Twitter Based Sentiment Analysis of Each Presidential Candidate Using Long Short-Term Memory

Dhruval Shah, Yanyan Li, Ahmad Hadaegh
International Journal of Computer Science and Security (IJCSS), Volume (15) : Issue (3) : 2021 87
ISSN: 1985-1553, https://www.cscjournals.org/journals/IJCSS/description.php
Twitter Based Sentiment Analysis of Each Presidential
Candidate Using Long Short-Term Memory
Dhruval Shah shah042@cougars.csusm.edu
Computer Science and Information Systems,
California State University San Marcos
San Marcos, CA, 92096, USA
Yanyan Li yali@csusm.edu
Ahmad Hadaegh ahadaegh@csusm.edu
Abstract
In the era of technology and internet, people use online social media services like Twitter,
Instagram, Facebook, Reddit, etc. to express their emotions. The idea behind this paper is to
understand people’s emotion on Twitter and their opinion towards Presidential Election 2020. We
collected 1.2 million tweets in total with keyword like “RealDonaldTrump”, “JoeBiden”,
“Election2020” and other election related keywords using Twitter API and then processed them
with natural language processing toolkit. A Bidirectional Long Short-Term Memory (BiLSTM)
model has been trained and we have achieved 93.45% accuracy on our test dataset. We then
used our trained model to perform sentiment analysis on the rest of our dataset. With the
sentiment analysis results and comparison with 2016 Presidential Election, we have made
predictions on who could win the US Presidential Election in 2020 with pre-election twitter data.
We have also analyzed the impact of COVID-19 on people’s sentiment about the election.
Keywords: Sentiment Analysis, LSTM, Deep Learning, Natural Language Processing, Data
Mining.
1. INTRODUCTION
The US presidential election is one of the most important events occurring in 2020 in the United
States. Political candidates’ tweet about their campaign and their views on various topics and so
do the daily twitter users. Due to the advancements in technology, many people have started
using platforms like Twitter and Reddit to discuss and share their views on the news and hot
topics around the globe. Everyday users post millions of tweets and express their emotions or
views on the platform [1]. Performing data analysis on the tweets can be used by the ruling party
and opposition party to decide their election strategy. As we are collecting real-time tweets, we
can find the live response of public towards the election. Thus, collecting tweets and performing
sentimental analysis on them would give us the predictions of who could win the US General
Election 2020. In this paper we discuss our approach of using tweets related to US Presidential
Elections in 2016 and 2020 and perform sentimental analysis on both. We then try to find the
comparisons between the two and make predictions for 2020 Election based on sentimental
analysis.
There are many techniques available to do sentiment analysis. For this work, we are going to use
Recurrent Neural Network based Long-short Term Memory (LSTM) algorithm. To better

understand how LSTM works and how we can use it for our twitter data, we referenced Tang Y.
and Liu J.’s work on “Gated Recurrent Units for Airline Sentiment Analysis of Twitter Data” [2],
After that we trained LSTM model using labelled tweets for the prediction and sentiment analysis.
In section 4, we will discuss about this approach in more detail under Methodology. Section 2
introduces the background of the US election and Twitter where we will discuss the first use of
twitter for election campaigning and its impacts on results. In section 3, we will discuss prior work
that is emoji-based sentimental analysis on US Elections and why it is not accurate. In section 4,
we will discuss about the implementation and methodology of our work in detail. We labelled the
data by using VADER API and got 93.45% accuracy on our test data. In this section, we will
discuss about this in more detail. In section 5, we will do analysis of our resulted data and
subsequently, we will analyze the results of the 2016's and 2020's election. We will take
reference of Donna Ladkin's work [3] to understand why Clinton lost the election even though she
had more positive sentiment than Trump in 2016. In section 6, we will conclude why our work is
more accurate than the previous works and predict the chances of who wins the presidential
candidate in 2020.
2. BACKGROUND
Social media is a platform where people express their views and opinions on various issues.
Twitter currently receives around 500 million tweets every day (As per the data from August
2013) in which people mostly share their opinions about trending news or headlines [4]. Political
parties hire experts who do sentiment analysis of these tweets and help them to decide their
strategies for the election campaign. One way to predict the possibility of a candidate winning the
election is by doing the sentiment analysis on the tweets.
In 2008, Barack Obama used Twitter to promote his candidacy for the presidential elections [5].
Obama's campaign employed 100 individuals to run his digital presence. This made a huge
impact on the election results. Since then, before each election, parties hire research associates
to do public sentiment analysis to conduct election manifesto. Through this way, they can
understand what their voters want from the next president. People from all over the world tweet
about the presidential candidates and the US elections in order to reflect their respective views.
Democratic party's presidential candidate Joe Biden and Republican party's presidential
candidate Donald Trump were the top ten political trending twitter topics on Twitter on August 12,
2020.
3. RELATED WORK
English is the highest spoken language in the world. Most research efforts are done in English
language for sentiment analysis. In 2009, A. Tumasjan, T. O Sprenger, P. G Sandner and I. M
Welpe did prediction with Twitter in 2009 and found that 50% users tweeted only one time in their
corpus and rest 50% made 90% of the tweets [6]. Their data set was of approx. 100K tweets
which is too ambiguous to predict for the large population of a country.
In the paper “A novel classification approach based on Naïve Bayes for Twitter sentiment
analysis”, they used positive and negative words file to predict the sentiment of a sentence with
Naïve Bayes classification. But since the number of positive and negative words are not always
equal, there are chances of getting Biased result. [7] To address this issue, they proposed two
ways. First one is by counting the number of positive and negative words in calculating weights
while seconds approach identifies significant words to predict the class of test document. By this
way they achieved 85.33 % accuracy. While it’s almost impossible to have all the positive words
in positive words file and negative words in negative words file, applying this model to our 10 M
tweets dataset may give less accuracy and there are chances of getting wrong sentiment for
many tweets.
Gautam, G., & Yadav, D proposed set of techniques of Machine Learning with Semantic analysis.
[8] They did performance comparison of Naïve Bayes, Maximum Entropy, Support Vector
Machine and Semantic Analysis (WordNet). And they got accuracy ranging between 88.2% to

