target Language Generation
WordNet Based Expansion Approach
Synonymy Expansion
 WordNet based expansion technique produces more synset
members: inactive, motionless, static for a source word still.
 Prior polarity scores are directly copied
Antonymy Expansion
 WordNet based expansion technique produces more sentiment
lexemes: ugly for a source word beautiful.
 Prior polarities are calculated as:
Tp=1-Sp
Tn=1-Sn
where Sp, Sn are the positivity and negativity score for the source
language (i.e, English) and Tp, Tn are the positivity and negativity
score for target languages
17
Continue….(5)
Target Language Generation
WordNet Based Expansion Approach
Hindi
 Hindi WordNet (Jha et al., 2001)
(http://www.cfilt.iitb.ac.in/wordnet/webhwn/) is a well
structured and manually compiled resource and is being
updated since last nine years.
 Almost 60% generated by this process
Bengali
 The Bengali (http://bn.asianwordnet.org/)
 It only contains 1775 noun synsets as reported in (Robkop
et al., 2010)
 Only 5% new lexicon entries have been generated in this
process
18
Continue….(6)
Target Language Generation
Antonymy Generation
Affix/Suffix
abX
misX
imX-exX
antiX
nonX
inX-exX
disX
unX
upX-downX
imX
illX
overX-underX
inX
rX-irX
Xless-Xful
malX
Word
Normal
Fortune
Im-plicit
Clockwise
Aligned
In-trovert
Interest
Biased
Up-hill
Possible
Legal
Overdone
Consistent
Regular
Harm-less
Function
Antonym
Ab-normal
Mis-fortune
Ex-plicit
Anti-clockwise
Non-aligned
Ex-trovert
Dis-interest
Un-biased
Down-hill
Im-possible
Il-legal
Under-done
In-consistent
Ir-regular
Harm-ful
Mal-function
About 8% of Bengali, 7% of Hindi and 11% of Telugu
SentiWordNet entries are generated in this process.
19
Continue….(7)
Target Language Generation
Corpus Based Approach
Language/culture specific words:
सहे रा (Sahera: A marriage-wear)
দুর্গ াপু জ া (Durgapujo: A festival of Bengal)
Technique
Generated sentiment Lexicon used a seed list
Tag-Set
SWP (Sentiment Word Positive)
SWN (Sentiment Word Negative)
Corpus
EILMT language specific corpus: approximately 10K
of sentences.
Model
Conditional Random Field (CRF)
An n-gram (n=4) sequence labeling model has been
used for the present task.
20
Limitations
Issues in Cross Lingual Projection

Sentiment score may not be equal to source language

Relative sentiment score is needed rather than absolute score

Language / Culture specific lexicons should be included

Sentiment score should be updated by time
21
Involving Human
Intelligence
WORLD INTERNET USAGE AND POPULATION STATISTICS
World Regions
Population
( 2010 Est.)
Africa
1,013,779,050
4,514,400
110,931,700
10.9 %
2,357.3 %
5.6 %
Asia
3,834,792,852
114,304,000
825,094,396
21.5 %
621.8 %
42.0 %
Europe
813,319,511
105,096,093
475,069,448
58.4 %
352.0 %
24.2 %
Middle East
212,336,924
3,284,800
63,240,946
29.8 %
1,825.3 %
3.2 %
North America
344,124,450
108,096,800
266,224,500
77.4 %
146.3 %
13.5 %
Latin
America/Caribbean
592,556,972
18,068,919
204,689,836
34.5 %
1,032.8 %
10.4 %
Oceania / Australia
34,700,201
7,620,480
21,263,990
61.3 %
179.0 %
1.1 %
6,845,609,960
360,985,492
1,966,514,816
28.7 %
444.8 %
100.0 %
WORLD TOTAL
Internet Users
Dec. 31, 2000
Internet Users
Latest Data
Penetration
(% Population)
Growth
2000-2010
Users %
of Table
22
Dr. Sentiment
Q1
23
Dr. Sentiment
Q2
Word
Positivity
Negativity
Good
0.625
0.0
Better
0.875
0.0
Best
0.980
0.0
24
Dr. Sentiment
Q3
25
Dr. Sentiment
Q4
26
Sentiment…Un-Explored
Dimensions
Geo-Spatial
Blue in Islam: In verse 20:102 of the Qur’an, the word ‫زرق‬
zurq (plural of azraq 'blue') is used metaphorically for evil
doers whose eyes are glazed with fear
27
Sentiment…Un-Explored
Dimensions
Age-Wise Senti-Mentality
28
Sentiment…Un-Explored
Dimensions
Gender-Specific Senti-Mentality
29
Expected Impact of the
Resources
Resources are useful in multiple aspect
Mono-Lingual Sentiment/Opinion/Emotion Analysis task
Generated language specific SentiWordNet(s) could be
expanded by other proposed methods (Dictionary, WordNet,
Antonym and Corpus Based Approach)
The other dimensions
Geospatial Information retrieval
Personalized search
Recommender System etc
Stylometry: A writer’s Senti-Mentality
Plagiarism: Spamming Technique: Geo-Spatial and User
Perspective
30
The Road Ahead

