A Survey on Sentiment Analysis and Opinion Mining.pdf

Figure 1(a). Visualization of summary of opinions on a car
Figure 1(b). Visual comparison of two cars.
A Survey on Sentiment Analysis and Opinion
Mining Techniques
Amandeep Kaur
University Institute of Engineering and Technology, Chandigarh, India
Email: amandeepk.cql@gmail.com
Vishal Gupta
University Institute of Engineering and Technology, Chandigarh, India
Email: vishal@pu.ac.in
Abstract—Sentiment Analysis (SA), an application of
Natural Language processing (NLP), has been witnessed a
blooming interest over the past decade. It is also known as
opinion mining, mood extraction and emotion analysis. The
basic in opinion mining is classifying the polarity of text in
terms of positive (good), negative (bad) or neutral (surprise).
Mood Extraction automates the decision making performed
by human. It is the important aspect for capturing public
opinion about product preferences, marketing campaigns,
political movements, social events and company strategies.
In addition to sentiment analysis for English and other
European languages, this task is applied on various Indian
languages like Bengali, Hindi, Telugu and Malayalam. This
paper describes the survey on main approaches for
performing sentiment extraction.
Index Terms— Natural Languages processing, Sentiment
Analysis, Indian languages.
I. INTRODUCTION
“Sentiment analysis or opinion mining refers to the
application of natural language processing,
computational linguistics and text analytics to identify
and extract subjective information in source
materials”(Source:Wikipedia). Opinion mining/sentiment
analysis is a multidisciplinary and multifaceted Artificial
intelligence problem. Its aim is to minimize the gap
between human and computer. Thus, it is collection of
human intelligence and electronic intelligence for mining
the text and classifying user sentiments, likes, dislikes
and wishes. The user generated content is available in
various forms such as web logs, reviews, news,
discussion forums. Web 2.0 & 3.0 has provided a
platform to share the feelings and views about the
products and services. The basic of this problem can be
better explained using the following review by a user
about a car:
“I bought a car a few days ago. It had a comfortable
suspension set-up but it did not provide stable and safe
feel. It delivers good fuel economy but does feel lethargic
engine. It has better quality interiors which look sporty.”
The above review is analyzed for opinion mining and
extracted views are visualized in Fig 1(a). Similarly
comparison of two or more can also be evaluated as
represented in Fig 1(b).
Social network revolution plays a crucial role in
gathering information containing public opinion. To
obtain subjective and factual information from this
information, public opinions are extracted. Thus, it is the
process to predict hidden information about user’s
intensions, likeliness and taste. These social networking
sites generate enormous data up to terabytes per week.
The statistics mentioned in Table I shows the rate at
which amount of user generated data is increasing:
JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 5, NO. 4, NOVEMBER 2013 367
© 2013 ACADEMY PUBLISHER
doi:10.4304/jetwi.5.4.367-371

