IRJET- Twitter Sentimental Analysis for Predicting Election Result using LSTM Neural Network

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 3665
Twitter Sentimental Analysis for Predicting Election Result
using LSTM Neural Network
Dipak Gaikar1, Ganesh Sapare2, Akanksha Vishwakarma3, Apurva Parkar4
1 Asst. Professor, Dept. of Computer Engineering, Rajiv Gandhi Institute of Technology, Mumbai, Maharashtra
2 B.E. student, Dept. of Computer Engineering, Rajiv Gandhi Institute of Technology, Mumbai, Maharashtra
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - In recent years, social media has emerged as a
powerful widespread technology and caused a huge impact
on the public debate and communication in the society.
More recently, micro-blogging services (e.g., twitter) and
social network sites are presumed to have the potential for
increasing political participation. Among the most popular
social media sites, Twitter serves as an ideal platform for
users to share not only information in general but also
political opinions publicly through their networks, political
institutions have also begun to use this media for the
purpose of entering into direct dialogs with citizens and
encouraging more political discussions. So we build a model
that can analyze these data and extract sentiment that can
help us determine the outcome of the election. The process
consists methods such as extraction of tweets from twitter
using API, data cleaning to get exact data, training the
LSTM (Long Short Term Memory) classifier using labelled
dataset and testing it to perform sentimental analysis for
classification and then representation of result. Further, a
comparison is made among the candidates over the type of
sentiment by table and bar graph.
Key Words: machine learning, twitter, social media,
prediction, recurrent neural networks, sentiment analysis,
embedding, keras
1. INTRODUCTION
Over the last decade with the arrival of social media, the
efforts to determine people’s point of view over a
particular event or a topic have garnered a wide research
interest in natural language processing (NLP) and thus
introduced “sentiment analysis.” [1] Many social
networking websites and micro blogging websites in
today’s world has become the biggest web destinations for
people to communicate with each other, to express their
perspective about products or movies, share their daily
life experience and present their opinion about real time
and upcoming events, such as sports or movies, etc.
Analysis of public’s sentiment requires a huge data. For
achieving a large, diverse dataset of current public opinion
or sentiments, Twitter could be used as a valuable
resource that lets the users to send and read small text
messages called “Tweets” [1]. Basically twitter allows
users to post brief and quick real-time updates regarding
various activities like sharing, forwarding and replying
messages quickly which allows the quick spread of news
or information. The wide use of hash tags also makes it
easy to search for tweets dealing with a specific subject,
thus making it quite a convenient way to gather data.
Using sentiment analysis for predicting an election’s result
is practically challenging to train a classification model to
perform sentiment analysis on tweet streams for a
dynamic event such as an election [7]. Implementing this
model have certain key challenges such as changes in the
topics of conversation and the people about whom social
media posts express opinions. [5] For doing this, we first
created a LSTM classifier (positive versus negative versus
neutral) which is a modified version of RNN (Recurrent
Neural Network) for analyzing opinions about different
election candidates as expressed in the tweets. We then
train our model for each candidate separately. The
inspiration for this separation comes from our observation
that the same tweet on an issue can be positive for one
candidate while negative for another. In fact, a tweet’s
sentiment is candidate- dependent. To train the model we
used over a 15,000 labelled tweets which are labelled as
positive, negative and neutral. In the project. using twitter
API, we extracted 40000 real time tweets from Jan 2019 to
Mar 2019 relating to names of Indian political parties such
as #BJP, #Congress, #TMC, #BSP, #Chowkidar,
#BJP4India, # ChowkidarChorHai, etc.
The rest of the paper is organized as follows: Section II
presents the objectives for implementing the system.
Section III provides a related work on sentiment analysis
lately. Section IV explains the proposed model. Section V
describes the methodology to determine the sentiments
associated with different candidates. Section VI is for
visualization and results and finally, section VII is based
around conclusions.
2. OBJECTIVES
 To predict the popularity of a candidate of a political
party and therefore extrapolate their chances of

