Prediction of Election Result by Enhanced

Sentiment Analysis on Twitter Data using

Classifier Ensemble Approach
Rincy Jose Varghese S Chooralil
Department of Computer Science and Engineering Department of Computer Science and Engineering
Rajagiri School of Engineering and technology Rajagiri School of Engineering and technology
Ernakulam, India Ernakulam, India
[email protected] [email protected]

Abstract—Sentiment analysis is the computational users can share their thoughts and opinions about
study of opinions, sentiments, evaluations, attitudes, different events and people on the micro-blogging
views and emotions expressed in text. It refers to platform. Therefore, Twitter is considered as a rich
a classification problem where the main focus is to
predict the polarity of words and then classify them source of information for sentiment analysis.
into positive or negative sentiment. Sentiment anal- Sentiment analysis can be considered as the use
ysis over Twitter offers people a fast and effective of natural language processing, text analysis and
way to measure the public’s feelings towards their computational linguistics to identify and extract
party and politicians. The primary issue in previous sentiment information in source materials. Gener-
sentiment analysis techniques is the determination
of the most appropriate classifier for a given classi- ally, sentiment analysis aims to find the attitude
fication problem. If one classifier is chosen from the of a writer with respect to some relevant topic or
available classifiers, then there is no surety in the the overall contextual polarity of a document.The
best performance on unseen data. So to reduce the main task in sentiment analysis is classifying the
risk of selecting an inappropriate classifier, we are polarity of a given text at the document, sentence,
combining the outputs of a set of classifiers. Thus in
this paper, we use an approach that automatically or feature level[1] whether the expressed opinion
classifies the sentiment of tweets by combining ma- in a document, a sentence or an feature is positive,
chine learning classifiers with lexicon based classifier. negative, or neutral.The accuracy of a sentiment
The new combination of classifiers are SentiWordNet analysis is based on how well it agrees with
classifier, naive bayes classifier and hidden markov human judgements. This can be measured by using
model classifier. Here positivity or negativity of each
tweet is determined by using the majority voting precision and recall[2].
principle on the result of these three classifiers. In this paper, we introduces an accurate sen-
Thus we were used this sentiment classifier for timent classifier by combining machine learning
finding political sentiment from real time tweets. classifiers with lexicon based classifier for finding
Thus we have got an improved accuracy in sentiment
political sentiment and sentiment towards new
analysis using classifier ensemble Approach. Our
method also uses negation handling and word sense released movies from real time tweets. Section 2
disambiguation to achieve high accuracy. contains detailed study of the method. Section 3
Index Terms—Sentiment Analysis; WordNet; Sen- includes implementation details and results. Sec-
tiWordNet; Word Sense Disambiguation; Machine tion 4 is the conclusion.
Learning Methods; ensemble Approach.
II. M ETHODOLOGY
I. I NTRODUCTION
The system mainly deals with the tweets ex-
Currently, twitter is becoming one of the most traction and sentiment classification. Here we
popular micro-blogging platforms. Millions of use a sentiment classifier by combining machine
learning classifiers with lexicon based classifier. the extracted tweet. Then it removes all the private
The classifiers using are SentiWordNet classifier, usernames identified by @ user. Then it removes
naive bayes classifier and hidden markov model all the Hash tags identified by the # symbol and
classifier.Our proposed system consists of mainly all the special characters. Refined tweets are then
six modules. They are classified using classification scheme.
1) Data acquisition Negation handling is one of the factors that
2) Pre-processing contributed significantly to the accuracy of our
3) Sentiment Classification using SentiWord- classifiers. A major problem occurring during the
Net sentiment classification is in the negation han-
4) Sentiment Classification using Naive Bayes dling. Since here we use each word as feature,
5) Sentiment Classification using HMM the word “win” in the phrase “not win” will
6) Sentiment Classification using ensemble be contributing to positive sentiment rather than
Approach. negative sentiment .This will leads to the errors in
Block diagram of proposed architecture for sen- classification. This type of error is due to the pres-
timent classification is shown in Figure 1. This ence of “not” and this is not taken into account. To
classifier is used to find political sentiment from solve this problem we applied a simple algorithm
extracted tweets in real time manner. for handling negations using state variables and
bootstrapping. We built on the idea of using an
alternate representation of negated forms[4]. This
algorithm stores the negation state using a state
variable. It transforms a word followed by a nt
or not into “not ”+ word form. Whenever the
negation state variable is set, the words read are
treated as “not ”+ word. When a punctuation
mark is encountered or when there is double
negation, the state variable will reset.We have
applied negation handling to our three classifiers
separately for accurate classification.

