International Conference on Inventive Communication and Computational Technologies

(ICICCT 2017)

Sentiment Analysis of Product Reviews:

A Review
Shivaprasad T K Jyothi Shetty
Dept. of CSE Dept. of CSE
NMAM Institute of Technology Nitte NMAM Institute of Technology Nitte
Nitte,Karkala,India Nitte,Karkala,India
[email protected] [email protected]

Abstract: Now a day's internet is the most valuable used to make good impactful decision. Sentiment
source of learning, getting ideas, reviews for a product analysis and sentiment classification are the two
or a service. Everyday millions of reviews are generated methodologies used in opinion mining. Whereas both
in the internet about a product, person or a place. have its own independent features, but sometimes it
Because of their huge number and size it is very
difficult to handle and understand such reviews.
may be used interchangeably. Sentiment
Sentiment analysis is such a research area which classification indicates the sentiment orientation by
understands and extracts the opinion from the given assigning the class labels to the document or
review and the analysis process includes natural segment. Sentiment orientation is a kind of text
language processing (NLP), computational linguistics, classification that classifies text data based on the
text analytics and classifying the polarity of the opinion. sentiment orientation of opinion. Sentiment
In the field of sentiment analysis there are many orientation indicates the polarity of the opinion either
algorithms exist to tackle NLP problems. Each true or false based on subjectivity [1]. Subjective
algorithm is used by several applications. In this paper analysis is a process of identifying whether the given
we have shown the taxonomy of various sentiment
text or reviews data is subjective or objective in
analysis methods. This paper also shows that Support
vector machine (SVM) gives high accuracy compared nature.
to Naïve bayes and maximum entropy methods. In this paper several sentiment analysis methods
have been discussed. Even though we find several
Keywords-Sentiment analysis, Opinion mining, Product papers proposed by different researchers in this area,
reviews, Natural language processing. there is a need to make sentiment analysis more
accurate and easy to understand. Sentiment analysis
is extremely useful in various situations. But it is
I. INTRODUCTION very difficult process because of the complexity
involved in the human language. It has got several
Sentiment analysis is a process of extracting and variants like grammatical, cultural etc. Humans can
understanding the sentiments defined in the text easily interpret statements like “My order been
document. The explosion of data in the various social delayed. Very good”. But it may be difficult for the
media channels like twitter, facebook, and linkdin machine to understand. Similarly word “thin” may be
has given consumer new way of expressing their taken as positive with respect to laptop but it may be
opinion on a particular product, person and places. negative when it comes to apartment wall. So to give
The user opinion is always in the form of textual the correct decision sentiment analysis must be
information. Per day millions of textual message data sometime more business specific.
is sent over social media or online shopping website.
Investigating and analyzing the sentiment of the II. LEVEL B ASED SENTIMENT ANALYSIS
opinion is a very critical task to perform. The NLP
with artificial intelligence capability and text In general the process of sentiment analysis is
analytics are used to determine whether the divided into 3 levels.
sentiment of the opinion is positive, negative and
Level Based Sentiment Classification
neutral. The opinion mining and sentiment analysis is
doesn’t depend on any particular domain or platform.
Document Level (Turney et al, 2002)
It spreads to all the social media networks,
healthcare, management, economy and many more
Sentence level (V.S Jagtap et al, 2013)
and also it is very useful for the growth of many
companies and organizations. Sentiment analysis is Aspect level (Zhang et al, 2012)
also providing a business intelligence which can be
Fig 1: Different levels of Sentiment Classification

