Emotions-Based Disaster Tweets Classification: Real or Fake

MARCO ALFONSE1,2, MARIAM GAWICH2

1Computer Science Department, Faculty of Computer and Information Sciences,

Ain Shams University,

Abbaseya, Cairo,

EGYPT

2Laboratoire Interdisciplinaire de l'Université Française d'Égypte (UFEID LAB),

Université Française d'Égypte,

21 Ismailia Desert Road Ville Shorouk, Cairo,

EGYPT

Abstract: - Social media platforms are considered interactive communication channels between governments,

civil society organizations, and the public. During disaster occurrences, social media platforms play a crucial

role such as the alertness of people towards the disaster occurrence, its risks, and consequences. They are used

as tools to spread real updated information rapidly related to the disaster. Furthermore, social media platforms

can facilitate the mobilization of volunteers as well as the organization of campaign donations after the disaster

occurrence. Nevertheless, the benefits of social media platforms can be a double-edged sword through the

dissemination of unreal information such as rumors or fake disasters. Unfortunately, the public can easily

believe unreal information due to the anxiety that they experienced during the occurrence of a past real disaster.

This paper presents a model to distinguish between the fake disaster tweets and the real ones. The

implementation of this model is established twice; the first implementation involves the use of Machine

Learning with the traditional Natural Language Processing techniques on the disaster dataset provided by

Kaggle, and the second implementation involves using the emotions that are extracted from the tweets in the

classification process. The proposed model achieves an accuracy of 88,34% without the usage of the emotion

extraction module while it achieves an accuracy of 89,39 % with the inclusion of the emotion extraction

module.

Key-Words: - Artificial Intelligence, Machine Learning, Natural Language Processing, Knowledge Discovery,

Sentiment Analysis, Fake Disasters Tweets.

Received: July 28, 2022. Revised: August 29, 2023. Accepted: October 1, 2023. Published: October 18, 2023.

1 Introduction

According to the United Nations Office for Disaster

Risk Reduction (UNDRR), [1], a disaster is defined

as a severe disturbance of a society or a community

during a dangerous phenomena occurrence. The

dangerous phenomenon can provoke at least one of

the following consequences: human life, loss,

economic downturn, environmental loss, and

material loss. Without any doubt, social media

platforms such as Facebook and Twitter play a

crucial role during disaster occurrence and disaster

management after its occurrence due to their ability

to speed up related information dissemination.

These platforms, [2], can provide users with disaster

awareness before or during its occurrence. They can

facilitate the collection of financial support and raise

awareness about the need for donations.

Furthermore, the use of social platforms helps

people discover their relatives’ status and locations

during the disaster or after the disaster occurrence.

On the other hand, social media platforms can have

a negative influence on society due to the

dissemination of false information, [3]. Social media

platforms allow any user to create and share

information in terms of tweet messages or posts,

regardless of their validation. The other users can

believe the false information and share it because

they can’t distinguish whether the information is

true or fake. Fake disaster tweets can have a

negative impact on humans, the government, and

the economy. They can affect the health of humans

due to the increase in anxiety that can provoke

human life loss, especially for those who suffer

from diabetes and other harmful diseases. The

dissemination of fake disaster tweets can provoke

panic among investors, which may lead to the

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

313

Volume 20, 2023

destruction of their businesses. Furthermore, fake

disaster tweets can have a negative influence on

public trust in the government. Examples of popular

fake tweets disseminated on social media platforms

are fake news concerning COVID-19 remedies, [4],

and the spread of false tweets concerning a pizza

shop that led to a shooting incident in the USA, [5].

To eliminate the dissemination of fake disaster news

on social media platforms, tweets, and Facebook

posts should be analyzed and verified automatically.

For this reason, Machine Learning (ML) and

Natural Language Processing (NLP) techniques, [2],

[6], [7], [8], are applied to classify whether the

tweets involve a true disaster occurrence or not. The

verification and analysis of the tweets face many

challenges, such as the interpretation of the informal

language and its transformation into formal

language, the interpretation of slang terms, the

emoticons, etc. Another challenge encountered in

tweet analysis is sentiment extraction and its

classification. Most of the researchers ignored the

introduction of sentiment analysis of tweets and its

influence on the performance of the classification

results. They only applied ML and NLP techniques

to tackle the disaster tweet classification.