Basic SentiWordNet has been developed for 56 languages
Languages
Afrikaans
Bulgarian
Dutch
German
Irish
Malay
Russian
Thai
Albanian
Catalan
Estonian
Greek
Italian
Maltese
Serbian
Turkish
Arabic
Chinese
Filipino
Haitian
Japanese
Norwegian
Slovak
Ukrainian
Armenian
Croatian
Finnish
Hebrew
Korean
Persian
Slovenian
Urdu
Azerbaijani
Creole
French
Hungarian
Latvian
Polish
Spanish
Vietnamese
Basque
Czech
Galician
Icelandic
Lithuanian
Portuguese
Swahili
Welsh
Belarusian
Danish
Georgian
Indonesian
Macedonian
Romanian
Swedish
Yiddish
A. Das and S. Bandyopadhyay. Towards The Global SentiWordNet, In the
Workshop on Model and Measurement of Meaning (M3), PACLIC 24, November 4,
Sendai, Japan, 2010. (Accepted)
31
References

Resources
I. A. Das and S. Bandyopadhyay. Towards The Global SentiWordNet, In the
Workshop on Model and Measurement of Meaning (M3), PACLIC 24,
November 4, Sendai, Japan, 2010.
II. A. Das and S. Bandyopadhyay. SentiWordNet for Indian Languages, In
the 8th Workshop on Asian Language Resources (ALR), August 21-22,
Beijing, China, 2010.
III. A. Das and S. Bandyopadhyay. SentiWordNet for Bangla, In Knowledge
Sharing Event-4: Task 2: Building Electronic Dictionary , February 23rd-24th,
2010, Mysore.
32
Sentiment / Subjectivity Detection

Solution Architecture Explored
 Rule-Based
 Machine Learning
 Hybrid
 Adaptive Genetic Algorithm: Multiple Objective
Optimization, The Evolutionary Technique to Detect
Sentiment
 Adaptive Genetic Algorithm: Multiple Objective Optimization
technique yielded all other techniques

