CYBERBULLYING MITIGATION IN SOCIAL NETWORKS

Shankara Gowda S R1, Tarun S R2, Sushmitha R3, Swathi D4

1
Department of Information Science and Engineering, Don Bosco Institute of Technology,Bangalore, Karnataka,
India.
2
Department of Information Science and Engineering, Don Bosco Institute of Technology,Bangalore, Karnataka,
India.
3
Department of Information Science and Engineering, Don Bosco Institute of Technology,Bangalore, Karnataka,
India.
4
Department of Information Science and Engineering, Don Bosco Institute of Technology,Bangalore, Karnataka,
India.

*Corresponding Author: Shankara Gowda S R Email:[email protected]

Abstract:
Detecting Cyberbullying is essential owing to its harmful impact on mental health, potentially causing
depression and low self-esteem, necessitating the advancement of automated tools for intervention. The
project focusses on hate speech detection in social media ,employing a hybrid approach of natural
language processing and ensemble machine learning .Data undergoes pre-processing steps like
stemming ,token splitting, character removal, and infection elimination .Various classifier , such as
support vector machine, decision tree , random forest ,relevance vector machine and Naïve Bayes are
utilized .The objective is to create a language independent model capable of classifying code mixed post
into hate speech, offensive language ,or non-hate speech. The ensemble technique enhances
cyberbullying detection. The fine-tuned resulted with highest F measure of 72.42 %. Our study
determined that transfer learning embedded as the optional approach for achieving enhanced
performance with reduced effort, due to elimination of feature engineering and resampling requirements.

Key Words:Cyberbullying Detection, Distil Bert, machine learning, pre-trained language models
(PLMs), Transfer learning, Toxicity Features, AmiCa Dataset, LIWC, empath.
1. Introduction

This project endeavors to tackle the persistent and conservative issues of hate crimes with our society
shining a spotlight on the increasing influential role of online communication channels in the realization
process. Notably, individuals implemented in the recent hate related terror attacks frequently exhibited
tribal history of discriminating the hate infused content across various online channels. The detection of
hate speech proves to a formidable challenge as, further compounded by the absence of universally
agreed upon. Technicians, leading to subjective interpretations based on the individual perspectives.

Existing studies have made in proposing solutions for the automatic detection of hate speech, employing
a range of feature engineering techniques and machine learning algorithms. However, our critical gap
exists within this body of work, as it lacks a comprehensive comparative assessment of the performance.
Of distinct feature engineering techniques and machine learning algorithms. In response to this gap, our
proposed system leverages natural language processing techniques and adapts an ensemble machine
learning approach, amalgamating various classification techniques to enhance the accuracy of hate
speech detection. The primary objective of our project is multifaceted in composing the minimization of
misclassification rates, the development of efficient system for precise hate speech. Detection, the
creation of user-friendly desktop application capable of categorizing user comments, and the execution
of an ensemble machine learning approach to argument classification. The overarching purpose is to
contribute a desktop UI application for hate speech classification, with a specific focus on real time
identification of hate speech within the tweets on the Twitter platform.

Addressing the issue of cyber bullying within our communities, necessities are proactive approach,
emphasizing the critical role of timely detection for effective prevention and reduction. These studies
make a noteworthy contribution by leveraging linguistic processing and surprise learning techniques to
automate the identification of cyberbullying incidents. The research delivers into examination of features
and methodologies aimed at modeling the automated detection of cyber bullying within

Textual content. These findings significantly improve the performance of test classification for cyber
bullying events, paving the way for improved solutions that online communication channels can utilize
in their efforts to combat cyber bullying.

Section I lays the groundwork by offering an overview of the study, stressing the significance of
automated cyber bullying detection, the motives behind the research, and highlighting the significance of
applying data science knowledge in this content. Section II provides A comprehensive overview of
recent works in the Reclaim of Cyber Bullying detection. Section III details the methodology and
process flow, offering insights into the tools employed and the system requirements during the research.
Section IV unveils the experimental results. This chapter not only benchmarks against previous studies,
but also develops into a detailed discussion of the findings. Full Finally, Section VI encapsulates the
research by summarizing the adopted approach, addressing the encountered limitations, and laying the
groundwork for future directions.

