Cyber Aggression and Cyberbullying Identification on

Social Networks

Vincenzo Gattulli

, Donato Impedovo

, Giuseppe Pirlo

and Lucia Sarcinella

Digital Innovation Srl, Via Edoardo Orabona, 4 (c/o Dipartimento di Informatica), 70125 Bari, Italy

Department of Computer Science, University of Studies of Bari “Aldo Moro”, Via Edoardo Orabona, 4, 70125 Bari, Italy

Keywords: Cyberbullying, Artificial Intelligence, Social Network, Cyber Aggression, Twitter, Machine Learning.

Abstract: Bullying includes aggression, harassment, and discrimination. The phenomenon has widespread with the great

diffusion of many social networks. Thus, the cyber aggression iteration turns into a more serious problem

called Cyberbullying. In this work an automatic identification system built up on the most performing set of

techniques available in literature is presented. Textual comments of various Italian Twitter posts have been

processed to identify the aggressive phenomenon. The challenge has been also identifying aggressive profiles

who repeat their malicious work on social networks. Two different experiments have been performed with

the aim of the detection of Cyber Aggression and Cyberbullying. The best results were obtained by the

Random Forest classifier, trained on an ad-hoc Dataset that contemplates a series of comments extracted from

Twitter and tagged manually. The system currently presented is an excellent tool to counter the phenomenon

of Cyberbullying, but there are certainly many improvements to be made to improve the performance of the

system.

1 INTRODUCTION

Social Networks are progressively exposed to

harmful issues including Cyber Aggression and

Cyberbullying. Cyber Aggression refers to aggressive

online behaviour using digital media content (text,

images, videos, etc.) to cause harm to another person.

Cyberbullying is defined as "An aggressive

intentional act by an individual or a group of

individuals, using electronic forms of contact,

repeated over time against a victim who cannot easily

defend himself" (Dredge et al., 2014).

This work deals with the automatic recognition of

Cyber Aggression (detected in textual comments) and

cyberbully profiling (Cyber Aggression repeated by a

certain user in multiple posts over time). The

experiments aim to prevent the phenomenon of

Cyberbullying on Social Networks. In this study

Italian behavioral patterns will be studied, recovering

recurrent patterns in the formulation of Italian

sentences or in the typology of attacks with the aim

of collecting and labeling comments and creating a

dataset called "Italian Aggressive Dataset". The

attention is mainly focused on the kind of language

used by the attacker considering vocabularies of

expressions and words belonging to the vulgar jargon.

Next, each word within the sentence is weighted

according to its negative, neutral, or positive value

along with a large set of other characteristics. The

main contributions of this work are:

 An ensemble of features with a comparison of

different classification models.

 The creation of a vocabulary of Italian words

considering four types of categories: Bad Word,

Second Person, Threats, Bulling Terms. These

dictionaries contain some of the most common

terms in Italian, used to verbally attack and

offend someone. The definition of these words

was made by viewing countless comments under

the posts of famous singers and politicians.

 "Aggressive Italian Dataset": Creation and

labeling of a balanced Italian dataset composed

of aggressive and non-aggressive comments,

extracted from the social platform Twitter

named.

The rest of the work is organized as follows.

Section 2 will illustrate the state of the art. Section 3

will describe the software design and implementation

methodology. Section 4 will describe the "Italian

Aggressive Dataset". The results of the

experimentation are provided in Section 5. Finally,

Section 6 concludes the document.

644

Gattulli, V., Impedovo, D., Pirlo, G. and Sarcinella, L.

Cyber Aggression and Cyberbullying Identiﬁcation on Social Networks.

DOI: 10.5220/0010877600003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 644-651

ISBN: 978-989-758-549-4; ISSN: 2184-4313

2 RELATED WORK

This document focuses on textual data being the most

widespread data in social media having the aim in

identifying aggression and pattern could be referred to

cyberbullies in their mild stage (Shah et al., 2021).

