Available online at www.sciencedirect.

com
ScienceDirect
ScienceDirect
Procedia Computer Science 00 (2019) 000–000
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2019) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Computer Science 165 (2019) 356–362

INTERNATIONAL CONFERENCE ON RECENT TRENDS IN ADVANCED COMPUTING

INTERNATIONAL CONFERENCE2019,
ON RECENT
ICRTAC TRENDS
2019 IN ADVANCED COMPUTING
2019, ICRTAC 2019
An Efficient Sentiment Analysis Approach for Product Review
An Efficient Sentiment
usingAnalysis
TurneyApproach
Algorithmfor Product Review
using Turney
P.Rajesh Algorithm
Kanna, P.Pandiaraja*
P.Rajesh
M.Kumarasamy CollegeKanna, P.Pandiaraja*
of Engineering, Karur – 639113, India
M.Kumarasamy College of Engineering, Karur – 639113, India

Abstract
Abstract
Sentiment analysis can be done by means of Classification and its most important tasks are text categorization, tone recognition,
image classification
Sentiment analysis can etc.beM ostlybythe
done extant
means of methods of supervised
Classification classification
and its most important are
tasksbased on categorization,
are text traditional statistics, which can
tone recognition,
provideclassification
image ideal results. etc.
The Mmain
ostlyaim
theis extant
to increase the accuracy
methods and toclassification
of supervised report the manufacturer
are based on about the negatives
traditional of the
statistics, product.
which can
The major
provide problem
ideal results.isThe
categorization
main aim isof to sentiment polarity,
increase the accuracy which
and is
to the problem
report of sentimentabout
the manufacturer analysis. There areoftwo
the negatives the levels
product.of
categorization
The major problemand they are Review-level
is categorization Categorization
of sentiment polarity, and
whichSentence-level
is the problemCategorization. Categorization
of sentiment analysis. There are of two
review-level
levels o f
becomes arduous
categorization andwhen
they we are attempt to classify
Review-level the reviewsand
Categorization respect with their Categorization.
Sentence-level specific rating Categorization
related to star -scaled. Second,
of review-level
Review-level
becomes Categorization
arduous has a drawback
when we attempt in Implicit
to classify -levelrespect
the reviews sentimentwithanalysis. M ostlyrating
their specific SVM ,related
Naïve to
Bayesian and Decision
star -scaled. Second,
Tree are mainly
Review-level used to improve
Categorization has a the efficiency
drawback of classification.
in Implicit Amazon
-level sentiment DatasetM is
analysis. used
ostly as ,Dataset
SVM in proposed
Naïve Bayesian andsystem
Decisionto
improve
Tree are the accuracy
mainly used ofto Turney
improvealgorithm. Semantic
the efficiency Orientation (SO)
of classification. Amazonwith Dataset
Point wise M utual
is used as Information yields good
Dataset in proposed results
system to
than otherthe
improve classification
accuracy ofmethods. The reviewSemantic
Turney algorithm. level gets subjected as
Orientation (SO)positive value,wise
with Point on acquaintance of positive
M utual Information average
yields goodSO. On
results
the other
than otherhand, the reviewmethods.
classification level acquires a negative
The review level level in accordance
gets subjected with attainment
as positive value, on of negative average
acquaintance SO. average SO. On
of positive
the other hand, the review level acquires a negative level in accordance with attainment of negative average SO.
© 2019 The Authors. Published by Elsevier B.V.
© 2019
© 2019
This The
The
is an Authors.
accessPublished
Authors.
open Published by
article underby Elsevier
Elsevier B.V.
B.V.
the CC BY-NC-ND license (http://creativecommons.org/licenses/by -nc-nd/4.0/)
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
This is an open
Peer-review access article under the CC BY-NC-ND license the
(http://creativecommons.org/licenses/by -nc-nd/4.0/)
Peer-review under responsibility of the scientific committee ofofthe
under responsibility of scientific committee INTERNATIONAL
INTERNATIONAL CONFERENCE
CONFERENCE ONON RECENT
RECENT TRENDS
TRENDS IN
IN ADVANCED
Peer-review under COM
ADVANCED COMPUTING 2019. PUTING
responsibility 2019
of the scientific committee of the INTERNATIONAL CONFERENCE ON RECENT TRENDS
IN ADVANCED COM PUTING 2019
Keywords: Pointwise Mutual Information;Semantic Orientation; Sentimental Analysis
Keywords: Pointwise Mutual Information;Semantic Orientation; Sentimental Analysis

