Ontology and AI Integration for Real-Time Detection of

Cyberbullying Among University Students

Khaliq Ahmed

, Ashley Mathew

and Shajina Anand

Department of Math and Computer Science, Seton Hall University, New Jersey, U.S.A.

Keywords: Cybersecurity, BERT, NLP, Cyberbullying, Ontology, Graph – Based Ontology, Youth, Online Safety.

Abstract: With the increasing of the internet, smartphones, and social media, nearly everyone is a potential target for

cyberbullying. Our research introduces an AI-driven approach to detect and address cyberbullying among

college students, with a focus on its impact on mental health. We developed a context-specific ontology,

drawing from real-time data, publicly available data, surveys, academic literature, and social media

interactions to categorize information into domains such as victims, causes, types, environments, impacts, and

responses. We collected real-time data from college students through surveys, interviews, and social media,

leveraging advanced NLP (Natural Language Processing) techniques and BERT for accurate and efficient

detection. By integrating this ontology with AI, our system dynamically adapts to emerging cyberbullying

patterns, offering more precise detection and response strategies. Experimental results show that the proposed

model achieves 96.2% accuracy, with 95.8% precision, 95.5% recall, and an F1-score of 95.6%. This

performance surpasses traditional methods, emphasizing its capability to identify both explicit and implicit

forms of abusive behavior. The approach not only introduces a tailored ontology for college students' unique

social dynamics but also offers solutions to evolving cyberbullying trends. This research significantly

enhances online safety and fosters a healthier digital environment for university students use.

1 INTRODUCTION

The widespread use of social media has significantly

increased instances of cyberbullying, with serious

implications for students' well-being and academic

performance. To address this, researchers have

developed automated methods to detect

cyberbullying and create safer online environments.

Deep learning techniques, particularly transformer

models like BERT and DistilBERT, have shown

significant promise, outperforming traditional

methods by leveraging large datasets (Teng &

Varathan, 2023). BERT-based models have also

enhanced Aspect Target Sentiment Classification

(ATSC), effectively identifying relationships

between targets and sentiments in cyberbullying

incidents (Chen et al., 2023). Advances in cross-

platform detection methods, such as adversarial

learning, have improved monitoring across multiple

social media platforms, increasing detection

https://orcid.org/0009-0000-7945-8476

https://orcid.org/0009-0005-0876-0703

https://orcid.org/0000-0001-6721-1150

flexibility (Yi & Zubiaga, 2022). While traditional

approaches often focus on explicit language cues,

they can miss subtle or covert abusive behaviors.

BERT captures contextual nuances effectively but

integrating it with graph-based ontologies further

enhances detection by mapping relationships between

abusive behaviors and associated concepts (Rogers et

al., 2020). Our study combines BERT's contextual

understanding with a hierarchical ontology

framework, enabling precise detection of both

explicit and implicit behaviors. Achieving 96.2%

detection accuracy, our system surpasses traditional

and state-of-the-art methods, offering a dynamic

solution to improve online safety for college students.

2 LITERATURE REVIEW

Recent studies have focused on improving

cyberbullying detection using advanced AI language

Ahmed, K., Mathew, A. and Anand, S.

Ontology and AI Integration for Real-Time Detection of Cyberbullying Among University Students.

DOI: 10.5220/0013213500003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 709-716

ISBN: 978-989-758-749-8; ISSN: 2184-4992

709

models. For example, Alrowais et al. (2024)

developed an upgraded RoBERTa model, called

RoBERTaNET, which uses GloVe word embeddings

to detect cyberbullying tweets with 95% accuracy.

While it shows high performance, it demands

significant computing power, making widespread

adoption challenging, especially in developing

countries. Similarly, Ogunleye and Dharmaraj (2023)

introduced a new dataset named D2 to enhance

RoBERTa’s detection capability. This approach

provides better accuracy and resists skewed class

problems but requires a large dataset, which limits its

use in environments with limited data availability.

