Hate Speech Detection Using Social Media Discourse: A Multilingual Approach with Large Language Model

Abstract

Online social networks (OSN) and microblogging websites are attracting Internet users and have revolutionized how we communicate with individuals, share their feelings, and exchange ideas across the world with ease. In the extensive age of social media, there is increasing online hate speech, which can provoke violence and contribute to societal division. Hate speech based on race, gender, or religion puts those affected at risk of mental health problems and exacerbates social problems. While current protocols have reduced overt hate speech, subtler forms known as implicit hate speech have emerged, making detection more challenging. This study focuses on hate speech detection using social media discourse, by creating a comprehensive multilingual dataset [25] in Urdu and English and applied multiple machine learning, deep learning, transfer learning, and Large Language model models, such as GPT-3.5 Turbo. By comparing GPT-3.5 Turbo, we identified the effectiveness of large language models in detecting both explicit and implicit forms of hate speech. Our analysis underscores the potential of automated classification systems to reduce reliance on human intervention and to promote constructive online discourse. Our proposed methodology achieved the highest accuracy of 0.91, and achieved the highest performance improvement of 5.81% over transformer models such as BERT. This research adds to the growing body of work on multilingual natural language processing (NLP) and offers insights for reducing hate speech and fostering respectful communication across diverse communities.

Full text

321 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Carlos Aguilar-Ibañez et al.
https://africanjournalofbiomedicalresearch.com/index.php/AJBR
Afr. J. Biomed. Res. Vol. 28(2s) (February 2025); 321-328
Research Article
Hate Speech Detection Using Social Media Discourse: A
Multilingual Approach with Large Language Model
Muhammad Ahmad1, Muhammad Usman1, Sulaiman Khan1, Muhammad
Muzamil2, Ameer Hamza2, Muhammad Jalal2, Ildar Batyrshin1, and Carlos Aguilar-
Ibañez1*
1Centro de Investigación en Computación, Instituto Politécnico Nacional (CIC- PN), Mexico City 07738,
Mexico
2Department of Artificial Intelligence, Computer Science and Software Engineering, The Islamia University
of Bahawalpur, 63100, Pakistan
3School of Informatics and Robotics, Institute of Arts and Culture, Lahore 54000, Pakistan
*Corresponding Author: Carlos Aguilar-Ibañez1
1Centro de Investigación en Computación, Instituto Politécnico Nacional (CIC- PN), Mexico City 07738,
Mexico
Abstract
Online social networks (OSN) and microblogging websites are attracting Internet users and have revolutionized how we
communicate with individuals, share their feelings, and exchange ideas across the world with ease. In the extensive age
of social media, there is increasing online hate speech, which can provoke violence and contribute to societal division.
Hate speech based on race, gender, or religion puts those affected at risk of mental health problems and exacerbates social
problems. While current protocols have reduced overt hate speech, subtler forms known as implicit hate speech have
emerged, making detection more challenging. This study focuses on hate speech detection using social media discourse,
by creating a comprehensive multilingual dataset [25] in Urdu and English and applied multiple machine learning, deep
learning, transfer learning, and Large Language model models, such as GPT-3.5 Turbo. By comparing GPT-3.5 Turbo,
we identified the effectiveness of large language models in detecting both explicit and implicit forms of hate speech. Our
analysis underscores the potential of automated classification systems to reduce reliance on human intervention and to
promote constructive online discourse. Our proposed methodology achieved the highest accuracy of 0.91, and achieved
the highest performance improvement of 5.81% over transformer models such as BERT. This research adds to the growing
body of work on multilingual natural language processing (NLP) and offers insights for reducing hate speech and fostering
respectful communication across diverse communities.
Keywords: LLM, GPT, Machine learning, CNN, BERT algorithm, Social media, SVM
Received: 04/01/2025 Accepted: 24/01/2025
DOI: https://doi.org/10.53555/AJBR.v28i2S.6805
© 2025 The Author(s).
This article has been published under the terms of Creative Commons Attribution-Noncommercial 4.0 International
License (CC BY-NC 4.0), which permits noncommercial unrestricted use, distribution, and reproduction in any medium,
provided that the following statement is provided. “This article has been published in the African Journal of Biomedical
Research”
1. Introduction
The extensive use of social media has led to increase in
online hate speech, which undermines constructive
public discourse and, in some cases, may incite violence
and extremism [1], [2]. Hate speech, often targeting
people based on characteristics such as race, gender, or
religion, fosters discrimination and hostility, leading to
negative impacts on victims’ mental health [3] and