TABLE I.
Statistics of user generated data
Facebook Twitter Google
850 million users
250 million photos
2.7 billion likes
465 million
accounts
175 million tweets
90 million users
675000 users per
day
Popular approaches used for sentiment analysis are:
Popular approaches used for sentiment analysis are:
• Subjective lexicon – is a list of words where
each word is assigned a score that indicates
nature of word in terms of positive, negative or
objective.
• Using N-Gram modeling- for given training
data , we make a N-Gram model(uni-gram, bi-
gram, tri-gram or combination of these)for
classification.
• Machine learning – perform the supervised or
semi- supervised learning by extracting the
features from the text and learn the model.
II. PROCESS OF SENTIMENT ANALYSIS FOR TEXT
The process of sentiment analysis is divided into five
steps [11]: Process of Sentiment Analysis for Text
(Lexicon Generation), Subjectivity Detection, Sentiment
polarity Detection, Sentiment Structurization, Sentiment
Summarization-Visualization-Tracking.
A. Process of Sentiment Analysis for Text (Lexicon
Generation)
In this phase, sentiment lexicon is created to acquire
the knowledge about sentiments. According to previous
studies, prior polarity should be attached at each lexicon
level. To develop SentiWordNet(s), Manual and
Automated processes have been attempted for multiple
languages.
Related Work: (Stone, 1966) Philip Stones developed
General Inquirer system, was the first milestone for
extracting textual sentiment. It was based on the manual
database containing set of positive or negative
orientations and the input words are compared with
database to identify their class such as positive, negative,
feel, pleasure [5].
(Brill, 1994) Brill Tagger depicted the semantic
orientation for verbs, adverb, noun and adjective. After
extracting these phrases, PMI algorithm (Turney, 2002) is
applied to identify their semantic polarity [31][7].
(Hatzivassiloglou et al., 1997) Hatzivassiloglou was
the first to develop empirical method of building
sentiment lexicon for adjectives. The key point is based
on the nature of conjunctive joining the adjectives. A log-
linear regression model is provided with 82% accuracy
[8].
(Turney, 2002) For the classification of positive and
negative opinion, Peter Turney proposed the idea of
Thumbs Up and Thumbs Down. For better problem
formalization, there was the necessity of an automated
system, which could be employed for electronic
documents. For consecutive words and their polarity,
Turney came up with an algorithm to extract Point wise
Mutual Information(PMI).Experiments were conducted
on movie review corpus and polarity is referred to as
"thumbs up" for positive and "thumbs down" for negative
[7].
(Pang et al.,2002) Pang build sentiment lexicon for
movie reviews to indicate positive and negative opinion.
This system motivated the other machine learning
approaches like Support Vector Machine, Maximum
Entropy and Naive Bayes[10].
(Kamps et al., 2004) Kamps, Marx, Mikken and Rijke
tried to identify subjectivity of adjectives in Word Net. In
this research, they classified adjectives into four major
classes and used base words (to measure relative distance)
depending on the class. For class Feeling their base
words were “happy” and “sad”, for class Competition
their base words were “pass” and “fail”, etc. Based on
this idea, they gathered a total of 1608 words in all four
classes with average accuracy of 67.18% for English [24].
(Gamon et al., 2005) proposed similar method as by
(Turney, 2002).Machine Learning based technique is
used with input of some seed words. This classifier is
based on assumption that the words with same polarity
co-occur in one sentence but words with different polarity
cannot [11].
(Read, 2005) have stated three different problems in
the area of sentiment classification: Time, Domain and
Topic dependency of sentiment orientation. It has been
experimented that associative polarity score varies with
time [12].
(Denecke, 2009) introduced uses of SentiWordNet in
terms of prior polarity scores. The author proposed two
methods: rule-based and machine learning based.
Accuracy of rule-based is 74% which is less than 82%
accuracy of machine learning based. Finally, it is
concluded that there need more sophisticated techniques
of NLP for better accuracy [13].
(Mohammad et al., 2009) proposed a technique to
increase the scope of sentiment lexicon. It includes the
identification of individual words as well as multi-word
expressions with the support of a thesaurus and a list of
affixes. The technique can be implemented by two
methods: antonymy generation and Thesaurus based.
Hand-crafted rules are used for antonymy generation.
Thesaurus method is based on the seed word list which
means if a paragraph has more negative seed words than
the positive ones, then paragraph is marked as negative
[14].
(Mohammad and Turney, 2010) developed Amazon
Mechanical Turk, an online service by Amazon, to gain
human annotation of emotion lexicon. But there was the
need of high quality annotations. Various validations are
provided so that erroneous and random annotations are
rejected, discouraged and re-annotated. Its output
provides 2081 tagged words with an average tagging of
4.75 tags per word [10].
B. Subjectivity Detection
Sentiment analysis categorizes the text at the level of
subjective and objective nature. Subjectivity means the
368 JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 5, NO. 4, NOVEMBER 2013
?2013 ACADEMY PUBLISHER