Many researchers have worked on textual comments

collected on Social Networks. Amali et. al (Ishara

Amali & Jayalal, 2020) with the aim of determining

insults (profane words) in the comments within the

tweets, five rules were taken into consideration (Ishara

Amali & Jayalal, 2020): i) percentage of bad words

within the tweet, ii) combination of first-person

pronoun, bad word and a second person pronoun, iii)

combination of second person pronoun with a bad

word combination of third person pronoun with a bad

word, iv) combination of first-person pronoun, bad

word and a third person pronoun. Selected comments

were successively classified adopting SVM, K-

Nearest Neighbors (KNN) and Naïve Bayes (NB): the

SVM with RBF kernels scored better than others

reporting a 91% f1-score.

Chatzakou et al. (Chatzakou et al., 2017)

processed 1.6 million tweets collected over 3 months

of conversations. In this case user-based features, text-

based features and network-based features were

considered. User-based features had the aim of

describing the general user’s behaviour (e.g., bully,

and generic aggressors are faster than normal users in

posting activity), text-based features were referred to

uppercases, specific word embedding and to the

positive/negative sentiment in short text. The

network-based features were aimed to evaluate

popularity, reciprocity, power difference and

influence of users within the group. RF classifier was

able to perform 90% of AUC (Chatzakou et al., 2017).

Raza et al. (Raza et al., 2020) developed a model with

LR, RF and NB algorithms to identify if a particular

comment is an insult, threat, or a hate message, with

Voting and AdaBoost classifiers. Supervised machine

learning with LR achieved 82.7% accuracy. With the

voting classifier, an accuracy of 84.4% was observed

(Raza et al., 2020). Shtovba et al. (Shtovba et al.,

2019) found syntactic dependencies in comments, i.e.,

relationships with proper nouns, personal pronouns,

possessive pronouns, etc. Three features were

highlighted that greatly improve detection: the number

of dependencies with proper names in the singular, the

number of dependencies that contain profanity, and

the number of dependencies between personal

pronouns and profanities. The data used comes from

the Kaggle contest "Toxic Comment Classification

Challenge (Large number of Wikipedia comments)".

An DT classifier is used (Shtovba et al., 2019).

Dwivedi et al. (Kumari et al., 2021b) present a

deep learning-based model (LSTM network) detecting

different levels of aggression (direct, indirect and no

aggression) in social media posts in a bilingual

scenario. Datasets from Facebook and Twitter with

bilingual (English and Hindi) data were used (Kumari

et al., 2021b). Sentiment description has been also

considered evaluating comments i) contain remarks,

critic, sarcasm, etc., ii) referred to specific topics (e.g.,

politics, crimes, race, sex, etc.), iii) containing swear

words. In this case three different classifiers were

adopted: KNN, SVM, and LR. The best performances

were achieved by the SVM with linear kernel

reporting 86% on accuracy and recall and 84% of f1

score (Chen et al., 2017). The automatic detection of

cyberbullying can be exploited considering

psychological features of users, including

personalities, feelings, and emotions. User

personalities can be determined using the Big Five

model (openness to experience, conscientiousness,

extraversion, agreeableness, and neuroticism) (Costa

& McCrae, 1992) (John & Srivastava, n.d.) and Dark

Triad (narcissism, Machiavellianism, and

psychopathy) which specifically refers to malevolent

qualities (Paulhus & Williams, 2002) (Goodboy &

Martin, 2015). Many scientific papers have used

hybrid approaches, correlating images (Dentamaro et

al., 2021) and post comments under the same images.

Singh et al. (Kumari et al., 2021a) present a model

based on Convolutional Neural Network (CNN) and

Binary Particle Swarm Optimization (BPSO) to

classify social media posts containing images with

associated textual comments into non-aggressive,

medium aggressive and highly aggressive classes. The

proposed model with optimized features and Random

Forest classifier (Dentamaro et al., 2020) achieves a

weighted F1-Score of 0.74 (Kumari et al., 2021a).

