See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361656399

A Deep Learning Approach for Public Sentiment Analysis in COVID-19 Pandemic

Preprint · April 2022

CITATIONS READS
0 84

4 authors:

G.M. Sakhawat Hossain Sayed Asaduzzaman

Rangamati Science and Technology University, Rangamati, Bangladesh University of North Dakota
3 PUBLICATIONS   0 CITATIONS    4 PUBLICATIONS   0 CITATIONS   

SEE PROFILE SEE PROFILE

Md Mynoddin Iqbal H. Sarker

Rangamati Science and Technology University Chittagong University of Engineering & Technology
5 PUBLICATIONS   0 CITATIONS    183 PUBLICATIONS   3,031 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Context-Aware Rule Learning for Intelligent Systems View project

Information Security (Digital Watermarking and Authentication) View project

All content following this page was uploaded by G.M. Sakhawat Hossain on 01 July 2022.

The user has requested enhancement of the downloaded file.

 A Deep Learning Approach for Public Sentiment
Analysis in COVID-19 Pandemic
*

G.M. Sakhawat Hossain Sayed Asaduzzaman

Department of Computer Science and Engineering Department of Computer Science and Engineering
Rangamati Science and Technology University Rangamati Science and Technology University
Rangamati, Bangladesh Rangamati, Bangladesh
[email protected] [email protected]

Md Mynoddin Iqbal H. Sarker

Department of Computer Science and Engineering Department of Computer Science and Engineering
Rangamati Science and Technology University Chittagong University of Engineering & Technology
Rangamati, Bangladesh Chattogram, Bangladesh
[email protected] [email protected]

