International Journal of Science and Research (IJSR)

ISSN (Online): 2319-7064

Impact Factor (2012): 3.358

Sentiment Analysis on Movie Reviews Based on

Combined Approach
Anurag Mulkalwar1, Kavita Kelkar2
1

Research Scholar, Dept. of Computer Engineering, K.J.Somaiya College of Engineering, University of Mumbai
VidyaVihar, Mumbai, India

Associate Professor, Dept. of Computer Engineering, K.J.Somaiya College of Engineering, University of Mumbai
VidyaVihar, Mumbai, India

Abstract: Sentiment Analysis is one of the most important area of concern towards classifying sentiments from any given textual
information. Sentiment analysis can be performed on textual data by using various machine learning methods like supervised learning
or unsupervised learning. But no any individual method is sufficient to classify sentiments with that much precision. In this paper we
are going to propose new approach to perform sentiment analysis on movie review called as Combined Approach. It uses two separate
classifier Support Vector Machine (SVM) and Hidden Markov Model (HMM). Then it combines results of these classifier using
classifier combine rule.

Keywords: Sentiment Analysis Combined Approach, HMM, SVM.

1. Introduction

know about each aspect without going throw each review

process like sentiment analysis is really useful.

Sentiment analysis is sub division of Data Mining for

detection of human behavioral contain from textual data.
Now today World is moving toward complete digitalization.
We can see that massive growth in user of internet from all
age group. By considering this as a benefit for business
promotion each business wants to take command on internet
platform. Looking at growth of users and their attraction
towards internet so many business start promoting their
products facilities on internet. It is not stick to only single
business but varies from promotion of movies, electronic
products, online marketing and also for social networking
like twitter, facebook. Each of this marketing portal start
providing some space for user to leave their reviews
regarding particular thing. If we look at any site which is
providing information related movies like IMDB we can see
lots of comments from visitors over there. Its not only about
movie reviews but it is all about each and every site which
wants take reviews from their visitors. Reviews are helpful
contain for any business to know what is exact response of
their reviewer. But if we what is exact response of their
reviewer. This is because of huge amount of reviews. Recent
days film industry is in boom. Turnover of this industry is in
millions or billions. This is because of huge amount of
reviews.
Recent day film industry is in boom. Turnover of this
industry is in millions or billions. Each weak number of
movies are released. Numerous websites provide
information about movies. Movie lovers always want watch
good movie for which they spend money. After watching
movie they always want to leave reviews or before going for
movie wants to know comments on that movie. Beside of
reading all reviews it will be always better to know how
many there are positive or negative reviews. Movie review
contains talk about various aspect like acting, direction,
story, film quality. But if we look carefully it is not possible
for any human analyzer to read each and every review and to
decide what reviewer saying either positive or negative. To

Paper ID: 0201413361

Sentiment analysis techniques take care of such situations.

Sentiment classifier looks for sentiment from text data and
classify that in their appropriate class positive, negative or
neutral. Sentiment analysis process get completed in various
steps. This process start with gathering of reviews in text
form., then collected data get cleaned and form unit for
processing called as token. Required features are get
extracted from tokens and all features are pass to the
sentiment classifier. Sentiment classifier is a core part of
sentiment analysis process.
Sentiment Analysis is possible at various levels like Aspect
level, sentence level and document level.
Aspect level: classification based on sentiment related to
aspect.
Sentence level: classification based on sentiment present
in particular sentence.
Document level: classification based on sentiment present
in whole document.
It is most important thing in sentiment analysis processes
which classifier you are using for sentiment classification. If
classifier is not working properly then main intension of
sentiment analysis will never get achieved. Mainly for
sentiment classification machine learning techniques are
used. Supervised learning or unsupervised learning methods
are preferably used for classification. But no single
classifiers efficiently manage to classify sentiment correctly.
So to improve classification technique we are going to use
two classifiers SVM and HMM.

2. Related Work
As per mention in [1] Sentiment analysis applies natural
language processing techniques and computational
linguistics to extract information about sentiments expressed
by users about any subject. It shows that there are number of
number of approaches are present to do sentiment analysis.