Kumari et al. (Kumari & Singh, 2021a) present textual

features extracted using a three-layer parallel

convolutional neural network. The image and text

features are then combined to obtain a hybrid feature

set that is further optimized using a binary firefly

optimization algorithm (Kumari & Singh, 2021a).

Finally, Singh et al. (Kumari & Singh, 2021b) present

a pre-trained VGG-16 network and a convolutional

neural network to extract features from images and

text, respectively. These features are further optimized

using a genetic algorithm to increase the efficiency of

the whole system. The proposed model achieves an F1

score of 78% (Kumari & Singh, 2021b). The hybrid

approach was not considered due to both the lack of

datasets and the poor performance reported in the read

article.

Cyber Aggression and Cyberbullying Identiﬁcation on Social Networks

645

3 METHODS

The proposed approach is organized in a pipeline

made-up of three stages:

A. Post selection and test comments extraction;

B. Feature engineering;

C. Classification and Metrics.

This paper proposes two different experiments.

The first one aims to identify Cyber Aggression, the

second one aims to identify Cyberbullying. The first

experiment identifies aggression from user comments.

In case an aggression is identified by the classifier and

there are multiple aggressions on multiple posts by the

same user, then that user could be flagged as an

aggressive profile (bully), thus giving rise to the

second phase. The system is designed not only to run

experiments as described in this paper, but also to be

able to work online. For the Training phase, the dataset

created in this study called the Aggressive Italian

Dataset is used. For the testing phase, an additional

1000 different comments were extracted from

different Twitter posts for each of the four celebrities

that we will discuss in the next subchapter. The

importance of identifying the different posts is related

to the problem of identifying cyberbullies stalking the

victim. The comments selected for the Test phase

were manually labeled, and the feature extraction

phase was performed for each comment, as well as for

the Training. Twitter comments from the Test phase

extracted are in Italian, dated November-December

2020. During the period considered, each post

contained approximately 100/150 comments (6

Twitter posts). In summary, the "Aggressive Italian

Dataset" containing 3028 comments was used for the

Training phase. As a Test, 1000 comments of different

posts were extracted for each famous person. Figure 1

illustrates the phases of the experiment.

Figure 1: General scheme of the system.

A. Post Selection and Test Comments Extraction

Famous people with large audiences and many

followers clearly attract both supporters and "haters".

Four famous Italian people who suffer acts of Cyber

Aggression have been considered in this work. Many

users carry out verbal aggression by commenting

under each post, also attacking private life, behavior

very similar to a Cyberbullying action. Profiles here

considered are: Achille Lauro (Italian singer/rapper),

Fabio Rovazzi (Italian singer and youtuber), Matteo

Renzi and Giuseppe Conte (Italian politicians). The

period selected for posts is between November-

December 2020. In this period, Italy was in a

government crisis and a coronavirus pandemic.

B. Feature Engineering

In the feature extraction phase, nine features are

considered:

Number of Negative Words (BW). This feature

has been implemented by means of a "BadWord"

vocabulary containing 540 negative words extremely

vulgar used for aggressive purposes, offenses, and

humiliations. The use of regular expressions has

allowed to identify also negative words written

grammatically incorrect, all attached or with spaces

(Ex. Assssshole  asshole) (Ishara Amali & Jayalal,

2020).

Number of "non/no"(NN). The use of "no/not"

within a sentence completely changes the meaning of

the sentence from positive to negative or vice versa.

Furthermore, the presence of a large number of

“no/not” can underline the controversy of the

comment.

Uppercase (U). This is a Boolean value

indicating whether the comment is capitalized or not.

In computer jargon, uppercase comment is about

shouting something. So, it can be interpreted as an

aggression against someone (Chatzakou et al., 2017).