Abstract—Sentiment analysis is a process of extracting opinions December 2019 in Wuhan, China. The whole world seemed to
into the positive, negative, or neutral categories from a pool of be collapsed at the initial phase of the COVID-19 virus when
text using Natural Language Processing (NLP). In the recent era, World Health Organization (WHO) detected and declared the
our society is swiftly moving towards virtual platforms by joining
virtual communities. Social media such as Facebook, Twitter, infectious disease as a pandemic on March 11, 2020. Almost
WhatsApp, etc are playing a very vital role in developing virtual all of the countries across the world went to the lock-down
communities. A pandemic situation like COVID-19 accelerated situation to stop the spreading of the virus and took their safety
people’s involvement in social sites to express their concerns or measures. As a result, all the people needed to stay at home
views regarding crucial issues. Mining public sentiment from to keep themselves safe from this ferocious virus.
these social sites especially from Twitter will help various orga-
nizations to understand the people’s thoughts about the COVID- Social sites such as Facebook, Twitter, Youtube, WhatsApp,
19 pandemic and to take necessary steps as well. To analyze the etc are already very famous throughout the mass people over
public sentiment from COVID-19 tweets is the main objective of the last decade. As these are popular media to express people’s
our study. We proposed a deep learning architecture based on opinions, thoughts, reactions, and discussions, during the lock-
Bidirectional Gated Recurrent Unit (BiGRU) to accomplish our down situation people became more active to share their views
objective. We developed two different corpora from unlabelled
and labeled COVID-19 tweets and use the unlabelled corpus to regarding the ongoing crisis. Most of the people tried to keep
build an improved labeled corpus. Our proposed architecture the positive intent about the pandemic, while on the contrary
draws a better accuracy of 87% on the improved labeled corpus others were spreading fear as well as fake content, views, or
for mining public sentiment from COVID-19 tweets. tweets all over the social media. So, it becomes an utmost
Index Terms—Sentiment Analysis, Deep Learning, BiGRU, necessity to identify the sentiment of the mass people during
Word2Vec
the pandemic situation. This type of work with intelligent
I. I NTRODUCTION services can help all the organizations like financial, health,
educational, transportation, restaurants, tourism as well as
A pandemic can be defined as an epidemic that spreads
government to understand their people’s views and to take
over international borders, frequently impacting individuals all
necessary actions [1]. Researchers have been working hard to
over the world [3]. Sickness or disease is not a pandemic just
determine public sentiment in a pandemic situation. Sentiment
because it is common worldwide or kills a lot of people; it
analysis is a part of Natural Language Processing (NLP) which
has to be infectious as well. Throughout history, humankind
can be identified using various machine learning and deep
had suffered from a large number of ferocious pandemics
learning approaches [4], [5]. In the context of computing, it
of diseases such as black death, HIV/AIDS, Asian flu, tu-
can also be considered as a part of the broad area of data
berculosis, cholera, Spanish flu, SARS, etc. causing deaths
science with textual analytics [2].
of millions of people across the globe. In recent years, the
Researchers frequently used COVID-19 tweets to perform
world is experiencing a new pandemic of a disease caused
the sentiment classification task about the pandemic situation
by a virus called COVID-19 which was appeared in late
[6]. This paper aims to mine public sentiment in a pandemic
situation. To do so, firstly we developed two different types of
corpora such as unlabelled and labeled corpus from the unla- Aware Dictionary for Sentiment Reasoning (VADER), a text
belled and labeled COVID-19 tweets. Secondly, we trained sentiment analysis model was used by ( A.J.Nair et al. 2021)
our labeled corpus with the help of the unlabelled corpus for mining public opinion from COVID-19 tweets to help out
to build an improved labeled corpus. Thirdly, we proposed public health organizations, as well as government officials
a deep learning-based architecture using Bidirectional Gated [12]. In [4], [13]–[17], the authors had used Multinomial NB,
Recurrent Unit (BiGRU) to perform sentiment analysis both Decision Tree (DT), Logistic Regression, Support Vector Ma-
from the labeled and improved labeled corpus. And, finally, chine (SVM), Random Forest Classifier with different features
we measured the accuracy, precision, recall, and f1-score of like weighted TF-IDF and n-gram to investigate sentiment
our proposed model. analysis from the tweets related to COVID-19 pandemic to
The rest of the paper is organized as follows. Section mine public anxiety.
II describes a brief overview of the related works for the Deep learning approaches [24] are also used heavily in the
sentiment analysis from COVID-19 tweets. In section III, the text classification domain over the few years successfully. To
proposed methodology is explained whereas the results and analyze the tweets regarding natural disasters like earthquakes