Volume 3 Issue 7, July 2014

www.ijsr.net

Licensed Under Creative Commons Attribution CC BY

1739

International Journal of Science and Research (IJSR)

ISSN (Online): 2319-7064
Impact Factor (2012): 3.358

Sentiment classification can be possible at three levels like

document; sentence and feature level complete discussion is
given in [2]. Sentiment analysis has their sentiment
classification techniques lexical approach and machine
learning approach. Lexical approach and its improved
approach is explained in [3] which shows improvement of
basic lexical approach for topic detection. Same ways there
are machine learning approaches are also used for sentiment
classification [4] like SVM, Nave Bayes, Maximum
Entropy Approach. It also shows comparison between these
approaches it concludes that SVM gives batter classification
of sentiment. In 2009 FAN Na , CAI Wan-dong, ZHAO Yu
propose A Method based on Generation Models for
Analyzing Sentiment-Topic in Texts[5]. As per this method
it firstly sentiment and topic of training texts are labeled by
hand and sentiment models and topic models are established
and secondly compute the Kullback-Leibler divergence
between a testing text and sentiment models in order to
determine sentiment of the text. Traditional approaches of
text sentiment analysis typically work at a particular level,
such as phrase, sentence or document level, which might not
be suitable for the documents with too few or too many
words. Considering every level analysis has its own
advantages, a combination model may achieve better
performance. In this [6], a novel combined model based on
phrase and sentence levels analyses and a discussion on the
complementation of different levels analyses are presented.
Sentiment analysis seeks to characterize opinionated or
evaluative aspects of natural language text thus helping
people to discover valuable information from large amounts
of unstructured data. A new methodology for sentiment
analysis called proximity-based sentiment analysis. It is a
different approach, by considering a new set of features
based on word proximities in a written text. It [7] propose
three proximity-based features, namely, proximity
distribution, mutual information between proximity types,
and proximity patterns. Experimental results show that
proximity based sentiment analysis is able to extract
sentiments from a specific domain, with performance
comparable to the state-of-the-art. Sentiment analysis of
product reviews has recently become very popular in Web
text mining, natural language processing and computational
linguistics research. Automated analysis of the sentiments
presented in online consumer feedbacks can facilitate both
organizations business strategy development and individual
consumers comparison shopping. The main contribution of
this paper [8] is the illustration of a novel feature-level
sentiment analysis mechanism which is underpinned by a
fuzzy domain sentiment ontology tree extraction algorithm.
The proposed mechanism can automatically construct fuzzy
domain ontology tree (FDSOT) based on the product
reviews, including the extraction of sentiment words,
product features and the relations among features. Here
product features (or features) mean product components and
attributes. [9]This paper goes beyond sentiment
classification by focusing on techniques that could detect the
topics that are highly correlated with the positive and
negative opinions. Such techniques, when coupled with
sentiment classification, can help the business analysts to
understand both the overall sentiment scope as well as the
drivers behind the sentiment. [10] This paper shows an
empirical study to apply classification-based sentiment
analysis on online reviews with multiple dimensions using

Paper ID: 0201413361

natural language processing techniques. The aim of this

study is to find the most influential part-of-speech on the
sentimental analysis and the performance of the multidimensional classification methods. [11] this paper focus on
various feature selection methods currently present in action.
This paper also propose new approach of feature selection it
also shows how feature selection is important in sentiment
analysis. [12] Paper shows result of sentiment classification
by using MI as a feature extraction method and various
machine learning classifiers.

3. Methodology
3.1 Preprocessing
In this module we clean incoming textual review for ease of
classification. This module consists of various processing
tasks like tokenization, stop-word removal and
lemmatization.
Tokenization: Tokenization process start with taking input
as raw text stream which is in our case one of the review
regarding movie. Task of this step is to form tokens of
continued text stream. Logic behind implementation of this
sub module is firstly convert whole uppercase in to
lowercases and then process each word until found
whitespace if whitespace found then consider it as a token.
Stop-word Removal: Fully tokenized input of
Tokenization sub module taken for finding unwanted
words. Stop word are those words which are the parts of
sentence but doesnt contain any sense.
Lemmatization: In English language single word has
many forms. It causes increase no. of features. To avoid
needs to cut down each word in to its root form.
Lemmatization processes transform each word in to its
original form.
Synonyms Finder: This step finds synonyms of each
word and generate proper set of tokens. It helps to reduce
number of tokens.