1. Introduction
1. Introduction
Now-a-days Sentiment analysis is used frequently to analyze the customer feedback. Customer opinion is more
Now-a-days
important for theSentiment analysis
success of is usedInfrequently
the product. olden daystopeople
analyze thereviews
hear customer feedback.
about Customer
the product opinion
and then, theyisdecide
more
important for the success of the product. In olden days people hear reviews about the product and then, they decide

* Corresponding author. : T el.: +91-9385851167

E-mail address:sppandiaraja@gmail.com
* Corresponding author. : T el.: +91-9385851167
E-mail address:sppandiaraja@gmail.com
1877-0509© 2019 T he Authors. Published by Elsevier B.V.
T his is an open
1877-0509© access
2019 T hearticle under
Authors. the CC by
Published BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Elsevier B.V.
Peer-review
T under
his is an open responsibility
access of the CC
article under scientific committee
BY-NC-ND of the
license INTERNATIONAL CONFERENCE ON RECENT T RENDS IN
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
ADVANCED COMPUTING
Peer-review under 2019
responsibility of the scientific committee of the INTERNATIONAL CONFERENCE ON RECENT T RENDS IN
ADVANCED COMPUTING 2019

1877-0509 © 2019 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the scientific committee of the INTERNATIONAL CONFERENCE ON RECENT TRENDS IN
ADVANCED COMPUTING 2019.
10.1016/j.procs.2020.01.038
 P. Rajesh Kanna et al. / Procedia Computer Science 165 (2019) 356–362 357
2 P.Rajesh Kanna , P.Pandiaraja / Procedia Computer Science 00 (2019) 000–000

the quality of the product [1]. This sentiment analysis is extensively used in social media. Natural Language
Processing (NLP) gets employed to work out the concept of sentiment analysis also entitled as opinion min ing or
emotion AI as it‘s far and wide notion that eventually gets identified as say-so aspect of customers‘. Datasets utilized
are manipulated on the basis of textual analysis procedure as well as Co mputational linguistics in order to excavate
intrinsic information available [2].
There are two types of sentences: Subject ive and Objective sentences. Mostly subjective sentence contains more
sentiment than objective sentences. For examp le, Object ive sentence: Milk is white and Subject ive sentence: Milk
tastes good. The subjectivity of words depends on the context of the words . Objective docu ment may also contain
subjective sentence. To overcome these difficulties, we implement proposed system with Turney algorith m which
also increases the accuracy of the reviews. Here, we use Recall, F -measure and Precision parameters as metrics to
assess the accuracy. Precision gives good accuracy than others. Finally, the system calcu lates the Polarity of the
reviews. If the reviews contain more negatives, our system will recommend the manufacturer to recover the
negatives in their products during their forthcoming manufacturing process to increase their sales [6,7]. Th is paper
presents a classification algorith m in order to classify the reviews as either negative or positive. This method takes
the input of A mazon dataset and gives the output as positive or negative review wh ich is g iven by the user. First, the
dataset is pre-processed then feature is extracted using POS Tagging with a count value. For the most portion of
technically necessitated information can be ch iefly acquired fro m ad jective and adverb segment. These two
compositions in parts of speech comparatively serves better than other segments available in a sentence. Point wise
Mutual Informat ion (PMI) parameter acco mpanied with SO is assessed by means of imp ly ing the phrase extorted
[10]. This evaluation merely imparts the correlation existing among words being processed so far. Latterly, those
processed linguistic terms are semantically categorized and classified by means of the review attained. A Positive
review gets evolved on attainment of a positive average semantic orientation. Otherwise, it gets hold for a negative
semantic orientation.