Sentence subjectivity: An objective sentence expresses some factual
information about the world, while a subjective sentence expresses some
personal feelings or beliefs.
Example: Type: Film Review, Film Name: Deep Blue Sea, Holder:
Arbitrary-outside of theatre
Oh, This is blue!
Is this statement an objective or subjective statement?
• “blue” is not a evaluative expression
• Among different cultures with a different colour scheme (blue;
positive or negative?)
Example: Type: Comment, Holder: Governor of WB, Issue:
Nandigram.
Governor said the government should keep
patience.
Is this statement an objective or subjective statement?
• “keep patience” regarding what?
• How to determine Governor’s comment is important?
•
Subjectivity is a social norms
•
Subjectivity knowledge is pragmatic
•
A prior knowledge always help to
identify Subjectivity
•
A rule-based approach
Use Themes and Ontology as pragmatic
knowledge
SentiWordNet (Bengali): a prior polarity lexicon
•
Features
•
•
•
•
•
•
•
•
•
•
•
•
Frequency
Average Distribution
Functional Word
Positional Aspect
Theme Identification
Ontology List
Stemming Cluster
Part of Speech
Chunk
SentiWordNet (Bengali)
Features
Overall
Performance
incremented by
80
Stemming
Cluster
4.05%
Part of
Speech
3.62%
Chunk
4.07%
30
Functional
Word
1.88%
20
SentiWordNet
(Bengali)
5.02%
Ontology List
3.66%
70
60
Base-Line
50
POS-Chunk
40
Ontology
Position
Distribution
10
0
English
Bengali
Feature wise System Performance
Domain
Precision
Recall
NEWS
72.16%
76.00%
BLOG
74.60%
80.40%
Overall System Performance
Observations
• Subjectivity detection is trivial for blog corpus rather than for news corpus
• Performance incremented by 2% only from rule-based system using CRF
technique with the same feature set
GBML used to identify automatically best feature set based on the principle of
natural selection and survival of the fittest.
The identified fittest feature set is then optimized locally and global optimization
is then obtained by multi-objective optimization technique.
The local optimization identify the best range of feature values of a particular
feature.
The Global optimization technique identifies the best ranges of values of given
multiple feature.
Types
Lexico-Syntactic
Syntactic
Discourse Level
Features
POS
SentiWordNet
Frequency
Stemming
Chunk Label
Dependency Parsing
Title of the Document
First Paragraph
Average Distribution
Theme Word
Experimentally Best Identified Feature Set
GAs are characterized by the five basic components as follows
I. Chromosome representation for the feasible solutions to the optimization
problem.
II. Initial population of the feasible solutions.
III. A fitness function that evaluates each solution.
IV. Genetic operators that generate a new population from the existing
population.
V. Control parameters such as population size, probability of genetic
operators, number of generation etc.
f s   i 0 fi
N
Where f s is the resultant subjectivity f i function, to be calculated and is the
ith feature function. If the present model is represented in a vector space
model then the above function could be re-written as:




f s  fi . fi 1 . fi  2 ...... f n
This equation specifies what is known as the dot product between vectors.
The GBML provides the facility to search in the Pareto-optimal set of
possible features.
 x  p y   i  xi  yi   i  xi  yi 
To make the Pareto optimality mathematically more rigorous, we state that
a feature vector x is partially less than feature vector y, symbolically x<p y,
when the following condition hold.
Imperialism/NNP is/VBZ the/DT source/NN of/IN war/NN and/CC the/DT
disturber/NN of/IN peace/NN.
NNP
1
VBZ
12
DT
6
NN
2
Features
POS
SentiWordNet
Frequency
Stemming
Chunk Label
Dependency Parsing
Title of the Document
First Paragraph
Average Distribution
Theme Word
IN
18
NN
2
CC
4
DT
6
Real-Values
1-21 (Bengali)/1-45 (English)
-1 to +1
0 or 1
1 to 17176/ 1 to 1235
1-11 (Bengali) / 1-21 (English)
1-30 (Bengali) / 1-55 (English)
Varies document wise
Varies document wise
Varies document wise
Varies document wise
NN
2
IN
18
NN
2
Crossover
The intuition behind crossover is the exploration of new solutions and
exploitation of old solutions. GAs construct a better solution by mixing the
good characteristic of chromosomes together.
Mutation
Mutation involves the modification of the values of each gene of a solution with some
probability (mutation probability).
For example, in the following chromosome a random mutation occurs at position 10.
Result:
101111110111101
101111110011101
The following cost-to-fitness transformation is commonly used with GAs.
f  x   Cmax  g  x  when g  x   Cmax
= 0 Otherwise
There are variety of ways to choose the coefficient . may be taken as an
input coefficient, as the largest g value observed thus far, as the largest g
value in the current population, or the largest of the last k generation.
There is some problems with negative utility function as in the particular case it
occurs during the fitness calculation of n number of features fitness evaluation. To
overcome this, we simply transform fitness according to the equation:
f  x   u  x   Cmin When u  x   Cmin  0
= 0 Otherwise
Languages
English
Bengali
Domain
Precision
Recall
MPQA
90.22%
96.01%
IMDB
93.00%
98.55%
NEWS
87.65%
89.06%
BLOG
90.6%
92.40%
Experimental Result of Subjectivity Detection by Genetic Algorithm
References