Figure 1: Preprocessing
3.2 Slang Words and Smiley Processing Unit
Smiley is mostly preferable type for giving comments. Wide
range of users or viewers use smiley along with text for
commenting. And it is really necessary task to find out such

Volume 3 Issue 7, July 2014

www.ijsr.net

Licensed Under Creative Commons Attribution CC BY

1740

International Journal of Science and Research (IJSR)

ISSN (Online): 2319-7064
Impact Factor (2012): 3.358

smiley from textual review and analyze it. Most of the time
users write comment by using slang words. By simple
looking at the slang word human can understand its meaning
but it is complicated task for classifier to understand such
tokens. So slangs are another challenge for proper sentiment
classification. This Slang words and Smiley processing
module take care of find out such slang words and return its
appropriate meaning to classifier. Same way slang words are
getting converted in to its meaningful form. For doing that
Slang words and preprocessing module use predefined
database to recover exact meaning of that smiley or slang.
3.3 Feature Extraction
Before sending all words to classifier minimize the whole
bag of word. Because when we use supervised learning
method for classification and if the size of feature set gets
increased then there are chances of accurate classification of
particular review by classifier get decreases. This violation
of supervised classifier from its goal is because of over
fitting. To avoid this situation it is compulsory task to
minimize size of feature set. Appropriate feature extraction
method can handle this case in efficient manner. We are
going to use Mutual Information (MI) feature extraction
method to extract useful features that will helps in accurate
classification to classifier. MI calculates mutual
independence of two random variables. In our case random
variables are nothing but features. W is a token from bag of
word and C is a desired class of classification from class set.
,

MI(W,C)= log

M= (I, E, T, O, S)
where I initial probabilities, E output symbol emission
probabilities, T states transition probabilities, O set of output
symbols and S set of states. So vector representation of
HMM.
3.5 Support Vector Machine (SVM : Classifier II)
Support Vector Machines are highly effective rather than
traditional test classification methods. SVMs are proven best
technique over Maximum Entropy and Nave Bayes Method.
SVMs has various types but of which Linear SVM is
consider as a suitable type for text classification. Simple idea
behind the SVM in two-category classification is to find
hyperplane which separates two classes with the help of
vector . It not only separates the document vectors but

maximize the margin that separates document vector.

Where is given by solving dual problem. with value

of is greater than zero are consider as Support Vectors.

SVM finds proper support vector which will separate two
categories exactly.
3.6. Classifier Combination Rules
As above mention our two classifier, one is SVM and HMM
are combine with classification rule as describe in following
section. There are many possible method for combination of
two different classifier. Consider, text classification problem
where text is assign to one of the possible class label Lk (Lk
= f1, f2,.fm). Let assume that we have N classifiers
denoted as Ck(k=1, 2, N). Input sample Xk is assign
to every classifier Ck and output measure in form of
posterior probability vector represent as follows,
| , , |
Where, | denote the probability of classifier when X
labeled as fi. Here, consider j type of label is assign to z test
data in given combination rules.
Sum Rule:
Label fi is assign to z when following condition occur.

arg max /

Figure 2: Architecture
3.4 Hidden Markov Model (HMM: Classifier I)
Hidden Markov Model is a well known structure of
probabilistic automata. It deals with number of states, and
associated transition probability. In our case number of
states are nothing but set of class {Positive and Negative}.
Transition probabilities are probability of observation by
which transitions are takes place. Hidden Markov Model for
text classification is depend on vector

Paper ID: 0201413361

Major Voting Rule:

Shortly, each classifier directly assign label to test data then
major voting rule assign the final label to test data by
selecting major vote given by all the classifier for that test
data. When different label got the same number of maximum
counts, then randomly class label is selected among them.
The mathematical formula for major voting rule is given as
follow.

Where,

Volume 3 Issue 7, July 2014

www.ijsr.net

max

1

0

Licensed Under Creative Commons Attribution CC BY

1741

International Journal of Science and Research (IJSR)

ISSN (Online): 2319-7064
Impact Factor (2012): 3.358

Max Rule:
Max rule is applied over information provided by probability
of Pj(fi|xk).Max rule is always winner in major voting
class. It is expressed by following function.

arg maxmax

Min and mean rule also use for combination of two

classifier. Min rule can be derive by the last result of max
rule. Mean rule can be obtain by max rule also by taking

mean of probability function mean .