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 298

Authorized licensed use limited to: University of Exeter. Downloaded on February 15,2025 at 00:37:13 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)
A. Document level analysis: Document level review or opinions to bipolar classes such as positive
sentiment analysis determines the overall opinion of or negative. Binary sentiment classification
the document. Apart from the sentiment orientation of classifies reviews or opinions by using multi-point
the individual sentences it classifies the sentiment rating scale. Multi-point rating scale uses the rating
expressed by the whole document. The classification inference. Using this rating inference class labels are
is expressed by either positive or negative sentiment assigned as scalar rating such as 1 to 5 "stars". The
[2]. This level analysis is helpful only if the document detailed description of binary approach is given in
related to a single entity. Because it expresses opinion section 4.
on a single entity (E.g. Product, Person). Thus it is
not applicable to the documents which contain the
comparison of the multiple entities. III. SENTIMENT ANALYSIS P ROCESS

B. Sentence level analysis: The sentence level The basic sentiment analysis framework consists of
sentiment analysis is also called as subjectivity following steps. Review collection process, Data
classification [3]. It distinguishes subjective preparation, Review analysis and Sentiment
information from the objective information. It classification.
considers each sentence as a separate unit and it
predicts that the sentence must contain only one A. Data Preparation: Data preparation is a process
opinion. It iterates each sentence and determines of collecting the reviews of a particular product from
whether the sentiment orientation of the sentence is internet sources or any online shopping websites.
positive, negative or neutral. The collected data may be either unstructured. There
are many publically available datasets which provide
C. Entity and Aspect level analysis: Entity and the huge collections of reviews which can be are
aspect level sentiment analysis captures the mixture used in sentiment analysis process. Sometimes the
of emotion from the review sentence. It performs set of reviews contains unwanted information such as
fine-grained sentiment analysis and it differentiates HTML tags, URL information etc. Removing such
what actually user wants and do not-want. One main useless information is done at the stage of
feature of aspect level sentiment analysis is that, it preprocessing of reviews.
directly looks into the opinion instead of paragraph,
sentences, phrases and document. The goal is to find B. Review Analysis: Review Analysis step analyze
sentiments on entities and their aspects. Meant to say the special features of the reviews and then identifies
that aspect level sentiment analysis concentrate on the interesting information in the review including
sentiment (positive or negative) and a target opinion. Review analysis step first applies various
(opinion). computation tasks to the review and then extracts the
opinion and features of the product. Two mainly
used methods in review analysis are POS tagging and
A sentence or a document may contain a
Negation tagging.
combination of positive and negative opinions.

x Sentiment level opinion mining is

performed by two tasks i.e. subjective and
objective.

Objective: I purchased a new Android

Smartphone few months ago.

Subjective: It is such a nice Gadget.

x The polarity of the opinion is determined by

using the subjective sentence.

Positive: It is such a nice Gadget.

Negative: It has poor radio signal

reception quality.

In a document level sentiment analysis it determines

the overall document and a sentence. The classes of
the opinion are determined by polarity. i.e. either
positive or negative. The binary classification
method is used for this purpose. Most of the existing
sentiment analysis algorithm uses the binary
classification methods. Mean to say that they assign Fig 2: Sentiment Analysis process Model

