International Conference on Inventive Communication and Computational Technologies

(ICICCT 2017)

Comparative Study of Machine Learning

Techniques in Sentimental Analysis

Bhavitha B K Anisha P Rodrigues Dr. Niranjan N Chiplunkar

Department of CSE Department of CSE Department of CSE
NMAMIT, Nitte NMAMIT, Nitte NMAMIT, Nitte
[email protected] [email protected] [email protected]

Abstract—Sentimental Analysis is reference to the task of sentences. In sentence-level the basic step is to recognize the
Natural Language Processing to determine whether a text sentence as objective or subjective. Suppose sentence is
contains subjective information and what information it subjective, it will decide whether it express negative or a
expresses i.e., whether the attitude behind the text is positive, positive opinion. In aspect-level analysis it aims to categorize
negative or neutral. This paper focuses on the several machine
the sentiment in respect of particular entities.
learning techniques which are used in analyzing the sentiments
and in opinion mining. Sentimental analysis with the blend of Generally, there are two approaches in sentimental
machine learning could be useful in predicting the product analysis. One is by considering symbolic methods and other
reviews and consumer attitude towards to newly launched one by machine learning method. In symbolic learning
product. This paper presents a detail survey of various machine technique, which is categorized according to some learning
learning techniques and then compared with their accuracy, strategies such as learning from analogy, discovery, examples
advantages and limitations of each technique. On comparing we and from root learning. In machine learning technique it uses
get 85% of accuracy by using supervised machine learning unsupervised learning, weakly supervised learning and
technique which is higher than that of unsupervised learning supervised learning. Along with lexicon based and linguistic
techniques.
method, machine learning will be considered as one of the
Keywords— Sentimental analysis, Classifiers, Supervised
learning, Unsupervised learning, SVM; mainly used approach in sentiment classification. The Fig.1
shows the sentiment classification techniques in detail.
I. INTRODUCTION 1.1 Machine Learning Approach
Sentimental Analysis is interpreted as determining the notion In artificial intelligence, machine learning is one of its
of people about distinct existence. Nowadays people are used subsections which are proceeding with algorithm that let
to review the comments and posts on the product which are systems to understand. In machine learning technique it uses
known as opinion, emotion, feeling, attitude, thoughts or unsupervised learning, weakly supervised learning and
behavior of the user. Sentimental Analysis is a method for supervised learning.
identifying the ways in which sentiment is expressed in texts.
Sentimental analysis attempts to divine the posture or notion 1.1.1 Supervised Learning
of a keynoter or author, or author against assertive field or an Supervised machine learning technique associate with
object. There are many claims in sentiment analysis. First is the use of a marked feature set to retain some classification
that, a viewpoint which is treated as positive in one case and function and includes learning of function from the
will be taken as negative in another case. The next claim is experiment along with its input and output. Supervised
that usually people don’t consider their viewpoint in same learning is task of assuming a function labeled trained data
form. Almost of all reviews incorporate with both positive as set. Training data set includes set of training examples; each
well as negative remarks, which can be feasible by and every example consists of couple of an input data as well
interpreting the sentences each at a time. Finding the opinion as expected output.
sites and monitoring them on the web is somewhat difficult.
So there will be a need of robotic opinion mining as well as a 1.1.2 Weakly-Supervised and Unsupervised Learning.
summarization system. In practical these supervised methods cannot be always
In sentiment analysis there are three classification levels: used, because it needs labeled corpora but they are not
document-level classification, sentence-level classification available all time. Another option for machine learning is
and feature-level sentiment analysis. In document-level weakly-supervised and unsupervised methods which do not
classification the main intention is to classify an opinion in require pre-tagged data. Weakly supervised learning consists
the whole document as positive and negative. It speculates of large set of unlabeled data and small set of labeled data.
entire document as a single unit. The aim of sentence-level Unsupervised method includes learning device for the input
analysis is to categorize emotion expressed in respective

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 216

 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)

classifier are, it is easy to understand and to interpret. This

classifier requires small data preparation. But these concepts
can create complicated trees that do not generalized easily.

