2014 International Conference on Intelligent Computing Applications

Sentiment Analysis On Unstructured Review

Mrs.R.Nithya1 Dr.D.Maheswari2
School of Computer Studies(UG) School of Computer Studies(PG)
RVS College of Arts and Science RVS College of Arts and Science
Sulur,India Sulur,India
[email protected] [email protected]

Abstract— Sentiment analysis mainly focuses on subjectivity comments, reviews, discussions, news, feedback or tweets,
and polarity detection. Today consumers make buying decision about a product, policy, person or topic. To be specific,
based on the customer’s review that is available in some of the
opinion mining can be defined as a sub discipline of
online shopping sites like shopclues, fabfurnish, pepperfry,
flipkart etc. There are also some of the specific websites which computational linguistics that focuses on extracting people’s
discuss about positive and negative facts of those products that opinion form the web. It analyses from a given piece of text
comes to market like reevoo, buzzillions, bizarte, amazon etc. about; which part is opinion expressing; who wrote the
Hence this type of analysis are socially very needed for sellers to opinion; what is being commented. Sentiment analysis, on the
undergo market analysis, branding, product penetration, market other hand is about determining the subjectivity, polarity like
segmentation and so on. Here, the proposed paper classifies the positive, negative or neutral and polarity strength. Opinion can
most identified features using supervised learning method Naïve be fetched in two different ways. One is of questionnaire
Bayes and determined their positive, negative and neutral where the questions and its answers will be very relevant o
polarity distribution.
product and its feature. So it is easy to make score and finalize
Keywords— opinion mining; feature extraction; sentiment the outcome whereas unstructured review that may usually
Classification include feedback in the form of text and images from various
social monitoring tools and online shopping sites. In market
I. INTRODUCTION each product may be introduced on the basis of some latest
Social media are popularly known as ‘democracy’s features they hold and they can either uplift or downsize the
pipeline’, ‘an amplifier of unfiltered emotion’, ‘an organism demand of that product.
with a million tongues and twice as many eyes’ and as ‘a
virtual megaphone with a global reach’. Recent surveys on In the preliminary stage, this paper starts with finding the
media by research firm Social Bakers and Semiocast a Paris features available in one of the electronic product device
states that; 75% of web users in India are below the age of Samsung Tablet PC. Researchers have reported lots of
35years, 42% smartphone users in India use device to access approaches towards feature extraction and they are broadly
news, nearly 72% netizens lives in urban areas, nearly 52% classified as two types like supervised and unsupervised. In
internet users connect to web via a mobile phone and about this proposal I focus on identifying features from unstructured
1.5Lakh new internet users added every month in India. reviews and clustered it manually. Then the polarity of each
Furthermore, Forrester estimates that Indians spent around clustered comments is determined to undergo further
$1.6 billion online on retail e-commerce sites in 2012. By classification using supervised method Naive Bayes for prior
2016 it can either extend upto $8.8 billion. So that the online distribution of each featured class attribute towards positivity
shopping sites are engaging with their consumers on the and negativity.
emotional front as well as fulfilling their need for information
in order to indicate that they are not limited to satisfy only on II. RELATED WORKS
their functional needs. Most of the consumers do not make Sentiment analysis for the emotional preference of online
buying decision immediately on shopping website by placing comments has gained great achievement since it was raised up
an order. Instead they make purchase decisions as they move by Pang etal.2002 [1] and studied in-depth. A common type of
in and out of TV and print commercials, a friend’s opinion summarization is Aspect-based Opinion
recommendation on social media, product information online, Summarization. It contains aspect feature identification,
product reviews on a trusted blog and the best deals in their sentiment prediction, and summary generation. Hu and Liu et
local store. They are totally engaged across all the places al [2][3] attempt to find features by using NLP-base
where they are about to access information they need and then techniques, they perform POS tagging and generate n-grams,
move to final step of ordering. Opinion mining or Sentiment for sentiment prediction, they choose some seed sentiment
analysis is an important sub discipline of Data mining and words. Popescu and Etzioni et al [4] investigated the same
Natural Language Processing which deals with building a problem. Their algorithm requires that the product class
system that explores the user’s opinions made in blog spots, is1597. The algorithm only reckon noun/noun phrase as the