discussion are analyzed in section IV. Finally, we conclude and pandemics, Multi-Layer Perceptron (MLP) network was
this paper in section V. proposed by the authors in [18]. They measured the per-
formance of the proposed network on the original COVID-
II. R ELATED W ORKS 19 dataset and obtained a classification accuracy of 83%.
Public sentiment analysis is a very important task especially To point out the shortcomings of the proposed network and
in a situation like the COVID-19 pandemic to mine people’s to explain the behavior, the authors used Local Interpretable
opinions, their demands, and their reactions. There are several Model-Agnostic Explanations (LIME) [18]. A deep learning
research works regarding sentiment analysis in the pandemic classifier was developed in [19] to extract the sentiment from
situation that had already been conducted throughout the 226,668 tweets related to the pandemic situation and obtained
COVID-19 pandemic. Mainly different machine learning and an admissible classification accuracy of 81%. Deep Long
deep learning techniques were used to perform sentiment Short-Term Memory(LSTM), a recurrent neural network along
analysis from COVID-19 tweets. To analyze COVID-19 re- with FastText word embedding technique was applied in [20]
lated tweets for extracting sentiments, a clustering-based clas- to examine the responses of the citizens from the cross-cultural
sification model named TClustVID was designed by (Md. community to the novel coronavirus. The authors used the
Shahriare Satu et al. 2021) [7]. They applied their model on the sentiment140 dataset for their experimental analysis and got
COVID-19 Twitter dataset and got a higher accuracy than the a permissible accuracy of 82.4%. Artificial neural network
state-of-the-art techniques of clustering criteria. An ensemble- (ANN) and LSTM techniques were executed to analyze the
based machine learning model can be used for sentiment global sentiment about the coronavirus pandemic in [21]. For
classification from COVID-19 tweets. Three ensemble ma- the LSTM network, the authors gained comparatively higher
chine learning models such as voting classifier (VC), stacking accuracy (84.5%) than ANN (76%).
classifier (SC), and bagging classifier (BC) were proposed Bidirectional Encoder Representations from Transformers
in [8] for analyzing sentiment from 12k tweets which were (BERT) which is an efficient text classification model, were
gathered from the United Kingdom (UK). The authors in [8] applied by the authors in [22] for multiclass classification of
got a better performance from stacking classifier (SC) with an coronavirus tweets to determine the public opinion about the
f1-score of 83.5%, followed by the voting (VC) and bagging pandemic situation. The model showed an admissible perfor-
classifier (BC) with 83.3% and 83.2% f1-score respectively. mance by achieving 85% accuracy on the classification task.
Naive Bayes (NB) and Logistic Regression (LR), two well- Though several research works were conducted to extract the
known machine learning approaches were applied in [9] for people’s thoughts and their opinion regarding the pandemic,
classifying coronavirus tweets of different lengths. The authors most of these works did not use any pretrained corpus of
observed a strong accuracy of 91% only for classifying the COVID-19 tweets which may improve the performance of
shorter tweets using the Naive Bayes (NB) method, while their proposed architecture. In this paper, we mainly aim
on the contrary, they obtained an accuracy of 74% using to develop such a type of corpus and use that corpus to
the Logistic Regression (LR) method. However, both of these pretrain our experimental dataset and apply it to our proposed
approaches showed comparatively less effective performance architecture for the sentiment analysis of the tweets in a
for the longer tweets classification [9]. Support Vector Ma- pandemic situation. In the following section, we introduce the
chine (SVM), Random Forest (RF), Extra Tree Classifier, proposed methodology used in this research work.
Gradient Boosting Machine, etc supervised machine learning
approaches were executed with different feature techniques III. M ETHODOLOGY
e.g. TF-IDF and bag-of-words (BoW) for US-based sentiment Fig. 1 depicts the methodology proposed in this paper. To
analysis [10]. Linear SVC, Perceptron, Passive Aggressive carry out our experiment, we used two types of COVID-
Classifier, and Ada Boost Classifier were tested with unigram, 19 tweets dataset, one was unlabelled containing 7,59,848
bigram, and trigram features for sentiment extraction from COVID-19 tweets and the other one was labeled containing
more than 72k COVID-19 related tweets in [11]. Valence 36,623 COVID-19 tweets. Both of them were collected from
the Kaggle platform and the first one was used only to build punctuation marks, emojis, non-alphabetic characters, etc. Pre-
a corpus for COVID-19 tweets. The second one was used processing the text data is a mandatory task for any type of text
to conduct our research work. As from the Fig. 1 we can classification or sentiment analysis problem. We performed the
see that the first step of our proposed architecture is data task of data cleansing for our datasets in the following way:
cleansing for both of the datasets used in this paper. The • Tokenization: Tokenization can be defined as a process