89.9% accuracy. The accuracy is quite good if we use their approach but this approach will fail to
predict correct sentiment sometimes as the word order is not preserved.
Sharma, P. & Moh, [9] did prediction of Indian Election Using Sentiment Analysis on Hindi Twitter.
For this work, they used three different algorithms which are Naïve Bayes, SVM and Dictionary
Based. They got 62.1%, 78.4% and 34% accuracy respectively.
B. Joyce and J. Deng gathered tweets from 2016's presidential election based on specific words
like Hillary Clinton and Donald Trump [10]. They have used Twitter's streaming API for tweets
collections. The keywords they included were democratic and republican and the full name of
the top political candidates like Donald Trump. Since they used Twitter's official tweets
streaming API, they collected around 79 million tweets by two month's 10,000 unique users. Then
they filtered 1.9 million tweets that contain emojis. After deleting retweets, the data set further
depleted to 783K tweets. With the Emojis Sentiment Ranking, they determine each tweet's
emotional orientation. For that, they extracted the emoji from the sentence and then converted it
into Unicode. Then applied regex on Unicode and found score from the customize list of 522
unique emoji characters. This score will determine emotional orientation or sentiment of the
sentence. To build a sentiment classifier, they used a multinomial Naïve Bayes classifier.
Since not every sentence contains an emoji, only judging the result based on emojis will not be
accurate. There may be some tweets where the user has added an emoji as a sarcastic tone.
Also, word ordering matters while doing sentiment analysis. For example, one sentence is I have
to read this book and another sentence is I have this book to read. Both examples have different
meaning. This simple example is one of the examples where we can say word ordering matters
while doing sentiment analysis.
To overcome the above problem, in our work we developed a model that will try to analyze the
sentiment based on word number. Deep learning approaches like Long Short-Term Memory
(LSTM) is one of the Recurrent Neural Network approaches where model is trained with the
word’s ordering whereas in the already existing work, they only considered emojis while doing the
analysis. Due to the above-mentioned reasons, LSTM seems to be the perfect fit for doing
sentiment analysis. So, we will be using LSTM model for our analysis.
4. METHODOLOGY
This section discusses about our methodology. To start with, we downloaded Twitter’s 55,000
prelabelled tweets. For training, we will used random 40K tweets and the rest of 15K tweets for
testing purposes. But these tweets were not sufficient to train the ML model. As the part of next
process, we will collect tweets data of Joe Biden and Donald Trump. For this, we used following
keywords: “JoeBiden”, “DonaldTrump”, “BidenHarris”, “US Election 2020”, “TrumpPence”. We
collected around 1.25M tweets for the period of October 2019 to July 2020. Figure 1 shows our
proposed system workflow.
4.1. System Overview
With the tweets collected, we did data cleaning and data preprocessing. In Data Preprocessing,
we check if sentence language is English or not. If the language is different then we remove
those sentences from our dataset. In data cleaning, we remove unwanted texts, URLs, language
slangs, emojis, hashtags, mentions, and retweets. After data preprocessing and data cleaning,
we were left with almost 1.2M tweets. We randomly took 55K tweets from this dataset. We
labelled these tweets using VADER sentiment API [11]. Once these 55K tweets were labelled, we
divided 40K tweets as training and rest 15K tweets for testing to make training dataset big
enough for model training. We passed these 80K tweets data to train our LSTM model. Before
passing tweets to the model, we implemented embedding layer. We used Word2Vc model which
initializes random weights and learns to embed all words in the dataset. Once embedding is
done, we use LSTM model for RNN layers. To prevent our model from overfitting, we used
dropout techniques. It drops irrelevant information from the network as they don’t contribute to the