Teng and Varathan (2023) used transfer learning

with DistilBERT to enhance detection, incorporating

psycholinguistic factors, but achieved only 64.8% on

the F-measure for logistic regression, indicating more

work is needed for diverse social media content. The

XP-CB model by Yi and Zubiaga (2022) uses

adversarial learning to improve cross-platform

detection, but it has high processing requirements,

limiting its scalability. Sen et al. (2024) combined

BERT with CNN and MLP, achieving 87.2% to

92.3% accuracy, outperforming other machine

learning methods. However, its complexity makes

real-time deployment challenging. Ejaz et al. (2024)

developed a dataset that covers multiple aspects of

cyberbullying, such as violence, repetition, and peer-

to-peer interaction. This makes it more flexible for

researchers but lacks detailed performance metrics. In

a separate study, Chow et al. (2023) found that BERT

achieved the highest accuracy (96%), slightly

outperforming Bi-LSTM (95%) and Bi-GRU (94%)

in detecting cyberbullying on tweets.

El Koshiry et al. (2024) used a CNN-BiLSTM

model with Focal Loss and GloVe embeddings,

achieving a 99% accuracy rate. However, the model

struggled with recall, indicating a need for further

improvements in capturing all instances of

cyberbullying. Lastly, Kaur and Saini (2023)

conducted a scient metric analysis of AI applications

for cyberbullying detection, highlighting trends,

contributions, and future research directions in this

field, but without evaluating specific model

performances. Ontology-based approaches (e.g.,

Gencoglu, 2020) provide structured domain

knowledge, enabling better categorization and

semantic understanding but often lack the ability to

process implicit language patterns effectively.

Transformer-based models, such as BERT and its

variants (Chen et al., 2023; Yi & Zubiaga, 2022),

excel in capturing linguistic nuances but struggle with

representing complex relationships between

concepts. Despite the strengths of these approaches,

they are generally applied independently, leaving a

gap in combining these techniques for tasks like

cyberbullying detection.

As of now, there is no existing research that

combines graph-based ontologies with BERT or

similar transformer models for cyberbullying

detection. Our work addresses this gap, providing an

opportunity to develop a dual-layered approach that

integrates the contextual understanding of BERT with

the hierarchical structuring capabilities of ontologies

to enhance both detection accuracy and adaptability

to evolving abusive behaviors.

3 METHODOLOGY

3.1 Data Description

The dataset used in this study consists of messages

labeled as either ‘cyberbullying’ or ‘not

cyberbullying.’ Data was collected from three main

sources: real-time input from university students

through surveys, direct interactions, and virtual

interviews; publicly available datasets from online

platforms; and web-scraped data from public forums

to capture diverse language patterns. The dataset

contains two primary columns: ‘message_text’,

which includes user-generated content from social

media, and ‘cyberbullying_type’, indicating whether

the content qualifies as cyberbullying. A majority of

the messages are labeled as ‘cyberbullying,’ while a

smaller portion is labeled as ‘not cyberbullying’.

3.2 Data Collection

Participants for this study were voluntarily recruited

from the university campus. After signing a consent

form, they completed a survey on Qualtrics where

they shared their experiences with cyberbullying,

detailing its impact on their mental health and

academic performance. Eligible participants were 18

years or older, currently enrolled in either bachelor's

or master’s programs, and had experienced or

witnessed cyberbullying within the past year. Those

interested in further participation engaged in 30-

minute virtual interviews via MS Teams to provide

deeper insights into their personal experiences.