data cleansing step includes performing tokenization, lower of isolating the smaller units from a chunk of text called
case conversion, punctuation marks removal, non-alphabetic tokens. Tokens usually consist of characters, words, or
characters removal, stopwords removal, and padding of the text subwords. For instance, if a tweet is like ” stay home
data for both of the unlabelled and labeled datasets. To build an stay safe”, we will get tokens like ’stay’, ’home’, ’stay’,
unlabelled corpus and a labeled corpus from the unlabelled and ’safe’ after tokenization. We performed tokenization for
labeled datasets respectively is our next step. The following each tweet of our unlabelled and labeled datasets.
step is to develop an improved labeled corpus by training • Lower Case Conversion: To preprocess the tweets, we
the labeled corpus with the help of the unlabelled corpus. converted all the tweets of our datasets into the lower
Our proposed Bidirectional Gated Recurrent Unit (BiGRU) case in order to maintain uniformity.
based deep learning approach is developed in the next step. • Punctuation Marks Removal: For separating sentences
We carried out two types of experiments, one was to classify or phrases, punctuation marks are used. Period(.), semi-
the COVID-19 tweets into the positive or negative category colon(;), comma(,), dash (-), or question mark(?) etc
from the labeled corpus and the second one was the same are used as common punctuation marks. In the case of
task as the first one except we used improved labeled corpus tweet preprocessing as punctuation marks do not have any
for the classification. To do so, labeled corpus and improved significant impact, we removed them from our datasets.
labeled corpus were split into train, validation, and test sets. • Non-alphabetic Characters Removal: Most of the tweets
Our BiGRU based deep learning architecture was trained and have a few number of non-alphabetic characters such as
validated according to these train and validation sets and used emoticons, emojis, or symbols etc. For cleansing tweets,
for sentiment classification and prediction from COVID-19 these non-alphabetic characters must be removed because
tweets. Finally, the performance of our proposed model was there will be no big deal on the tweets classification.
measured in terms of precision, recall, f1-score, and accuracy. • Stopwords Removal: Stopwords are words that may have
In the remaining part of the section, the details of our proposed no meaning e.g. (conjunctions, prepositions, etc) and
methodology are described. they are used frequently in the context of classification.
They provide no essential information to determine the
classification category of a tweet. For instance, stopwords
like ’a’, ’the’, ’in’, ’to’ etc were removed from the token
list of our datasets.
• Padding: As all of the tweets of our datasets did not have
the same sentence length, we performed padding to make
all the input of the same length.
B. Developing Corpus
In Natural Language Processing (NLP), corpus development
is a mandatory task that may be defined as an organized set
of machine-readable texts. At the end of the data cleansing
step, we developed two individual corpora namely unlabelled
and labeled corpus for our unlabelled and labeled datasets re-
spectively. We used a well-known word embedding technique
called Word2Vec to build our corpora. By using the word
embedding technique, a word can be represented into a real-
valued vector. The representation of words is done in such a
way so that the same meaningful words are represented with
almost look-alike word vectors. For instance, ’supermarket’
and ’shop’ will have closer word vectors. Word2Vec is such a
Fig. 1. Proposed methodology for sentiment analysis from COVID-19 tweets.
type of word embedding technique to generate word vectors.
It is a simple neural network with only one hidden layer.
A. Data Cleansing Word2Vec model uses the back-propagation technique during
Data cleansing refers to the technique of removing un- training for adjusting its weights and to reduce the loss
necessary elements or noises from the data. In the case of function. The hidden weights are the word vectors that are
Natural Language Processing (NLP), data cleansing means generated after the completion of the training of the Word2Vec
to preprocess text data by eliminating noises like stopwords, model.
 For developing unlabelled corpus we used vector size =
100, window = 5, and min count = 2 in our Word2Vec model.
The same hyperparameters were also applied to generate the
labeled corpus in the Word2Vec model. Once both of these
were developed, we trained our labeled corpus with the help
of the unlabelled corpus to build an improved labeled corpus.
Later, we split both of the labeled and improved labeled corpus
into train, validation, and test sets for training, validating, and
testing our proposed model to classify COVID-19 tweets into
the positive or negative category and for sentiment prediction.
In the following subsection, we introduce our proposed archi-
tecture for sentiment analysis from COVID-19 tweets.
C. Proposed Model Fig. 2. BiGRU architecture.
Our proposed model is based on Bidirectional Gated Re-
current Unit (BiGRU) which is a Recurrent Neural Net-
work (RNN) based architecture and an extension of the
Gated Recurrent Unit (GRU) [24]. Fig. 2 shows the basic
architecture of Bidirectional Gated Recurrent Unit (BiGRU),