2. Related Work

Durgesh K imp lemented three different types of classification techniques such as Naïve Bayesian, Random Forest
and Support Vector Machine (SVM). Random Forest yields better results than other two methods. The author uses
Twitter dataset to classify the review. He selected four subsets from a co mplete set namely A, B, C and D that
possibly comprises of 300, 3,000, 30,000 and 300,000 vector count correspondingly. F1 scores are obtained with
these sets of vector. For Rev iew-level categorization F1 scores obtained is fairly low over 0.7 3. There is also a
problem of neutral sentence, where judgment is difficult [3].
Peter D imp lemented Turney the unsupervised classified algorith m for thumbs up and thumbs down semantic
orientation. He conducted experiments with 410 rev iews of E-opinions which g ives an average accuracy of 74%. For
movie reviews it is difficult to predict and it produces an average accuracy of about 66%. 80% to 84% of average
accuracy is achieved for t ravel reviews. Th is paper talks of E-opinion reviews of classificat ion using thumbs up and
thumbs down algorithm [4].

3. Classification Methods

SVM, Naïve Bayesian, Decision tree are well known classification methods. SVM is used for text categorization.
Hyper plane is represented using Vector. Hyper plane separates the vectors from one class to another class. By this
way classification is done in SVM. In Naïve Bayesian we use Bayes rule to classify the text. It is also used for text
categorization. Decision tree can be easily done using J48 classifier in Weka package. A ll these give good results of
accuracy [5].

3.1 SVM

The practice of classification can be proficiently carried out by means emp loying Support Vector Machine (SVM )
methodology on the informat ion available. Eventually, all these data available on the whole is set apart as linear and
non-linear segregations. Furthermo re, linear kernel (an optimal hyper plane that segregates linearly) is been searched
358 P. Rajesh Kanna et al. / Procedia Computer Science 165 (2019) 356–362
P.Rajesh Kanna , P.Pandiaraja / Procedia Computer Science 00 (2019) 000–000 3

for on achiev ing a robust classification accuracy after emp loying SVM methodology [6, 21]. Th is linear kernel
optimally serves as a decision boundary for isolating a specific set of information available in a class from those
available devoid of any correlation. Mathematically the aforement ioned hyper plane gets designated as, W . X+b = 0.
Here, the vector weight gets specified as W that perhaps ranges from, w1 , w2 ,...,wn . Train ing of information can be
accomplished by means of manipulating tuple X acco mpanied by a scalar component ‗b‘. Furthermore, minimized
value of ||W|| gets evaluated by means of computing
n

 y x
i 1
i i i (1)

where support vectors (Xi ) labels are particularized on yi and numerical factors gets specified as α i .
n
y i =1 then w x
i 1
i i 1 (2)

n
if y i =-1 then w x
i 1
i i 1 (3)

On unproductive segregation by means of emp loying linear manner, SVM transmutes the procedure of classificat ion
by utilizing non-linear way of mapping. Certainly this procedural alteration co mpletely reveals a higher d imension
form fro m the data being processed. For such a profound form of procedural transformation a kernel function called
Gaussian Radial Basis Function (RBF) is absolutely utilized in order to extort most robust form of results [7,8].

 
 y Xi X j 2
K(X i , X j )  e /2 (4)

Where Xi are support vectors, γ is a free parameter and Xj are testing tuples.

3.2 Naive Bayes

D confines to be the dataset that comprises features that scales up to n -dimensional form of feature vector
specified as X=x1 ,x2 ,..,xn , that are to be trained on the basis of characterized tuple set. Here, ―n‖ accounts for
dimensions specified on every particu lar tuple. Initially the classifier imposed concentrates on ‗m‘ nu mber of classes
ranging fro m C1 ,C2 ,...,C m [9]. The tuple X of data considered is liab le to get categorized into any class C i after
classification when, P(Ci |X) > P(Cj |X), where i,j∈[1,m]and i≠j. Furthermore,
n
P(C i | X )   P( xi | C i ) (5)
i 1