winning the election by utilizing sentimental analysis
of social media data.
 Identify the class to which the tweet belongs based on
the sentiment of the tweet.
 Perform an efficient pre-processing unit to make data
suitable for analysis.
 Obtain the labelled dataset to train the classifier
model.
 Create a LSTM model to perform classification process
by training the model with training data to fit and test
to model to evaluate the test data.
 Evaluate the results obtained using bar graphs and pie
charts and thus comparing it with the result of actual
election.
3. RELATED WORK
In [1] they performed data set creation by first collecting
data using twitter streaming API and stored them in CSV
file, then pre-processing is done to remove special
characters and URLs, followed by data labelling which is
done manually using hash-tag labelling and then using
VADER tool which is lexicon and rule based sentiment
analysis tool. Here [2] they obtained the live twitter feeds
using twitter streaming API tool and processed the data to
remove URLs, #tags and special characters. Negation
handling is applied to differentiate the meaning of words
and sentiment classification is done using various
approaches such as sentiWordNet, Naïve Bayes, HMM and
ensemble approach. In [3] they fetched raw tweets in
Hindi language using twitter archiver. Pre-processing of
tweets is done and algorithms are applied to calculate the
polarity of tweets, then sentimental analysis & prediction
using NB, SVM and dictionary based. Here [4] they
propose a unified CNN-RNN framework for multi-label
image classification. The CNN-RNN framework learns a
joint image- label embedding to characterize the semantic
label dependency as well as the image-label relevance. For
which they conclude the experimental results on public
benchmark datasets demonstrate of system achieves
better performance than the state-of-the-art multi-label
classification models. In [5] they use the multi-task
learning framework to jointly learn across multiple related
tasks. The recurrent neural network has three different
mechanisms of sharing information to model text with
task-specific and shared layers. The entire network is
trained jointly on all these tasks. Experiments on four
benchmark text classification tasks show that the
proposed models can improve the performance of a task
with the help of other related tasks. In [6] they Proposed
pattern based approach for sentiment quantification in
twitter. They defined two metrics to measure the
correctness of sentiment detection and proved that
sentiment quantification can be more meaningful task
than the regular multi-class classification.
4. PROPOSED MODEL
Proposed architecture for sentiment classification is
shown in Figure 1. The system deals with the tweets
extraction and sentiment classification. It consists of
following modules.
1. Data collection
2. Preprocessing
3. Train the classifier
4. Sentiment Classification
5. Data visualization
4.1 Data collection
Accessing tweets from Twitter is the primary
requirement for building a dataset to get processed and
extract information. [8] Twitter allows its users to retrieve
real time streaming tweets by using twitter API. We
propose to use the python library Tweepy which has the
API to extract the tweets through authenticating
connection with Twitter server. While collecting tweets
we filter out retweets. The tweets are then stored in csv
file for further processing.
4.2 Data Preprocessing
The data extracted from twitter contains lot of special
characters and unnecessary data which we not require [1].
If data is not processed beforehand, it could affect the
accuracy as well as performance of the network down the
lane. So it is very important to process this data before
training. We need to get rid of all the links, URLs and @
tags. Pre-processing also includes removal of stop words
from the text to make analysis easier.
4.3 Train the classifier
To train the classifier model we will be using a labelled
dataset in which every single tweet is labelled as positive
or negative based on sentiment [5]. The dataset is input to
the LSTM model in which the text input is embedded and
operations are performed to fit the model. One of the most
common problem in training deep neural network is over-
fitting, because the model performs exceptionally well on
training data but poorly on test data. This happens as our
classifier algorithm tries to fit every data point in the input
even if they represent some randomly sampled noise.
Once the model is trained we can determine its accuracy
by testing it to classify the sentiments of test data.
4.4 Sentiment Classification
Sentiment Analysis is the most common text
classification technique that analyses an incoming
message and tells whether the underlying sentiment is
positive, negative or neutral. After the model is trained we