where, X0 ,X1 ....Xi represents the inputs and Y0 ,Y1 ....Yi , (i=
0,1,2,..), represents the corresponding outputs from BiGRU
layer respectively. Each layer of BiGRU architecture consists
of sevreral GRUs. GRU can solve the vanishing or exploding
gradients problem of traditional RNNs [23]. Unlike other RNN
architectures, the internal cell state is not maintained by GRU.
It has three types of gates which are update gate, reset gate, and
current memory gate. Update gate determines how much of the
previous information must be passed along into the future. On
the other hand, the output gate determines how much previous
information needs to be forgotten. The current memory gate
is a sub-part of the reset gate, mainly responsible for reducing
the effect of previous information onto the current information
Fig. 3. Proposed BiGRU architecture.
which is being processed for the future. The main goal of
BiGRU is to examine a certain sequence from front to back
and back to front. As a result, a context is constructed by the
labeled COVID-19 tweets dataset consists of 36,623 tweets
network for each character of the text depending on its pasts
among them, 19592 tweets were positive and 17,031 were
and its future. The BiGRU network appears to be identical
negative and mainly used for sentiment classification. In the
to its unidirectional counterpart. The proposed architecture of
labeled dataset positive COVID-19 tweets were labeled as
this paper is shown in Fig. 3. The model consists of an input
1 whereas the negative ones were labeled as 0. As there is
layer, followed by an embedding layer, a BiGRU layer, and
a slight imbalance between positive and negative tweets, to
a dense layer. We used ’sigmoid’ as an activation function at
maintain the balance we executed the oversampling technique.
the dense layer and also used a dropout layer of 20% to avoid
The summary of the labeled dataset after data cleansing is
the overfitting problem. The total number of parameters of our
described in table I in terms of the total number of words,
model was 4,822,093, among them 25,793 were trainable and
mean of the words in a tweet, standard deviation, minimum
the rest were non-trainable parameters. Once the model was
and maximum number of words in a tweet.
built, our next step was to compile and evaluate the model and
to use that model for sentiment classification and prediction
from the COVID-19 tweets as well. In the next section, we B. Model Comilation and Evaluation
discuss the results and the evaluation of our model used for
sentiment analysis from the COVID-19 tweets. To compile our proposed model, we used ’adam’ as an opti-
mizer, ’binary crossentropy’ as a loss function, and ’accuracy’
IV. R ESULTS AND D ISCUSSION
as metrics. We executed our proposed model with a batch size
A. Dataset Description = 128 and for 50 epochs. Accuracy, precision, recall, and f1-
To perform sentiment analysis from COVID-19 tweets, score are measured to evaluate the model performance. Accu-
we used two types of datasets i.e unlabelled and labeled racy may be defined as the percentage of correct predictions
COVID-19 tweets from Kaggle as mentioned earlier. The for the test data. It can be measured by dividing the number
 TABLE I From Fig. 4, it can be seen that our model obtained an
COVID-19 LABELED TWEETS DATASET DESCRIPTION (N UMBER OF accuracy of 79% with f1-score of 78% and 80% for classifying
WORDS )
both positive and negative tweets respectively. On the other
Total Mean˜ Std Min 25% 50% 75% Max hand, from Fig. 5, we can see that our model achieved an
44954 20.818793 7.100794 1 15 21 26 50 accuracy of 87% with f1-score of 86% and 87% for classifying
both positive and negative tweets respectively. So it is clear
that the performance of our proposed architecture on the
of correct predictions by the number of overall predictions. improved labeled corpus is significantly higher than on the
CorrectP redictions labeled corpus. The reason behind that is we developed an
Accuracy = improved labeled corpus with the help of a pretrained corpus
OverallP redictions
(unlabelled corpus). Because of that, we got more accurate
Precision can be defined as the fraction of true positives and
word vectors and the model performed well in identifying
the sum of true positives and false positives.
the word sequences and eventually classifying them into the
T rueP ositives positive or negative category.
P recision =
T rueP ositives + F alseP ositives Fig. 6, 7 depict the train and validation accuracy graph of
Recall can be defined as the fraction of true positives and the our proposed model on the labeled corpus and the improved la-
sum of true positives and false negatives. beled corpus in classifying the COVID-19 tweets respectively.
The graphs explain for each epoch how the model achieved
T rueP ositives the training and validation accuracy.
Recall =
T rueP ositives + F alseN egatives
F1-score is a function of precision and recall.
P recision ∗ Recall
F 1 − score = 2 ∗
P recision + Recall
We executed our proposed model in two types of scenarios.
In the first case, we applied our model on the labeled corpus
to classify COVID-19 tweets into the positive or negative
category. In the second case, we applied our model on the
improved labeled corpus to do the same task as the first one.
Fig. 4 shows the performance of the model for the first case
whereas Fig. 5 shows the performance of the model for the
second case.