The objective of this research is to highlight the

role of sentiment analysis in improving the

performance of ML algorithms to differentiate

between fake and real disaster tweets. For this

reason, a model was proposed for disaster tweet

classification based on ML, NLP, and sentiment

analysis. This model has many benefits, it can

prevent the spread of false information during the

disaster occurrence, and the model can facilitate the

support provided to individuals suffering in the face

of disasters by providing accurate information to the

humanitarian movements, in addition to upholding

human dignity during disasters.

This paper is organized as follows: section 2

presents the related work; section 3 demonstrates

the proposed model for disaster classification; and

section 4 presents the results and discussion. Section

5 contains the conclusion.

2 Related Work

In, [9], the author provided a classification model to

identify which tweets are real and which are fake.

The proposed model has several stages including,

preprocessing the input data using many techniques

such as count vector, Term Frequency Inverse

Document Frequency (TF-IDF), a continuous bag of

words, and a skip-gram vector. The author built a

classifier network based on BERT that includes six

layers. In this network, he modified some of the

hyperparameters of the BERT to minimize the loss.

These parameters include random state split,

dropout and learning rates, batch size, and finally

the optimizer. The data set used in this experiment is

given by Kaggle which contains 7613 records used

for training and 3263 records used for testing. It is a

binary dataset that has two labels, not disaster and

disaster. The author applied some of the data

cleansing steps for each tweet such as normalization

and removing all emails, URLs, HTML tags,

emoticons, abbreviations, stop-words, special

characters, and punctuation. The author compared

his classification performance with some of the

traditional machine learning techniques, and he

proved that the BERT-based model outperforms

other techniques having an overall F1-score equal to

0.8867. In, [10], the authors proposed a

classification model that can identify the real tweets

from the fake ones based on BERT (Bidirectional

Encoder Representations from Transformers). The

authors added a dropout layer and another dense

layer with a ReLU (rectified linear unit) activation

function to the utilized BERT model. The authors

used a dataset downloaded from Kaggle that

contains 10873 comments, of which 57.03% are not

real disasters. Each tweet has an identifier, the text

of the tweet, location, keywords, and target. The

cleaning of the data is done by removing URLs,

HTML tags, special characters, duplicates, special

characters, and emails. The data are converted into

vectors using TF-IDF, and the linear SVC is utilized

as a classifier. The validation accuracy obtained by

the proposed model is 79%.

In, [11], the authors propose a methodology that

depends on machine learning methods to classify

disaster tweets. They proposed two classifiers,

which are support vector machine and naive bayes,

for their classification system as these two

algorithms are frequently used for tweet

classification, according to their literature review.

They also proposed the inclusion of emoticons in

the classification system for better recognition of

disaster tweets. They proposed extracting the tweet

using an API named Twitter Streaming API.

Preprocessing tweets is to be applied to remove

noise, repetition, and any unwanted elements from

the collected tweets. The authors proposed using the

LSTM recurrent neural network to consider

emoticons in their classification system, as they

mentioned that most models usually remove

emoticons from the data in their classification

systems. The authors mentioned that they are going

to implement this model and compare the results of

the proposed algorithms.

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

314

Volume 20, 2023

In, [12], the authors provided a model for the

analysis of disaster tweets. The author proposed a

model to identify the informative tweet from the

other tweets. Their methodology is based on the

following stages: preprocessing, feature extraction,

and classification. The preprocessing involves

converting all tweets to lowercase letters, removing

hashtags, punctuation, URLs, digits, and stop words,

removing words that are of length 2, and making all

tweets the same length by padding techniques. The

feature extraction is performed through the

embedding of words using the TF-IDF technique.

The authors used the linear SVC as a classification

algorithm. They evaluated their method using the F1

score metric, which gives 0.72.