use it to classify the dataset which is extracted from
twitter and stored in a csv file. Sentiment values are
assigned to words that describe the positive, negative and
neutral attitude of the speaker.
4.5 Data visualization
The final step of this process is to take in the classified
tweets and generate pie chart, group bar graph and table
to visualize the results. The most frequent words in the
dataset can be used to generate word cloud.
Fig -1: Proposed model
5. METHODOLOGY
We will use Recurrent Neural Networks, and in particular
LSTMs, to perform sentimental analysis. The model will
use a labelled dataset for training and will classify the test
data which is collected from twitter. The test data contains
the tweets which is pre-processed and then converted into
vectors. These vectors is given as input to classifier.
5.1 Dataset Description
The dataset consists of 15 thousand tweets which are
classified as positive, negative and neutral tweets. The
positive tweets are labelled as 1, negative tweets are
labelled as -1 and neutral tweets as 0. Dataset are
structured in a way that each line contains only one tweet.
Moreover, all tweets in the provided dataset have already
been pre-processed so that all words (tokens) are
separated by a single whitespace and the smileys (labels)
have been removed. All the links and unwanted data is
also removed. Thus, all the tweets are stored in csv file
with UTF-8 encoding.
5.2 Word Embedding
Twitter tweets contain a lot of relevant data to generate
insights, which can also be used to generate political data.
These tweets cannot be processed as it is, because it is
difficult for machine learning algorithms to process string
inputs. So we need to transform them into numerical or
vector format. Word embedding are in fact a class of
techniques where individual words are represented as
real-valued vectors in a vector space. Each word is
mapped to a particular vector and the vector values are
trained in a way that resembles a neural network. keras
offers an embedding layer that can be used for processing
text data. This layer is initially assigned with random
weights and will learn an embedding for all of the words in
the training data set. Also this format can be used to find
semantic relationships. Embedding layer is a flexible layer
that can be either used to learn a word embedding that
can be saved and used in another model later or we can
include it in a deep learning model where learning takes
place with the model.
Three important arguments need to be specified for
using keras Embedding which are input dimensions,
which gives the size of vocabulary in our text data, output
dimensions which give the size of vector space in where
our words will be embedded, and input length, that is
length of input sequences.
For e.g.:
eval = Embedding (150, 30, input length=45)
5.3 Neural Network
Neural network models have been demonstrated to be
capable of achieving significant performance in text
classification. [5] LSTM is a simple recurrent neural
network. RNN analyzes the sentences word by word and
stores the semantics of all the previous sentences in a
fixed-sized hidden layer. The advantage of RNN is the
ability to better store the information. LSTM’s are
improvement over RNN that can learn and remember
long-term dependencies.
5.3.1 Recurrent Neural Network (RNN)
Recurrent Neural Networks understands each word
based on the understanding of previous words. They are
networks with loops in them, allowing information to
persist. In Recurrent Neural Network the output from
previous step is given as an input to the current step. [6]
In traditional neural networks, all the inputs and outputs
are independent of each other, but sometimes it is
required to predict the next word of a sentence, for that
previous words are required and hence there is a need to
remember the previous words. To overcome the
drawback, RNN is used which has solved this issue with
the help of a Hidden Layer. The important feature of RNN