Based on this classifier combination rules HMM and SVM

model greatly outperform. Sum rule give the more accuracy
than any other combination rule.

4. Conclusion

In this paper we propose new approach called combined

approach to classify text reviews based on sentiment present
in that reviews. With the help of two classifier and classifier
combination rules it is possible to improve expected
classification results. We also propose way of handling slang
words and smiley. It will overall causes good sentiment
classification with higher accuracy.

5. Future Work
Currently we are working towards development of such
system which will parse through all reviews regarding
particular movie from various sites and it will classify all the
reviews based on sentiments present in reviews. We plan to
design such a model which will perform best classification
on reviews from all domains. We plan to apply this
technique on audio sentiment classification.

References
[1] Sowmya Kamath S, Anusha Bagalkotkar, Ashesh
Khandelwal, Shivam Pandey, Kumari Poornima,
Sentiment
Analysis
Based
Approaches
for
Understanding User Context in Web Content,
International Conference on Communication Systems
and Network Technologies, 2013
[2] Bing Liu, Sentiment Analysis and Opinion Mining,
Morgan & Claypool Publishers, May 2012.
[3] Keke Cai*, Scott Spangler!, Ying Chen!, Li Zhang,
Leveraging Sentiment Analysis for Topic Detection
IEEE/WIC/ACM International Conference on Web
Intelligence and Intelligent Agent Technology ,2008.
[4] Bo Pang and Lillian Lee, Shivakumar Vaithyanathan,
Thumbs up? Sentiment Classification using Machine
Learning Techniques, Proceedings of the Conference
on Empirical Methods in Natural Language Processing
(EMNLP), Philadelphia, July 2002.
[5] FAN Na , CAI Wan-dong, ZHAO Yu, A Method
based on Generation Models for Analyzing SentimentTopic in Texts IEEE, 2009.
[6] Si Li, Hao Zhang, Weiran Xu, Guang Chen and Jun
Guo, Exploiting Combined Multi-level Model for
Document
Sentiment
Analysis,
International
Conference on Pattern Recognition , 2010

Paper ID: 0201413361

[7] S.M.Shamimul Hasan, Donald A. Adjeroh, ProximityBased Sentiment Analysis, IEEE , 2011.
[8] Lizhen Liu, Xinhui Nie, Hanshi Wang, Toward a
Fuzzy Domain Sentiment Ontology Tree for Sentiment
Analysis 5th International Congress on Image and
Signal Processing (CISP) ,2012.
[9] Mostafa Karamibekr, Ali A. Ghorbani, Sentiment
Analysis of Social Issues, International Conference on
Social Informatics , 2012
[10] Samatcha Thanangthanakij, Eakasit Pacharawongsakda,
Nattapong Tongtep, Pakinee immanee, Thanaruk
Theeramunkong, An Empirical Study on MultiDimensional Sentiment Analysis from User Service
Reviews, Seventh International Conference on
Knowledge, Information and Creativity Support
Systems, 2012.
[11] Peter Koncz and Jan Paralic, An approach to feature
selection for sentiment analysis, 15th International
Conference on Intelligent Engineering Systems, 2011.
[12] Wang Zuhui Jiang Wei, Online Reviews Sentiment
Analysis Applying Mutual Information, 9th
International Conference on Fuzzy Systems and
Knowledge Discovery, 2012.

Author Profile
Mr Anurag V. Mulkalwar is B.E.(Computer) From
University of Mumbai. Currently pursuing his
M.E.(computer) from University of Mumbai,
Maharashtra, India.
Mrs Kavita M Kelkar is working as Associate
Professor in Department of Computer Engineering,
KJSCE, Vidya Vihar, Mumbai. She has been working
as faculty at KJSCE since June 2000. Mrs Kavita is B
E (Computer) from University of Pune and M E(IT) from
University of Mumbai. She is currently pursuing her Ph.D from
University of Mumbai, Maharashtra, India

Volume 3 Issue 7, July 2014

www.ijsr.net

Licensed Under Creative Commons Attribution CC BY

1742