In, [13], the authors applied many machine

learning and deep learning techniques to categorize

disaster tweets. They applied their techniques to a

data set related to cyclone AMPHAN and

NISARGA, and they collected these tweets using

Tweepy. They applied different data preprocessing

methods including the removal of hashtags,

mentions, and URLs, replacing each line and tab

break with a space, converting each emoticon into

positive and negative, and finally applying

lemmatization. For the feature extraction

techniques, they applied the glove embedding with

the deep learning techniques and count vectorizer,

TF-IDF, n-gram, word, and character level for the

machine learning techniques. They applied many

classifiers, including a classifier based on BERT, bi-

directional LSTM, TextCNN, SVC, XG-Boost,

logistic regression, SGD, linear SVC, random forest,

KNN, AdaBoost, decision tree and Gaussian Naive

Bayes. They achieved accuracies that range from

0.51 to 0.72 while categorizing the data into five

categories; ‘Important Help Related’, ‘Informative’,

‘Damage and Casualty related’, ‘Emotional’, and

‘Irrelevant’. On the other hand, they achieved

accuracies that range from 0.56 to 0.8 while

categorizing the data into four categories,

‘Important for Disaster Managers’, ‘Important for

Public’, ‘Important for Both’, and ‘Others’. The

authors showed that the BERT-based classifier

outperforms machine learning and deep learning

techniques in both cases.

3 Proposed Model For Tweet Analysis

The idea behind the proposed model of disaster

tweet classification, which indicates whether the

disaster is real or not, is based on the sentiment

analysis (polarity) of each tweet.

The proposed system involves five phases. The

first phase is dataset selection, where a benchmark

dataset is selected to examine the proposed model.

The second phase is dataset preprocessing, which is

necessary for cleaning and preparing the data for the

next phases as the quality of the data has a great

impact on the classification performance. The third

phase is emotions extraction, where the dataset is

enriched by a new feature extracted from the tweets.

This newly added feature enhances the performance

of our proposed model. The fourth phase is feature

selection, where the input features are converted

into a format that can be understood by the applied

classifier in the next phase. The final phase is

classification, where the input records are given a

label by a machine learning classifier, either real or

fake. Figure 1 demonstrates the five phases of the

proposed model. The model is implemented using

Python Language, [14], and the R program, [15].

3.1 Data Selection

The dataset is provided by, [16]. It involves the

following features: the id which is a unique

identifier for the tweet, the keyword that describes

the disaster, the location where the tweet was

posted, the text (the tweet itself) and the target

(decision class label) that determines whether the

disaster is real or not. The dataset contains 11371

records.

3.2 Data Preprocessing

The preprocessing of the Tweet disaster dataset is

established through the removal of records that have

no values (NaN values) using the Panda library,

[17]. The disaster dataset contains 7953 records

after the removal of the missing value.

Preprocessing also involves noun phrase detection

using the TextBlob library, [18], word

lemmatization using WordnetLemmatizer, [19], and

applying regular expressions to clean the dataset

using the RegEx library, [20]. All the previous text

processing operations are applied to the Tweet text.

Therefore, each Tweet text provided in the disaster

dataset is processed and added as a new feature to

the dataset.

3.3 Emotions Extraction

In this stage, the newly added feature to the dataset

is analyzed using the syuzhet package, [21],

of the R program to identify the emotions in the

text. These emotions include fear, anticipation, joy,

anger, sadness, disgust, surprise, and trust. The

Syuzhet package assigns a score for the positivity or

negativity of the processed tweet.

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

315

Volume 20, 2023

Fig. 1: The Proposed Model for Tweets Disasters Classification

Each Tweet is assigned a score of positivity and

negativity. If the positive score of a tweet is greater

than the negativity score, the polarity of the tweet

will be positive, and vice versa. If the tweet has a

positive score that is equal to a negative score, its

polarity will be neutral. Furthermore, the polarity

will be added to the initial dataset and considered as

a new feature called “tweet polarity” that will be

taken into consideration in the application of the

machine learning algorithms.

Consider the following tweet concerning an

accident that occurred in Covina as an example to

demonstrate how the syuzhet package works: "How

could you leave when I gave you my all �" The

syuzhet assigns the following score for each one of

the eight emotions: 0 for anger, 0 for anticipation, 0

for disgust, 0 for fear, 0 for joy, 1 for sadness, 1 for

surprise, and 0 for trust. The syuzhet assigns 1 for

the negativity of this tweet and 0 for its positivity.