Fig. 6. Train and validation accuracy graph of proposed model (labeled

corpus).

Fig. 4. Performance of our proposed model on labeled corpus.

Fig. 7. Train and validation accuracy graph of proposed model (improved

labeled corpus).
Fig. 5. Performance of our proposed model on improved labeled corpus.
 V. C ONCLUSION [11] Kamal Gulati, S. Saravana Kumar, Raja Sarath Kumar Boddu, Ke-
tan Sarvakar, Dilip Kumar Sharma, M.Z.M. Nomani, ”Compara-
tive analysis of machine learning-based classification models us-
Sentiment analysis can be an important tool to determine ing sentiment classification of tweets related to COVID-19 pan-
the views or opinions of mass people expressed on social demic,” Materials Today: Proceedings, 2021, ISSN 2214-7853,
sites regarding various crucial issues like the COVID-19 https://doi.org/10.1016/j.matpr.2021.04.364.
[12] A. J. Nair, V. G and A. Vinayak, ”Comparative study of Twitter
pandemic to identify the people’s expression is positive or Sentiment On COVID - 19 Tweets,” 2021 5th International Conference
negative. In this paper, we mainly focused to mine public on Computing Methodologies and Communication (ICCMC), 2021, pp.
sentiment in coronavirus pandemic from COVID-19 tweets. 1773-1778, doi: 10.1109/ICCMC51019.2021.9418320.
[13] J. Nguyen and R. Chaturvedi, ”Quarantine Quibbles: A Sentiment Anal-
We built an unlabelled and labeled corpus from unlabelled ysis of COVID-19 Tweets,” 2020 11th IEEE Annual Information Tech-
and labeled COVID-19 tweets and developed an improved nology, Electronics and Mobile Communication Conference (IEMCON),
corpus by training the labeled corpus using the unlabelled 2020, pp. 0346-0350, doi: 10.1109/IEMCON51383.2020.9284875.
[14] Kaur C., Sharma A., ”COVID-19 Sentimental Analysis Using Machine
corpus. We measured the performance of our proposed deep Learning Techniques,” In: Panigrahi C.R., Pati B., Pattanayak B.K.,
learning approach (BiGRU based) on both the labeled and Amic S., Li KC. (eds) Progress in Advanced Computing and Intelligent
improved labeled corpus. Our proposed architecture shows Engineering. Advances in Intelligent Systems and Computing, 2021,
vol 1299. Springer, Singapore. https://doi.org/10.1007/978-981-33-4299-
better performance on improved labeled corpus by obtaining 6 13
87% accuracy, whereas on the labeled corpus an accuracy of [15] S. Banik, A. Ghosh, S. Banik and A. Mukherjee, ”Classification of
79%. Our future work will be to develop an efficient aspect- COVID19 Tweets based on Sentimental Analysis,” 2021 International
Conference on Computer Communication and Informatics (ICCCI),
based multiclass sentiment classification model using a deep 2021, pp. 1-7, doi: 10.1109/ICCCI50826.2021.9402540.
learning approach to mine public opinion. [16] Imamah and F. H. Rachman, ”Twitter Sentiment Analysis of Covid-
19 Using Term Weighting TF-IDF And Logistic Regresion,” 2020 6th
Information Technology International Seminar (ITIS), 2020, pp. 238-
242, doi: 10.1109/ITIS50118.2020.9320958.
R EFERENCES [17] Tribedi S., Biswas A., Ghosh S.K., Ghosh A.,”Machine Learning Based
Anxiety Prediction of General Public from Tweets During COVID-
[1] Sarker, Iqbal H. ”Ai-based modeling: Techniques, applications and 19,” In: Nayak J., Naik B., Abraham A. (eds) Understanding COVID-
research issues towards automation, intelligent and smart systems.” SN 19: The Role of Computational Intelligence. Studies in Computational
Computer Science 3.2 (2022): 1-20. Intelligence, 2022, vol 963. Springer, Cham. https://doi.org/10.1007/978-
3-030-74761-9 13
[2] Sarker, Iqbal H. ”Data science and analytics: an overview from data-
[18] Shivam Behl, Aman Rao, Sahil Aggarwal, Sakshi Chadha, H.S.
driven smart computing, decision-making and applications perspective.”
Pannu, ”Twitter for disaster relief through sentiment analysis for
SN Computer Science 2.5 (2021): 1-22.