Additionally, we used existing datasets such as the

Cyberbullying Detection Dataset on Twitter (2023),

Instagram Cyberbullying Dataset (2022), and the

OLID Dataset (2020). These datasets helped refine

the ontology model and enhance the accuracy of AI-

based detection algorithms.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

710

3.3 Data Preprocessing

The initial stage of data preprocessing and critical

component of Natural Language Processing (NLP)

involved cleaning and preparing the raw text for

analysis. Text preprocessing included tokenization,

where sentences were split into individual words or

sub words to facilitate analysis. For example, the text

“You are so annoying!” was tokenized into ["You",

"are", "so", "annoying", "!"]. All text was converted

to lowercase to ensure consistency and reduce

duplication, so that “annoying” and “Annoying”

would be treated the same. Additionally, stop words

like “and” “the,” and “is”, were removed as they did

not add any significant meaning. Special characters,

HTML tags, and other extraneous symbols were

eliminated to ensure only relevant content remained

for analysis. Figures 1 and 2 provide a comprehensive

view of the dataset characteristics detailing

distributions of cyberbullying types.

Figure 1: University data collection.

Figure 2: Cyberbullying Types.

3.4 Feature Extraction Using Bert

In the context of detecting cyberbullying, accurately

interpreting the text often filled with slang,

misspellings, or context-specific terms is crucial.

BERT is used for feature extraction, given its ability

to understand complex language patterns. The

tokenization process is a fundamental step in

preparing data for BERT, converting the input text

into a format BERT uses Byte-Level Byte-Pair

Encoding (BPE) for tokenization, breaking words

into sub words to accommodate slang, rare words,

and spelling errors commonly seen in user-generated

content. that the model can process.

3.5 Graph Based Ontology

This section outlines the creation of a graph-based

ontology designed to categorize and structure

cyberbullying behaviors for better detection and

analysis. By integrating domain knowledge with real-

world examples, the ontology addresses limitations of

existing models. Combining BERT's contextual

understanding with ontology-based reasoning

enhances the detection of explicit and implicit forms

of cyberbullying. The ontology comprises two key

components:

• T-Box (Terminological Box): Represents the

schema, including categories (𝐶) and their

hierarchical relationships (𝑅).

• A-Box (Assertional Box): Contains specific

instances (I) derived from raw data (D),

representing real-world examples of

cyberbullying.

3.6 Building Ontology

Map each Feature fj directly to a Category:

For each identified feature related to cyberbullying

behavior, we assign it to a specific category 𝐶.

Example: If f

m+1 = “using derogatory terms”, then:

m+1 à “ Insults”

If f

m+2 = creating fake profiles”, then:

m+2 à “ Impersonation”

Instance Mapping to Categories

Instances (I) identified from raw data are linked to

appropriate categories:

If i

k+1=“sending mean messages”, then: g(ik+1)

=“Harassment”

If i

k+2=“spreading false rumor”, then: g(ik+2)

=“spreading rumors”

Constructing the Ontology Graph

The ontology graph (𝐺) is built with vertices (𝑉)

representing categories and instances, and edges (𝐸)

representing relationships:

V = C ∪ I

E={(Spreading Rumors,Gossiping about

someone),( Impersonation, Creating fake

profiles)}

Ontology and AI Integration for Real-Time Detection of Cyberbullying Among University Students

711

Example: If “Spreading Rumors” includes

“Gossiping about someone,” then

𝐸 = Spreading Rumors, Gossiping someone)

Define relationship types for Ontology

Define hierarchical relationships for the new

categories:

Hierarchy: ℎ(“Cyberbullying”) =

{“Insults”. “Using derogatory terms”}

Associations: 𝑟: 𝐼 à 𝐶, e.g.,

“Creating fake profiles” à Impersonation.

Semantic Validation

Semantic validation ensures consistency between

categories and relationships. Categories and instances

are vectorized, and semantic similarity measures

(SSS) are applied:

𝑆𝑐



,𝑐



=

𝑉𝑒𝑐

(

𝑐





∗𝑉𝑒𝑐𝑐







𝑉𝑒𝑐

(

𝑐









𝑉𝑒𝑐𝑐







For example:

• S(“Spreading Rumors,” ”Gossiping about some

one”): Measures closeness and adjusts

categories if needed.