2.2 Linear Classifier

In linear classifier, for classifying input vectors to classes
they use linear decision margins. There are many types of
linear classifiers. Support vector machine is one of them. This
classifier provides a good linear scatters between various
classes.
2.2.1 Neural Network
Neural network includes numerous neurons in which this
neuron is its elemental unit. Multilayer neural network were
used with non-linear margins. The results of the neuron in the
previous layer will be given as input for the next layer. In this
Figure 1 Sentiment classification techniques. type of classifier training of data set is more complicated,
because the faults must be back-propagated for various
and there is no expected output values are given. Some layers.
examples for unsupervised learning approach are cluster
analysis, expectation-maximization algorithms. These 2.2.2 Support Vector Machine
algorithms use Dictionary based approach to compile Support Vector Machine (SVM) is known as the best
sentimental text. A dictionary contains antonyms and classifier that provides the most accurate results in speech
synonyms for every word. So this approach is used to find classification problems. They achieved by creating a
seed sentimental word according to antonym and synonym hyperplane with maximal Euclidean distance for the nearest
arrangement of a dictionary. A small set of words is initially trained examples. Support Vector Machine hyperplane are
collected with known positive or negative coordination. This completely resolved by a comparatively minute subset of the
iterating procedure completed until there is no new words are trained data sets which are treated as support vectors. The
found. remaining training data sets have no access on the qualified
classifier. So for the purpose of text classification, the
1.2 Lexicon Based Approach classifier SVMs have been applied successfully and also used
In lexicon based method it supports a lexicon to achieve in different sequence processing application. SVMs are used
sentiment classification through weighting and counting in hypertext and text classification since they do not require
sentiment associated words has to be calculated and labeled. labeled training data set.
To assemble the viewpoint list there are three major methods 2.3 Rule Based Classifier
are considered: dictionary-based method, corpus-based
method and the manual opinion approach. As the name indicates in rule based classifiers, data set is
designed along with a group of rules. In rules left hand side
indicates the condition of aspect set and right hand indicates
II. Sentiment Classification Based on Machine the class label.
Learning Methods 2.4 Probabilistic Classifier
These classifiers use various forms for categorization.
In machine learning technique it uses unsupervised This variety of forms takes each and every class as part of
learning, weakly supervised learning and supervised learning.
that mixture. All various elements are the productive model
2.1 Decision Tree Classifier in which it gives the probability of inspecting a distinct word
In Decision Tree classifier, the interior nodes were marked for that element. These classifiers are also known as
with features and edges that are leaving the node were named generative classifiers. Some of the probabilistic classifiers are
as trial on the data set weight. Leaves in the tree are named Naïve Bayes, Bayesian Network and Maximum Entropy.
by categorization. This categories whole document by
starting at the root of the tree and moving successfully down
2.4.1 Naïve Bayes Classifier
through its branches till a leaf node is reached. Learning in
A Naive Bayesian classifier is one of the familiar
decision tree adopts a decision tree classifier as an anticipated supervised learning techniques which are frequently used for
model in which it maps information of an item to conclusions classification purpose. Their classifier is named as naive
of that item’s expected value. In decision tree large amount of since it considers the contingency that are actually linked are
input can be figure out by using authoritative computing not depending on the further. Calculation of whole document
assets in finite time. The main advantages of decision tree

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 217

 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)

feasibility would be the substance in aggregation of all the

feasibility report of single word in the file. These Naïve
Bayesian classifiers were frequently applied in sentiment # Machine Advantage Disadvantage
categorization since they are having lower computing power Learning
Classifier
when comparing to the other approach but independence
assumptions will provide inaccurate results. 1 KNN It is simple and also It requires more time
used for multiclass to categorize when
categorization of huge number data are
2.4.2 Bayesian Network document. inclined.
The main disadvantage of Naïve Bayes classifier is its
Takes lot of memory
independent assumption of aspects in data sets. This for running a process
assumption is the reason for start of using Bayesian Network.
This Bayesian network is directed non-cyclic graph where 2 Decision Tree This is very fast in It has problem that it is
learning data set. difficult handle data
nodes correspond to variables and those edges are correspond with noisy data
to conditional independency. In text classification Bayesian Easy for
Network is not usually used since it is expensive in understanding Over fitting of data
purpose
computation.