text contains opinion and objectivity means text contains
no opinion but contains some fact. In precise form,
Subjectivity can be explained as the Topical Relevant
Opinionated Sentiment [9].Genetic Algorithm (Das, 2011)
achieved a good success for the subjectivity detection for
Multiple Objective Optimization [27].
Some example-
1. Subjective- The car is comfortable. (This sentence
expresses the feeling which is an opinion; hence it is of
subjective nature)
2. Objective- Maruti launched a new car. (This
sentence contains a fact).
Related Work: (Wiebe, 2000) defined the concept of
subjectivity in an information retrival perspective which
explains the two genres subjective and objective [9].
(Aue and Gamon, 2005) told that subjectivity
identification is a context dependent and domain
dependent problem which replaces the earlier myth of
using sentiwordnet or subjectivity word list etc. as prior
knowledge database [17].
(Das and Bandyopadhyay, 2009) explained the
techniques for subjectivity based on Rule-based, Machine
learning and Hybrid phenomenon [2].
The idea of collecting subjectivity clues helped in the
subjectivity detection. This collection includes entries of
adjectives (Hatzivassiloglou and Mackeown, 1997) and
verbs(Wiebe,2000) and n-grams(Dave et.al, 2003)
[8][9][18].
The detail of sentiment analysis and subjectivity
detection is given by Wiebe in 1990 [16].
Methods of identification of polarity are explained in
(Aue and Gamon,2005) [17].
Some algorithms like Support Vector Machine (SVM),
Conditional Random Field(CRF) (Zhao et. al,2008) have
been used for clustering of opinions of same type[6].
C. Sentiment Polarity Detection
The sentiment polarity detection means classifying the
sentiments into semantic classes(Turney et.al.,2002) such
as positive, negative or neutral or other emotion classes
like anger, sad, happy, surprise [7].
SentiWordNet is most popular to be used as polarity
lexicon. Another technique used for polarity detection is
Network Overlap Technique [27]. In this, contextual prior
polarity is assigned to each sentiment word.
Related Work: Since the last few years, Tweet Feel
http://www.tweetfeel.com) and Twitter Sentiment
Analysis Tool(http://twittersentiment.appspot.com/) are
available. To satisfy the end users, level of research
should be raised [19].
(Cambria et al., 2011) developed a new paradigm
known as Sentic Computing. This research is based on a
common sense and emotion representation. It have been
used for short texts to infer emotional states over the web
[20].
Concept Net, a semantic network was introduced with
approx 10000 concepts and more than 72000 features
extracted from Open mind corpus. In the sentic
computing, four dimensions are taken as basis to classify
the affective states: Sensitivity, Attention, Pleasantness
and Aptitude.
D. Sentiment Structurization
Sentiment Analysis explained till now is not sufficient
to satisfy the needs of end user, because the latter is not
interested in binary output in terms of positive or negative
but interested in aspectual sentiment classification.
Aspectual can be explained as relative information. For
example, a social worker may be interested to know the
change in the society before and after implementation of
his scheme. So, a sentiment analysis system should be
understand and identify the aspectual sentiments present
in the text.
For this problem, sentiment structurization technique
has been proposed by Das(Das 2010).This technique is
based on 5W (Why, Where, When, What, Who).The
drawback of 5Ws is that it may lead to label bias problem.
To solve this Problem Maximum Entropy Model
(MEMM) was introduced.
Related Work: (Bethard et. al., 2006) have introduced
the automatic identification of opinions from question
answering.
(Bloom et al., 2007) describes Appraisal Theory
(Martin and white ,2005).The system classifies the
opinions into three types: affect, appreciation or judgment.
(Yi et.al. 2006) introduced e sentiment analyzer for
online text documents.
(Zhou et. al., 2006) have introduced the architecture
for blogosphere to get the summarized text.
E. Sentiment Summarization-Visualization-Tracking
One of the main needs of end users is the aggregation
of data. After the Literature survey, following
summarization attempts are found:
• Polarity wise (Hu, 2004),(Yi and Niblack,
2005),(Das and Chen, 2007)
• Topic wise(Yi et al.,2003),(Pang and Lee,
2004) ,(Zhou ,2006)
Visualization and Tracking is the last phase of
sentiment analysis which is most important to satisfy the
needs of end users. In this phase, visual sentiments are
generated which are further tracked with polarity wise
graph according to some dimension or combination of
dimensions. The final graph for tracking is created with a
timeline.
III. SENTIMENT ANALYSIS FOR INDIAN LANGUAGES
There is comparatively less research has been done for
Indian languages.
(Das and Bandyopadhyay, 2010) suggested a
computational technique for developing
SentiWordNet(Bengali) using English-Bengali bilingual
dictionary and English Sentiment Lexicons [21].
(Das and Bandyopadhyay, 2010)- The author
introduced four approaches to predict the polarity of a
word. In the First strategy; an interactive game is
provided which identify the polarity of the words. In the
Second strategy, a bi-lingual dictionary is developed for