2.RELATED WORKS

This segment is structured into three sub-categories. These sub-categories provide an overview of the
historical research in the reclaim of cyberbullying detection, conventional algorithms, attention
mechanism methodologies.
Table 1. Different Bullying Categories

A. DETECTION OF CYBERBULLYING USING CONVENTIONAL MACHINE LEARNING

METHODS

In the reclaim of cyberbullying detection, researchers have extensively explored a range of machine
learning techniques to effectively identify instances of online harassment. Among the tried and methods
tested and proven are traditional approaches such as support vector Machines, (SVM) and Navie Bayes,
which have been specifically employed in the analysis of the textual data for the purpose of recognizing
cyber bullying, as demonstrated in the study conducted by Ychuang. In 2017. The importance of these
studies lies in their role as foundational pillars, laying the groundwork for understanding the intricate
challenges associated with classifying nuanced and contact dependent language within the context of
online interactions. By utilizing the traditional algorithms like SVM and Navie Bayes, these approaches
contribute to ongoing efforts aimed at developing robust cyberbullying detection systems. These effects
are crucial given the dynamic and ever evolving nature of online communication, where identifying
subtle forms of harassment. Require sophisticated tools that can navigate the complexities of language
nuisance and contextual intricacies.

B. CYBERBULLYING IDENTIFICATION USING DEEP LERANING APPROACHES

In recent times, attention has shifted towards prominent machine learning techniques in the field of
cyberbullying detection A.Bozyigit and S.Utku and E.Nasibov [19] expanded upon the semantic
enhanced marginalized denoising autoencoder, transforming it into a robust tool for uncovering hidden
features within cyber bullying posts and acquiring a discriminative text representation. Utku et al.
introduced the innovative pronunciation based on volitional net neural network designed to process
phenomenon notably we enhanced outperformed other neural network models commercial casing and
ability to handle spelling errors without compromising pronunciations accuracy. Kumar and Sach Deva
crafted a hybrid declaring framework known as Bi- GRU attention caps net. This framework combined
by GRUS self-attention and with a capsule network effectively capturing both semantic representation
and spatial information from textual content and social media for cyberbullying detection. Sriniket et al.
In the domain of identifying cyberbullying, there has been a recent search in the interest and exploration
of deep learning techniques. Cutting edge developments in these fields have led researchers to leverage
neural networks for enhanced cyberbullying detection, notably the works conducted by
M.Ptaszynski,P.Lempa et al. 2019 and D.Nguyen,M.Liakata et al 2020[11]have developed into a
utilization of deep learning models, comma specifically Convolutional Neural networks, CNN and
Recurrent Neural Networks are involved Their objective has been to harass the capabilities of these
neural networks to capture and analyze the contextual information comma, thereby reverting the
accuracy of cyberbullying detection full stop.

In similar terms, recent research, I’ve embraced advanced neural network architectures to better identify
instances of cyberbullying. Studies by Wang et al. And Lee et al. Have specifically explored
convolutional and recurrent neural networks, respectively, aiming to leverage these sophisticated
methods. For improved accuracy in detecting cyber bullying. Furthermore, the introduction of long-term
short memory.Long short-term memory(LSTM) networks have expanded their exploration which
subsequent research, such as that conducted by J.Bhagya and PS.Deepthi ET Al in 2021[29],
highlighting the effectiveness of LSTMs in modeling temporal dependencies within cyberbullying Text.
These demonstrates. An evolving landscape in which side advanced machine learning approaches are
growing and are applied to address the dynamic and intricate nature of cyber bullying detection.

C.DETECTION OF CYBERBULLYING USING LONG SHORT-TERM MEMORY APPROACH

Long short-term memory networks have gained traction in tasks involving natural language processing,
mainly because of their capabilities to grasp long term dependencies in the sequential data. Noteworthy
research in sentiment analysis and text classification conducted by. Y Fang, and Yang in 2015[21], along
with most recent work by Sepp Fang in 2021[21], provides A theoretical foundation. Leverging the
utilization of LSTMs in the content of cyberbullying detection. The distinctive strength of LSTM lies in
their ability to return and process information over extended sequences, rendering them suitable for
discerning the dynamic and evolving nature of online interactions.