Positive/negative Weight of the Comment

(PW/NW). This feature includes two values: a

positive and negative weight of the comment within

the range [0,1]. To do this, the relative synset and

weight of each word was extracted, using WordNet

and SentiWordNet (M et al., 2017) (Rendalkar &

Chandankhede, 2018) and then averaged for both

positive and negative weights. The average value was

chosen to take into account the length of the comment

and therefore the number of words.

Use of the Second Person (SP). It is a Boolean

value indicating the presence or absence of a second

singular or plural form in the comment. This feature

is important because attacks are often accompanied

using the second person, thus targeting a specific

person. This feature was extracted through a specially

created dictionary containing 24 words, including

verbs and pronouns referring to the second person, for

example words ending with "tu" or "ti" ("you" in

English) (Shtovba et al., 2019).

Presence of Threats (TR), instigation to violence

or suicide. A Boolean value indicates the presence of

threats, violence, or instigation to suicide within the

comments. Many negative comments are accompa-

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

646

nied using profane language or threats such as "I kill

you" or instigations to suicide "thrown from a

bridge", all expressions used only in aggressive

contexts. Even in this case, these expressions have

been identified using a specifically devoted

vocabulary containing 314 violent or instigating

expressions (Raza et al., 2020).

Presence of Bulling Terms (KW). A Boolean value

indicates the presence of the so-called keywords of

cyberbullying, or insults, used to injure or attack a

person (e.g., idiot, stupid, ...), but also target words

which in themselves do not take on a negative

meaning, but in some contexts, such as that of

cyberbullying, they can be used equally to insult (e.g.,

clown, whale, garbage, ...). In this case a vocabulary

containing 359 terms identified as insults and possible

insults has been created.

Comment Length (L). This feature represents the

length of the comment in terms of words. In fact, it

has been observed that most negative comments are

made up of a few words, usually no more than three.

The choice of these nine features was dictated by

both the state of the art and a careful analysis made

on the real comments of the people of Twitter. It has

been found that the language on the web is rough, full

of expressions and words belonging to the vulgar

jargon that leaves little room for misunderstanding.

The Italian language has many identical terms that are

used when verbally attacking someone. This led to the

creation of a list of these words, creating a veritable

dictionary of profane words. The weight of profane

words does not have to decree with certainty the

negativity of the sentence, for this reason another

feature has been devised that considers the weight

that a word can have in a sentence, both in negative

and in positive. Again, it was noted that some

negative comments were capitalized, as if to simulate

a higher tone of voice. This has led to thinking of a

way to keep track of this particularity. Another

feature that was highlighted is the presence of

negation in aggressive comments, in fact in many

cases it was noted that the aggression sessions started

with the word "no / non" to contradict the victim.

Again, the presence of the second person, an example

would be "TI uccido" (I kill you), "DEVI morire"

(You must die). As an enrichment of the vocabulary

on "profane words", two other vocabularies have

been defined with expressions very close to

aggressive juvenile language. The first is defined as

expressions of incitement to violence with the

purpose of wishing someone's death. The second are

defined as expressions linked to juvenile and

offensive language, closely linked to pokes and

assonances with animals in a derogatory way.

Aggression in both bullying and in a more general

context embraces these themes which have been

gradually considered and applied to the extraction of

each individual comment.

C. Classification and Matrics

The classification of the comments has been carried

out using four supervised classification algorithms:

SVM with linear kernel, RF, MLP and DT. The

problem of classification has been considered here as

a two class one: Aggressive comments and Non-

Aggressive ones. In this work the SVM kernel is linear

because it works well for text classification (Malmasi

& Zampieri, 2018) (Davidson et al., n.d.). In this work

the maximum RF depth has been set at 10, and the

number of estimators is set at 1800 (Islam et al., 2019)

(Chatzakou et al., 2017). In this work the MLP alpha

parameter has been set equal to 0.05, hidden layer

levels equal to (25, 20) and learning rate equal to 0,001

(Ramchoun et al., 2016). The parameters considered

were tested as best after a Greed Search approach.