Consequently, the sentiment analysis result of this

tweet is negative. Table 1 summarizes the results of

the emotion extraction phase.

Table 1. The experimental Results of the Emotion

Extraction Phase

Tweet Polarity

Number of tweet texts

Positive

2030

Negative

3295

Neutral

2628

The sentiment or emotion analysis of the

processed tweets demonstrates that 25% of tweets

are positive, 41% of tweets are considered negative

tweets, and 33% of tweets are neutral. The disaster

dataset is populated by the new feature “tweet

polarity “generated by the Syuzhet package using

the R program.

3.4 Features Selection

This stage is applied twice, as we performed two

experiments. In the first experiment, the tweet

represents the independent variable, and the target

column is used as the output class. Concerning the

second experiment, the tweet and its polarity that

were detected in the previous phase are taken as

independent variables while the target column is

used as the output class. For both experiments, two

feature extraction techniques are applied: the TF-

IDF vectorizer, [22], and the Count Vectorizer, [23].

The TF-IDF is used to determine the term relevance

and its occurrence frequency in the tweets. The

Count Vectorizer is applied to switch the tweets and

their polarity to numeric values to be processed by

the classification algorithms.

3.5 Classification

For the two experiments we performed, several

supervised machine-learning algorithms were

applied to the disaster dataset. The applied

supervised machine learning algorithms are:

Multinomial Naive Bayes (MNB), Logistic

Regression (LR), Linear Support Vector

Classification (SVC), K Nearest Neighbors (KNN),

Decision Tree (DT), Random Forest (RF), Gradient

Boosting (GB), Stochastic Gradient Descent (SGD),

Multi-Layer Perceptron (MLP), and Bagging

Classifier (BC). The objective of this phase is to

identify the new tweet as a fake disaster tweet or a

real disaster one.

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

316

Volume 20, 2023

3.5.1 Multinomial Naive Bayes (MNB)

It is a type of the Naïve Bayesian algorithm, [24],

that is widely applied for text classification. The

MNB estimates the likelihood of the word

occurrence with a class, regardless of the word

position in the text. The Laplace smoothing

technique is employed by the applied classifier to

prevent division by zero while calculating

probabilities. Also, it presumes that the prior

probability distribution for the data is uniform.

3.5.2 Logistic Regression (LR)

It comes from the statistics domain. Its objective,

[25], is to discover the link between the output

(class) and the numerical values (input). It involves

the use of the sigmoid function to determine the

class label. Additionally, it implies the use of the L2

regularization technique, which makes the weights

down towards zero to prevent the over-fitting issue.

3.5.3 Linear Support Vector Classification

(SVC)

It originated from the Support Vector Machine

algorithm; it is suitable to be applied for multi-

classification or single classification. The SVC,

[26], generates the hyperplanes through an

optimization method. A hyperplane is a decision

boundary that splits the input data according to the

classes. The support vectors are the data points that

are closest to the hyperplane. The SVC implies the

use of kernels to easily find the hyperplane, which

can separate different classes.

3.5.4 K Nearest Neighbors (KNN)

Being a non-parametric technique, the KNN, [27],

does not require any restrictions on the distribution

of the data. The classification is established by

computing the distance (the Euclidean distance)

between the test datum input and the training datum

output while taking the customized K value into

account. The number of neighbors (K) used by the

applied KNN classifier is three. All records in the

neighborhood are equally weighted.

3.5.5 Decision Tree (DT)

It is a non-parametric classifier, [28], that doesn't

rely on a mathematical model. It is widely applied in

several domains such as loan approval and disease

diagnosis. Its concept is based on the modeling of a

tree that represents the data structure used to classify

new cases. In the decision tree, the case is stated in

the matter of features/attributes that can be textual

or numerical values. According to the decision rules

that are inferred from the feature values, the

algorithm can predict the output (the decision class).

3.5.6 Random Forest

The random forest algorithm, [29], [30], can be used

for classification and regression. Its main idea is to

build many decision trees that can be trained on

samples of the dataset and find the class label based

on the majority of the decisions of these trees.