3.3 Decision tree

All classified form o f informat ion can be capably signified by means of utilizing decision tree that suitably keeps
apart those components comprises of features acquired out of finite distinct domains. Classified features of every
distinct domain are personified into d iscrete classes. Every single feature observed, probably that different fro m one
another serves to be the internal nodes of this decision tree [10, 11, 12]. All those other data coming under the
similar do main also gets appended to the former one. Finally, all sorts of data get labeled under a specific class that
possibly got classified as leaf node or any of the internal nodes belonging to the decision tree. Moreover,
amalgamation of arith metical as well as co mputing methodologies compactly yielded a robust methodical procedure
named as decision tree. Henceforth, this structure can resourcefully systematize those data being processed by it in a
categorical manner [13, 14, 16].The information to be categorized progresses as,

(x, Y)=(x1 ,x2 ,x3 ,……,xk ,Y) (6)

 P. Rajesh Kanna et al. / Procedia Computer Science 165 (2019) 356–362 359
4 P.Rajesh Kanna , P.Pandiaraja / Procedia Computer Science 00 (2019) 000–000

The elements available in (6) gets designated as, reliant variable that is also named as target variable Y. x gets
specified as input variable that typically encompasses x1 , x2 , x3 ….xn [15, 17, 18].

4. Proposed Algorithm

Our proposed algorithm is based on turney algorithm. The flo w chart shown in Fig.1. depicts about proposed
process of our algorithm. Firstly, review data that is A mazon Dataset is given as input dataset. PoS Tagging is done
to extract feature and then semantic o rientation is performed. Nature of rev iew utterly relies upon the outcome of
average value of SO parameter. Polarity gets indicated as positive only when average becomes positive.

Fig.1. Flow diagram for proposed system

4.1 Selection of Phrase and feature phrase extraction

The first step in classification is pre -processing which removes unwanted data and stop words. The stop word
removes based on stop word list file. Stop words do not affect the meaning of the sentence. Some of the stop words
―a‖, ―the‖, ―of‖ which reduces the corpus size without losing any information. There are 8 parts of speech form
which two words are extracted (adjective and adverb). So me words can have two different meaning at two different
places. For examp le, ―great book‖ for book review is positive and ―great flop‖ for movie review is negative. Tab le 1
shows the Pos tag and their meaning.

T able 1. Pos Tags and their meaning

PoS Tagging Meaning
JJ Adjective
RB, RBR, or RBS Adverb
NN, NNS Noun
VB, VBD, VBN or VBG Verb
CC Cardinal Number

Tag patterns are extracted fro m two word phrase from the review. Semantic Orientation is done with these kinds
of patterns.
 P.Rajesh Kanna , P.Pandiaraja / Procedia Computer Science 00 (2019) 000–000 5
360 P. Rajesh Kanna et al. / Procedia Computer Science 165 (2019) 356–362

T able 2. T wo-Word Phrase

First Word Second Word Third Word
JJ NN or NNS Anything
RB, RBR, or RBS JJ Not NN or NNS
JJ JJ Not NN or NNS
RB, RBR, or RBS VB, VBD, or VBN Anything
NN, NNS JJ Not NN or NNS

4.2 Seed words

We should find the seed words and we want to pass it to the algorith m to find the semantic orientation. If the
phrase occurs along with negative seed word, the phrase is negative. If the phrase occurs along with positive seed
word, the phrase is positive.

T able 3. Seed Words

Positive Seed Words Excellent, Best, Good
Negative Seed Words Poor, Bad, Hate, Terrible

4.3 PMI and SO

The Point wise correlative information between two words (extracted word and seed word) can be given as

 p ( word1 & word 2 ) 

PMI ( word1 , word 2 )  log 2  
 p ( word1 ) p ( word 2 )  (7)

Here, p(word1,word 2) gives the probability of the word co -occur. Here, p(word1) and p(word2) g ives the
probability of the word occurs independently. The ratio between p(word1,word 2) and p(word1) p(word 2) gives the
degree of statistical dependence between those words. After the calculat ion of PMI we have to calculate the semantic
orientation. Thus, SO is calculated as

SO(phrase) = PMI(phrase, {positive paradigms}) - PMI(phrase,{negative paradigms}) (8)

If the phrase is with positive seed word, then that phrase is considered as positive. If the phrase is with negative
seed word, then the phrase is considered as negative. PMI -IR issues query to search engine to find the number of hits
(matching document). The AltaVista NEA R operator is used to search documents which is mo re efficient than AND
operator. From equation (2) and (3) we can derive this equation with NEAR operator [10].