English and Indian Languages. In the third strategy, word
net expansion is done using antonym and synonym
relations. In the fourth approach, a pre-annotated corpus
is used for learning [1].
(Das and Bandyopadhyay, 2010)- developed the
method for tagging using the Bengali words.
Classification of words is performed into six emotion
classes (happy, sad, surprise, fear, disgust, anger)
according to three categories of intensities (low, general
and high) [22].
(Draya et.al., 2009) performed blog sentiment analysis
to extract domain specific adjectives [23].
(Joshi, et.al. 2010) used two lexical resources: English-
Hindi Word Net Linking and English SentiWordNet and
created H-SWN(Hindi-SentiWordNet) [28].
(Kim and Hovy,2004) Kim and Hovy did the research
work for Hindi Language but their work is restricted to
synonyms [24].
(Narayan, et.al., 2002) Hindi Subjective Lexicon and
hindi WordNet is used for the identification of semantic
orientation of adjectives and adverbs [25].
(Pang, et.al. 2002) sentiment classification is done at
document level using syntactic approach of N-Grams.
This method is used to perform machine learning [10].
(Rao and Ravichandran, 2009) performed the
classification of bi-polar nature [26].
(Turney, 2002) Turney used semantic mining for
binary classification and also did research on part of
speech (POS) information. It is document level and
review level sentiment analysis [7].
IV. CONCLUSION AND FUTURE WORK
A. Conclusion
Sentiment Analysis has lead to development of better
products and good business management. This research
area has provided more importance to the mass opinion
instead of word-of-mouth.
In the conclusion, it has been proved that coverage
expansion is good by using automatic processes where as
prior polarity assignment is credible by using manual
methods. SentiWordNet has been successfully generated
for Hindi, Telugu and Bengali and global SentiWordNet
has been generated for 57 languages. The 5W structure is
more acceptable solution across domains. The success of
Genetic Algorithm can be estimated by the fact that the
system based on this algorithm has highest performance
till date for Bengali and English.
For Indian languages, scarcity of resources has become
the biggest issue. Research is going on for building
subjective lexicon and datasets for Indian languages.
B. Future Work
As SentiWordNet has been generated for various
languages, there can be further research on cross-lingual
sentiment sense mapping. It is necessary to update the
prior polarity scores according to various dimensions.
The future research can be to develop web service API so
that latest prior polarity scores can be accessed. The
concept of Artificial intelligence can be used for further
research that can mimic human biological mechanism. In
the future, Event Tracking can also be implemented using
the concept of 5W structure.
There are 22 official languages and 13 languages
having more than 10 million speakers in India. Research