Training multiple trees prevents the model from

over-fitting and enhances the classification

accuracy. The applied algorithm utilizes 10 different

trees and applies the Gini measure to determine the

split quality.

3.5.7 Gradient Boosting (GB)

This classifier, [31], utilizes some models (weak

models) to produce a strong one. Each weak model

is a one-split-point decision tree. During training,

the decision tree that produces the minimal error is

added to the other weak models. The applied

algorithm utilizes 100 boosting stages which makes

sure that there is no overfitting. The quality of the

split is measured using the mean squared error

proposed by Friedman.

3.5.8 Stochastic Gradient Descent (SGD)

It is an optimization algorithm, [32], where each

iteration uses a single training data sample, and the

weights are adjusted based on the gradient descent

value. The SGD runs multiple times until it

minimizes the loss function as much as possible.

The applied model uses SGD learning to create a

regularized linear model (linear Support Vector

Machine). Each record's gradient of the loss is

computed, and the model updating is executed at a

decreasing learning rate.

3.5.9 Multi-Layer Perceptron (MLP)

This algorithm, [33], is a type of artificial neural

network that can be used for classification and

involves at least three layers: one for input, one is

hidden, and the last one is the output layer. The

hidden layers can be one or more. This network is

trained for many epochs until the error rate reaches

an acceptable value. The applied algorithm contains

2 hidden layers and each hidden layer has 10

neurons. The applied activation function is ReLU

and the value of alpha is 0.0001. The batch size is

200 and the learning rate is 0.001. The algorithm is

trained for 200 epochs.

3.5.10 Bagging Classifier (BG)

It is an ensemble classifier, [34], [35], where

samples of the data set are used to train the model in

a random order, and the final output is determined

by the mean value of the whole data set. This

classifier uses a base estimator such as a Support

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

317

Volume 20, 2023

Vector Machine or K-nearest neighbors to predict

the suitable class label. The applied algorithm uses

the k nearest neighbor algorithm where the distance

is calculated using the Euclidean distance.

4 Results and Discussion

The implementation of the proposed model is

performed using the Python language through the

Google Colab, [36], and the use of the R program.

The data preprocessing phase, feature extraction

phase as well and machine learning phase are

implemented through Colab whereas the R program

is used to implement the emotion extraction phase.

The disaster dataset is split into 70% training and

30% testing. The proposed model has been

implemented twice. The first implementation

involves the data preprocessing phase, the feature

selection phase, and the classification phase. The

second implementation is based on the data

preprocessing phase, the feature selection phase, the

emotion analysis phase, and the classification phase.

Table 2 demonstrates the results of the classification

without the inclusion of the emotion analysis phase.

Table 3 demonstrates the results of the classification

with the inclusion of the emotion extraction phase.

Table 2. The Model Performance Without Using

Emotion during the Classification Process

Concerning the first implementation of the

proposed model, the classification takes into

consideration the processed tweets and the target

(decision class) without applying the emotion

extraction phase. The LR achieves the highest

accuracy of 88.34%, which is almost the same as the

Multinomial Naïve Bayes (MNB) algorithm with an

accuracy of 88.30%. Furthermore, the linear

Support Vector Classification (SVC) achieves

accuracy above 85%, which is better than other

classifiers.

Table 3. The Model Performance With The Usage

Of The Emotion Analysis During The Classification

Process

Concerning the second implementation of the

proposed model, the classification takes into

consideration the processed tweets, tweet polarity

(a new feature), and the target (the decision class).

The accuracy of most of the applied classifiers with

the inclusion of the emotion extraction phase is

enhanced compared to the first implementation

results. The increased accuracy scores of the SVC,

Gradient Boosting (GB), and Bagging Classifier

(BC) reflect the impact of emotion extraction on the

performance of the classification. The LR achieves

the highest accuracy, 89.39%. In addition, the MNB

and the SVC have achieved an approximately

similar accuracy rate, which is 88%.

Table 4 presents a comparison between our

proposed model and the available literature.