Subjectivity
I. A. Das and S. Bandyopadhyay. Subjectivity Detection using Genetic
Algorithm. In the 1st Workshop on Computational Approaches to
Subjectivity and Sentiment Analysis (WASSA10), Lisbon, Portugal, August
16-20, 2010.
II. A.Das and S. Bandyopadhyay. Subjectivity Detection in English and
Bengali: A CRF‐based Approach, In Proceeding of ICON 2009, December
14th-17th, 2009, Hyderabad.
III. A. Das and S. Bandyopadhyay. Theme Detection an Exploration of
Opinion Subjectivity, In Proceeding of Affective Computing & Intelligent
Interaction (ACII) 2009.
IV. A. Das and S. Bandyopadhyay. Extracting Opinion Statements from
Bengali Text Documents through Theme Detection, In Proceeding of 17th
International Conference on Computing (CIC-09), Mexico City, Mexico.
48
Sentiment Polarity Detection:
Inspired by Structural Human
Intelligence

Explored Solution Architectures
 Rule-Based
 Machine Learning
 Structural
Observations
• Sentence level polarity is simply integration of
phrase level polarity.
• Word level polarity depends on four POS
categories.
• Polarity mainly depends on syntactical modifiermodified relations.
• Dependency relations-vital clue.
• Negative word-direct clue.
Identifies polarity at phrase level
Using Support Vector Machine (SVM)
Features
• Part Of Speech (POS)
• Chunk label
• Functional Word
• SentiWordNet (Bengali) orientation
• Stemming Cluster
• Negative Word Feature
• Dependency Tree feature
•
Solution Architecture
 Rule Based
 Statistical
 Malt
 CRF (Conditional Random Field)
 Hybrid
 Statistical + Post-process
Features
• Part Of Speech (POS)
• Chunk label
• Root Word
• Vibhakti
• Morph Features
• Animate-Inanimate gazetteer
• Verb argument structure
Dependency Tree Feature
Accuracy
Precision
Recall
Overall
70.04%
63.02%
Positive
56.59%
52.89%
Negative
75.57%
65.87%
Results on Polarity Identification
References

Polarity
I. A. Das and S. Bandyopadhyay. Phrase-level Polarity Identification for
Bengali, In International Journal of Computational Linguistics and
Applications (IJCLA), Vol. 1, No. 1-2, Jan-Dec 2010, ISSN 0976-0962,
Pages 169-182.
II. A. Das and S. Bandyopadhyay. Opinion-Polarity Identification in Bengali,
In ICCPOL 2010, California, USA.
55
Sentiment Structurization:
Rational Human Intelligence

Proposed Generalized Structurization using 5W
Who, What, When, Where, Why
Overview
 SRL
Contemporary Theories
 Motivation
Paninian Karaka Theory
Fillmore’s Case Grammar
 The
Proposed Concept of 5W
 Resource Acquisition
 Challenges
 System Overview
Statistical Model – MEMM
Rule Based Post Processing
57
SRL Contemporary Theories
 SRL
extensively studied for English
 SRL has not been studied for Indian
Languages
Hindi SRL-Genitive Marker (Sharma at al., 2009)
 Domain
Specific
FROM_DESTINATION
TO_DESTINATION
DEPARTURE_TIME etc.
 Verb