In simpler terms, LSTM’s renders them especially suitable effectiveness in understanding the intricate
patterns and relationships present in the sequence of data. And evolving online interaction requires a
framework that can effectively capture and consider the context over extended periods. The theoretical
underpinnings from sentiment analysis and text classification studies provide a solid basis for leveraging
LTMS in addressing the nuanced challenges of identification of cyberbullying instances in the ever-
changing landscape of online communication.

D.DETECTION OF CYBERBULLING USING ATTENTION MECHANISM APPROACH

Researchers have delved into the integration of attention mechanisms within neural network
architectures to improve cyberbullying detection these. Retention mechanisms empower models to
concentrate on specific words or phrases, offering valuable insights into linguistics. Cues that play a role
in identifying instances of cyberbullying, as demonstrated in the work by J.Cai,J.Li,W.li and J.Zang
al[14]. In 2014. Grasping the functionalities of these attention mechanisms is essential for enhancing the
interpretability of models and instilling confidence in the cyberbullying detection process.

In simpler terms, attention mechanism has been incorporated into neural network designs to allow
models to focus on particular words or phrases that are indicative of cyber bullying. These approaches
are exemplified by Hs.Lee,Ju.Park[20]. Research in 2018, helps uncover the linguistic patterns that
contribute for discerning cyber bullying instances. Fully understanding how these attention mechanisms
operate is crucial because it is not only improving the model’s ability to explain its decision-making
process, but also fosters trust in the overall cyber bullying detection system.

3.METHODOLOGY

Schematic representation of the cyber bullying detection framework, focusing on binary text
classification, illustrates figure 1. The text classification task is delayed into 666 key stages later
collection in the case of this study using a secondary data screen on the compact transformation,
characteristic extraction, attribute choice, model advancement performance evaluation. The process
commences with data preparation and text preprocessing diverging into two distinct paths to congruently
address the task through both conventional machine logging and transfer learning approaches the
culmination involves the evaluation of the models. In the conventional machine learning approach
additional attention was devoted to feature engineering, a crucial step in converting the textile input into
quantifiable features before integration into the model. The adoption of transfer learning approaches and
involves leveraging pre trained language models PLMs the subsequent subsections provide in the
insights into the intricacies of each stage within the proposed methodology.

A. DATASET COLLECTION
Acquired for the project is publicly available but presents an imbalance in class distribution. To address
this, both under sampling and oversampling techniques will be employed to balance the data set given
the challenges in finding a sub. Suitable code-mixed data set, effects will be directed towards the
exploration, addition of data values, and implementation of data argumentation methods to ensure the
data set meets the projects requirements. The data set cycle minded the literature reviewer shows from
12 tablets origins encompasses seven online communication channels like twitter Instagram from spring
ask FM Myspace for my YouTube wine and read it an online collaborative platform and news website
Yahoo News ,these instances of cyberbullying have been observed on these platforms to witness
instances of cyber bullying, making them integral for constructing data such dedicated to server bridging
detection some of the examples of offensive comments extracted from these data sources included
source the count of positive and adverse samples, the proportion of positive versus negative samples
focused area example cybercrime heads reach cyber aggression etc. availability and relevant references
twitter renowned the leading plaforms in terms of social media for cyber bullying occurrences, features
prominently in the reviewed literature with 13 data sets of varying sizes from the collection methods,
and annotation techniques public Twitter Api was employed for collecting tweets. Annotators always
equals reliable expertise and a comprehensive memory to cyber physics of considerable success versus
command reformation defense, insert, sexual content, and threats additionally the data sets maintain the
structured annotation process with detailed guidelines, setting a depart from another available data sets
lacking similar information.

B.DATA PREPROCESSING

Once the data set is obtained, the next step involves preprocessing the data. Techniques such as case
folding, data cleaning, text normalization, tokenization, and stop word removal will be applied. These
methods collectively. Aim for improving the quality and consistency of the textual data.