Classifier

Accuracy

Performance

Class

Precision

Recall

F1-Score

MNB

0.8830678960603521

0.97

0.89

0.93

0.54

0.81

0.65

0.8834870075440067

0.97

0.90

0.93

0.55

0.81

0.65

SVC

0.8696563285834031

0.94

0.90

0.92

0.57

0.72

0.64

KNN

(K is 3)

0.8310980720871752

0.99

0.83

0.90

0.20

0.81

0.32

0.8461860854987426

0.93

0.89

0.91

0.53

0.64

0.58

0.8725901089689857

0.99

0.87

0.93

0.39

0.93

0.55

0.8461860854987426

0.99

0.84

0.91

0.26

0.88

0.41

SGD

0.8654652137468567

0.94

0.90

0.92

0.59

0.69

0.63

MLP

0.8616932103939648

0.93

0.90

0.91

0.60

0.67

0.63

0.8088851634534786

1.00

0.81

0.89

0.04

1.00

0.08

Classifier

Accuracy

Performance

Class

Precision

Recall

F1-Score

MNB

0.8822296730930428

0.97

0.90

0.93

0.51

0.78

0.61

0.893964794635373

0.97

0.91

0.94

0.56

0.80

0.66

SVC

0.8818105616093881

0.94

0.92

0.93

0.61

0.71

0.66

KNN

(K is 3)

0.8445096395641241

0.99

0.84

0.91

0.19

0.86

0.31

0.8524727577535625

0.93

0.89

0.91

0.51

0.62

0.56

0.8755238893545683

0.99

0.87

0.93

0.36

0.92

0.51

0.8608549874266554

0.99

0.86

0.92

0.28

0.89

0.43

SGD

0.8763621123218777

0.94

0.91

0.93

0.58

0.70

0.63

MLP

0.863788767812238

0.92

0.63

0.8243922883487007

1.00

0.82

0.90

0.05

1.00

0.09

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

318

Volume 20, 2023

Table 4. A Comparison between the Proposed

Model and Other Models

Model

Methodology

Performance

[9]

BERT

F1-score=88%

[10]

BERT

Accuracy=79%

[11]

Support Vector Machine,

Naive Bayes, and LSTM

Implementation

[12]

linear SVC

F1-score=0.72

[13]

BERT

Accuracy=72%

The

proposed

model

NLP+

Emotion Analysis +

Machine Learning

Accuracy=89,39%

As shown in Table 4, the overall performance of

the applied classifiers on the disaster dataset is

better than the previous research models, as

including the emotion analysis phase in addition to

applying the NLP techniques to the data in the

preprocessing phase has a salient positive impact on

the performance of the proposed model.

5 Conclusion

The disaster tweets can have a great impact on the

decision-making of many people and organizations.

Therefore, it is necessary to have an automated

model that can identify the real disaster tweets from

the fake ones. Most existing classification models

don’t take into consideration the emotional analysis

in the classification process. This paper provides a

new classification model that considers emotion

analysis during the classification process. What

distinguishes the proposed model from other models

is the introduction of an emotion extraction phase

that gives the ability to discover new knowledge

from the tweets. Furthermore, the enrichment of the

disaster dataset with the new feature (tweet polarity)

has a positive impact on the performance of

classifiers. The model achieves an accuracy rate of

89.39%, which is greater than the accuracy of

existing models. The proposed model doesn’t

consider the processing of the emotions’ icons

(emoticons), which may lead to lower performance

if the model is provided with a tweet that contains a

lot of emoticons. For this reason, in future work, the

authors are going to include another phase to handle

the emoticons, which can result higher accuracy

score.

References:

[1] United Nations Office for Disaster Risk

Reduction (UNDRR),

https://www.undrr.org/terminology/disaster

(Accessed Date: Jul. 3 2023).

[2] V. Linardos, M. Drakaki, P. Tzionas, and Y.

Karnavas, “Machine learning in disaster

management: Recent developments in

methods and applications,” Machine Learning

and Knowledge Extraction, vol. 4, no. 2, pp.

446–473, 2022. doi:10.3390/make4020020

[3] S. Madichetty, S. Muthukumarasamy, and P.