Four parameters were considered to evaluate the

system performance: Accuracy, Precision (P), Recall

(R), F1-score (F1) (Prastowo et al., 2019).

4 DATASET

The "Aggressive Italian Dataset" consists of Italian

comments extracted from Twitter, both Aggressive

and Non-Aggressive and contains 3028 comments.

Comments were divided between 1514 aggressive

and 1514 non-aggressive. The dataset was carefully

balanced keeping the same number of aggressive and

non-aggressive comments, labelled (T) with the

manual procedure explained below. Each comment

was analyzed by ten people, each person categorized

the comment as aggressive and non-aggressive

through their attitude towards the issue.

Finally, the most frequent classes were assigned

to each of the comments. Aggression was understood

as any form of aggression that hurt the sensibilities of

the person being attacked. The content of the

comments did not have to contain a profane word, but

a verbal attack that could hurt any person receiving

that message. While, about comments classified as

non-aggressive, those comments that did not go to

hurt the sensitivity of others were considered. After

labeling, statistically it was noted that the people in

question agreed because the selected comments carry

little ambiguity. Many aggressive comments

registered feelings of violence and aggression

account a particular person. If the dataset will be

extended and shared the labeling part will be better

specified.

Cyber Aggression and Cyberbullying Identiﬁcation on Social Networks

647

Table 1: Evaluation of the comments of the last six posts by Achille Lauro.

SVM DTRFMLP

chille Lauro P R F1 P R F1 P R F1 P R F1

-aggressive 0.98 0.88 0.93 0.94 0.86 0.90 0.99 0.91 0.95 0.96 0.91 0.94

Aggressive 0.70 0.94 0.81 0.64 0.83 0.72 0.77 0.98 0.86 0.75 0.75 0.81

Accuracy 0.90 0.85 0.93 0.90

Table 2: Evaluation of the comments of the six posts of Fabio Rovazzi.

SVM DT RF MLP

abio Rovazzi P R F1 P R F1 P R F1 P R F1

-aggressive 0.94 0.84 0.89 0.89 0.78 0.84 0.98 0.83 0.90 0.92 0.86 0.89

Aggressive 0.75 0.90 0.82 0.66 0.82 0.73 0.75 0.97 0.85 0.76 0.87 0.81

Accuracy

0.86 0.80 0.88 0.86

Table 3: Evaluation of the comments of the last six posts by Matteo Renzi.

SVM DT RF MLP

atteo Renzi P R F1 P R F1 P R F1 P R F1

-aggressive 0.98 0.95 0.97 0.98 0.95 0.96 0.99 0.98 0.98 0.98 0.96 0.97

Aggressive 0.74 0.89 0.81 0.71 0.84 0.77 0.85 0.95 0.90 0.76 0.88 0.82

Accuracy 0.94 0.94 0.97 0.95

Table 4: Evaluation of the comments of the last six posts by Giuseppe Conte.

SVM DT RF MLP

Giuseppe Conte P R F1 P R F1 P R F1 P R F1

-aggressive 0.93 0.84 0.89 0.89 0.77 0.83 0.96 0.85 0.90 0.90 0.85 0.88

Aggressive 0.73 0.87 0.79 0.64 0.81 0.71 0.75 0.92 0.82 0.82 0.84 0.78

Accuracy 0.85 0.80 0.87 0.84

Table 5: Extract of the table containing the profiles of cyberbullies.

Twitter Users nAC nC nAC

Lauro

nAC

Rovazzi

nAC

Renzi

nAC

Conte

Peppe*** 7 12 7 0 0 0

paoloG*** 6 8 0 3 3 0

nonseic*** 6 8 6 0 0 0

peso*** 6 8 0 6 0 0

bettav*** 5 5 0 0 5 0

marcoLu** 4 9 0 0 4 0

alessand_* 4 4 0 4 0 0

fatazu**** 4 13 0 0 0 4

Mart*** 4 4 0 4 0 0

… … … … … … …

Table 6: Extract of table containing the profiles of the victims.