SO(phrase)=log 2  hits( phraseNEAR" excellent " )hits" poor"  (9)
 hits( phraseNEAR" poor" )hits(" excellent " ) 

p_query = (good OR best OR….OR superior)

n_query = (bad OR poor OR…..OR inferior)
This can give positive semantic orientation or negative semantic orientation.

4.4 Sentiment Classification

After finding semantic orientation we have to find the average semantic orientation. If the central semantic
orientation is negative, the review is negative and the review is recommended review. In case there is an attainment
of a negatively polarized average SO then the review does not serves to be an optional review and will not be
encouraged to be induced anymore [19,20].
 P. Rajesh Kanna et al. / Procedia Computer Science 165 (2019) 356–362 361
6 P.Rajesh Kanna , P.Pandiaraja / Procedia Computer Science 00 (2019) 000–000

5. Experiments

Experiments are done with 5000 labeled positive and negative product reviews. It consists of reviews of the
products. There are 2500 labeled positive reviews and there are 2500 labeled negative reviews. Turney algorithm
gives better results.

T able 4. Performance Analysis of differ Classifier Algorithms

Algorithm TP Rate FP Rate Precision Recall F-Measure RO C
Turne y 0.781 0.213 0.788 0.787 0.787 0.871
De cision Tre e 0.607 0.393 0.607 0.607 0.607 0.611
SVM 0.696 0.304 0.731 0.696 0.684 0.696
Naive Baye sian 0.781 0.212 0.781 0.782 0.780 0.870

6. Results

The Table 4. shows the performance analysis of different classifiers wh ich is measured using precision, recall, f -
measure and roc where turney classifier yields the good accuracy (0.871) than other classifiers like Decision tree,
Naïve Bayes and SVM for movie review dataset and its graphical representation is been shown in Fig.2.. Decision
tree yields good accuracy with roc (0.611). SVM y ields g ood accuracy with precision (0.731). Naïve Bayes yields
good accuracy with roc (0.870).

4
3
Naive Bayesian
2
SVM
1
0 Decision Tree
Turney

Fig.2. Performance Analysis different classifier Conclusion

Thus, the classifiers Decision tree, SVM, Naïve Bayes yields less accuracy we go for Semantic Orientation is
used to classify the reviews. Semantic Orientation with PMI increases the performance of the classifier. SO-PMI is
simp le, unsupervised, simp le to imp lement and it doesn‘t restrict to ad jectives, but it requires a large corpus. The
PMI algorith m has three simple steps: first is to extracts two -word phrase containing adjective and adverbs. Second
is to find Semantic Orientation of phrases and third is to take average of all SO and assign sentiment as positive or
negative. Turney classifier algorith m yields good performance measure than the other three algorith ms for the
Benchmark Dataset. Turney classifier classifies accurately upto ~78% of accuracy wh ere the other three algorith ms
classifies upto ~69%.

7. Future Work

The errors and the negatives of the product can be mailed auto matically to the manufacturer to rectify their erro rs.
They can also increase their sales in the market. In future work, to imp ro ve the accuracy we can use semantic
orientation with other features in a combined manner as like a combination in a supervised classification algorithm.
362 P. Rajesh Kanna et al. / Procedia Computer Science 165 (2019) 356–362
P.Rajesh Kanna , P.Pandiaraja / Procedia Computer Science 00 (2019) 000–000 7