978-1-4799-3966-4/14 $31.00 © 2014 IEEE 367

DOI 10.1109/ICICA.2014.81

Authorized licensed use limited to: LIVERPOOL JOHN MOORES UNIVERSITY. Downloaded on April 13,2022 at 12:43:54 UTC from IEEE Xplore. Restrictions apply.
 candidate features. It determines whether a noun/noun phrase using Brill Tagger. Most of the adjective words bear
is a feature by computing the Point-wise Mutual Information sentiment, so they are highlighted in Fig. 2. Those feature
(PMI) score between the phrase and class discriminators, words are visualized using TagCrowd in Fig. 3.
e.g.,“of xx”, “xx has”, “xx comes with”, etc., where xx is a
product class. But it calculates the PMI by searching the Web.
Querying the Web is time-consuming. Khairrullah khan et al
[5] has suggested that Brill Tagger or CST tagger can be used
to identify which category of words can be features. Hsiang
Hui Let et al [6] recommends the observation that there are
relations between the product features or aspects and opinion
words. Thelwall, M., Buckley et al [7] has given some
confidence that SentiStrength is a robust algorithm for
sentiment strength detection on social web data and is Fig. 1. A portion of comments given POS
recommended for applications in which exploiting only direct
affective terms is important. Alekh Agarwal et al [8] tried to
focus on adjectival word that increase the polarity score and
gained accuracy of about 61.1% compared to non-adjectival
word of 55.93%. Hai-bing ma et al [9] suggests a typical
approach first to identify k positive words (such as excellent,
awesome, fine) and k negative words (such as bad, poor).
Later to get the sentiment weight of a word, we should
subtract the associated weight with k negative words, These
2k words are often selected by experts. This is a kind of Fig. 2. Short list of sentiment Bearing words are highlighted
supervised learning algorithm where 2k words have to been
taken for further classification. Dipali V.Talele et al [10] used
Naïve bayes classification with tf-idf for summarizing review
and its accuracy is 47.8% which is higher than of SVM with
27.0%.
III. PROPOSED WORK
A. Feature Extraction Fig. 3. TagCrowd used for visualizing feature and its bearing words
Feature Extraction and polarity detection is one of the very
interesting as well as difficult tasks in opinion mining. C. Product Aspects
Sentiment strength detection is one which predicts the strength TextStat is a freely available which can be used for pattern
of positive or negative sentiment within a text. We tried a very
extraction. The aspect words are nothing but the noun or noun
common approach for sentiment analysis by selecting a
phrases like display, fabrication, response, screen, accessories,
machine learning algorithm and a method of extracting features
from texts and then train the classifier with a human-coded applications, batterylife, speed, weight, size, price, cost,
corpus. Corpus is a large collection of texts. It is a body of navigation, connectivity with its number of frequency are
written or spoken material upon which a linguistic analysis is retrieved. From which most occurring aspect words are taken
based. The features are usually words that can undergo further and they are clustered manually which is shortlisted in
stemming or part-of-speech tagged words. TABLE I.
B. Steps in Preprocessing TABLE I. A PORTION OF FEATURE-LIST
The term stemming refers to the reduction of words to their Features
roots. Porter’s stemming algorithm can be used to remove Screen/di
accessories/applic Batterylife/sp
splay/tou weight/size price/cost
stop words. Brill Tagger, Tree Tagger, CST Tagger are the ch screen
ations eed
tool used for annotating text with part-of-speech (POS). POS Good a.
No apps installed Excellent
quality of takes too long
also called grammatical tagging is the process of marking up a fabricatio
for office and
other applications.
to charge
Slightly heavy. value for
money
n
word in a corpus as corresponding to a particular part of Touch Limited video
Available at
speech, based on both its definition, as well as its adjacent and screen formats available battery life Solid but not
affordable
very for viewing movies acceptable heavy.
related words in a phrase, sentence or paragraph. A parser reactive etc.
rate
Great tablet
processes input sentences according to the productions of a Stunning
Movies, pages &
not superb Very nice to and good
content looks
grammar, and builds one or more constituent structures that display
good!
battery hold value for
money.
conform to the grammar. The assumption here is that positive Have nothing A bit more
words will tend to co-occur with other positive words more Wonderfu Lack of to compare It is not expensive
l make. accessories. the battery compact than other 10
than with negative words, and vice-versa. Fig. 1, shows a part life with. inch tablet
a.
of sample sentence which has undergone stemming, POS Sample set of comments that are clustered under common feature-set manually by human effort