is going on these languages but successful results are
obtained in few languages such as Bengali, Hindi and
Malayalam. There are many languages which are
unexplored. Multilingual dictionary is available for
English and 11 Indian languages (Hindi, English, Marathi,
Bengali, Gujarati, Oriya, Malayalam, Urdu, Punjabi,
Tamil, and Telugu).In future; subjective lexicon can be
developed for the unexplored languages which does not
have a word net. The basic resources like parsers, named
entity recognizers, morphological analyzers, and part of
speech tagger need to be improved to reach the state of
accuracy.
REFERENCES
[1] A. Das and S. Bandyopadhay, “SentiWordNet for Indian
languages,” Asian Federation for Natural Language
Processing, China, pp. 56–63, August 2010.
[2] A. Das and S. Bandyopadhay, “Subjectivity Detection in
English and Bengali: A CRF-based Approach,” In
Proceedings of the 7th International Conference on Natural
Language Processing, Macmillan 2009.
[3] H. Tang, S. Tan and X. Cheng, “ A survey on sentiment
detection of reviews”, In Proceedings of the Expert
Systems with Applications 36, Elsevier Ltd., Beijing, 2009
[4] N. Mohandas, J.P. Nair and G.V, “ Domain Specific
Sentence Level Mood Extraction from Malayalam Text”,
In Proceedings of the International Conference on
advances in Computing and Communications, IEEE, 2012.
[5] P.J. Stone, “The General Inquirer: A Computer Approach
to Content Analysis”, The MIT Press, 1966.
[6] J. Zhao, K. Liu and G. Wang, “Adding Redundant Features
for CRFs- based Sentence Sentiment Classication” In
Proceedings of the Conference on Empirical Methods in
Natural Language Processing, pp.117-126, 2008.
[7] P. Turney, “Thumbs up or thumbs down? Semantic
orientation Applied to Unsupervised Classification of
Reviews” In Proceedings of the Association for
Computational Linguistics, pp.417-424, Philadelphia, 2002.
[8] V. Hatzivassiloglou and K. R. McKeown, “Predicting the
Semantic Orientation of Adjectives”, In Proceedings of the
35th
Annual Meeting of the ACL and 8th
Conference of the
European Chapter of the ACL, pp.174-181, Madrid, 1997.
[9] J. M. Wiebe, “ Learning Subjective Adjectives from
Corpora”, In Proceedings of the 17th
National Conference
on Artificial Intelligence and 12th
Conference on
Innovative Applications of Artificial Intelligence, pp. 735-
740, Mento Park, 2000.
[10] B. Pang, L. Lee, and S. Vaithyanathan, “Thumbs up?
Sentiment Classification using Machine Learning
Techniques” In Proceedings of the Empirical Methods on
Natural Language Processing, pp. 79-86, Pennsylvania,
2002
[11] M. Gamon, A. Aue, S. Corston-Oliver and E. Ringger,
“Pulse: Miming Customer Opinions from Free Text”, In
Proceedings of the International Symposium on Intelligent
Data Analysis, pp. 121-132, 2005.
[12] J. Read, “Using Emotions to Reduce Dependency in
Machine Learning Techniques for Sentiment
370 JOURNAL OF EMERGING TECHNOLOGIES IN WEB INTELLIGENCE, VOL. 5, NO. 4, NOVEMBER 2013
?2013 ACADEMY PUBLISHER