2.4.3 Maximum Entropy 3 Naïve Bayesian Simple and work Performs very poorly
Maximum Entropy classifier is parameterized by a weight well with textual as when feature set is
set that are used to associate with the joint-future, well as numerical highly correlated.
data.
accomplished by a trained data set by encoding it. This It gives relatively low
Maximum Entropy classifier appear with the group of Easy to implement classification
classifiers such as log-linear and exponential classifier, as its performance for large
Computationally data set.
job is done by deriving some data sets against the input cheap
binding them directly and the result will be treated as its Independent
assumption of attribute
exponent. may lead to inaccurate
result.
2.5 K- Nearest Neighbor Classifier
K-Nearest Neighbor is a unsupervised learning
algorithm for text classification. In this algorithm the entity is 4 Support Vector High accuracy even Problems in
classified with various trained data set along with their Machine with large data set representing document
into numerical vector
nearest distance against each entity. The advantage with this Works well with
algorithm is its simplicity in text categorization. It also works many number of
dimensions
well with multi-class text classification. The main drawback
of KNN is it necessitate with large amount of time for No over fitting
categorizing entities where huge data set are inclined.

In table 1 it shows the comparative observation Table 1: Comparison between machine learning methods
between different machine learning techniques.

In paper [6], [7], and [8] authors proposed aspect based

III. Literature Survey sentimental analysis. In the paper [6], during experiment four
From the past set of years, many articles, papers and data sets were used to test the SVM model. Authors have
books have been written on sentimental analysis. At the same compared Maximum Entropy classifier method for feature
time some researchers focus more on specific burden like extraction with SVM method and they have concluded that
finding the subjectivity expression, subjectivity clues, SVM Method superior in terms of recall and precision rates.
subjective sentence, topics, and sentiments of words and In paper [7], authors proposed different approach which
extracting sources of opinions, while others target is on bunch up the benefits of Senti-WordNet, dependency parsing,
assigning sentiments to whole document. All analyzers of and co reference resolutions are well organized for the
sentiment analysis have adapted several methods to purpose of sentimental analysis. This was done by using
automatically predict the expression, sentiments of words or a Support Vector Machine classifier. The paper [8] presents the
document. The data set for sentimental analysis considered comparison between most likely used approaches for
are movie, product review or social media data from the sentiments like Naive Bayes, Max Entropy, Boosted trees and
source of internet. They use pattern based approaches, Random Forest algorithms. By using Random Forest
Natural Language Processing (NLP) as well as machine Classifier in sentimental classification it apparently shows
learning techniques. that the result is obtained with greater accuracy and
performance. And also this classifier is easy to understand

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 218

 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)