C. FEATURE EXTRACTION

Performing feature extraction from data and convert a textual data into a matrix of features to perform
this feature extraction the N-Gram model word embedding techniques of NLP are used.

D.DATA SPLITTING

In the Data Splitting phase, the collected and preprocessed data will be partitioned into training and
testing sets. These ensures that the model is trained using the subset of the data and evaluated on
another, facilitating an unbiased assessment of its performance.

E. TRAINING MODEL

For model training, and ensemble machine learning approach will be adopted. Single classifiers such as
WKNR, RVM, Navie Bayes, among others, will be evaluated to identify the best. Algorithm for the
classification. These selected models will then be combined to create an optimal predictive model.
Voting based classifier, specifically soft voting, will be implemented to aggregate predictions from each
base model. The final prediction will be determined by the majority of the voting, representing an
ensemble decision.

F. VALIDATION

The validation phase involves assessing the performance of the classification model. Each step, from
data collection and preprocessing for the feature extraction and model training, will be rigorously
validated for improcing the reliability and efficiency of the developed cyber bullying detection system.
Validation will involve through testing on separate data sets to measure the model’s accuracy, precision,
recall, and others relevant metricsLogistic Regression Formula : P(y=1|x)=1/(1+e^-(w^Tx+b).
 FIGURE 1. System Architecture

G.FEATURES

Engineering is pivotal in the effectiveness of machine learning models. The identifying crucial features
are generating new ones from existing data. Drawing insights from the literature, nuclear features
derived from text have been justified, including textile features, sentiment, and emotional. He does,
word embedding, psychologist features, personality traits, features, topic modeling, and customized
dictionary lists. Within this investigation, all these features were incorporated with the expectation of
personality traits and topic modeling the decision to exclude.

1)TEXTUAL CONTENT

Text Patrick for statistically represented by Kali’s narrating essential metric, short word and character
frequency, soft words, digits, uppercase words, pronunciation, emojis, and average word length parts of
speech. Talks and named entity recognition labels were also included using spacy. A bag of words was
generated with the count vectorizer employing both word level and character level up to quadruple
grams. This approach enhances the capture of rare words and addresses type common in online
platforms. Each unique N gram served as an attribute for model training, aiming to improve machine
learning performance.

2)SENTIMENT AND EMOTIONAL FEATURES

Since Python. Packages, various sentiments, and emotions related features were obtained from the text.
The text block package are just the polarity and subjectivity of the text, generating sentiments goes.
Different tools including Vader by Bhutto and Gilbert Afi. Nm by measles, and Piscean Timent. Where
employee to ensure diverse perspectives as these tools rely on distinct dictionaries. The. Nrc Lex
package facilitated the extraction of eight fundamental emotional related metrics from the next, covering
anger, fear, anticipation, trust, surprise, sadness, joy, and disgust. Developed in accordance with the
national research councils in Canada affect lexicon introduced by Mohammed and Turney, the NRCLex
package enriched the emotion analysis aspect.

3)STATIC WORD EMBEDDING

While Magos words this commonly used, it lacks context. As it reached each word independently.
Word, embedded, such as word to web and gloves, capturing semantic meal with representing words as
vectors WordToVec predicts words based on context, one clear considers word 4 accurate frequency
glove, trained on common crawl, Wikipedia, and Twitter, handles rare words well fast and extension of
word to web encodes words into N grand characters. This study uses pretrained word vectors from fast
text leveraging one million sized vectors trained on Wikipedia 2017, UMBC And new sites.

4)PYSCHOLINGUISTIC FEATURES

The study is the pioneer in utilizing the latest version of the LIWC 2022 cyber bullying detection tool,
featuring 117 attributes. Additionally, the research incorporates the empath library by Fast et al, text
analysis tool akin to LIWC. Impact. Derives categories from a concise set of seed terms and offers text
analysis for approximately two hundred category sources.

Table 2. Feature grouping for Cyberbullying Detection.