COVID-19 and natural hazard crises,” International Journal of Dis-
[3] Porta, Miquel, ed, (2008). ”Dictionary of Epidemiology,” Oxford Uni- aster Risk Reduction, 2021, Volume 55, 102101, ISSN 2212-4209,
versity Press. p. 179. ISBN 978-0-19-531449-6. Retrieved 14 September https://doi.org/10.1016/j.ijdrr.2021.102101.
2012. [19] Koyel Chakraborty, Surbhi Bhatia, Siddhartha Bhattacharyya, Jan
[4] H. Krishnan, P. G, A. Poosari, A. Jayaraj, C. Thomas and G. M. Joy, Platos, Rajib Bag, Aboul Ella Hassanien, ”Sentiment Analysis of
”Machine Learning based Sentiment Analysis of Coronavirus Disease COVID-19 tweets by Deep Learning Classifiers—A study to show
Related Twitter Data,” 2021 2nd International Conference on Secure how popularity is affecting accuracy in social media,” Applied Soft
Cyber Computing and Communications (ICSCCC), 2021, pp. 459-464, Computing, 2020, Volume 97, Part A, 106754, ISSN 1568-4946,
doi: 10.1109/ICSCCC51823.2021.9478145. https://doi.org/10.1016/j.asoc.2020.106754.
[5] T. T. Mengistie and D. Kumar, ”Deep Learning Based Sentiment [20] A. S. Imran, S. M. Daudpota, Z. Kastrati and R. Batra, ”Cross-Cultural
Analysis On COVID-19 Public Reviews,” 2021 International Conference Polarity and Emotion Detection Using Sentiment Analysis and Deep
on Artificial Intelligence in Information and Communication (ICAIIC), Learning on COVID-19 Related Tweets,” in IEEE Access, vol. 8, pp.
2021, pp. 444-449, doi: 10.1109/ICAIIC51459.2021.9415191. 181074-181090, 2020, doi: 10.1109/ACCESS.2020.3027350.
[6] U. Naseem, I. Razzak, M. Khushi, P. W. Eklund and J. Kim, [21] Muvazima Mansoor, Kirthika Gurumurthy, Anantharam R U, V R Badri
”COVIDSenti: A Large-Scale Benchmark Twitter Data Set for COVID- Prasad, ”Global Sentiment Analysis Of COVID-19 Tweets Over Time,”
19 Sentiment Analysis,” in IEEE Transactions on Computational arXiv:2010.14234, Nov, 2020
Social Systems, vol. 8, no. 4, pp. 1003-1015, Aug. 2021, doi: [22] U. N. Wisesty, R. Rismala, W. Munggana and A. Purwari-
10.1109/TCSS.2021.3051189. anti, ”Comparative Study of Covid-19 Tweets Sentiment Classifica-
[7] Md. Shahriare Satu, Md. Imran Khan, Mufti Mahmud, Shaha- tion Methods,” 2021 9th International Conference on Information
dat Uddin, Matthew A. Summers, Julian M.W. Quinn, Moham- and Communication Technology (ICoICT), 2021, pp. 588-593, doi:
mad Ali Moni, ”TClustVID: A novel machine learning classifica- 10.1109/ICoICT52021.2021.9527533.
tion model to investigate topics and sentiment in COVID-19 tweets,” [23] Chung J, Gülçehre Ç, Cho K, Bengio Y, ”Empirical evaluation of gated
Knowledge-Based Systems, Volume 226, 2021, 107126, ISSN 0950- recurrent neural networks on sequence modeling,” CoRR abs/1412.3555,
7051, https://doi.org/10.1016/j.knosys.2021.107126. http://arxiv .org/abs/1412.3555, 2014.
[8] Rahman M.M., Islam M.N., ”Exploring the Performance of Ensem- [24] Sarker, Iqbal H. ”Deep learning: a comprehensive overview on tech-
ble Machine Learning Classifiers for Sentiment Analysis of COVID- niques, taxonomy, applications and research directions.” SN Computer
19 Tweets,” In: Shakya S., Balas V.E., Kamolphiwong S., Du KL. Science 2.6 (2021): 1-20.
(eds) Sentimental Analysis and Deep Learning. Advances in Intelli-
gent Systems and Computing, 2022, vol 1408. Springer, Singapore.
https://doi.org/10.1007/978-981-16-5157-1 30
[9] J. Samuel, G. G. M. N. Ali, M. M. Rahman, E. Esawi, and Y. Samuel,
“COVID-19 Public Sentiment Insights and Machine Learning for Tweets
Classification,” Information, vol. 11, no. 6, p. 314, Jun. 2020.
[10] R. Khan, F. Rustam, K. Kanwal, A. Mehmood and G. S. Choi,
”US Based COVID-19 Tweets Sentiment Analysis Using TextBlob
and Supervised Machine Learning Algorithms,” 2021 International
Conference on Artificial Intelligence (ICAI), 2021, pp. 1-8, doi:
10.1109/ICAI52203.2021.9445207.

View publication stats