RF Model Number of aggressive

comments predicted

Number of actual

aggressive

comments

Aggressive comments

classification error percentage

Achille Lauro 123 129 5%

Fabio Rovazzi 224 229 2%

Matteo Renzi 332 342 3%

Giuseppe Conte 302 329 8%

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

648

5 EXPERIMENTS

5.1 First Experiment

In the first experiments are revealed multiple

observations. The results for the different characters

are very similar to each other, but the RF model

achieved the best results considering the Accuracy

value for each of the four famous characters. In fact,

the results of the RF model regarding the

identification of comments without Cyber

Aggression, P is between 96-99%, R between 83-98%

and F1-score values between 90-98%. As for the

identification of aggressive comments, the RF

achieves an P ranging from 75% to 84%, R values

ranging from 92% to 98%, F1 score values ranging

from 82% to 90%. The results are shown in Tables 1,

2, 3 and 4. In general, however, the accuracy of the

entire system is very high, with values ranging from

80% to 98%. The cases in which the accuracy

assumes very low values, are within the posts in

which there were several ironic comments or

offensive words hidden by an incorrect use of

grammar, not identified by the system. All capitalized

comments that do not present aggression could be

identified as non-aggressive, or situations in which

the aggression does not refer to the person in the post,

but to another of similar social class. Hashtags that

may contain vulgar and offensive slogans are not

recognized.

5.2 Second Experiment

Based on considerations previously reported, a basic

and easy to implement approach to reveal a tendency

or a propensity to a trait of cyberbullying is to

evaluate recurrence of aggressions. To the aim the

total number of comments considered as Cyber

Aggression by the system have been enumerated.

Two tables were created. In the first, all the users who

have been identified for a Cyber Aggression through

their comments are stored, and for each of them the

total number of comments, the number of comments

identified as Cyber Aggression and the related posts

are marked. With this first table it is therefore

possible to identify the cyberbully, detecting who has

many cyber-attacks in general or aimed at a specific

victim. An example would be the user named

"paoloG****" who wrote 8 comments, and more than

half were classified as Cyber Aggressive. In addition,

three were addressed on several posts to the politician

Matteo Renzi and three were addressed to the singer

Fabio Rovazzi. This means that this user does not

attack a single social category but attacks more than

a semantically different social category. A simpler

example could be the user "mart****" who made four

comments in several posts by Fabio Rovazzi, all

classified as Cyber Aggressive (see Table 5).

The second table (see Table 6), on the other hand,

shows the results of the Random Forest. It can be seen

that for all four characters, a 10% detection error was

made out of 1000 standard comments in the Test

phase (see Table 6). In the tables, the wording "nC"

identifies the number of comments made by the

particular user, while the wording "nAC" identifies

the aggressive comments predicted by the system.

6 CONCLUSIONS

In this work, the problem of Cyber Aggression related

to Cyberbullying was considered. The results

obtained show that the identification of aggressive

comments is done with a good degree of accuracy.

Different classification schemes were compared, and

Random Forest (RF) was found to be the one that

achieved the highest accuracy for all the different

cases considered here. The next step was the

identification of Cyberbullying sessions by tracking

users who posted comments classified as aggressive.

Through this analysis it was possible to obtain the

profiles of the pages most prone to this type of

phenomenon, being able to monitor these victims, to

report the situation to the competent authorities. In

addition, in this work have been considered some

innovative Feature Engineering phases and some

state-of-the-art ones, that have allowed together with

the creation of the "Aggressive Italian Dataset" in

Italian, the possibility to identify common patterns in

the Italian culture that could identify an aggression.

In addition to the macro dataset, innovative sub

vocabularies were created that also allowed the

identification of verbal aggression. This procedure

could be performed online, in a smaller context such

as schools to prevent cyberbullying. These datasets

could only be shared if the article is accepted.