C. Sentiment Classification using SentiWordNet

Sentiment classification is done on twitter
data using SentiWordNet(SWN) and WordNet.
Concept of word sense disambiguation is used
Fig. 1. Proposed Architecture
for accurate classification[5]. WordNet is lexical
Tweets extracted in real time manner is serve as database for the English language that groups
input to pre-processing module and then they are English word into set of synonyms called synset.
further classified as positive, negative or neutral. SentiWordNet is an extension of WordNet that
assigns to each synset of WordNet three sentiment
A. Data acquisition numerical scores, positivity, negativity and objec-
To obtain the Twitter feeds in continuous fash- tivity.
ion Twitter streaming API tool is used[3]. It allows WordNet lexical relations are not always a good
real time access to publicly available data on Twit- indicator of polarity detection. Synonyms may
ter. These tweets serve as input to pre-processing have different polarity based on the part of speech
module and then they are further classified as of the word in that sentence. This can be solved
positive, negative or neutral. by using sense-tagged word lists.For that here we
introduce Sentiment Classifier using Word Sense
B. Pre-processing Disambiguation which is based WordNet. SWN
It first identifies the presence of URL using a classifier assigns different sentiment weights to
regular expression and removes all the URLs from different words. It also depends on the how the
word is being used in the sentence i.e. identifica- current probability and the probability given the
tion of “part of speech” for the word is necessary next word to be introduced will cause a transition.
to be classified by SWN classifier. The Markov assumption takes into consideration
the states preceding and succeeding the current
D. Sentiment Classification using Naive Bayes states.
The Naive Bayes classifier is the simplest and F. Sentiment Classification using Ensemble Ap-
most commonly used classifier. Naive Bayes proach
classification model computes the posterior
we introduce an approach that automatically
probability of a class, based on the distribution
classifies the sentiment of tweets by combining
of the words in the document. It relies on very
machine learning classifiers with lexicon based
simple representation of document as Bag of
classifier[6]. Thus we are taking advantages of
words. The model works with the bag of words
these three classifiers( SentiWordNet classifier,
feature extraction which ignores the position of
naive bayes classifier and hidden markov model
the word in the document. It uses Bayes Theorem
classifier) for accurate classification of political
to predict the probability that a given feature set
data. Here positivity or negativity of each tweet is
belongs to a particular label. For twitter sentiment
determined by using the majority voting principle
analysis bigrams from the twitter data are used
on the result of these three classifiers. Senti-
as features on Naive Bayes .It Classifies tweets
WordNet classifier uses opinion lexical resource
into positive and negative labels.
SentiWordNet and WordNet along with Word
Sense Disambiguation for accurate classification
E. Sentiment Classification using HMM of tweets which is extracted in real time manner.
Thus this classifier considers the most suitable
Our sentiment tagging system makes use of the sense of word in its context. Other two classifiers
Viterbi forward backward algorithm to traverse are working based on the training data. Thus we
the states: where each state represents a possible have developed an accurate sentiment classifier for
prediction of the sentiment over the context tra- finding political sentiment from real time tweets.
versed by the algorithm. In the continuity senti-
ment tags are predicted as we move forward with III. I MPLEMENTATION D ETAILS
the algorithm. The Viterbi algorithm is a search We extracted tweets about arvind kejriwal and
algorithm that avoids the polynomial expansion kiran bedi for 3 weeks during the Delhi election
of a breadth first search by trimming the search days. Then we applied sentiment analysis on these
tree at each level using the best ”m” Maximum tweets using classifier ensemble Approach.Tool
Likelihood Estimates (MLE) where ”m” repre- used for Data Acquisition is Twitter Streaming
sents the number of tags of the following word. API. Using Twitter Streaming API, it retrieved
The HMM models make use of two kinds of related tweets for a given query string in real time
probabilities to keep account of the state of the manner and it had been added to .csv file. Then
sentence sentiment for the current: an emission pre-process the extracted tweets and refined tweets
probability keeps track of the sentiment tag given are classified using classification scheme.
the word and its frequency of occurrence in the We extracted 12000 tweets about Arvind Kejri-
training data. The second probability is known as wal for 3 weeks during the Delhi election days.
the transition probability accounts for the current Here we uses three classifiers for accurate classi-
state of the system given the state that the system fication of political data. Here positivity/negativity
was in previously. The advantages of this is that of each tweet is determined by using the majority
if there exists enough evidence to incite belief voting principle on the result of these three clas-
that the system is no longer analysing a positive, sifiers.
negative or neutral sentence, a necessary transition The figure 2 comparing the probability of pos-
may be made to a better or more probable state. itivity of real time extracted tweets about Arvind
The decision for this transition is decided taking Kejriwal and Kiran Bedi for 3 weeks during the
into context the previous state of the system, the Delhi election days.
 TABLE I
ACCURACY OF DIFFERENT CLASSIFIERS