Specific
EATER
EATEN
58
Continue….(1)
SRL Contemporary Theories
 PropBank
(Approx 11K Semantic Role Labels)
Verb Specific Semantic Roles
 Agent, patient or theme etc.
More general semantic roles
 FrameNet
(Approx 7K Semantic Role Labels)
Verb frame specific semantics roles
Frame-to-frame semantic relations
Inheritance, Perspective_on, Subframe, Precedes,
Inchoative_of, Causative_of and Using.
 VerbNet
(Approx 23 Semantic Role Labels)
Thematic roles specific semantic relationship between
a predicate and its arguments.
agent, patient, theme (From PropBank), experiencer,
stimulus, instrument, location, source, goal, recipient,
benefactive etc.
59
Continue….(2)
SRL Contemporary Theories
 Conclusion
No Adequate semantic role Labels exists!
Across various domains
Across various languages
60
Overview
 SRL
Contemporary Theories
 Motivation
Paninian Karaka Theory
Fillmore’s Case Grammar
 The
Proposed Concept of 5W
 Resource Acquisition
 Challenges
 System Overview
Statistical Model – MEMM
Rule Based Post Processing
61
Motivation
 Historical
Panini’s Karaka Theory
First Grammarian
300-600 BC Astadhayi
Still Influential
 Syntactic-Semantic
Relationship
Karta: central to the action of the verb
Karma: the one most desired by the karta
Karana: instrument, essential for the action to
take place
Sampradaan: recipient of the action
Apaadaan: movement away from a source
Adhikarana: location of the action
62
Continue….(1)
Motivation
 Fillmore’s Case Grammar
 He posited the following preliminary
list of cases,
noting however that ‘Additional cases will surely be
needed’
Agent: The typically animate perceived instigator of the
action.
Instrument: Inanimate force or object causally involved in
the action or state.
Dative: The animate being affected by the state or action.
Factitive: The object or being resulting from the action or
state.
Locative: The location or time-spatial orientation of the state
or action.
Objective: The semantically most neutral case conceivably
the concept should be limited to things which are affected by
the action or state.
63
Overview
 SRL
Contemporary Theories
 Motivation
Paninian Karaka Theory
Fillmore’s Case Grammar
 The
Proposed Concept of 5W
 Resource Acquisition
 Challenges
 System Overview
Statistical Model – MEMM
Rule Based Post Processing
64
The Proposed Concept of 5W
The 5W task seeks to extract the semantic
information of nouns in a natural language sentence
by distilling it into the answers to the 5W questions:
Who, What, When, Where and Why.
Who? Who was involved?
What? What happened?
When? When did it take place?
Where? Where did it take place?
Why? Why did it happen?
65
Overview
 SRL
Contemporary Theories
 Motivation
Paninian Karaka Theory
Fillmore’s Case Grammar
 The
Proposed Concept of 5W
 Resource Acquisition
 Challenges
 System Overview
Statistical Model – MEMM
Rule Based Post Processing
66
Continue….(1)
Resource Acquisition
 Annotation
Madhabilata (was keeping) her (wrist watch) then (on the table) as she (was about
to sleep).
What
When
Why
Who
Where
মাধবীলতা (শ াবব ববল) তখন (হাবতর ঘড়ি) খু বল শেড়ববল রাখড়িল.
67
Continue….(2)
Resource Acquisition
Annotation
Tag
Annotators X and Y Agree percentage
Who
What
When
Where
Why
88.45%
64.66%
76.45%
75.23%
56.23%
68
Overview
 SRL
Contemporary Theories
 Motivation
Paninian Karaka Theory
Fillmore’s Case Grammar
 The
Proposed Concept of 5W
 Resource Acquisition
 Challenges
 System Overview
Statistical Model – MEMM
Rule Based Post Processing
69
Challenges
 Irregular
Occurrences
Tags
Who
What
When
Where
Why
Percentage
What
When
Where
Why
Overall
58.56%
73.34%
78.01%
28.33%
73.50%
Who
When
Where
Why
Overall
58.56%
62.89%
70.63%
64.91%
64.23%
Who
What
Where
Why
Overall
73.34%
62.89%
48.63%
23.66%
57.23%
Who
What
When
Why
Overall
78.0%
70.63%
48.63%
12.02%
68.65%
Who
What
When
Where
Overall
28.33%
64.91%
23.66%
12.02%
32.00%
70
Overview
 SRL
Contemporary Theories
 Motivation
Paninian Karaka Theory
Fillmore’s Case Grammar
 The
Proposed Concept of 5W
 Resource Acquisition
 Challenges
 System Overview
Statistical Model – MEMM
Rule Based Post Processing
71
Proposed System
 Machine
Learning
Maximum Entropy (MEMM)
CONLL 2005 SRL Shared Task
8 system used MEMM among 19
First two highest performing system used MEMM
 Rule-Based
Post Processing
Heterogeneous problem structure
Label-biased problem of ML techniques
Bengali is Morpho-syntactically rich
Avoid the limitations of available tools
72
Proposed System
 Machine
Learning
Feature Engineering
Features
Lexico-Syntactic
POS
Verb
Morphological
Noun
Types
Syntactic
Root Word
Gender
Number
Peson
Case
Root Word
Modality
Head Noun
Chunk Label
Dependency Relation
73
Experimental Result