process of enhancing our model accuracy. We have used dense layer in this model which
connects every input with every output using weights. We have used Relu activation function as it
helps complex relationship in the data to be captured by the model. Relu is more efficient than
Tanh or Sigmoid activation function because Relu activates certain number of neurons.
FIGURE 1: Project process workflow.
To test our model, we passed the rest 15K pre labelled tweets from twitter and 15K labelled
tweets from VADER API. We got 93.45 % test accuracy which is better than previous works [10].
We passed rest of tweets to this model for labelling. Once these tweets were labelled, we
counted the number of positive and negative sentiment tweets. We also passed 2016’s tweets
data and labeled them. Once the data from both the elections was labelled, we counted number
of positive and negative tweets in these to do the analysis.
We have divided the process into 4 major sections below. Data Gathering, Data Processing,
Model Training and Testing on test data. All of these steps are very crucial for training the model.
The model must be fed with correctly labeled data in order to generate the most accurate results.
To give a broader overview, here is the overall process. Using Get Old Tweets (GOT3) API [12]
we will collect the tweets. The raw data might be very informal and unstructured. It may contain
hyperlinks, emojis, mentions, and retweets. To convert it into trainable data, data cleaning is
required. Data cleaning will be performed on tweets that were collected from October 1, 2019 to
July 31, 2020 and also October 1, 2015 to November 7, 2016. During the data processing,

hashtags, mentions, emojis, links (photo, video, or gifs) and retweets get removed. After that, the
data is divided into 80-20 % for training and testing respectively. Once we have the data ready,
we use it to feed into the model to train and then test its accuracy with the test data set.
4.2. Data Gathering
Initially, we start collecting tweets using Tweepy streaming API. Since February 1, 2018, standard
users cannot access past tweets for more than 7 days. So, we used GOT3 API (Get Old Tweets).
GOT3 is a GitHub repository which uses URLlib for fetching tweets from Twitter's advance
search. Like this, we collected tweets from October 1, 2019 to July 31, 2020, and October 1, 2015
to November 7, 2016. We randomly selected 1.25M tweets from this time period and their geo-
graphical location as United States. By doing data preprocessing, we checked each tweet and
filtered out non-English tweets. For this, we used Python’s Langdetect library. Then we did data
cleaning. In this, we filtered out tweets which only contained emojis, images, gifs, or video. Of
1,250,000 tweets, we obtained 1,200,718 tweets that were containing tweets and may contain
images, video, or gifs.
For 2020's data, we used keywords such as "US Elections" with the twitter usernames
"realDonaldTrump" and "JoeBiden". Then we collected 2016's tweets for presidential candidates
Donald Trump and Hillary Clinton using the same GOT3 API. We collected tweets containing
keywords like US Elections, realDonaldTrump, HillaryClinton. Some of the keywords and
hashtags we used for data gathering are US Elections, US Elections 2020, TrumpPence,
republican election, presidential elections, election 2020, donald trump for 2020, BidenHarris. And
we also collected election related tweets from the twitter handles JoeBiden, POTUS,
DonaldTrump, KamalaHarris, Mike_Pence. In the data gathering, we only collected tweets from
United States. We did this by filtering tweets by its location (in our case it’s United States).
Example 1: @JoeBiden VP Biden I wish you wouldn't debate Trump unless he: Debate #1
Trump shows taxes Debate #2 Tells Putin stop killing our soldiers Debate #3 Put sanctions on
Putin for 2016 election hacks & I wish you'd publicly challenge him with this. #Biden/Rice 2020
Example 2: Best @realDonaldTrump political ad I've seen yet. @TheDemocrats. will attempt just
about anything to impact the 2020 election. Pure evil. cc @Scavino45 @parscale
Example1 and Example 2 are the sample tweets from the training dataset and the words in bold
are those words which will would cause problem while training model. So, we need to remove
these words, special characters, URLs and images from the tweets. We remove them during
Data Processing.
4.3. Data Processing
As the language used on Twitter is informal and unstructured, directly using this data will result in
failure of our model to understand some words and it will decrease model accuracy. Thus, we
performed the following steps to each tweet we collected.
1. Converted all tweets to lowercase as some text mining algorithms are case sensitive.
2. Browsed through all the tweets and replaced words like - no, not, never, cannot, don't,
doesn't with not.
3. Removed repeated words and slang from the sentence and cleaned it.
For example, the sentence is Hiiii! Good to see u then we will remove Hiii and replace it with a
word Hi and we will replace word u with you. So, after processing the sentence we will get Hi!
Good to see you. We did this using python's nltk tokenizing. After data processing our two
example tweets become the following:

Processed Example 1: i wish you would not debate trump unless he debates 1 trump shows
taxes debate 2 tells putin stop killing our soldiers debate 3 put sanctions on putin for 2016
election hacks i wish you would publicly challenge him with this biden/rice 2020.
Processed Example 2: best realdonaldtrump political ad I have seen yet. thedemocrats will
attempt just about anything to impact the 2020 election. pure evil. cc scavino45 parscale
As we have converted tweets to the trainable tweets, now we can proceed with the model
training.
4.4. Model Training
Model training in machine language is the process of feeding an ML algorithm with data to help
identify and learn good values for all attributes involved [13]. There are several types of machine
learning models, of which the most common ones are supervised and unsupervised learning [14].
We will train the model using Long Short-Term Memory (LSTM) algorithm. LSTM is an artificial
recurrent neural network (RNN) architecture used in the field of deep learning. Unlike standard
feed forward neural networks, LSTM has feedback connections. It can not only process single
data points but also entire sequences of data. As we have used Bidirectional LSTM, will
represent forward and
< t >
will represent backward activation.
Unlike standard feedforward neural networks, LSTM has feedback connections. It can not only
process single data points (such as images), but also entire sequences of data [15].
FIGURE 2: Recurrent Neural Network [16].
In a traditional neural network, all inputs and outputs are independent of each other but while
doing sentiment analysis input word sequence also matters. In Figure 2 we show the architecture
of RNN. Hence, we need to remember the previous work while guessing the next word. So, RNN
remembers every information upon time [17]. RNN converts individual activation into dependent
activation by providing the same weights to each layer which decreases the increasing
parameters and memorizing each previous output by giving each output to the next hidden layer.
So now, these hidden layers can be joined together so that the weights and bias of all the hidden
layers is the same as a single recurrent layer.

FIGURE 3: Long Short-Term Memory [18].
As shown in Figure 3. Bidirectional recurrent networks are just the same as putting two
independent RNNs together. This Bi-LSTM will have two networks, forward and backward
information sequences at every step. This will run input in two ways, one is from future to past
and another is from past to future. The only difference from bidirectional to unidirectional is that it
also runs backward, and you can preserve information from the future using two hidden states.
There are some activation functions available like relu, tanh, softmax. We have used relu
activation function.
Tuning Parameters Bi-LSTM Model
Learning Rate 0.01
Dropout 0.5
Embed Size 64
Vocab Size 1000
No Filters 256
Max Length 280
TABLE 1: Model parameters.
We passed English tweets labeled positive or negative to the model. The parameters used in the
model are presented in Table 1. Once the training is completed, we saved the model using
model.save() which will help us to label the data more rapidly and use the already saved data and
there will be no need of saving the data again.
4.5. Testing
Model testing is the process of checking the correctness of the model on the pre-labeled data. To
test the model, we passed the pre-labeled data to the saved model. Once the test data is passed
to the BiLSTM, the number of parameters is calculated. The number of parameters is the
connection between layers and the biases in every layer. The vocab size and max_length is fixed
which are 1000 and 280 respectively.
After 25 epochs, the model accuracy is 93.45 % which means that out of 100 sentences, the
model will correctly label 93 sentences as positive or negative sentiments. BiLSTM keeps records
of previous words and also the next words. For example,
Sentence 1: The person carrying black suitcase is Mr. Mosbey.
Sentence 2: They were shocked when they saw the person carrying water in the desert!
In the above sentences, we cannot say what will be the next word after carrying is black suite
case or water. It depends on the context of a sentence. As word orderings are preserved, model
accuracy will increase.