Classifier Accuracy (in % )

SentiWordNet 21.05
Naive Bayes 69.92
HMM 64.06
classifier using ensemble approach 71.48

IV. C ONCLUSION
In this work, we have Implemented a real time
Fig. 2. political sentiment analysis from tweets using classifier twitter sentiment analyser using classifier ensem-
ensemble approach ble approach. Here it combines machine learning
classifiers with lexicon based classifier. Thus we
are taking advantages of these three classifiers
From this graph, it is clear that public sentiment such as SentiWordNet classifier, naive bayes clas-
towards Arvind Kejriwal is higher than that of sifier and hidden markov model classifier for accu-
Kiran Bedi. rate classification of political data. Thus we were
We extracted tweets about two new released developed a novel accurate sentiment classifier for
movies,namely Baahubali and Bajrangi Bhaijaan finding political sentiment from real time tweets.
for 3 weeks. Then we applied sentiment anal- Then Compared political sentiment towards two
ysis on these tweets using classifier ensemble politicians by plotting graphs using results of
approach.Then we got a graph in the following sentiment analysis on extracted twitter data. Again
figure 3. this classifier was used to Compare two new re-
leased movies by the sentiment analysis on twitter
data. Finally,we were analysed improvement in
the accuracy of our classifier compared to the
individual ones.
R EFERENCES
[1] N. D. Valakunde, Dr. M. S. Patwardhan, Multi-Aspect
and Multi-Class Based Document Sentiment Analysis of
Educational Data Catering Accreditation Process IEEE,
International Conference on Cloud and Ubiquitous Com-
puting and Emerging Technologies ,2013.
[2] Bing Liu, Sentiment Analysis and Opinion Mining Mor-
gan and Claypool Publishers, May 2012.
[3] A. Bifet, G. Holmes, B. Pfahringer, MOA-TweetReader:
real-time analysis in twitter streaming data LNCS
Fig. 3. sentiment analysis on new released movies using 6926,Springer-Verlag, Berlin Heidelberg, 2011, pp. 4660.
classifier ensemble approach [4] Vivek Narayanan1, Ishan Arora2, Arjun Bhatia3, Fast and
accurate sentiment classification using an enhanced Naive
Bayes model 2012.
From this graph,it is clear that public sen- [5] Pulkit Kathuria ,Kiyoaki Shirai, Example Based Word
Sense Disambiguation towards Reading Assistant System
timent towards film ’Baahubali’ is higher than The Association for Natural Language Processing,2012.
that of ’Bajrangi Bhaijaan’ upto the end of first [6] Farhan Hassan Khan, Saba Bashir, Usman Qamar, TOM:
week. During the second week and third week Twitter opinion mining framework using hybrid classifica-
tion scheme Decision Support Systems 57 (2014) 245257,
’Bajrangi Bhaijaan’ has higher public sentiment 2013 Elsevier B.V.
than ’Baahubali’.The accuracy improvement in
this sentiment analysis method is shown in Table1.
While considering accuracy, it is observed that
classification using ensemble Approach has higher
accuracy compared to the individual classifiers.