Jayadev, “Multi-modal classification of

Twitter data during disasters for humanitarian

response,” Journal of Ambient Intelligence

and Humanized Computing, vol. 12, no. 11,

pp. 10223–10237, 2021. doi:10.1007/s12652-

020-02791-5

[4] M. N. Alenezi and Z. M. Alqenaei, “Machine

learning in detecting COVID-19

misinformation on Twitter,” Future Internet,

vol. 13, no. 10, p. 244, 2021.

doi:10.3390/fi13100244

[5] C. Kang and A. Goldman, “In Washington

Pizzeria attack, fake news brought real guns,”

The New York Times,

https://www.nytimes.com/2016/12/05/busines

s/media/comet-ping-pong-pizza-shooting-

fake-news-consequences.html (Accessed

Date: July 9 2023).

[6] A. Kruspe, J. Kersten, and F. Klan, “Review

article: Detection of actionable tweets in crisis

events,” Natural Hazards and Earth System

Sciences, vol. 21, no. 6, pp. 1825-1845, 2021.

doi:10.5194/nhess-21-1825-2021

[7] M. Zhang, “Applications of deep learning in

news text classification,” Scientific

Programming, vol. 2021, pp. 1–9, 2021.

doi:10.1155/2021/6095354

[8] D. Gupta, N. Narang, M. m, and M. Sood,

“Disaster tweets classification,” SSRN

Electronic Journal, 2022.

doi:10.2139/ssrn.4020410

[9] A. D.Le."Disaster Tweets Classification using

BERT-Based Language Model". arXiv

preprint arXiv:2202.00795. 2022

[10] A. K. Ningsih and A. I. Hadiana, “Disaster

tweets classification in disaster response using

bidirectional encoder representations from

transformer (Bert),” IOP Conference Series:

Materials Science and Engineering, vol. 1115,

no. 1, p. 012032, 2021. doi:10.1088/1757-

899x/1115/1/012032

[11] K. Asinthara, M. Jayan, and L. Jacob,

“Classification of disaster tweets using

machine learning and Deep Learning

Techniques,” 2022 International Conference

on Trends in Quantum Computing and

Emerging Business Technologies (TQCEBT),

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

319

Volume 20, 2023

2022.

doi:10.1109/tqcebt54229.2022.10041629

[12] A. Kumar and S. Reddy, “A detailed analysis

on disaster tweet analysis using Deep

Learning Techniques: DTWEET,” 2022

Fourth International Conference on Emerging

Research in Electronics, Computer Science

and Technology (ICERECT), 2022.

doi:10.1109/icerect56837.2022.10060783

[13] S. Sharma, S. Basu, N. K. Kushwaha, A. N.

Kumar, and P. K. Dalela, “Categorizing

disaster tweets into actionable classes for

disaster managers: An empirical analysis on

Cyclone Data,” 2021 International Conference

on Electrical, Computer, Communications and

Mechatronics Engineering (ICECCME),

2021.

doi:10.1109/iceccme52200.2021.9591063

[14] Python Language- "Welcome to Python.org",

Python.org, 2022. [Online]. Available:

https://www.python.org/. (Accessed Date: 22

July 2023).

[15] R Program - “The R project for statistical

computing,” R, https://www.r-project.org/

(Accessed Date: 24 July 2023).

[16] Disaster tweets- Kaggle,

https://www.kaggle.com/datasets/vstepanenko

/disaster-tweets (Accessed Date: 22 July

2023).

[17] Pandas - Python Data Analysis Library.

Pandas.pydata.org. (2022). (Retrieved on: 24

July 2023), from https://pandas.pydata.org/.

[18] S. Loria, P. Keen, M. Honniba, R.

Yankovsky, D. Karesh, E. Dempsey,

“Textblob: simplified text processing.”,

Secondary TextBlob: simplified text

processing, 3, 2014.

[19] WordnetLemmatizer, NLTK,

https://www.nltk.org/_modules/nltk/stem/wor

dnet.html (Accessed Date: 24 July 2023).

[20] J. Friedl, Mastering Regular Expressions, 3rd

Edition. [S.l.]: O'Reilly Media, Inc., 2006.

[21] M. Jockers, "Extracts Sentiment and

Sentiment-Derived Plot Arcs from Text [R

package syuzhet version 1.0.6]", Cran.r-

project.org, 2022. [Online]. Available:

https://cran.r-

project.org/web/packages/syuzhet/index.html.