Name Positive % Negative %
Trump 52.04 47.96
Biden 59.54 40.46
TABLE 2: Sentiment analysis result for 2020 presidential election related tweets.
5. ANALYSIS
For 2020 US Presidential Election predictions, we passed tweets of Trump and Biden that we
have collected from October 2019. As shown in Table 2, we got 52% positive sentiment for
Trump and 59.5% positive sentiment for Biden. For simplicity, we removed neutral tweets from
the consideration and kept only positive or negative sentiment tweets. To have a better
understanding of the sentiment, we collected tweets before and after the Covid-19 pandemic for
2020.
We used the same keywords and twitter handles to collect tweets for before and after Covid-19.
We took tweets from October 2019 to March 2020 as before Covid-19 tweets and March 2020 to
July 2020 as after Covid-19 tweets. The reason behind doing before and after Covid-19 is to
analyze how much sentiment has changed in people towards Trump and Biden. Many strategists
believe that Trump Administration has failed to stop the Covid-19 in the United States. This
means that people are not happy with the Trump administration which can be analyzed by the
data we collected.
Trump (Before) 48.81 51.18
Biden (Before) 54.54 45.36
Trump (After) 49.10 50.90
Biden (After) 66.67 33.33
TABLE 3: Sentiment analysis result before and after Covid-19.
As shown in Table 3, there is a slight increase in people's positive sentiment for Trump. Whereas
for Biden, people's positive sentiment for him has increased by 12%. After collecting this data, we
collected 2016's data and tried finding some correlation between both the elections.
Trump 55.73 44.27
Clinton 78.26 21.74
TABLE 4: Sentiment analysis result for 2016 presidential election related tweets.
We have collected 2016's tweet data. The results we got is very appalling. After a significant
amount of training and testing, as shown in Table 4 we got 55.73 % positivity for Trump and
78.26 % positivity for Clinton. Usually, positivity denotes that people are in favor of making Clinton
the President. But she could not make it and instead Trump won the election to become the
President.
To understand this in more detail, we took reference of Donna Ladkin's work [3] where it states
the reason why Clinton lost the election even after winning the popular vote by 2.8 million votes in

2016. Trump won the election because Electoral college votes in the States in which he won the
popular votes exceeded that of those held by the States in which Clinton won the popular votes.
This is the fifth time when a candidate has won by the popular votes but later lost because the
Electoral college didn't vote for them.
After comparing both the election data, people have more positive sentiment towards Biden
compared to Trump in 2020. So, the possibility of winning the election is higher for Biden as
opposed to Trump provided that the Electoral college votes to go for Biden.
6. CONCLUSION AND FUTURE WORK
We have compared the previous work on Machine Learning based Lexicon sentiment analysis
with our work on positive/negative words-based Machine Learning algorithm. Though
positive/negative words-based Machine Learning algorithm should work well, it failed to give
correct sentiment in many cases. So, instead of labelling tweets with positive and negative words,
we used libraries to label the tweets and trained model using LSTM. This approach gave better
accuracy and reduced the labelling time of each tweet. After that we wanted to find the reason as
to why Clinton lost the election in 2016 to Trump despite having much positive sentiment
compared to Trump. We discussed and concluded that though public sentiment was much
positive for Clinton, the electoral votes didn't go for Clinton in 2016 and thus, she lost the election.
In 2020 elections, public sentiment is more positive for Biden compared to Trump. So, if the
electoral votes also go for Biden, then he could win the 2020 US General Election.
According to the article [19], there are more republican twitter users to democrats. So, there are
chances of getting biased result. So, in order to get more accurate result, we should consider
other social media sources like Reddit, Facebook, Blogposts, chat rooms and email campaigns.
For these social medias too, our model can be used to predict sentiment for each candidate. To
use our model, the data should be in text form (Images, URLs should be removed). And before
passing data to the model, the data should be cleaned, structured and labelled.
For future work, we can use official Twitter API to get the more data. GetOldTweets API will fetch
very less tweets compare to Twitter Premium APIs. Also, one can use retweets, likes, and share
count to weight on other parameters. Also, we can consider keeping track of trending tweets and
the social media account getting more tweets in order to consider the current popularity of a
candidate. These parameters can help us understand and predict sentiment in better way. For
future elections result predictions, our model can be used for sentiment analysis as we have
labelled data using VADER API and our model accuracy is 93.45%.
7. REFERENCES
[1] Hao, M., Rohrdantz, C., Janetzko, H., Dayal, U., Keim, D. A., Haug, L.-E., & Hsu, M.-C.
(2011). Visual sentiment analysis on twitter data streams. 2011 IEEE Conference on Visual
Analytics Science and Technology (VAST). doi:10.1109/vast.2011.6102472.
[2] Tang Y. and Liu J., Gated Recurrent Units for Airline Sentiment Analysis of Twitter Data,
Technical Report, Stanford University, 2011.
[3] Ladkin, D. (2017). How did that happen? Making sense of the 2016 US presidential election
result through the lens of the ‘leadership moment.’ Leadership, 13(4), 393–412.
[4] Twitter Usage Statistics, https://www.internetlivestats.com/twitter-statistics/, accessed: 2020-
08-08.
[5] Social Media on the Campaign Trail: Barack Obama and Donald Trump, accessed on
August 8 2020, https://contentgroup.com.au/2017/09/social-media-campaign-trail-obama-
trump/.