4.CHALLENGE IN CYBERBULLING DETECTION

The current body of research underscores the difficulties associated with imbalanced data sets in the
context of cyberbullying detection. Notably, Fortuna et al (2018) Stress the significance of mitigating
biases introduced by such imbalances. As model performance may be disproportionately influenced by
the majority class , this is done. Recent discussions in literature, as highlighted by Tarzako et al (2017),
place considerable emphasis on ethical considerations in this domain. Factors such as fairness and
transparency are deemed crucial points.

A) LANGUAGE CHALLENGE

The study of cyber bullying is still in the early stages within the research community. For instance,
sentences like the picture use and annoyed me so much and I don’t want to contact you anymore. Pause
a challenge for straightforward cyber bullying classification as negativity alone may not suffice for
analysis. Occasionally, positive messages may carry sarcastic undertones. Cyber guilt is complex
because of its subjective and subtle nature. Furthermore, in today’s rapidly evolving technological
landscape language, especially among adolescents under ghosts, swift changes, impacting the efficacy of
keywords as attributes in identifying in cyberbullying detection. Therefore, additional factors may be
necessary to accurately enable such messages as cyberbullying.

B) DATASET CHALLENGE

Another obstacle within the reclaim of cyberbullying identification lies in the data set. Gathering data
from social media feels challenging due to privacy concerns, across social media platforms typically do
not openly disclose data. Consequently, essential information, such as a list of friends may be
inaccessible. Additionally, annotating, or labeling data is a formidable task. Demanding the participation
of experts in categorizing the corpus, as emphasized by previous studies. Sharing data sets Aman
Prudential Resources would constitute a valuable contribution to academia.

C) DATA REPRESENTATION CHALLENGE

Many researchers focus solely on bullying words and telecommunication but extracting content-based
features presence its own set of challenges. If user accounts lack information like gender or age, the
effectiveness within cyberbullying identification may suffer. However, some studies delve into
analyzing language patterns to estimate the age range of users. Identifying words in the corpus
associated with age can be time consuming. For instance, the word study may be indicative of users aged
13 to 18. Establishing a robust cyberbullying detection system or application is no simple task, given the
environment of human behavior and the elusive nature of cyberbullying in its context.

D) IMBALANCED CYBERBULLYING DATASETS

Dealing with imbalanced cyberbullying data sets presented a notable challenge, given the substantial
disproportion between cyberbullying instances and non cyber bullying instances. The scarcity of
cyberbullying instances having the potential to impact the model’s effectiveness. It is recommended that
upcoming research endeavors delve into methodologies such as data Augmentation or the
implementation of weighted loss functions to effectively tackle this imbalance and enhance model
performance.

E) REALM OF TEMPORAL DYNAMICS

In the realm of temporal dynamics, although the LSTM effectively capture temporal dependencies, it is
crucial to acknowledge that the models performance might still be susceptible to the rapid evolution of
language trends and contextual shifts. To address this, a proactive approach involving continuous
monitoring and periodic retraining is suggested. This adaptive strategy ensures that the model remains
attuned to emerging cyber bullying patterns, thereby maintaining its relevance and efficacy overtime.

5. IMPROVING THE EFFECTIVENESS OF CYBERBULLING DETECTION

Expanding the capabilities within the realm of identifying cyberbullying detection models could be
achieved through the integration of multi-modal data involving not only textual information, but also
images and videos. By incorporating a variety of data types, researchers aim to create a more holistic
and nonce understanding of online interactions across different platforms. This app wrote recognizes that
cyber bullying may manifest in various forms and the inclusion of multimodal data allows the model to
capture a broader spectrum of potentially detrimental content.

In terms of model transparency and user trust, it is imperative to focus on improving the explainability of
the models’ decisions This involves developing methods to articulate how and why the model classify
certain instances as cyber bullying. Providing interpretable explanations not only enhances user
understanding, but also contributes to the ethical deployment of such models.
Ethical considerations, including the detection and mitigation of biases, are crucial elements in the
development of cyberbullying detection systems. Addressing potential biases ensures that the models
protections are fair and unbiased across diverse user groups. This is particularly important task biases in
ai models can perpetuate existing inequalities and impact certain demographic groups This
proportionately.

Therefore, ongoing research efforts should prioritize ethical considerations to uphold the principles of
fairness and inclusivity in the cyberbullying detection.