is it has hidden state, which remembers some information
about a sequence.
Fig -2: Recurrent Neural Network (RNN)
In figure 2, a chunk of neural network, A, looks at some
input xt and outputs a value ht. The information is to be
passed along the loop from one step of the network to the
next [8]. A recurrent neural network can be viewed as
multiple copies of the same network, each passing a
message to a successor. This chain-like nature reveals that
recurrent neural networks are related to sequences and
lists. They form natural architecture of neural network to
use for such data. RNN has got an incredible success in
solving a variety of problems: speech recognition,
language modelling, translation, image captioning, etc.
A usual RNN has a short-term memory, so it can’t
remember past data for a long period of time [8]. For large
datasets where we have long texts we can’t use RNN. This
problem can be solved with the help of LSTM.
5.3.2 Long short term memory (LSTM)
Long Short-Term Memory networks are an extension
for recurrent neural networks, which extends their
memory. It is very necessary to learn from important
experiences that have very long time gaps in between.
They are capable of learning long-term dependencies.
LSTM’s work extremely well on a large variety of
problems, and therefore are now widely used. They are
specially designed to avoid the long-term dependency
problem. To remember information for long periods of
time is practically their default behavior.
LSTMs have same chain like structure, but the
repeating module has a different structure. Instead of
having a single neural network layer, there are four,
interacting in a very special way [4]. The figure 3 below
shows the internal structure of each cell in LSTM.
Fig -3: Long Short Term Memory (LSTM)
The key to LSTMs is the cell state, the horizontal line
running through the top of the diagram [4]. The cell state
is kind of like a conveyor belt. It runs straight down the
entire chain, with only some minor linear interactions. It’s
very easy for information to just flow along it unchanged.
The LSTM does have the ability to remove or add
information to the cell state, carefully regulated by
structures called gates. Gates are a way to optionally let
information through. They are composed out of a sigmoid
neural net layer and a pointwise multiplication operation.
The sigmoid layer outputs numbers between zero and one,
describing how much of each component should be let
through. A value of zero means “let nothing through,”
while a value of one means “let everything through!”
LSTM’s enable RNN’s to remember their inputs over a
long period of time [8]. This is because LSTM’s contain
their information in a memory that is much like the
memory of a computer because the LSTM can read, write
and delete information from its memory.
An LSTM has three of these gates, to protect and
control the cell state [4]. The first step in LSTM is to decide
what information is going to be thrown away from the cell
state. This decision is made by a sigmoid layer called the
“forget gate layer.” It is represented as 𝜎. It then looks at
ht-1 (output of previous layer) and xt (input), and outputs a
number between 0 and 1 for each number in the cell state
Ct-1 as shown in eq (1). 1 represents “completely keep this”
while a 0 represents “completely get rid of this.”
𝐶𝑡 = tanℎ ( 𝑤𝑐 · [ℎ 𝑡−1, 𝑥𝑡] + 𝑏 𝐶) (1)
The next step is to decide what new information we’re
going to store in the cell state. This has two parts. First, a
sigmoid layer called the “input gate layer” decides which
values we’ll update. Next, a tanh layer creates a vector of
new candidate values, Ct that could be added to the state.
In the next step, we’ll combine these two to create an
update to the state. Following are the equations:
𝑓𝑡 = 𝜎( 𝑤 𝑓. [ℎ 𝑡−1, 𝑥𝑡] + 𝑏 𝑓) (2)
𝑖𝑡 = 𝜎( 𝑤𝑡. [ℎ 𝑡−1, 𝑥𝑡] + 𝑏𝑖) (3)
It now updates the old cell state, Ct-1, into the new cell
state Ct. the previous steps already decided what to do. It
then multiplies the old state by ft, forgetting the things it
decided to forget earlier. Then we add it*Ct as shown in eq
(2) and (3). This is the new candidate values, scaled by
how much we decided to update each state value.
Finally, we need to decide what we’re going to output.
This output will be based on our cell state, but will be a
filtered version. First, we run a sigmoid layer which
decides what parts of the cell state we’re going to output.
Then, we put the cell state through tanh (to push the
values to be between −1 and 1) and multiply it by the
output of the sigmoid gate, so that we only output the
parts we decided to.

In the case of the language model, this is where we’d
actually drop the information about the old subject’s
information and add the new information using Ot and ht
eq (4) and (5) as we decided in the previous steps. The
equations are:
𝑂𝑡 (4)
(5)
6. RESULTS
Once the classification of tweets is done it is stored in a
csv file labelled as positive, negative and neutral. These
data can be visualized using bar graphs, pie charts, tables,
word cloud, etc. There are various methods that are
available in python and its libraries. The number of tweets
which are extracted for a particular party may not be the
same as for another party. Also the number of positive,
negative and neutral tweets may be different for all the
candidates, so we cannot directly conclude the winner
based on the number of positive tweets, since the data set
count may be biased towards a particular candidate.
Consider a scenario where 10,000 tweets for Congress are
mined out of which only 3,000 are positive and 7,000
tweets are mined for BJP out of which 4,000 are positive
then direct comparison of positive tweets would yield
incorrect results since the percentage of positive tweets
for BJP is much higher. We have to use the ratio of number
of positive to the total count, for which we have
considered total count of minimum 1000 tweets. The
table 1 shows the positive to total count ratio.
Table -1: PvT ratios of political parties
Party Positive Negative Neutral Total
Pos/Total
ratio
BJP 4216 2781 1083 8080 0.52
INC 3233 3492 1357 8082 0.40
BSP 2461 3792 1564 7817 0.31
TMC 2692 3582 1497 7771 0.34
The total positive, negative and neutral tweets are
compared and plotted using bar graph as in figure 4. It
shows the total number sentiments for each type for
different parties [7]. It will help us to decide the overall
score of sentiments for a particular party.
Tweets for the various parties such as BJP, Congress, BSP
and TMC are considered. Here it shows that for BJP out of
8080 tweets 4216 are positive whereas for Congress 3233
tweets are positive out of 8082. Similarly BSP and TMC
have 2461 and 2692 positive tweets out of 7817 and 7771
respectively. Here BJP has maximum positive to total
count ratio.
Fig -4: Tweets comparison using bar graphs
Similarly, we can also plot our data using pie chart which
will give us the complete idea of how much percentage of
tweets are positive and how much are negative for the
corresponding party. The pie charts in figure 5 and 6
below shows the percentage of positive, negative and
neutral tweets for congress and BJP.
Fig -5: Congress tweets
Fig -6: BJP tweets