Machine Learning (1)
Machine Learning + Rule Based (2)
Tag
Who
What
When
Where
Why
Systems Precision
Recall
F-measure
1
2
1
76.23%
79.56%
61.23%
64.33%
72.62%
51.34%
69.77%
75.93%
55.85%
2
65.45%
59.64%
62.41%
1
2
69.23%
73.35%
58.56%
65.96%
63.44%
69.45%
1
70.01%
60.00%
64.61%
2
1
77.66%
61.45%
69.66%
53.87%
73.44%
57.41%
2
63.50%
55.56%
59.26%
Avg F-Measure
1
62.22%
2
68.10%
74
References

Structurization
I. A. Das and S. Bandyopadhyay. Customized Opinion Summarization and
Visualization: 5Ws Dimensions, In the International Workshop on Topic
Feature Discovery and Opinion Mining (TFDOM 2010) Joint with the 10th
IEEE International Conference on Data Mining (ICDM10);, 13 December
2010, Sydney, Australia.
II. A. Das, A. Ghosh and S. Bandyopadhyay. Semantic Role Labeling for
Bengali Noun using 5Ws, In the International Conference on Natural
Language Processing and Knowledge Engineering (IEEE NLP-KE2010),
August 21-23, Beijing, China, 2010.
75
Sentiment Summary: Can a
Machine Aggregate Sentiment
as Human Intelligence?

Sentiment Summarization: the problem
 Single Doc?
 Multi-Doc?
 Extractive?
 Generative?
 Proper output format?

Explored Solution Architectures
 Topic-Opinion multi-doc Summary
 Visualization
 Tracking
The Topic-Opinion Summarizer
The present system follows a topic-sentiment model for sentiment
identification and aggregation. Topic-sentiment model is designed as
discourse level theme identification and the topic-sentiment
aggregation is achieved by theme clustering (k-means) and Document
level Theme Relational Graph representation. The Document Level
Theme Relational Graph is finally used for candidate summary
sentence selection by standard page rank algorithms used in
Information Retrieval (IR).
Annotation
 The agreement drops fast as the number of annotators increases

Inter-annotator agreement is better for theme words annotation rather than
candidate sentence identification for summary
 Discussion with annotators reveals that the psychology of annotators is to grasp as
many as possible theme words identification during annotation but the same groups of
annotators are more cautious during sentence identification for summary as they are very
conscious to find out the most concise set of sentences that best describe the opinionated snapshot
of any document.
Annotators X vs. Y X Vs. Z Y Vs. Z Avg
Percentage
82.64% 71.78% 80.47% 78.30%
All Agree
69.06%
Agreement of Annotators at Theme Words Level
Annotators
Percentage
All Agree
X vs. Y X Vs. Z Y Vs. Z Avg
73.87% 69.06% 60.44% 67.8%
58.66%
Agreement of Annotators at Sentence Level
Theme Identification
The Theme detection technique has been proposed to identify discourse level
most relevant topic-semantic nodes in terms of word or expressions using a
standard machine learning technique. The machine learning technique used
here is Conditional Random Field (CRF). The theme word detection is
defined as a sequence labeling problem. Depending upon the series of input
feature, each word is tagged as either Theme Word (TW) or Other (O).
Types
Lexico-Syntactic
Syntactic
Discourse Level
Features
POS
SentiWordNet
Frequency
Stemming
Chunk Label
Dependency Parsing Depth
Title of the Document
First Paragraph
Term Distribution
Collocation
Theme Clustering
The cluster hypothesis (Jardine and van Rijsbergen, 1971)
 A reasonable cluster is defined as the one that maximizes the withincluster document similarity and minimizes between-cluster similarities.
 Standard bottom-up soft k-means clustering technique
 Document represented as a theme matrix
cricket
hotel 
 election
A   parliament sachin vacation 
 governor soccer tourist 
   