368

Authorized licensed use limited to: LIVERPOOL JOHN MOORES UNIVERSITY. Downloaded on April 13,2022 at 12:43:54 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. LIST OF FEATURES WITH ITS SENTISTRENGTH VALUES D. Finding the polarity of opinionated sentence
battery SentiStrength is a lexicon-based classifier that uses
display accessories life weight cost additional linguistic information and rules to detect sentiment
3,-2a. 1,-1 1,-2 1,-2 3,-1 strength in short informal English text. For each text, the
1,-3 1,-1 1,-1 1,-1 2,-1 SentiStrength output is of two integers: 1 to 5 for positive
2,-1 1,-1 1,-1 2,-1 1,-3 sentiment strength and a separate score of 1 to 5 for negative
3,-1 1,-1 1,-1 1,-1 1,-2 sentiment strength. For instance, 0 indicates no emotion, 1
3,-1 1,-1 1,-1 1,-1 2,-1 indicates not positive, 2 indicates slightly positive, 3 indicates
4,-1 2,-1 1,-2 1,-3 1,-2 normal positive, 4 indicates positive and 5 indicates very
1,-2 3,-1 2,-1 3,-1 2,-1 positive. These scales are used because even short texts can
1,-1 2,-1 3,-1 2,-1 1,-3 contain both positivity and negativity.
1,-1 1,-3 3,-1 1,-3 4,-1
1,-1 1,-2 4,-1 4,-1 1,-2 IV. EXPERIMENTAL SETUP
2,-1 3,-1 1,-2 1,-2 1,-1 The experiment starts with the work flow diagram depicted in
3,-1 4,-1 1,-1 1,-1 2,-1 Fig. 4. drawn using Dia tool.
2,-1 1,-2 2,-1 2,-1 1,-3
1,-3 1,-1 1,-3 1,-3 1,-2
1,-2 1,-1 1,-2 1,-2 1,-2
2,-1 1,-1 2,-1 2,-1 1,-1
1,-2 1,-2 1,-2 1,-2 1,-1
1,-1 3,-1 2,-1 2,-1 1,-1
2,-1 2,-1 1,-3 1,-3 2,-1
1,-2 1,-3 1,-2 4,-1 1,-2
3,-2 1,-2 1,-2 1,-3 1,-1
1,-3 2,-1 2,-1 3,-1 1,-1
2,-1 3,-1 1,-3 2,-1 2,-1
3,-1 3,-1 4,-1 1,-3 1,-3
Fig. 4. Workflow diagram using Dia Tool
3,-1 4,-1 1,-2 4,-1 1,-2
4,-1 1,-2 1,-1 1,-2 3,-1
1,-2 1,-1 2,-1 1,-1 4,-1 A. Dataset
2,-1 2,-1 1,-3 2,-1 1,-2
Totally 575 reviews were taken from shopping sites. A
1,-3 1,-2 1,-2 1,-2 1,-1
snapshot of its sentistrength binary value listed in TABLE II.
3,-2 1,-2 1,-2 1,-3 1,-1
1,-2 3,-1 2,-1 3,-1 2,-1 B. Classification through Tanagra1.4
1,-1 2,-1 3,-1 2,-1 1,-3 Tanagra1.4 is free data mining software for academic and
1,-1 1,-3 3,-1 1,-3 4,-1 research purposes. It proposes several data mining methods
1,-1 1,-2 4,-1 4,-1 1,-2 from exploratory data analysis and machine learning. The
2,-1 3,-1 1,-2 1,-2 1,-1 main purpose of its project is to give researchers an easy-to-
3,-1 4,-1 1,-1 1,-1 2,-1 use data mining software. TANAGRA acts more as an
2,-1 1,-2 2,-1 2,-1 1,-3 experimental platform. Thus, Tanagra can be considered as a
1,-2 1,-1 1,-2 1,-2 1,-2 pedagogical tool for learning programming techniques to
2,-1 1,-1 2,-1 2,-1 1,-1 undergo Naïve bayes classification.
1,-2 1,-2 1,-2 1,-2 1,-1
1,-1 3,-1 2,-1 2,-1 1,-1 Step 1: Import dataset specified in above excel sheet which
2,-1 2,-1 1,-3 1,-3 2,-1 consist of sentiment value for each of 575 unstructured
1,-2 1,-3 1,-2 4,-1 1,-2 reviews detected using sentistrength tool.
3,-2 1,-2 1,-2 1,-3 1,-1 Step 2: Define status and set parameters as discrete binary
1,-3 2,-1 2,-1 3,-1 1,-1 values for all the most basic features that are identified.
2,-1 3,-1 1,-3 2,-1 2,-1 Step 3: From supervised learning method select Naïve bayes
3,-1 3,-1 4,-1 1,-3 1,-3 classifier and set to each of the defined status thru step 2.
3,-1 4,-1 1,-2 4,-1 1,-2
Step 4: Set the classification function to true so that the features
gets prior distribution to each of class attributes.
1,-2 1,-3 2,-1 1,-2 1,-2
1,-1 4,-1 3,-1 2,-1 3,-1 Fig. 5. to Fig. 9. Shows the Naïve bayes classification made
1,-1 1,-2 2,-1 1,-2 4,-1 through our Tanagra tool based each individual features.
2,-1 1,-1 1,-2 2,-1 1,-2
b.
Set of features for which sentiment binary values are determined using Sentistrength

369

Authorized licensed use limited to: LIVERPOOL JOHN MOORES UNIVERSITY. Downloaded on April 13,2022 at 12:43:54 UTC from IEEE Xplore. Restrictions apply.
 Fig. 5. Screenshot indicating prior distribution of feature- display.