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 299

Authorized licensed use limited to: University of Exeter. Downloaded on February 15,2025 at 00:37:13 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)
formulated as regression problem. This is happened
C. Sentiment Classification: The two major because each class corresponding to a different point
approaches used for classifying review are Sentiment of scale. More specifically each class may have its
orientation approach and machine learning approach. own vocabulary and meaning of the sentence will be
Sentiment orientation consists of two subtasks. First different depending on domain and situations.
task is to extract the opinions from the review.
While the second subtask is to determine the overall C. Contextual or Fuzzy approach: Contextual or
sentiment orientation of a sentence. The Sentiment fuzzy approach is used to find whether the review
Orientation contains two seed adjective, "excellent" text document is in neutral form or polar form.
and "poor". The machine learning approach depends Contextual approach is very useful when user unable
on the classification classes such as positive and to find the actual intension of people about the
negative. particular product. In simple this approach explains
about whether the review text documents belongs to
the polarity measures (positive, negative and neutral)
IV. P OLARITY B ASED SENTIMENT or it belongs to the neutral polarity (only neutral).
C LASSIFICATION And this creates dis-ambigutiy in polarity of polar
expression. Sometime, it may be a combination of
A typical approach to sentiment analysis is to begin the both approaches. In order to tackle the problem
with a lexicon of negative and positive words and of ambiguity the 4 way approach is used like
phrases. Identification of polarity from this lexicon positive, negative, both and neutral[6].
decides the performance of sentiment analysis.
Sentiment classification can also be used to analyze
the perspective of the different sources of data or V. SENTIMENT C LASSIFICATION T ECHNIQUE
information. Especially in the field of reviews the
numerical rating system makes the categorization Sentiment classification methods usually divided into
process into fine grained scales other than positive machine learning approach and lexicon based
and negative categories. This richer information can approach. Machine Learning Approach uses
be used for rank and compare the reviews linguistic methods and applies common ML
quantitatively. algorithms. The lexicon based approach depend on
sentiment lexicon and a set of commonly known and
a precompiled sentiment words. The lexicon based
Polarity Based Sentiment Analysis approach is sub divided into dictionary based
approach and corpus based approach. The corpus
Binary approach (Turney et al, 2004)
based approach contains two sub methods called
statistical method and semantic method.
Multi-level approach (Bo Pang et al, 2005
The sentiment classification method using machine
Contextual or Fuzzy approach (Teresa Learning approach is divided into supervised
Wilson et al, 2014) learning and unsupervised learning approaches. The
supervised method is a sophisticated approach where
Fig 3: Polarity Based Sentiment Analysis it uses only labeled training documents. There are
many supervised classifiers in literature. Most
A. Binary Approach: Binary approach is also common used classifiers are probabilistic classifiers
called as two class categorization process. Where the and Naive Bayes Classifiers. If supervised methods
sentiment classification is categorized into two unable to find the labeled training document then
classes namely positive and negative. Turney [4] unsupervised method is used to solve the problem.
proposed counting of positive and negative terms in a
product review to determine the polarity. Binary The lexicon based approach is used to discover the
approach is very useful for brief analysis for any type opinion lexicon which is used to analyze the text
of reviews. This approach is implemented by document. It includes two sub approaches. The
including the other metrics called negation words, dictionary based approach depends on the dictionary
intensifier and diminishes and it searches the antonyms and synonyms for the
seed words in the dictionary. The corpus based
B. Multi-class approach: Instead of using the binary approach started with list of opinion words and find
classification system multiclass approach is used to out some other opinion words in the large corpus to
convey the sentiment expressed by the piece of text retrieve semantic orientation.
document. Multi class method is also called as star
based inference approach where the classification
includes categorization process as high (5 stars) and VI. SENTIMENT ANALYSIS METHODS
low (1 star) inference [5]. We can easily categorize
the reviews based on these star rating. But the The main branches of Sentiment Analysis include
sentiment based multi-class approach naturally Machine learning approach and Lexicon based
978-1-5090-5297-4/17/$31.00 ©2017 IEEE 300
Authorized licensed use limited to: University of Exeter. Downloaded on February 15,2025 at 00:37:13 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)
approach. Each of these approaches includes sub- approach. The corpus based approach contains two
approaches. The Taxonomy of Sentiment analysis sub methods called statistical and semantic based
methodologies are listed below in graphical model. approaches.
We listed some techniques and procedures used by
these methodologies. The two important sentiment 2) Lexicon based approaches
classification methods are
Lexicon Based Approach
1) Machine learning approaches
Dictionary Based Approach
Machine Learning Approach
Shallow Parsing Based Sentiment
Supervised Learning Lexicon (Jayaraj Desai et al., 2015)

Phrase Extraction Methods (Lada

Sentence level Categorization (Xing
Fang et al., 2015) Banic et al., 2013)

Feature Specific Sentiment Classify Slack Based Extraction Methods

(S.Mukherjee et al., 2015) (Li Zhang et al., 2012)