s  qk , d j   qk . d j   wi ,k  wi , j  (1)


i 1
N


s  qk , d j  




 w w
 w   w
N
i 1
i ,k
i, j
N
2
N
2
i 1
i ,k
i 1
i ,k
 (2)
Continue……(1)
Theme Clustering
ID
1
1
1
1
2
2
2
3
3
3
3
3
3
3
Themes
প্র াসন (administration)
সু াসন (good-government)
সমাজ (Society)
আইন (Law)
গববষণা (Research)
কবলজ (College)
উচ্চড় ক্ষা (Higher Study)
শজহাড়ি (Jehadi)
মসড়জি (Mosque)
মু ারফ (Musharaf)
কাশ্মীর (Kashmir)
পাড়কস্তান (Pakistan)
নয়াড়িল্লী (New Delhi)
বর্ডার (Border)
1
0.63
0.58
0.58
0.55
0.11
0.15
0.12
0.13
0.05
0.05
0.03
0.06
0.12
0.08
2
0.12
0.11
0.12
0.14
0.59
0.55
0.66
0.05
0.01
0.01
0.01
0.02
0.04
0.03
3
0.04
0.06
0.03
0.08
0.02
0.01
0.01
0.58
0.86
0.86
0.93
0.82
0.65
0.79
Document Level Theme Relational Graph
 Document Graph G=<V,E> from a given source document .
 The input document d is parsed and split into a number of text fragments
(sentence) using sentence delimiters (Bengali sentence marker “।“, “?” or “!”).
 Cosine similarity for inter-document similarity
Summarization System
 Extractive opinion summarization system
 Extraction based on sentence importance in representing the shared
subtopic (cluster) is an important issue and it regulates the quality of the
output summary.
 Adaptive page rank algorithm (Page et al., 1998)has been used for
sentence selection among documents in the same cluster.
 The summation of edge score reflects the correlation measure between
two nodes
 Sentences are presented in original sequence for coherent summary
Candidate Sentence
মহম্মি আড়মবনর মবতা পড়লেবু ুবরার 'নবীনতম' সিসুবকও ড়কন্তু
বয়বসর ড়িক হইবত নবীন ভাবা কঠিন।
এবার ড়িন্তা আরওএকেু শবড় , কারণ এই মূ লুবৃ ড়ির ড়পিবন শেমন
শিব র ড়ভতবর ড়জড়নসপবের শজাগান কবম োওয়া আবি, শতমনই
আবি আন্তজডাড়তক বাজাবর মূ লুবৃ ড়ির প্রবণতা।
স্বাধীনতার পর ষাে বির গত হইল, এখনও প্রায় সকল সরকাড়র
পড়রকল্পনার ড়পিবন এই একটিই ভাবাি ড কাজ কবর: ড়বড়ভন্ন
শভােবুাঙ্কবক তুষ্ট কড়রয়া শেন শতন প্রকাবরণ ড়নবজবির িলীয় ড়িড়ত
ড়নড়িত করা।
IR Score
151
167
130
Summarization System
 Extractive opinion summarization system
 Extraction based on sentence importance in representing the shared
subtopic (cluster) is an important issue and it regulates the quality of the
output summary.
 Adaptive page rank algorithm (Page et al., 1998)has been used for
sentence selection among documents in the same cluster.
 The summation of edge score reflects the correlation measure between
two nodes
 Sentences are presented in original sequence for coherent summary
Candidate Sentence
মহম্মি আড়মবনর মবতা পড়লেবু ুবরার 'নবীনতম' সিসুবকও ড়কন্তু
বয়বসর ড়িক হইবত নবীন ভাবা কঠিন।
এবার ড়িন্তা আরওএকেু শবড় , কারণ এই মূ লুবৃ ড়ির ড়পিবন শেমন
শিব র ড়ভতবর ড়জড়নসপবের শজাগান কবম োওয়া আবি, শতমনই
আবি আন্তজডাড়তক বাজাবর মূ লুবৃ ড়ির প্রবণতা।
স্বাধীনতার পর ষাে বির গত হইল, এখনও প্রায় সকল সরকাড়র
পড়রকল্পনার ড়পিবন এই একটিই ভাবাি ড কাজ কবর: ড়বড়ভন্ন
শভােবুাঙ্কবক তুষ্ট কড়রয়া শেন শতন প্রকাবরণ ড়নবজবির িলীয় ড়িড়ত
ড়নড়িত করা।
IR Score
151
167
130
Evaluation
Evaluation of Theme Detection
Theme
Detection
Metrics
Precision
X
Y
Z
87.65% 85.06% 78.06%
Avg
83.60%
Recall
80.78% 76.06% 72.46%
76.44%
F-Score
84.07% 80.30% 75.16%
79.85%
Summarization
Evaluation on Summarization
Metrics
X
Y
Z
Avg
Precision
77.65%
67.22%
71.57% 72.15%
Recall
68.76%
64.53%
68.68% 67.32%
F-Score
72.94%
65.85%
70.10% 69.65%
Sentiment Visualization and Tracking
 Semantic Document Clustering
   