6. RESULTS AND DISCUSSIONS

The LSTM based cyberbullying detection model showcased strong performance across multiple
evaluation metrics. Having been trained on a diverse cyberbullying data set, the model attained an
impressive overall accuracy of 86%.. This signifies the effectiveness of the LSTM architecture in
capturing the temporal dependencies inherent in cyber bullying texts.

Furthermore, precision recall, as well as the F1 score were employed to provide a more non devaluation
of the models capabilities. Precision, denoting the ratio of true positive predictions to the total predicted
positives, stood at 89.. Recall, representing the models ability to identify all. Relevant instances,
achieved the value of 92. The F1 score,a balanced measure of precision and recall, reached 90. These
metrics collectively affirm the models accurate identification of cyber bullying instances.

In a comparision analysis against baseline models, including traditional machine learning approaches
and alternative deep learning architectures, the LSTM model demonstrated a notable superiority. This
highlights the importance of the LSTM’s capability to capture long term dependencies, a crucial factor
in effective cyber bullying detection.

Additionally, the LSTM model exhibited scalability and generalization during testing on larger data sets
and diverse types of cyber. Bullying instances. Its consistent accuracy across varying conditions,
underscores its adaptability to different online environments.

For enhanced interpretability and attention mechanism was incorporated into the LSTM architecture.
This mechanism facilitated a clearer understanding of the models decision making process by
visualizing attention weights. This a lab researchers to identify keywords and phrases that significantly
influenced the model’s predictions. They. Integration of the attention mechanism contributes to the
models transparency, promoting a better understanding of its reasoning and potentially fostering user
trust in its outcomes.

Table 3.Comparison of cyberbullying detection methods.

7. CONCLUSIONS

In summary, this project marks a significant advancement in leveraging cutting edge deep learning
techniques, particularly long short-term memory bracket LSTM networks for the identification of
cyberbullying. The robust effectiveness of the proposed model underscores the usage of LSTM in
capturing intricate temporal dependencies inherent in online communication. With an outstanding
overall accuracy of Linsert accuracy percentage, the model not only surpass traditional machine learning
approaches, but also exhibited resilience in handling diverse instances of cyber bullying.

The incorporation of an attention mechanism further enriched the model’s interpretability, revealing
crucial linguistic elements that influence its decision-making process. In the ever-evolving landscape of
online interactions. This research lays the groundwork for the development of more sophisticated and
adoptable cyber bullying detection systems. However, amidst these promising outcomes, the project also
highlights critical challenges that demand ongoing attention. Addressing the imbalance in cyber bullying
data sets remain a priority, and future research should explore strategies to mitigate potential biases
stemming from the scarcity of positive instances.

Additionally, the model sensitivity to temporal dynamics necessitates a proactive approach to

monitoring and retraining. Ensuring its sustained effectiveness in the face of evolving language
strengths. This has been done. The exploration of integrating multimodal data and the commitment to
enhancing model explainability and ethical considerations will be crucial in advancing the field of
cyberbullying detection. In essence, this project not only contributes valuable insights to the academic
discourse, but also issues a call for ongoing collaborative efforts to create safer and more inclusive
digital spaces.

REFERENCES
[1] B. Cagirkan and G. Bilek, ‘‘Cyberbullying among Turkish high school students,’’ Scandin. J. Psychol., vol. 62, no. 4, pp.
608–616, Aug. 2021, doi: 10.1111/sjop.12720.

[2]P. T. L. Chi, V. T. H. Lan, N. H. Ngan, and N. T. Linh, ‘‘Online time, experience of cyber bullying and practices to cope
with it among high school students in Hanoi,’’ Health Psychol. Open, vol. 7, no. 1, Jan. 2020, Art. no. 205510292093574,
doi: 10.1177/2055102920935747.