Inference and Reasoning over Ontology

Inference rules are implemented to enhance

detection:

• If i

j involves "using derogatory terms," infer it

as:

Φ(i

j) = "Insults”

• If i

k involves "Creating fake profiles," infer it as

"Impersonation."

The ontology was developed using Protege for

designing and visualizing the hierarchical structure,

while Owlready2 was employed to integrate the

ontology into the detection framework seamlessly.

Additionally, Python-based NLP tools were utilized

for preprocessing the raw data (D), extracting

relevant features (F), and identifying instances (I) of

cyberbullying behaviors.

Final Ontology Structure: The final ontology is

represented as:

O = (C, I, R, G, f, g)

where: C: Set of defined concepts, I: Set of identified

instances, R: Set of relationships, G: Ontology graph

structure, f, g: Functions mapping features and

instances to categories.

The ontology expands detection capabilities by

capturing complex relationships, improving

accuracy, and adapting dynamically to evolving

cyberbullying behaviors.

3.7 Ontology Reasoning and

Development

Ontology in computer science categorizes

cyberbullying behaviors based on social context

(Lembo et al., 2013), aligning data with domain-

specific categories for accurate detection. Our

approach adapts the ontology to new data,

categorizing emerging cyberbullying terms through

semantic relations. A hierarchical structure (Figure 3)

represents categories like insults, harassment, and

catfishing.

Arrows in Figure 3 denote two types of relationships:

• Hierarchical Relationships: Solid arrows

connect broader categories to subcategories, such

as linking "Cyberbullying" to "Insults" or

"Impersonation."

• Associative Relationships: Dashed arrows

connect co-occurring or conceptually related

terms, such as "Idiot" and "Ugly," reflecting

linguistic patterns from real-world data.

Instances like "sharing false information" under

Spreading Rumors and "creating fake profiles" under

Impersonation capture explicit and implicit

behaviors. BERT maps feature predefined ontology

categories, ensuring precise classification of abusive

content while capturing nuanced, context-dependent

meanings. The ontology dynamically refines

categories based on new data, making it a powerful

tool for detecting online abuse and supporting

students' mental health.

3.7.1 NLP with BERT

The core technology behind our project is Natural

Language Processing powered by BERT algorithm.

We use both to detect and understand cyberbullying

among college students. BERT is particularly well-

suited for this task due to its exceptional ability to

read and comprehend language contextually, much

like a human would. It also can process words in

relation to both preceding and succeeding words in a

sentence. This enables BERT to capture hidden

meanings, which is a crucial feature for detecting

cyberbullying, where harmful intent may be implicit

or context dependent (Lee et al., 2019).

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

712

Figure 3: Ontology Graph.

3.8 BERT’s Self-Attention Model

BERT processes text using Transformer architecture,

which enables it to understand the relationships

between cyberbullying words. The mathematical core

of BERT lies in the self- attention mechanism,

which computes relationships between words using

this formula:

𝐴𝑡𝑡𝑒𝑛𝑡𝑖𝑜𝑛 (𝑄, 𝐾, 𝑉) = 𝑠𝑜𝑓𝑡𝑚𝑎𝑥(

𝑄𝐾

𝐾

) V where,

√𝑑𝑘

• Q (query), K (key), and V (value) matrices are

taken from the input words or sentences.

• The SoftMax function is used to confirm the

values add up to 1. Allows BERT to focus on the

most relevant parts of the sentence specific to the

cyberbullying issue.

•

(





√



) Makes sure the attention is prioritized,

and each cyberbullying word is compared with

every other word to decide how much focus to

give each one (Rogers et al., 2020).

3.9 Integration of NLP with Graph-

Based Ontology

The proposed cyberbullying detection model

combines BERT’s NLP capabilities with a graph-

based ontology for semantic comprehension: The

algorithm starts by initializing the model with a pre-

trained BERT and loading a structured ontology tree.