Fig. 9. Screenshot indicating prior distribution of feature- cost.

C. Equations

Inorder to calculate the positive, negative and neutral

polarity percentage following is the formula to be used.

Pos % =( positive value’s count /  comments) X 100 (1)

Neg % =( negative value’s count /  comments) X 100 (2)
Neu % =( neutral value’s count /  comments) X 100 (3)
Fig. 6. Screenshot indicating prior distribution of feature- accessories.
TABLE III. LIST OF FEATURE AND ITS POLARITY DISTRIBUTION
% of % of
Features % of positive
negative neutral
distribution
distribution distribution
Displayc. 43.5
36.5 20
Accessories
42.6 44.4 13

Batterylife 46 35 19
weight 45.2 45.2 9.6
cost 32 40 28
c.
List of values indicating each feature and its percentage of polarity distribution

Fig. 7. Screenshot indicating prior distribution of feature- batterylife.

Fig. 10. Chart indicating each feature and its prior distribution

Fig. 8. Screenshot indicating prior distribution of feature- weight.

370

Authorized licensed use limited to: LIVERPOOL JOHN MOORES UNIVERSITY. Downloaded on April 13,2022 at 12:43:54 UTC from IEEE Xplore. Restrictions apply.
 V. CONCLUSION AND FUTURE WORK
Thus our supervised learning method has classified
the most identified features and their percentage of polarity
distribution using the formula (1), (2) and (3). The proposal
has visualized the result in a graphical format that is depicted
in Fig. 10. by incorporating the values listed in TABLE III. As
‘batterylife’ bears most positive value it can improve branding
and ‘cost’ bears very low positive value which indicates seller
to concentrate on reputation and product penetration. In future
this work can be extended by studying the various feature
selection methods and incorporating the best one in our
experiment.

ACKNOWLEDGMENT
Thanks to S.A.Sriranjani and D. Maheswari for their
valuable suggestions in choosing this broad area to undergo my
research.
REFERENCES

[1] Pang,L.Lee and S.Vaithyanathan,Thumbs up?: sentiment classification

using machine Learning techniques. In EMNLP' 02: Proceedings of the
ACL-02 conference on Empirical methods in natural language
processing. Association for Computational Linguistics, Morristown, NJ,
USA, 79-86,2002.
[2] M. Hu and B.Liu,Mining and summarizing customer reviews. In KDD'
04: Proceedings of the Tenth ACM SIGKDD international conference
on Knowledge discovery and data mining. ACM, New York, NY,
USA,168-177,2004.
[3] M. Hu and B. Liu,Mining opinion features in customer reviews. In
AAAI' 04: Proceedings of the 19th national conference on Artificial
Intelligence. AAAI Press,2004.son, B. Noble, and I.N. Sneddon, “On
certain integrals of Lipschitz-Hankel type involving products of Bessel
functions,” Phil. Trans. Roy. Soc. London, vol. A247, pp. 529-551,
April 1955. (references)
[4] Popescu, Ana-Maria and Oren, Etzioni, “Extracting product features and
opinions from reviews,” In Proceedings of EMNLP, 2005.
[5] Khairullah khan and Baharum B.Baharudin, Analysis of Syntactic
Patterns for Identification of Features from Unstructured Reviews,
4thInternational Conference on Intelligent and Advanced Systems, 2012.
[6] Hsiang Hui Lek and Danny C.C.Poo, Sentix: An Aspect and Domain
Sensitive Sentiment Lexicon, 24th Internation Conference on Tools with
Aritificial Intelligence, 2012.
[7] Thelwall, M., Buckley, K., & Paltoglou, G, Sentiment strength detection
for the social Web, preprint of an article published in the Journal of the
American Society for Information Science and Technology, 63(1), 163-
173, © copyright 2011 John Wiley & Sons, Inc.
[8] Alekh Agarwal and Pushpak Bhattacharyya, Augmenting WordNet with
Polarity Information on Adjectives, Petr Sojka, Key-Sun Choi,
Christiane Fellbaum, Piek Vossen (Eds.): GWC 2006, Proceedings, pp.
3–8. c Masaryk University, 2005.
[9] Hai-Bing Ma, Yi-Bing Geng, Jun-Rui Qiu, Analysis Of Three Methods
For Web-Based Opinion Mining, Proceedings of the 2011 International
Conference on Machine Learning and Cybernetics, Guilin, 10-13 July,
2011.
[10] Dipali V.Talele, Sonal Patil, Extracting and Analyzing Sentiments of the
Crowd Using Naive Bayes Classification, Asian Journal of Computer
Science and Information Technology, ISSN 2249 – 5126,2013.

371

Authorized licensed use limited to: LIVERPOOL JOHN MOORES UNIVERSITY. Downloaded on April 13,2022 at 12:43:54 UTC from IEEE Xplore. Restrictions apply.