Sentiment Analysis on Sub Corpus Based Approach

Sentence Level (Cacilia Zin et al.,
2011) Statistical Method

Feature Based Opinion

Sentiment Classification using
Extraction (Fermin L.Crur et al.,
supervised methods (P.Kalaivani, et
2013)
al., 2013)
Domain Specific Corpus
Unsupervised Learning
Approach (Antonio Moreno-Ortiz
Multi-Aspect Sentiment Analysis et al., 2015)
(Fu Xiang et al., 2012)
Specialized Vocabulary Approach
Sentence level Polarity Dovglas R Rice et al., 2013)
classification (Lanjun Zhou et al.,
2011) Semantic Method

Semantic Analysis Methods

Ant-Based Clustering Method (Emil
(Wenhao Zhang et al., 2012)
et al., 2015)
Deep Sentiment Analysis Method
(Isa Marks et al., 2012)
Fig 4: Machine learning approaches

In Fig 4, we listed some of the methods which are

used by the machine learning approaches. The main
branches of the machine learning methods are Fig 5: Lexicon based approaches
supervised learning and unsupervised learning. The
goal of the supervised learning method is to find the
desired output for the desired input. The mapping VII. FEATURE SELECTION METHODS
between the input and output is identified in
supervised learning method. In unsupervised method The most frequent feature selection methods are used
there is no suitable input for learning process. Each in the sentiment analysis is a lexicon based methods
input and desired output identification is done by the which needs manual annotations. While Statistical
algorithm itself. The table contains the list of methods are fully automatic. The feature selection
research done on the machine learning approach. In method treat text document as a Bag of Words
Fig 5, it contains the methods of lexicon based (BOW’s). Because of their simplicity BOW’s are
approaches. Lexicon based approaches uses text most commonly used in the feature selection
classification approach and it mainly aims to find the methods. The common use of BOW’s is stop words
semantic orientation of a document or a text. Main removal and stemming in sentiment analysis process.
branches of the lexicon based approaches are
dictionary based approach and corpus based
978-1-5090-5297-4/17/$31.00 ©2017 IEEE 301
Authorized licensed use limited to: University of Exeter. Downloaded on February 15,2025 at 00:37:13 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)
VIII. LITERATURE SURVEY classification approach which uses naïve bayes,
maximum entropy and support vector machine
Xing Fang et al., [7] represent a sentiment analysis machine learning classifiers. They have developed a
system for product reviews which uses the sentiment framework which processes the experiences of
polarity categorization process. The overall process customers of popular brands of drugs and cosmetic
is split into 3 phases. The classification methods products. Naïve bayes classifier are used to obtain
selected for categorization are naive Bayesian, baseline result for accessing other classifiers.
support vector machine and random forest. During
phase 1 the subjective content is extracted from the
review data and objective content is removed. After IX. ANALYSIS O F SOME IMPORTANT
the extraction perform the POS tagging to extracted METHODS
content. During phase 2 identify the sentiment phase
either negative of adjective (NOA) and negation of Researchers use sophisticated methods to predict the
verb (NOV). And perform the sentiment score sentiments of product features. The datasets of
computation for sentiment tokens. Using the review may crawl from online or it may be use
sentiment score formula feature vector for sentiment offline data. The classifiers play very important role
are generated. Finally in phase 3 they performed the in sentiment analysis. Classifiers work independent
sentiment polarity categorization. to all domains (product, movie or place). The widely
used classifier like naïve bayes (NB), support vector
Subhabrata mukherjee and pushpak Bhattacharyya et machine (SVM) and maximum entropy (Max-Ent)
al., [8] investigate the product review using feature ease the process of sentiment analysis. Naïve bayes
specific sentiment analysis. A dependency parsing classifiers are highly scalable and they are very
method is used to capture the relation between the popular method used in text categorization. With its
features and their associated opinions. They appropriate results it gives competition to the very
developed a system that extracts the potential advanced method called support vector machine.
features from review and clusters opinion expression Many researchers use naïve bayes for their first
describing each of the features. To achieve good preference and for better results they move on to
experimental result they used two datasets Lakkaraju other advanced classifiers. Support vector machine
et. al., [9] and Hu and Liu et. al., [10]. The proposed builds a set of hyper planes in a infinite dimensional
system performs average accuracy of 80.98% for space which can be used for classification and other
dataset 1 and 70% of accuracy for dataset 2. processes. SVM algorithms are widely used in
advanced classifier, and experimental results show
Jian jian and ping ji et. al., [11] presents a framework that SVM achieves high accuracy in classification
for the identification of product feature aspects and process as compared to any other methods.
consumer detailed reasons from online product Maximum entropy is suitable for finding the best
reviews. The proposed co-clustering algorithm probability distribution. Max-Ent is also widely used
provides a short summary about consumers concerns method in sentiment analysis. The graph shown in
on different aspects of product features. And reason fig (6) depicts the accuracy of support vector
of consumers to provide a concise description machine, naïve bayes and maximum entropy for the
regarding customer requirement for product different product datasets [14]. In the graph naïve
designers. The approach of conditional random fields bayes and maximum entropy shows moderate
is employed to detect aspects of product features and accuracy with small variation and support vector
detailed reasons jointly. machine shows the highest accuracy for all the
dataset.
Lada Banic and Ana mihanovic et al., [12]
investigate hotel review mining using machine
learning and natural language processing. They 102
develop a product evaluation system configuration 100
No. of Instances