[3]A. López-Martínez, J. A. García-Díaz, R. Valencia-García, and A. Ruiz-Martínez, ‘‘CyberDect. A novel approach for
cyberbullying detection on Twitter,’’ in Proc.Int. Conf. Technol. Innov., Guayaquil, Ecuador: Springer, 2019, pp. 109–121,
Doi: 10.1007/978-3-030-34989- 9_nine.[4]R. M. Kowalski and S. P. Limber, ‘‘Psychological, physical, and academic
correlates of cyberbullying and traditional bullying,’’J. Adolescent Health, vol. 53, no. 1, pp. S13–S20, Jul. 2013, doi:
10.1016/j.jadohealth.2012.09.018.

[5]Y.-C. Huang, ‘‘Comparison and contrast of Piaget and Vygotsky’s Theo- Ries,’’ in Proc. Adv. Social Sci., Educ.
Humanities Res., 2021, pp. 28–32, doi: 10.2991/assehr.k.210519.007.

[6]A. Anwar, D. M. H. Kee, and A. Ahmed, ‘‘Workplace cyberbullying and interpersonal deviance: Understanding the
mediating effect of silence and emotional exhaustion,’’ Cyberpsychology., Behave., Social Newt., vol. 23, no. 5, pp. 290–
296, May 2020, doi: 10.1089/cyber.2019.0407.

[7]D. M. H. Kee, M. A. L. Al-Anesi, and S. A. L. Al-Anesi, ‘‘Cyberbully- lying on social media under the influence of
COVID-19,’’ Global Bus. Organizational Excellence, vol. 41, no. 6, pp. 11–22, Sep. 2022, doi: 10.1002/joe.22175.

[8]I. Kwan, K. Dickson, M. Richardson, W. MacDowall, H. Burchett. Stansfield, G. Brunton, K. Sutcliffe, and J. Thomas,
‘‘Cyberbullying and children and young people’s mental health: A systematic map of systematic reviews,’’
Cyberpsychology., Behave., Social Netw., vol. 23, no. 2, pp. 72–82, Feb. 2020, doi: 10.1089/cyber.2019.0370.

[9]R. Garett, L. R. Lord, and S. D. Young, ‘‘Associations between social media and cyberbullying: A review of the
literature,’’ mHealth, vol. 2, p. 46, Dec. 2016, doi: 10.21037/mhealth.2016.12.01.

[10]M. Ptaszynski, F. Masui, Y. Kimura, R. Rzepka, and K. Araki, ‘‘Automatic extraction of harmful sentence patterns with
application in cyberbullying detection,’’ in Proc. Lang. Technol. Conf. Poznan, Poland: Springer, 2015, pp. 349–362, Doi:
10.1007/978-3-319-93782-3_twenty-five.
[11]M. Ptaszynski, P. Lempa, F. Masui, Y. Kimura, R. Rzepka, K. Araki, M. Wroczynski, and G. Leliwa, ‘‘"Brute-force
sentence pattern extortion from harmful messages for cyberbullying detection,’’’ J. Assoc. Inf. Syst., vol. 20, no. 8, pp. 1075–
1127, 2019.

[12]M. O. Raza, M. Memon, S. Bhatti, and R. Bux, ‘‘Detecting cyber- bullying in social commentary using supervised
machine learning,’’ in Proc. Future Inf. Commun. Conf. Cham, Switzerland: Springer, 2020, pp. 621–630.

[13]D. Nguyen, M. Liakata, S. Dedeo, J. Eisenstein, D. Mimno, R. Tromble, and J. Winters, ‘‘How we do things with words:
Analyzing text as social and cultural data,’’ Frontiers Artif. Intell., vol. 3, p. 62, Aug. 2020, doi: 10.3389/frai.2020.00062.

[14]J. Cai, J. Li, W. Li, and J. Wang, ‘‘Deep learning model used in text classification,’’ in Proc. 15th Int. Comput. Conf.
Wavelet Act. Media Technol. Inf. Process. (ICCWAMTIP), Dec. 2018, pp. 123–126, doi:
10.1109/ICCAMTIP.2018.8632592.

[15]D. G. Krutka, S. Manca, S. M. Galvin, C. Greenhow, M. J. Koehler, and E. Askari, ‘‘Teaching ‘against’ social media:
Confronting prob- Lems of profit in the curriculum,’’ Teachers College Rec., Voice Scholarship Educ., vol. 121, no. 14, pp.
1–42, Dec. 2019, doi: 10.1177/016146811912101410.