and the performance would be improved with a time period. its capability to change and to bring qualified design for exact
Because of aggregation of decision tree the accuracy was purpose together with its content. In the paper [14], authors
improved with higher rate. On behalf of it, the classifier proposed Naïve Bayesian classifier to analyze the sentences.
requires high processing power and training time. In the Their experimental result shows that Naive Bayesian
paper they conclude that, if accuracy has the first classifier model which has acceptable achievement for
consideration then Random Forest classifier must be distinct Social Network Site and for large data set in which it
preferred even though it uses high learning time. Due to consists of long comments.
lesser processing condition and small memory usage the
The challenges for automated analysis of tweets are (i) a
Naïve Bayesian classifier was applied. Alternatively the Max
single word are considered as subjective in one case and the
Entropy classifier is used because it requires smaller training
same word will be treated as objective in another case (ii)
time with large memory and processing time. From these
same sentence with different discipline (iii) sarcasm
papers we can conclude that support vector machine yields
sentences (iv) in some case a whole sentence will not be
higher accuracy in classification of product reviews. But
considered because only little part of the text gives the
authors have not dealt with sarcastic sentences and
complete contention (v) negative word can be expressed in
comparative sentences.
distinct way in contrast to words like never, no, not etc.
In the paper [9], [10] and [19], a movie review is analyzed Analyzing such contradiction is challenging. That means the
by linking machine learning application with Natural twitter analysis still more improved by considering these
Language Processing technique. In paper [9], authors applied challenges.
SVM and Naïve Bayes classifier in analyzing the movie
In the paper [15],[1],[5] and [16], authors discussed about
sentiments. By this categorization they conclude that linear
existing models for analyzing sentiments of unorganized data
Support Vector Machine outperforms the Naïve Bayesian in
which were posted on social media. Analyzing sentiments it
case of accuracy. In the paper [10], authors demonstrate how
doesn't consider objective sentences. Authors proposed
machine learning technique was used to understand the
approach for sentence classification or sentence of
Malayalam movie comment. For classifying the sentiments
documents. For this purpose [15][1] used SVM, Naive Bayes,
two machine learning approaches are used; they are Support
Part of Speech and SentiWordNet techniques. From the result
Vector Machine and CRF along with rule based approach. In
they conclude that machine learning classifier for instance
[19] , author compared two most frequently used supervised
Naïve Bayesian and Support Vector Machine yield the
machine learning approaches SVM and Naive Bayes for
highest efficiency. And also act as basic standard model for
sentiment classification of reviews. The result shows that
all classification. But lexicon approach is very aggressive in
SVM has misclassified more number of data points as
sentiment. For this problem deep learning approach was
compared to Naive Bayes and Naive Bayes approach
introduced. By comparing lexicon based approach with
outperformed the SVM when there are less number of
machine learning, the two classifiers such as SVM and Naive
reviews. Authors suggested that there will be a considerable
Bayes provides higher accurate values in classification.
scope of improving in the creation of corpus and effective
preprocessing and feature selection Researchers are still The paper [5],[16], depicts that for the purpose of
working for the automated analysis of score and rating of the sentiment analysis they use classifier of Support Vector
movie reviews. Machine (SVM) on the benchmark feature sets to scale the
sentiment classifier. To extract the classical features of data
In paper [11],[12],[13] and [14] authors describe the
set, weighting scheme like N-grams and other weighting
various tools used in sentimental analysis of twitter data.
scheme were applied. For selecting requested feature to the
Since the opinions in the twitter are heterogeneous, highly
classification they go into the Chi-Square weight feature. In
unstructured and along with these it includes positive,
the present method the structure involves preprocessing,
negative or neutral in different situation, it is important
aspect selection, aspect extraction and finally the data sets are
analyze the sentiments. In the paper [11] authors used
classified. Since SVM is having great potential to hold big
lexicon-based methods for classification but it requires small
data set, the text classification is done with good result. Other
effort in individual labeled text document. In the paper [12]
mentionable advantage is SVM is robust with sparse set of
they have shown the outline of recommended methods along
examples. N-gram, unigram and other weighting schemes are
with its most recent advancements in the same field. As a
input to the SVM classifier. Based on these weighting
result, authors concluded that unsupervised machine learning
schemes some standard data sets are routine to train the
techniques fails to provide better achievement in sentiment
classifier. In the experimental result it found that unigrams
classification than that of supervised learning. In paper [13],
outperforms bigram and n-gram model. To improve the
they describe the various tools used in sentimental analysis
accuracy in classification authors suggested using Chi-Square
and some approaches for text classification. In this method
aspect selection scheme.
they use hybrid approach which uses the aggregation of both
lexicon based and machine learning techniques. This In paper [17], the author presents comparative analysis of
compound approach then leads to obtain higher classification currently used techniques for sentiment analysis which
performance. The fundamental usage of machine learning is includes lexicon-based and machine learning techniques

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 219

 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)

along with cross-lingual and cross-domain method. Finally

they conclude that machine learning approach provides # Positive Negative Total
experimen
highest accuracy and lexicon-based approach are highly t
competitive and also manually needs more effort in labeling 1 50 50 100
document. 2 100 100 200
Table 2 shows machine learning approach SVM yields 3 150 150 300
highest accuracy as compared to Naïve Bayes and Senti- 4 200 200 400
WordNet. 5 400 400 800
6 550 550 1100
7 650 650 1300
TP FP FN TN Accurac
y 8 800 800 1600
Senti- 148 91 52 109 64.25% 9 900 900 1800
WordNet
10 1000 1000 2000
NB 156 81 44 119 68.75%
SVM 135 51 65 149 71.00% Table 4: Collected reviews

The accuracy obtained by using three algorithms are

Table 2: Performance comparison of learning methods shown in table 5. They have done 10 experiments for each
with Senti-WordNet approach. Result shows that even data is either small or large
SVM provides higher accuracy than NB and kNN.