which works on a KNIME sentiment analysis 98

framework. The sentiment analysis process includes 96 Naive
retrieving data from the database, dictionary 94
92 Byes
development and review scoring. They represents 90 Max-Ent
two evaluation system it evaluated each term or 88
phrase positive or negative. Cumulative information 86 SVM
about the number of positive and negative terms or
phrases was generated on the level of final
evaluation. In second evaluation they maintain a
grade from 1 to 5. In which 1 is referred as bad and 5
is referred as excellent.

Haruna isah and Paul Trundle et al., [13] represents a Fig 6: Accuracy of sentiment classifiers over different dataset
lexicon based and machine learning based sentiment
978-1-5090-5297-4/17/$31.00 ©2017 IEEE 302
Authorized licensed use limited to: University of Exeter. Downloaded on February 15,2025 at 00:37:13 UTC from IEEE Xplore. Restrictions apply.
 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)
X. CONCLUSION Coherence for the simultaneous discovery of latent
facts and associated sentiments",SIAM International
Conference on Data Mining (SDM),April2011.
In this paper it is observed that sentiment analysis or
opinion mining plays important role while making a [10] Minqing Hu and Bing Liu,"Miming and Summarizing
decision towards a particular product or a service. customer reviews", KDD 04: proceedings of the tenth
ACM SIGKDD international Conference on knowledge
But it is very important to consider certain quality discovery and data mining.
measures like helpfulness, usefulness and utility
while analyzing each review. In the literature survey [11] Jian Jin and Ping Ji,"Mining online productreviews to
there are many sophisticated methods explained identify consumers FineGrinedConcerns",IET,2015,Hong
Kong SAR,Chaina.
which defines the sentiment analysis with respect to
different aspects. [12] Lada Banic,Ana mihanovic,Marko Brakus,"Using Big
In future, more research work is needed to improving Data and Sentiment Analysis in Product Evaluation
the performance measures further. Sentiment "MIPRO,2013,Croatia.
analysis or opinion mining can be applied for any
[13] Haruna isah,Paul Trundle,Daneiel Neagu,"Social Media
new applications which follow data mining rules. Analysis for Product Safety and using Text Mining and
Although the techniques and algorithms used for SA",IET,2015,University of Bradford,UK
sentiment analysis are advancing fast and giving high
[14] R. RajKumar,V.P. Kallimani, Lam H. Lee, Dino Isa,”
quality results, lot of problems in this field of study
Text document processing with naïve bayes and support
remain unresolved and also it is hard to find the fake vector machine”, published by IEEE computer society.
review by reading. Sometimes fake reviews also seen
as good quality review and it was modified like no [15] Zirn C, Niepert M, Stuckenschmidt H, Strube M.
one can identify their actual intension. So fake Fine-grained sentiment analysis with structural features.
In: Presented at the 5th International Joint Conference on