[16]H. Rosa, N. Pereira, R. Ribeiro, P. C. Ferreira, J. P. Carvalho, S. Oliveira, L. Coheur, P. Paulino, A. M. V. Simão, and I.
Trancoso, ‘‘Automatic cyberbullying detection: A systematic review,’’ Comput. Hum. Behav., vol. 93, pp. 333–345, Apr.
2019, doi: 10.1016/j.chb.2018.12.021.

[18]S. Bharti, A. K. Yadav, M. Kumar, and D. Yadav, ‘‘Cyberbullying detection from tweets using deep learning,’’
Kybernetes, vol. 51, no. 9, pp. 2695–2711, Sep. 2022.

[19]A. Bozyiğit, S. Utku, and E. Nasibov, ‘‘Cyberbullying detection: Uti- lizing social media features,’’ Expert Syst. Appl.,
vol. 179, Oct. 2021, Art. no. 115001, doi: 10.1016/j.eswa.2021.115001.

[20]H.-S. Lee, H.-R. Lee, J.-U. Park, and Y.-S. Han, ‘‘An abusive text detection system based on enhanced abusive and non-
abusive word lists,’’ Decis. Support Syst., vol. 113, pp. 22–31, Sep. 2018, doi: 10.1016/j.dss.2018.06.009.

[21]Y. Fang, S. Yang, B. Zhao, and C. Huang, ‘‘Cyberbullying detection in social networks using bi-GRU with self-attention
mechanism,’’ Informa- tion, vol. 12, no. 4, p. 171, Apr. 2021, doi: 10.3390/info12040171.

[22]G. Jacobs, C. Van Hee, and V. Hoste, ‘‘Automatic classification of partici- pant roles in cyberbullying: Can we detect
victims, bullies, and bystanders in social media text?’’ Natural Lang. Eng., vol. 28, no. 2, pp. 141–166, Mar. 2022, doi:
10.1017/S135132492000056X.

[23]M. Gada, K. Damania, and S. Sankhe, ‘‘Cyberbullying detection using LSTM-CNN architecture and its applications,’’ in
Proc. Int. Conf. Comput. Commun. Informat. (ICCCI), Jan. 2021, pp. 1–6, doi: 10.1109/ICCCI50826.2021.9402412.

[24]H. H.-P. Vo, H. Trung Tran, and S. T. Luu, ‘‘Automatically detecting cyberbullying comments on online game forums,’’
in Proc. RIVF Int. Conf. Comput. Commun. Technol. (RIVF), Aug. 2021, pp. 1–5, doi: 10.1109/RIVF51545.2021.9642116.

[25]F. Elsafoury, S. Katsigiannis, Z. Pervez, and N. Ramzan, ‘‘When the timeline meets the pipeline: A survey on automated
cyberbully- ing detection,’’ IEEE Access, vol. 9, pp. 103541–103563, 2021, doi: 10.1109/ACCESS.2021.3098979.

[26]J. Howard and S. Ruder, ‘‘Universal language model fine-tuning for text classification,’’ in Proc. 56th Annu. Meeting
Assoc. Comput. Linguistics (Long Papers), vol. 1, 2018, pp. 328–339.

[27]R. Silva Barbon and A. T. Akabane, ‘‘Towards transfer learning techniques—BERT, DistilBert, BERTimbau, and
DistilBERTimbau for automatic text classification from different languages: A case study,’’ Sensors, vol. 22, no. 21, p. 8184,
Oct. 2022, doi: 10.3390/s22218184.

[28]J. Eronen, M. Ptaszynski, F. Masui, G. Leliwa, and M. Wroczynski, ‘‘Exploring the potential of feature density in
estimating machine learn- ing classifier performance with application to cyberbullying detection,’’ 2022, arXiv:2206.01949.

[29]J. Bhagya and P. S. Deepthi, Cyberbullying Detection on social media Using SVM (Inventive Systems and Control).
Singapore: Springer, 2021, pp. 17–27, Doi: 10.1007/978-981-16-1395-1_two.