# # reviews SVM (%) Naïve kNN

Below table 3 shows performance contrast between experim Bayes (%) (%)
different sentiment classification approaches. ent
1 50 60.07 56.03 64.02
Method Data set Accuracy 2 100 61.53 55.01 53.97
3 150 67.00 56.00 58.00
Machine SVM Movie 86.40% 4 200 70.50 61.27 57.77
learning reviews 5 400 77.50 65.63 62.12
CoTraining Twitter 82.52% 6 550 77.73 67.82 62.36
SVM 7 650 79.93 64.86 65.46
Deep Standard 80.70% 8 800 81.71 68.80 65.44
learning benchmark 9 900 81.61 71.33 67.44
Lexicon Corpus Product 74.00% 10 1000 81.45 75.55 68.70
based Reviews
Dictionary Amazon ---
Cross-lingual Ensemble Amazon 81.00% Table 5: Accuracy obtained after testing data set
Co-Train Amazon, 81.30%
IT168
EWGA IMDb >90% IV. Conclusion
moview
review This paper includes outline of current works that done on
CLMM MPQA, 83.02% sentimental classification and analysis. From the survey we
NTCIR,ISI can conclude that supervised learning methods like Naive
Cross-domain Active Book, 80% of Bayesian and Support Vector Machine are considered as
learning DVD, average
Thesaurus Electronics,
standard learning method. Support Vector Machine provides
SFA Kitchen excellent accuracy as compared to many other classifiers. In
terms of accuracy we concluded that with small feature set
Naive Bayes performs well, if large feature set is taken then
Table 3: Performance comparison of sentiment SVM will be the best choice. Lexical based approaches are
classification technique ideally aggressive because it requires manual work on
document. Maximum Entropy also performs better but it is
In paper [18], they have taken online movie reviews for suffered from over fitting. Many researches implemented
analyzing sentiments. For classification they used three opinion mining different techniques but still there is a need of
supervised learning approaches such as Naïve Bayes, SVM automated analysis which addresses all the challenges of
and kNN. Experimental results show that SVM method beat sentimental analysis simultaneously. A more innovative and
the kNN and Naïve Bayes approaches. Table 4 shows the effective techniques required to be invented which should
collected reviews for sentiment classification. overcome the current challenges like classification of indirect
opinions, comparative sentences and sarcastic sentences.

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 220

 International Conference on Inventive Communication and Computational Technologies
(ICICCT 2017)