References

[1] Andrea Esuli and Fabrizio Sebastiani, (2012) ,Determining the Semantic Orientation of Terms through Gloss Classification, In Proceedings
of the 14th ACM International Conference on Information and Knowledge Management (CIKM 2005), Bremen, DE, pp. 617-624.
[2] Callen Rain, (2012), Sentiment Analysis in Amazon Reviews Using Probabilistic Machine Learning,2012.
[3] Durgesh K. Srivastava and Lekha Bhambhu, (2010) , Data Classification using Support Vector Machine, Journal of Theoretical and Applied
Information Technology. 12(1):1-7 · February 2010.
[4] Peter D. T urney, (2002), Thumbs Up or Thumbs Down? Semant ic Orientation Applied to Unsupervised Classification of Reviews,
Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL), Philadelphia, pp. 417 -424.
[5] Peter D. T urney and Michael L.Littman, (2011), Measuring Praise and Criticism: Inference of Semantic Orientation from Association ,
National Research Council Canada. ACM Transactions on Information Systems 21(4):315—346.
[6] P.D. T urney and M.L. Littman, (2002), Unsupervised Learning of Semantic Orientation from a Hundred-Billion, Word Corpus - National
Research Council Canada, Institute for Information Technology. May 16, 2002.
[7] Sandhya Khanna and S1avita Shiwani, (2013), Subjectivity detection and Semantic orientation based Methods for Sentiment Analysis,
International Journal of Scientific & Engineering Research, 4(9), pp 868-873.
[8] Shivakumar Vaithyanathan, (2012), Thumbs up? Sentiment Classification using Machine Learning T echniques, Department of Computer
Science Cornell University Ithaca, NY 14853 USA.
[9] S.L.T ing, W.H. Ip and Albert H.C. Tsang, (2011) , Is Naïve Bayes a Good Classifier for Document Classification?, International Journal of
Software Engineering and its Applications 5(3).
[10] Sneha M Nakade and Sachin N Deshmukh,(2014), Finding Semantic Orientation of Reviews Using Unsupervised PMI Algorithm,
International Journal of Science and Research. 5( 2) ,pp 2101-2110.
[11] P.Pandiaraja and S.Parasuraman, (2015), Applying secure authentication scheme to protect DNS from rebinding attack using proxy, 2015
International Conference on Circuits, Power and Computing T echnologies [ICCPCT-2015] ,pp 1-6.
[12] S T hilagamani and N Shanthi, (2011), Object recognition based on image segmentation and clustering Journal of Computer Science 7(11)
pp 1741-1748.
[13] P.Rajesh Kanna and S.Keerthi, (2017), Location Based Image Retrieval System on Ranking User Clicks, Indian Journal of Natural
Sciences,8(47),pp 13426-13429.
[14] S.Deepika and P.Pandiaraja, (2013), Ensuring CIA triad for user data using collaborative filtering mechanism, 20 13 International
Conference on Information Communication and Embedded Systems (ICICES),pp 925 -928.
[15] P.Rajesh Kanna, S.Abirame and M.Madura, (2017), Language Translator for Images ,Indian Journal of Natural Sciences,8(47), pp 13421-
13425
[16] S Thilagamani and N Shanthi, (2010), Literature survey on enhancing cluster quality, International Journal on Computer Science and
Engineering 2 (6), pp 1999-2002.
[17] P.Rajesh Kanna and S.Keerthi, (2017), Automation of Lab with Attendance Monitoring Screen Capturing and Performance Analysis,
International Journal of Pure and Applied Mathematics, 118 (18 ),pp, 2765 -2770.
[18] P.Pandiaraja and S.Chitra, (2014), Protection of a webpage in realtime using proxy server with performance, International Journal of
Applied Engineering Research , 9(23), pp 23211- 23218.
[19] P.Rajesh Kanna, K.Sindhanaiselvan and M.K.Vijaymeena, (2017), A Defensive Mechanism based on PCA to Defend Denial of-Service
Attack, International Journal of Security and Its Applications 11(1) , pp.71-82.
[20] P Pandiaraja and J Manikandan, (2015), Web proxy based detection and protection mechanisms against client based HTTP attacks,
International Conference on Circuits, Power and Computing T echnologies [ICCPCT-2015] ,pp 1-6.
[21] S.Elanthiraiyan and P.Pandiaraja ,(2013), Interactive Detection and Classification of DDoS Attacks Using ESVM, Journal of Computer
Engineering, 9(4),pp 50-56.