This pie chart tells us the tweets extracted for BJP has
more positive tweets percentage than other parties.
The figure 7 shows below is the word cloud which
signifies the frequency of text or the importance of words.
Fig -7: Word cloud
6.1 Comparison with news channel survey
As another factor we have compared our results with
the online survey conducted by ABP-C voters and India
Today Survey for lok sabha elections 2019 .The following
figure 8 shows the opinion poll
Fig -8: ABP-C voters and India Today online survey
7. CONCLUSION AND FUTURE SCOPE
In this paper, we performed LSTM based sentiment
analyser which classifies the tweets based on its sentiment
value. The tweets considered are for Indian elections
2019. Tweets corresponding to a particular party is
extracted using twitter API which are then pre-processed
to remove unwanted and irrelevant data. The LSTM model
is built by training the classifier using labelled dataset.
Thus classification is done on the extracted tweets to label
them as per sentiments, which enables us to present the
comparison between the top parties for general elections
2019. Later the obtained results are represented using bar
graphs and pie charts. Then we compared political
sentiment towards different parties by plotting graphs
using results of sentiment analysis on extracted twitter
data. In the future, a multi lingual based sentiment
classification can be built to acquire tweets in different
languages and perform analysis. Also there could be many
other prospective areas to conduct this research in,
including the data from other big social media sites like
Facebook to increase the size of the data set.
REFERENCES
[1] Ramteke, Jyoti, Samarth Shah, Darshan Godhia, and
Aadil Shaikh. "Election result prediction using Twitter
sentiment analysis." In 2016 international conference
on inventive computation technologies (ICICT), vol. 1,
pp. 1-5. IEEE, 2016.
[2] Jose, Rincy, and Varghese S. Chooralil. "Prediction of
election result by enhanced sentiment analysis on
twitter data using classifier ensemble Approach." In
2016 international conference on data mining and
advanced computing (SAPIENCE), pp. 64-67. IEEE,
2016.
[3] Sharma, Parul, and Teng-Sheng Moh. "Prediction of
Indian election using sentiment analysis on Hindi
Twitter." In 2016 IEEE International Conference on
Big Data (Big Data), pp. 1966-1971. IEEE, 2016.
[4] Wang, Jiang, Yi Yang, Junhua Mao, Zhiheng Huang,
Chang Huang, and Wei Xu. "Cnn-rnn: A unified
framework for multi-label image classification." In
Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 2285-2294. 2016.
[5] Pengfei Liu, Xipeng Qiu, and Xuanjing Huang,
“Recurrent Neural Network for Text Classification
with Multi-Task Learning.” Twenty-Fifth International
Joint Conference on Artificial Intelligence, pp.2873-
2879, 2016.
[6] Bouazizi, Mondher, and Tomoaki Ohtsuki. "Sentiment
analysis in twitter: From classification to
quantification of sentiments within tweets." In 2016
IEEE Global Communications Conference
(GLOBECOM), pp. 1-6. IEEE, 2016.
[7] Wang, Hao, Dogan Can, Abe Kazemzadeh, François
Bar, and Shrikanth Narayanan. "A system for real-time
twitter sentiment analysis of 2012 us presidential
election cycle." In Proceedings of the ACL 2012
System Demonstrations, pp. 115-120. Association for
Computational Linguistics, 2012.
[8] http://colah.github.io/posts/2015-08Understanding-
LSTM
[9] http://towardsdatascience.com/another-twitter-
posentiment-analysis

IRJET- Twitter Sentimental Analysis for Predicting Election Result using LSTM Neural Network

More Related Content

What's hot (20)

Similar to IRJET- Twitter Sentimental Analysis for Predicting Election Result using LSTM Neural Network (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Twitter Sentimental Analysis for Predicting Election Result using LSTM Neural Network