review detection is another important field which Natural Language Processing (IJCNLP’11); 2011.
requires deep data mining techniques.
[16] Maks Isa, Vossen Piek. A lexicon model for deep
sentiment analysis and opinion mining applications.
REFERENCES [17] Cruz Fermı´n L, Troyano Jose´ A,Enrı´quez
Fernando, Javier Ortega F, Vallejo Carlo G.Long
[1] Bing Liu,"Exploring User Opinions in autonomy or long delay?’ The importance of domain in
Recommender Systems",Proceeding of the secondKDD opinion mining. Expert Syst Appl 2013.
workshop on Large ScaleRecommender System and the
Netflix Prize Competition",April 2012,LasVegas,USA. [18] Zhou L, Li B, Gao W, Wei Z, Wong K. Unsupervised
discovery of discourse relations for eliminating intra-
[2] Antonio Moreno-Ortiz, Javier Fernández-Cruz,” sentence polarity ambiguities. In: Presented at the 2001
Identifying polarity in financial texts for sentiment conference on Empirical Methods in Natural Language
analysis: a corpus-based approach”, 7th International Processing (EMNLP’11); 2011.
Conference on Corpus Linguistics: Current Work in
Corpus Linguistics: Working with Traditionally- [19] Emil Şt. Chifu , Tiberiu Şt. Leţia Viorica R. Chifu,”
conceived Corpora and Beyond (CILC 2015) Unsupervised Aspect Level Sentiment Analysis Using
Ant Clustering and Self-organizing Maps”,IEEE,2015
[3] Zhang Wenhao, Hua Xu, Wan Wei. Weakness finder:
find product weakness from Chinese reviews by using [20] Jayraj M. Desai, Swapnil R. Andhariya,” Sentiment
aspects based sentiment analysis. Expert Syst Appl 2012. analysis Approach to adapt a shallow parsing based
sentiment lexicon”, IEEE Sponsored 2nd International
[4] Peter D. Turney.” Thumbs Up or Thumbs Down? Conference on Innovations in Information Embedded and
Semantic Orientation Applied to Unsupervised Communication Systems, ICIIECS’15
Classification of Reviews”, Proceedings of the 40th
Annual Meeting of the Association for [21] Maks Isa, Vossen Piek. A lexicon model for deep
ComputationalLinguistics (ACL), Philadelphia, July sentiment analysis and opinion mining applications.
2002, pp. 417-424. Decis Support Syst 2012.

[5] Bo Pang and Lillian Lee. “Seeing stars Exploiting class

relationships for sentiment categorization with respect
to rating scales “. Proceedings of the ACL, 2005

[6] TheresaWilson, JanyceWiebe, Paul Hoffmann,”

Recognizing Contextual Polarity in Phrase-Level
Sentiment Analysis”, Advanced Research and
Development Activity (ARDA)

[7] Xing Fang ,Justin Zhan ,"Sentiment Analysis using

product review data",Springer: Journal of Big
data",2015,North Carolina A& T State
university,Greensboro,NC,USA.

[8] Subhabrata Mukherjee,Pushpak Bhattacharyya,"Feature

Specific Sentiment Analysis for product
Reviews",IET,2015,IIT Bombay.

[9] Himabindu Lakkaraju, Chiranjib Bhattacharyya,Indrajit

Bhattacharyya and Srujana Merugu,"Exploiting
978-1-5090-5297-4/17/$31.00 ©2017 IEEE 303
Authorized licensed use limited to: University of Exeter. Downloaded on February 15,2025 at 00:37:13 UTC from IEEE Xplore. Restrictions apply.