s  dk , d j   dk . d j   wi ,k  wi , j


i 1
N
Generated Clusters
5Ws
5W Opinion Constituents
Mamata Banerjee
CM
Gyaneswari Express
What
Derailment
24th May 2010
When
Misdnight
Jhargram
Where
Khemasoli
Maoist
Why
Bomb Blast
Who
Doc1
Doc2
Doc3 Doc4
Doc5
0.63
0.00
0.98
0.98
0.94
0.68
0.76
0.87
0.78
0.13
0.01
0.12
0.79
0.76
0.01
0.78
0.25
0.01
0.89
0.78
0.55
0.37
0.58
0.35
0.01
0.01
0.01
0.01
0.06
0.01
0.02
0.17
0.36
0.15
0.01
0.01
0.76
0.01
0.14
0.78
0.93
0.10
0.47
0.23
0.01
0.01
0.13
0.01
0.10
0.01
Dimension Wise Opinion Summary
and Visualization
 Real life users do not always require overall summary or visualization,
rather they seek for opinion changes of any “Who” during “When” and
depending upon “What” or “Where” and “Why”.
 A market surveyor from company A may have a need to find out the
changes in public opinion about their product X after release of product Y by
company B. An overall opinion summary on product X may miss some
valuable information as the market surveyor need. Another example: A voter
may be interested to find out the rate of change of public opinion about any
leader or any public event before and after of any election.
Dimension Wise Opinion Summary
and Visualization
A Snapshot of the Visualize-Tracking System
Evaluation
Tags
Who
What
When
Where
Why
Average Scores
What
When
Where
Why
Overall
3.20
3.30
3.30
2.50
3.08
Who
When
Where
Why
Overall
3.20
3.33
3.80
2.6
3.23
Who
What
Where
Why
Overall
3.30
3.33
2.0
2.5
3.00
Who
What
When
Why
Overall
3.30
3.80
2.0
2.0
2.77
Who
What
When
Where
Overall
2.50
2.6
2.5
2.0
2.40
References

Summarization, Visualization and Tracking
I. A. Das and S. Bandyopadhyay. Event-Sentiment Visual Tracking, In the
FALA 2010 "VI Jornadas en Tecnologia del Habla" , November 10-12, Vigo,
Spain, 2010. .
II. A. Das and S. Bandyopadhyay. Opinion Summarization in Bengali: A
Theme Network Model, In the Second IEEE International Conference on
Social Computing (SocialCom-2010), Minneapolis, USA, August 20-22,
2010.
III. A. Das and S. Bandyopadhyay. Topic-Based Bengali Opinion
Summarization, In the 23rd International Conference on Computational
Linguistics (COLING 2010), August 23-27, 2010, Beijing, China.
90
Conclusion
 Sentiment beyond Human Intelligence.
 Sentiment Innate Human Intelligence.
 Sentiment: the Psychology of Human Intelligence.
Thank You
Download
Related flashcards

Security

27 cards

Surgery

42 cards

Orthopedics

21 cards

Hygiene

25 cards

Medical treatments

33 cards

Create Flashcards