Text preprocessing involves tokenization and

embedding generation. These embeddings are then

matched with ontology-defined concepts to identify

relevant instances of abusive behavior. The system is

designed for continuous learning, logging new

instances and retraining itself periodically to remain

effective against evolving language dynamics.

Algorithm 1: Cyberbullying detection model.

3.10 Model Architecture and

Methodology Framework

The proposed cyberbullying detection framework

consists of four components: Data Collection, Data

Preprocessing, Ontology Integration, and AI/ML

1. Model Initialization

BERTLoad(BERT_pretrained,TrainingDataset)

If OntologyFile exists then:

Ontology  Load(OntologyFile)

OntologyTree 

ConstructOntologyTree(Ontology)

2. Text Preprocessing

If RelevantContent is not empty then:

Tokens  Tokenize(RelevantContent)

BERT_Embeddings 

GenerateEmbeddings(Tokens, BERT)

3. Contextual Analysis

If BERT_Embeddings are valid then:

MatchedConcepts 

MatchConcepts(OntologyTree,

BERT_Embeddings)

If MatchedConcepts are found then:

ContentAnalysis 

Integrate(MatchedConcepts,

ContentAnalysis)

4. Continuous Learning and Updates

LogInteraction(CyberbullyingType,

ContextualUnderstanding)

If LoggedData is sufficient for retraining

then:

BERT  RetrainModel(BERT,

LoggedData)

OntologyTree 

UpdateOntologyTree(LoggedData)

End procedure

Ontology and AI Integration for Real-Time Detection of Cyberbullying Among University Students

713

Integration. Data is sourced from public datasets,

surveys, and interviews, followed by cleaning,

normalization (e.g., lowercasing and punctuation

removal), and tokenization to ensure compatibility

with BERT. BERT generates contextual embeddings

that capture complex language nuances, which are

then aligned with a graph-based ontology. This

ontology defines abusive behaviors and their

relationships, mapping concepts like 'harassment' and

'impersonation' to improve detection accuracy for

both explicit and subtle forms of cyberbullying.

Figure 4: Overall Methodology framework.

The framework undergoes training with cross-

validation and hyperparameter tuning to enhance

performance. A continuous learning mechanism logs

interactions, periodically retrains the BERT model,

and updates the ontology to adapt to evolving trends

in abusive behavior. By integrating NLP, BERT, and

ontology-based reasoning, the system effectively

identifies and classifies various forms of online

abuse, contributing to safer digital environments

through real-time detection and mitigation.

4 EXPERIMENTAL SETUP

The evaluation of cyberbullying detection models

enhanced with a graph-based ontology shows notable

performance variations between Logistic Regression

(LR) and Random Forest (RF) across dimensions like

overall detection, ethnicity, gender, age, and religion-

related content. The experiment used 10,000 labelled

data points collected from public datasets, surveys,

and interviews with college students. The data was

split into training (70%), validation (15%), and

testing (15%) sets. Preprocessing included text

normalization, tokenization, and enrichment to

ensure BERT compatibility for analysing complex

language patterns. The ontology was created and

managed using Python’s owlready2 library to

integrate semantic reasoning into the framework.

4.1 Performance Comparison Models

Logistic Regression (LR) outperformed Random

Forest (RF) overall with an AUC of 0.62 versus 0.46,

as LR better captured cyberbullying characteristics.

However, RF excelled in detecting ethnicity- and

gender-related content (AUCs of 0.33 and 0.45,

respectively) due to its ability to model non-linear

patterns. LR performed slightly better for age-related

content (AUC 0.38), while both models were highly

effective for religion-related content (LR: 0.99, RF:

0.98).

Figure 5: Comparison for class other_cyberbullying.

Figure 6: Comparison for class not_cyberbullying.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

714

A BERT-based model integrated with NLP and

ontology achieved the best results, with 96.2%

accuracy, 95.8% precision, 95.5% recall, and an F1

score of 95.6%, highlighting its strength in detecting

both explicit and implicit abusive behaviors.