Regarding future developments, the detection of

Cyberbullying comments should be improved, as it

was noted that many slang forms, or even

grammatically incorrect, were not identified, so it is

possible to expand the vocabularies used. It is also

possible to make a deeper analysis by monitoring

victims and cyberbullies to understand the frequency

or reasons that lead to this phenomenon and thus

prevent them. Finally, it would be interesting to

greatly expand the dataset by adding comments to the

posts of other people, not necessarily famous, but

ordinary people.

Cyber Aggression and Cyberbullying Identiﬁcation on Social Networks

649

ACKNOWLEDGEMENTS

This work is supported by the Italian Ministry of Education,

University and Research within the PRIN2017 -

BullyBuster project - A framework for bullying and

cyberbullying action detection by computer vision and

artificial intelligence methods and algorithms.

REFERENCES

Chatzakou, D., Kourtellis, N., Blackburn, J., de Cristofaro,

E., Stringhini, G., & Vakali, A. (2017). Mean Birds:

Detecting Aggression and Bullying on Twitter. WebSci

2017 - Proceedings of the 2017 ACM Web Science

Conference, 13–22. https://arxiv.org/abs/1702.06877v3

Chen, J., Yan, S., & Wong, K. C. (2017). Aggressivity

detection on social network comments. ACM

International Conference Proceeding Series, Part

F127854, 103–107.

https://doi.org/10.1145/3059336.3059348

Costa, P. T., & McCrae, R. R. (1992). Four ways five

factors are basic. Personality and Individual

Differences, 13(6), 653–665.

https://doi.org/10.1016/0191-8869(92)90236-I

Davidson, T., Warmsley, D., Macy, M., & Weber, I. (n.d.).

Automated Hate Speech Detection and the Problem of

Offensive Language *. Retrieved September 14, 2021,

from www.facebook.

Dentamaro, V., Giglio, P., Impedovo, D., & Pirlo, G.

(2021). Benchmarking of Shallow Learning and Deep

Learning Techniques with Transfer Learning for

Neurodegenerative Disease Assessment Through

Handwriting. Lecture Notes in Computer Science

(Including Subseries Lecture Notes in Artificial

Intelligence and Lecture Notes in Bioinformatics),

12917 LNCS, 7–20. https://doi.org/10.1007/978-3-

030-86159-9_1

Dentamaro, V., Impedovo, D., & Pirlo, G. (2020). Gait

Analysis for Early Neurodegenerative Diseases

Classification through the Kinematic Theory of Rapid

Human Movements. IEEE Access, 8, 193966–193980.

https://doi.org/10.1109/ACCESS.2020.3032202

Dredge, R., Gleeson, J., & de La Piedad Garcia, X. (2014).

Presentation on Facebook and risk of cyberbullying

victimisation. Computers in Human Behavior, 40, 16–

22. https://doi.org/10.1016/J.CHB.2014.07.035

Goodboy, A. K., & Martin, M. M. (2015). The personality

profile of a cyberbully: Examining the Dark Triad.

Computers in Human Behavior, 49, 1–4.

https://doi.org/10.1016/J.CHB.2015.02.052

Ishara Amali, H. M. A., & Jayalal, S. (2020). Classification

of Cyberbullying Sinhala Language Comments on Social

Media. MERCon 2020 - 6th International

Multidisciplinary Moratuwa Engineering Research

Conference, Proceedings, 266–271.

https://doi.org/10.1109/MERCON50084.2020.9185209

Islam, M. Z., Liu, J., Li, J., Liu, L., & Kang, W. (2019). A

semantics aware random forest for text classification.

International Conference on Information and Knowledge

Management, Proceedings, 1061–1070.

https://doi.org/10.1145/3357384.3357891

John, O. P., & Srivastava, S. (n.d.). The Big-Five Trait

Taxonomy: History, Measurement, and Theoretical

Perspectives.