(Accessed Date: 24 July 2023).

[22] S. Tongman & N. Wattanakitrungroj,

“Classifying Positive or Negative Text Using

Features Based on Opinion Words and Term

Frequency-Inverse Document Frequency”. In

2018 5th International Conference on

Advanced Informatics: Concept Theory and

Applications (ICAICTA), pp. 159-164,

IEEE,2018.

[23] F. Pedregosa, G. Varoquaux, A. Gramfort, V.

Michel, B. Thirion, O. Grisel, M. Blondel, P.

Prettenhofer, R.Weiss, V.Dubourg,

J.Vanderplas, A.Passos, & D.Cournapeau,

“Scikit-learn: Machine learning in Python”,

the Journal of machine Learning research, vol.

12, pp.2825-2830, 2011

[24] A. M Kibriya, E. Frank, B. Pfahringer & G.

Holmes, 2004,” Multinomial naive bayes for

text categorization revisited”. In Australasian

Joint Conference on Artificial Intelligence,

pp. 488-499, Springer, Berlin, Heidelberg,

2004.

[25] J. Brownlee,” Master Machine Learning

Algorithms: discover how they work and

implement them from scratch”, Machine

Learning Mastery. Chapter 13: Logistic

Regression, pp.51-55, 2016a.

[26] J. Novakovic, & A. Veljović, “C-support

vector classification: Selection of kernel and

parameters in medical diagnosis”. In 2011

IEEE 9th International Symposium on

Intelligent Systems and Informatics, pp. 465-

470, 2011.

[27] J. Goldberger, S. Roweis, G. Hinton, and R.

Salakhutdinov,” Neighbourhood components

analysis”. In Advances in neural information

processing systems, pp. 513-520, 2005.

[28] T. Hastie, T. Robert, and F. Jerome,” The

elements of statistical learning: data mining,

inference, and prediction”. Springer Science

& Business Media, 2009.

[29] J. Brownlee,” Master Machine Learning

Algorithms: discover how they work and

implement them from scratch”, Machine

Learning Mastery. Chapter 28: Bagging and

Random Forest, pp.126-129, 2016b

[30] G. Biau, 2012. Analysis of a random forests

model. The Journal of Machine Learning

Research, vol 13, no.1, pp.1063-1095,2012.

[31] J. Friedman, "Greedy function approximation:

A gradient boosting machine.", The Annals of

Statistics, vol. 29, no. 5, pp. 1189-1232, 2001.

[32] L. Bottou,” Large-scale machine learning with

stochastic gradient descent”, In Proceedings

of 19th International Conference on

Computational Statistics (COMPSTAT'2010),

pp. 177-186, Physica-Verlag HD, 2010.

[33] H. Ramchoun, M. Amine, J. Idrissi, Y.

Ghanou and M. Ettaouil, "Multilayer

Perceptron: Architecture Optimization and

Training", International Journal of Interactive

Multimedia and Artificial Intelligence, vol. 4,

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

320

Volume 20, 2023

no. 1, pp. 26-30, 2016. Available:

10.9781/ijimai.2016.415.

[34] J. Brownlee, Master Machine Learning

Algorithms: discover how they work and

implement them from scratch. Machine

Learning Mastery. Chapter 28, pp.26-127,

2016c.

[35] L. Breiman, "Bagging predictors", Machine

Learning, vol. 24, no. 2, pp. 123-140, 1996.

Available: 10.1007/bf00058655.

[36] Google Colab,

https://colab.research.google.com/ (Accessed

Date: 24 July 2023).

Contribution of Individual Authors to the

Creation of a Scientific Article (Ghostwriting

Policy)

The authors equally contributed to the present

research, at all stages from the formulation of the

problem to the final findings and solution.

Sources of Funding for Research Presented in a

Scientific Article or Scientific Article Itself

No funding was received for conducting this study.

Conflict of Interest

The authors have no conflicts of interest to declare.

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the

Creative Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en

_US

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS

DOI: 10.37394/23209.2023.20.34

Marco Alfonse, Mariam Gawich

E-ISSN: 2224-3402

321

Volume 20, 2023