[6] A. Tumasjan, T. Sprenger, P. Sandner, and I. Welpe. Predicting elections with Twitter: What
140 characters reveal about political sentiment, Proceedings of the International AAAI
Conference on Web and Social Media (ICWSM), pp. 178–185, 2010.
[7] A novel classification approach based on Naïve Bayes for Twitter sentiment analysis.
(2017). KSII Transactions on Internet and Information Systems, 11 (6).
https://doi.org/10.3837/tiis.2017.06.011.
[8] Gautam, G., & Yadav, D. (2014). Sentiment analysis of twitter data using machine learning
approaches and semantic analysis. 2014 Seventh International Conference on
Contemporary Computing (IC3). doi:10.1109/ic3.2014.6897213.
[9] P. Sharma and T. Moh, "Prediction of Indian election using sentiment analysis on Hindi
Twitter," 2016 IEEE International Conference on Big Data (IEEE Big Data 2016), 2016, pp.
1966-1971, doi: 10.1109/BigData.2016.7840818.
[10] B. Joyce and J. Deng, "Sentiment analysis of tweets for the 2016 US presidential election,"
2017 IEEE MIT Undergraduate Research Technology Conference (URTC), Cambridge, MA,
2017, pp. 1-4, doi: 10.1109/URTC.2017.8284176.
[11] Vader Sentiment API, https://pypi.org/project/vaderSentiment/.
[12] Jefferson Henrique, Get Old Tweet in Python, https://github.com/Jefferson-
Henrique/GetOldTweets-python, accessed: 2020-08-08.
[13] What Is Model Training, https://oden.io/glossary/model-training/, accessed: 2020-08-08.
[14] Alloghani M., Al-Jumeily D., Mustafina J., Hussain A., Aljaaf A.J. (2020) A Systematic
Review on Supervised and Unsupervised Machine Learning Algorithms for Data Science. In:
Berry M., Mohamed A., Yap B. (eds) Supervised and Unsupervised Learning for Data
Science. Unsupervised and Semi-Supervised Learning. Springer, Cham.
https://doi.org/10.1007/978-3-030-22475-2_1.
[15] Long Short-Term Memory, https://en.wikipedia.org/wiki/Long_short-term_memory/.
[16] Khuong, Ben. RNN Figure. The Basics of Recurrent Neural Networks (RNNs). 11 June
2020, medium.com/towards-artificial-intelligence/whirlwind-tour-of-rnns-a11effb7808f.
[17] Khuong, Ben. RNN Definition. The Basics of Recurrent Neural Networks (RNNs). 11 June
2020, medium.com/towards-artificial-intelligence/whirlwind-tour-of-rnns-a11effb7808f.
[18] Aggarwal, Raghav, Bi-LSTM, published in 4 July 2019,
medium.com/@raghavaggarwal0089/bi-lstm-bc3d68da8bd0.
[19] How Democrats and Republicans Use Twitter, Pew Research Center, published in Oct 15,
2020, https://www.pewresearch.org/politics/2020/10/15/differences-in-how-democrats-and-
republicans-behave-on-twitter/ accessed: 2021-04-15.

Twitter Based Sentiment Analysis of Each Presidential Candidate Using Long Short-Term Memory

More Related Content

What's hot (16)

Similar to Twitter Based Sentiment Analysis of Each Presidential Candidate Using Long Short-Term Memory (20)

Recently uploaded (20)

Twitter Based Sentiment Analysis of Each Presidential Candidate Using Long Short-Term Memory