Classification”, In Proceedings of the Student Research
Workshop, pp. 43-48, Arbor, 2005.
[13] K. Denecke, “Are SentiWordNet Scores Suited For Multi-
Domain Sentiment Classification”, In Proceedings of the
4th
International Conference on Digital Information
Management, pp. 33-38, Ann Arbor, 2009.
[14] S. Mohammad, B. Dorr, and C. Dunne, “ Generating High-
Coverage Semantic Orientation Lexicons fom Overly
Marked Words and a Thesaurus”, In Proceedings of the
2009 Conference on Empirical Methods in Natural
Language Processing,pp. 599-608, 2009.
[15] S. Mohammad and P. Turney, “Emotions Evoked by
Common Words and Phrases: Using Mechanical Turk to
Create an Emotion Lexicon” In Proceedings of the
Workshop on Computational Approaches to Analysis and
Generation of Emotion, pp. 26-34, California, 2010.
[16] J. M. Wiebe, “Recognizing Subjective Sentences: A
Computational Investigation of Narrative Text”, Doctoral
Thesis. UMI Order Number: UMI Order No. GAX90-
22203., State University of New York, 1990.
[17] M. Gamon and A. Aue, “ Automatic Identification of
Sentiment Vocabulary: Exploiting Low Association with
Known Sentiment Terms”, In Proceedings of the
Workshop on Feature Engineering for Machine Learning
in Natural Language Processing, pp. 57-64, Ann Arbor,
2005.
[18] K. Dave, S. Lawrence and D. M. Pennock, “Mining the
Peanut Gallery: Opinion Extraction and Semantic
Classification of Product Reviews”, In Proceedings of 12th
International Conference on World Wide Web, pp. 519-
528, Hungary, 2003.
[19] B. Liu, “Sentiment Analysis: A Multi- Faceted Problem”,
In Proceeding of IEEE Intelligent Systems, pp. 76-80,
2010.
[20] E. Cambria, A. Hussain and C. Eckl, “ Taking Refuge in
Your Personal Sentic Corner”, In Proceeding of Workshop
on Sentiment Analysis where AI meets Psychology, pp.
35-43, Thailand, 2011.
[21] A. Das and S. Bandyopadhay, “SentiWordNet for Bangla”,
2010.
[22] A. Das and S. Bandyopadhay, “Labeling emotion in
Bengali blog corpus a fine grained tagging at sentence
level”, In Proceeding of 8th
Workshop on Asian Language
Resources, pp. 47-55, Beijing, 2010.
[23] G. Draya, M. Planti, A. Harb, P. Poncelet, M. Roche and F.
Trousset, “ Opinion Mining from Blogs”, In Proceeding of
International Journel of Computer Information Systems
and Industrial Management Applications, 2009
[24] S. M. Kim and E. Hovy, “Determining the sentiment of
opinions”, In Proceeding of COLING, pp. 1367-1373,
2004.
[25] D. Narayan, D. Chakrabarti, P. Pande and P. Bhattacharyya,
“ An experience in building the indo wordnet -a wordnet
for hindi”, In Proceeding of First International Conference
on Global WordNet, 2002.
[26] D. Rao and D. Ravichandan, “ Semi-supervised polarity
lexicon induction”, In Proceeding of 12 conference of the
European Chapter of the Association for Computational
Linguistics, pp. 675-682, USA, 2009.
[27] A. Das, “Opinion Extraction and Summarization from Text
Documents in Bengali”, Doctoral Thesis. UMI Order
Number: UMI Order No. GAX95-09398., Jadavpur
University, 2011.
[28] A. Joshi, A. R. Balamurali and P. Bhattacharyya, “A Fall-
back Strategy for Sentiment Analysis in Hindi: a Case
Study”, In Proceedings of the 8th ICON, 2010.
[29] Amitava Das, http://www.amitavadas.com/
[30] Janyce M. Wiebe,"Learning Subjective Adjectives from
Corpora",http://www.cs.columbia.edu/~vh/courses/Lexical
Semantics/Orientation/wiebe-aaai2000.pdf
[31] Brill Tagger, http://www.ling.gu.se/~lager/mogul/brill-
tagger/index.html.

A Survey on Sentiment Analysis and Opinion Mining.pdf

More Related Content

Similar to A Survey on Sentiment Analysis and Opinion Mining.pdf (20)

More from Mandy Brown (20)

Recently uploaded (20)

A Survey on Sentiment Analysis and Opinion Mining.pdf