Kumari, K., & Singh, J. P. (2021a). Multi-modal cyber-

aggression detection with feature optimization by

firefly algorithm. Multimedia Systems.

https://doi.org/10.1007/S00530-021-00785-7

Kumari, K., & Singh, J. P. (2021b). Identification of

cyberbullying on multi-modal social media posts using

genetic algorithm. Transactions on Emerging

Telecommunications Technologies, 32(2), e3907.

https://doi.org/10.1002/ETT.3907

Kumari, K., Singh, J. P., Dwivedi, Y. K., & Rana, N. P.

(2021a). Bilingual Cyber-aggression detection on

social media using LSTM autoencoder. Soft Computing

2021 25:14, 25(14), 8999–9012.

https://doi.org/10.1007/S00500-021-05817-Y

Kumari, K., Singh, J. P., Dwivedi, Y. K., & Rana, N. P.

(2021b). Multi-modal aggression identification using

Convolutional Neural Network and

Binary Particle Swarm Optimization. Future

Generation Computer Systems, 118, 187–197.

https://doi.org/10.1016/J.FUTURE.2021.01.014

M, R. v, Kumar, M. P., Raman, S. R., & Sridhar, R. (2017).

Emotion And Sarcasm Identification Of Posts From

Facebook Data Using A Hybrid Approach. Online)

Ictact Journal On Soft Computing, 2.

https://doi.org/10.21917/ijsc.2017.0197

Malmasi, S., & Zampieri, M. (2018). Challenges in

Discriminating Profanity from Hate Speech. Journal of

Experimental and Theoretical Artificial Intelligence,

30(2), 187–202. https://arxiv.org/abs/1803.05495v1

Paulhus, D. L., & Williams, K. M. (2002). The Dark Triad

of personality: Narcissism, Machiavellianism, and

psychopathy. Journal of Research in Personality, 36(6),

556–563. https://doi.org/10.1016/S0092-

6566(02)00505-6

Prastowo, E. Y., Endroyono, & Yuniarno, E. M. (2019).

Combining SentiStrength and Multilayer Perceptron in

Twitter Sentiment Classification. Proceedings - 2019

International Seminar on Intelligent Technology and Its

Application, ISITIA 2019, 381–386.

https://doi.org/10.1109/ISITIA.2019.8937134

Ramchoun, H., Amine, M., Idrissi, J., Ghanou, Y., &

Ettaouil, M. (2016). Multilayer Perceptron:

Architecture Optimization and Training.

International Journal of Interactive Multimedia

and Artificial Intelligence, 4(1), 26.

https://doi.org/10.9781/IJIMAI.2016.415

Raza, M. O., Memon, M., Bhatti, S., & Bux, R. (2020).

Detecting Cyberbullying in Social Commentary Using

Supervised Machine Learning. Advances in Intelligent

Systems and Computing, 1130 AISC, 621–630.

https://doi.org/10.1007/978-3-030-39442-4_45

Rendalkar, S., & Chandankhede, C. (2018). Sarcasm

Detection of Online Comments Using Emotion

Detection. Proceedings of the International Conference

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

650

on Inventive Research in Computing Applications,

ICIRCA 2018, 1244–1249.

https://doi.org/10.1109/ICIRCA.2018.8597368

Shah, D. K., Sanghvi, M. A., Mehta, R. P., Shah, P. S., &

Singh, A. (2021). Multilabel Toxic Comment

Classification Using Supervised Machine Learning

Algorithms. Lecture Notes in Networks and Systems,

141, 23–32. https://doi.org/10.1007/978-981-15-7106-

0_3

Shtovba, S., Shtovba, O., & Petrychko, M. (2019).

Detection of Social Network Toxic Comments with

Usage of Syntactic Dependencies in the Sentences.

Undefined.

Cyber Aggression and Cyberbullying Identiﬁcation on Social Networks

651