RoBERTa-based models (RoBERTa + CNN and

RoBERTa + GRU) also performed well, achieving

95.2% and 94.8% accuracy, respectively. The graph-

based ontology further improved detection accuracy

by understanding relationships between concepts,

making BERT with Ontology a superior solution for

identifying diverse abusive behaviors.

Figure 7: Graph Based Ontology - words classification.

Figure 8: New Proposed Approach.

5 CONCLUSIONS

This study introduces an effective approach for

detecting cyberbullying by integrating Natural

Language Processing (NLP) with the BERT model

and a Graph-Based Ontology framework. The system

achieved an accuracy of 96.2%, with precision, recall,

and an F1-score all exceeding 95%. These results

represent a significant improvement over traditional

methods, demonstrating the model’s ability to

accurately identify both clear and subtle forms of

cyberbullying. By using BERT’s advanced language

understanding for feature extraction along with the

structured insights provided by graph-based

ontologies, this approach effectively handles complex

language patterns often found in abusive behavior

online. The strong performance of the combined

BERT + Ontology model shows its capability to

detect nuanced instances of cyberbullying that other

models may overlook. Beyond improving detection

accuracy, this method has practical applications in

areas like social media monitoring and online safety

programs. It offers a comprehensive solution that

adapts to real- time language changes, making online

spaces safer.

Future work could further enhance this approach by

integrating additional language models and

expanding the ontology to cover emerging trends in

digital interactions. This ongoing development would

strengthen the system's ability to detect evolving

forms of cyberbullying, ultimately contributing to

more effective online safety measures.

REFERENCES

RRoBERTa Transformer Based Model for Cyberbullying

Detection with GloVe Features. IEEE Access.

Ogunleye, B., & Dharmaraj, B. (2023). The use of a large

language model for cyberbullying detection. Analytics,

2(3), 694-707.

Pericherla, S., & Ilavarasan, E. (2021). Performance

analysis of word embeddings for cyberbullying

detection. IOP Conference Series: Materials Science

and Engineering, 1085(1), IOP Publishing.

Emon, M. I. H., et al. (2022). Detection of Bangla hate

comments and cyberbullying in social media using NLP

and transformer models. In International Conference

on Advances in Computing and Data Sciences (pp. 1-

12). Cham: Springer International Publishing.

Pericherla, S., & Ilavarasan, E. (2024). Transformer

network-based word embeddings approach for

autonomous cyberbullying detection. International

Journal of Intelligent Unmanned Systems, 12(1), 154-

166.

Yani, M. A., & Maharani, W. (2023). Analyzing

Cyberbullying Negative Content on Twitter Social

Media with the RoBERTa Method. JINAV: Journal of

Information and Visualization, 4(1), 61-69.

Teng, T. H., & Varathan, K. D. (2023). Cyberbullying

detection in social networks: A comparison between

machine learning and transfer learning approaches.

IEEE Access, 11, 55533-55560.

Chen, C. K., Ramjee, S., & Wang, J. Aspect-Target

Sentiment Classification for Cyberbullying Detection.

Yi, P., & Zubiaga, A. (2022). Cyberbullying detection

across social media platforms via platform-aware

adversarial encoding. Proceedings of the International

AAAI Conference on Web and Social Media, 16, 123-

133.

Pericherla, S., & Egambaram, I. (2021). Cyberbullying

detection on multi-modal data using pre-trained deep

learning architectures. Ingeniería Solidaria, 17(3), 1-

20.

Gencoglu, O. (2020). Cyberbullying detection with fairness

constraints. IEEE Internet Computing, 25(1), 20-29.

Ontology and AI Integration for Real-Time Detection of Cyberbullying Among University Students

715

Hasan, M. T., et al. (2023). A review on deep-learning-

based cyberbullying detection. Future Internet, 15(5),

179.