REFERENCES
[1] Hailong Zhang, Wenyan Gan, Bo Jiang, “Machine Learning and [16] M. Thelwall, K. Buckley, and G. Paltoglou, “Sentiment in twitter
Lexicon based Methods for Sentiment Classification: A Survey”,978-1- events,” J. Am. Soc. Inf. Sci. Technol., vol. 62, no. 2, pp. 406–418, Feb.
4799-5727-9/14 $31.00 © 2014 IEEE. 2011.
[2] Walaa Medhat a,*, Ahmed Hassan b, Hoda Korashy, “Sentiment [17] Hailong Zhang, Wenyan Gan, Bo Jiang, “Machine Learning and
Lexicon based Methods for Sentiment Classification: A Survey”, 978-
analysis algorithms and applications: A survey”, Ain Shams
Engineering Journal (2014) 5, 1093–1113. 1-4799-5727-9/14 $31.00 © 2014 IEEE DOI 10.1109/WISA.2014.55.
[18] P.Kalaivani, Dr. K.L.Shunmuganathan, “Sentiment Classification Of
[3] Xing Fang* and Justin Zhan, “Sentiment analysis using product review Movie Reviews By Supervised Machine Learning Approaches”, ISSN :
data”, Fang and Zhan Journal of Big Data (2015) 2:5 DOI 0976-5166 Vol. 4 No.4 Aug-Sep 2013.
10.1186/s40537-015-0015-2
[19] Suchita V Wawre1, Sachin N Deshmukh2 , “Sentiment Classification
[4] Kaijie Guo, Liang Shi*, Weilong Ye, Xiang Li, “A Survey of Internet using Machine Learning Techniques”, International Journal of Science
Public Opinion Mining”, 978-1-4799-2030-3 /14/$31.00 ©2014 IEEE. and Research (IJSR) Volume 5 Issue 4, April 2016.
3/
[5] Nurulhuda Zainuddin, Ali Selamat, “Sentiment Analysis Using Support
Vector Machine”, 978-1-4799-4555-9/14/$31.00©2014 IEEE. [20] Walaa Medhat a,*, Ahmed Hassan b, Hoda Korashy, “Sentiment analy-
sis algorithms and applications: A survey”, Ain Shams Engineering
14/$31.00 ©2014 IEEE. METHODOLOG Journal (2014) 5, 1093–1113.
[6] Chuanming Yu, “Mining Product Features from Free-Text Customer
Reviews: An SVM-based Approach”, iCISE 2009 December 26-28, [21] Pratiksha Y. Pawar and S. H. Gawande, “A Comparative Study on
2009, Nanjing, China. Different Types of Approaches to Text Categorization”, International
[7] Raisa Varghese, Jayasree M, “Aspect Based Sentiment Analysis using Journal of Machine Learning and Computing, Vol. 2, No. 4, August
Support Vector Machine Classifier”, 978-1-4673-6217-7/13/$31.00_c 2012.
2013 IEEE
[8] Amit Gupte, Sourabh Joshi, Pratik Gadgul, Akshay Kadam, [22] S. B. Kotsiantis, “Supervised Machine Learning: A Review of Classifi-
“Comparative Study of Classification Algorithms used in Sentiment cation Techniques”, Informatica 31 (2007) 249-268 249.
Analysis”, International Journal of Computer Science and Information
Technologies, Vol. 5 (5) , 2014, 6261-6264.
[9] Gautami Tripathi1 and Naganna S, “ Feature Selection And
Classification Approach For Sentiment Analysis”,Machine Learning
and Applications: An International Journal (MLAIJ) Vol.2, No.2, June
2015.
[10] Deepu S. Nair, Jisha P. Jayan, Rajeev R.R, Elizabeth Sherly,
“Sentiment Analysis of Malayalam Film Review Using Machine
Learning Techniques”, 978-1-4799-8792-4/15/$31.00_c-2015-IEEE

[11] Vishal A. Kharde, S.S. Sonawane , “ Sentiment Analysis of Twitter

Data: A Survey of Techniques”,International Journal of Computer
Applications (0975 – 8887) Volume 139 – No.11, April 2016.
[12] Neha S. Joshi, Suhasini A. Itkat, “A Survey on Feature Level Sentiment
Analysis”, International Journal of Computer Science and Information
Technologies, Vol. 5 (4) , 2014, 5422-5425.
[13] Alessia D’Andrea, Fernando Ferri, Patrizia Grifoni, Tiziana Guzzo
,“Approaches, Tools and Applications for Sentiment Analysis
Implementation”,International Journal of Computer Applications (0975
– 8887) Volume 125 – No.3, September 2015.
[14] Shun Yoshida, Jun Kitazono, Seiichi Ozawa, Takahiro Sugawara,
Tatsuya Haga and Shogo Nakamura, “Sentiment Analysis for Various
SNS Media Using Naive Bayes Classifier and Its Application to
Flaming Detection”,978-1-4799-4540-5/14/$31.00 ©2014 IEEE .
[15] Jalaj S. Modha*, Prof & Head Gayatri S. Pandi, Sandip J. Modha, “
Automatic Sentiment Analysis for Unstructured Data”, IJARCSSE
Volume 3, Issue 12, December 2013.

978-1-5090-5297-4/17/$31.00 ©2017 IEEE 221