Pericherla, S., & Egambaram, P. (2021). Detecting

cyberbullying across multiple platforms using multi-

modal deep learning techniques. Journal of Computer

Vision and Multimedia Studies, 29(4), 102-120.

Gencoglu, O. (2020). Fairness-aware cyberbullying

detection using deep learning. In Proceedings of the

IEEE/ACM International Conference on AI Ethics and

Fairness (pp. 25-34). IEEE. https://doi.org/10.1109/

MIC.2020.3032461

Akhter, A., et al. (2023). A robust hybrid machine learning

model for Bengali cyberbullying detection in social

media. Natural Language Processing Journal, 4,

100027.

Islam, M. S., & Rafiq, R. I. (2023). Comparative analysis

of GPT models for detecting cyberbullying in social

media platforms threads. In Annual International

Conference on Information Management and Big Data

(pp. 245-257). Cham: Springer Nature Switzerland.

Ottosson, D. (2023). Cyberbullying detection on social

platforms using large language models. Journal of AI

Applications in Social Media, 35(1), 59-71.

Akinyemi, J. D., et al. (2023). Cyberbullying detection and

classification in social media texts using machine

learning techniques. In International Conference on

Computer Science, Engineering and Education

Applications (pp. 102-115). Springer Nature

Switzerland.

Sen, M., Masih, J., & Rajasekaran, R. (2024). From tweets

to insights: BERT-enhanced models for cyberbullying

detection. 2024 ASU International Conference in

Emerging Technologies for Sustainability and

Intelligent Systems (ICETSIS). IEEE.

Ferrer, R., Ali, K., & Hughes, C. (2024). Using AI-based

virtual companions to assist adolescents with autism in

recognizing and addressing cyberbullying. Sensors,

24(12), 3875.

Ejaz, N., Razi, F., & Choudhury, S. (2024). Towards

comprehensive cyberbullying detection: A dataset

incorporating aggressive texts, repetition, peerness, and

intent to harm. Computers in Human Behavior, 153,

108123.

Murshed, B. A. H., et al. (2022). DEA-RNN: A hybrid deep

learning approach for cyberbullying detection in

Twitter social media platform. IEEE Access, 10, 25857-

25871.

Chow, D. V., et al. (2023). Cyberbullying detection: An

investigation into natural language processing and

machine learning techniques. 2023 5th International

Conference on Cybernetics and Intelligent System

(ICORIS). IEEE.

Jadhav, R., et al. (2023). Cyber bullying and toxicity

detection using machine learning. 2023 3rd

International Conference on Pervasive Computing and

Social Networking (ICPCSN). IEEE.

El Koshiry, A. M., et al. (2024). Detecting cyberbullying

using deep learning techniques: A pre-trained GloVe

and focal loss technique. PeerJ Computer Science, 10,

e1961.

Kaur, M., & Saini, M. (2023). Role of artificial intelligence

in cyberbullying and cyberhate detection. 2023 14th

International Conference on Computing

Communication and Networking Technologies

(ICCCNT). IEEE.

Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H., &

Kang, J. (2019). BioBERT: A pre-trained biomedical

language representation model for biomedical text

mining.

Bioinformatics, 36(4), 1234–1240. https://doi.org/10.

1093/bioinformatics/btz682

Rogers, A., Kovaleva, O., & Rumshisky, A. (2020). A

primer in BERTology: What we know about how

BERT works. Transactions of the Association for

Computational Linguistics, 8, 842–866.

https://doi.org/10.1162/tacl_a_00349

Stenetorp, P., Pyysalo, S., Topić, G., Ohta, T., Ananiadou,

S., & Tsujii, J. (2012). brat: A web-based tool for NLP-

assisted text annotation. Conference of the European

Chapter of the Association for Computational

Linguistics, 102–107. http://pontus.stenetorp.se

/res/pdf/stenetorp2012brat.pdf

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

716