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
322 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
exacerbating social tensions and divisions [4]. While
stronger regulations have helped limit overt forms of
hate speech, they have also led to the emergence of more
indirect, subtle expressions of hate, known as “implicit
hate speech,” which is harder to detect and recognize
because of its nuanced nature [5], [6]. Thus, developing
effective automated systems to detect both explicit and
implicit forms of hate speech is a priority for researchers
and society alike [7].
Existing studies have mainly focused on hate speech
targeted at specific groups, often based on
characteristics like immigration status, gender [8],
religion [9] [10], and race [11]. However, as social
media platforms have become more popular, they have
also raised public awareness of political issues.
Supporters of political figures can now access real-time
updates on their activities and proposed policies,
increasing engagement across the board. Unfortunately,
this engagement sometimes prompts people with
polarized political views to use social media as a channel
for spreading hate speech against those with opposing
beliefs. Detecting hate speech is essential to prevent
potential violence and discrimination, whether from
those spreading it or toward those it targets. This
growing trend highlights the need for robust systems to
monitor and address hate speech in political discussions,
fostering a more respectful online environment.
Recent years have seen the rise of more efficient hate
speech detection systems owing to advancements in
natural language processing (NLP) and machine
learning [20]. With BERT and Roberta, as well as
multilingual embedding’s, modeling of complex
structures of languages and cross-lingual transfer for the
training of language models with generalization
capability and even for low-resource languages such as
Urdu becomes possible[6,7]. Furthermore, the
combination of transformer models with some other
classifiers, such as SVM and Random Forest, presents
avenues for improvement in detection systems in terms
of both efficiency and accuracy. [18, 19].
This study makes three significant contributions to the
field of hate speech detection. Firstly, we built a dataset
related hate speech using a semi supervised learning
annotation procedure in Urdu and English. Secondly we
employed joint multilingual techniques and combine
datasets into a single CSV file, which provide a valuable
resource for the future research. Thirdly, we used
preprocessing techniques. Fourthly, we utilized
advanced feature extraction methods such as TF-IDF,
FasText , Glove and Contextual Embedding’s using
Transformer models to capture both semantic and
syntactic features, enhancing the accuracy and context-
awareness in hate speech detection and lastly we apply
multiple machine learning , deep learning and
transformer models [21] to identify the best fit model
which gives the higher accuracy in this way we will
identify the most effective solution for our hate speech
detection task. These techniques, when applied
effectively, enable the model to accurately flag harmful
content and support efforts to reduce online hate speech.
The term multilingual refers to the structure of our
dataset, which integrates both English and Urdu data
into a single file. The data is organized in an alternating
pattern, where two consecutive rows contain English
text followed by two rows of Urdu text, and this
sequence is repeated throughout the dataset.
This study makes the following Contributions:
1. We applied the schema to develop a comprehensive
Hate speech detection dataset in Urdu and English
with 8700 samples and applied the joint multilingual
techniques as first time.
2. Conduct a comprehensive analysis and performance
evaluation using various deep learning and transfer
learning techniques and Large Language model,
along with comprehensive visualizations.
3. Propose, implement, and evaluate state-of-the-art
Large Language model such as GPT-3.5 Turbo
designed to automatically detect hate speech in
social media discourse, thereby reducing reliance on
human intervention through automated classification
systems
The remainder of this paper is organized as follows.
Section II outlines the literature survey. Section III
contains methodology and design. Section IV presents
results and analysis. Finally, Section V presents the
conclusions of the study.
2. Literature Survey
Yaosheng et al [12] proposed a comprehensive hate
speech detection dataset comprising 20,000 entries
across nine domains, addressing the lack of resources for
Chinese hate speech detection. They introduced a novel
Domain-enhanced Prompt Learning (DePL) method to
efficiently handle the complexities of domain specificity
and data scarcity. Their experimental results indicate
that this methodology achieves state-of-the-art
performance in both few-shot and full-scale detection
scenarios.
Wang et al [13] addresses the challenge of hate speech
proliferation in online environments, proposing a
method to develop a political hate speech lexicon and
train AI classifiers for detection. The authors collected a
Chinese hate speech dataset and implemented both deep
learning and lexicon-based approaches to enhance
detection capabilities. Their framework aims to balance
the need for effective hate speech detection while
preserving the freedom of speech online.
Alkomah et al [14] discussed a detail and comprehensive
review of detection of hate speech in textual the datasets,
they extract textual features, and applied machine
learning. They identified key themes across 138 relevant
studies, noting that many existing approaches do not
yield consistent results across different categories of
hate speech. The analysis revealed a tendency towards
using combined methods, particularly those that
integrate multiple deep learning models. Furthermore,
the review highlighted limitations in the available hate
speech datasets, noting that many are small and lack
reliability for various detection tasks. This study
ultimately aims to provide valuable insights and
empirical evidence regarding the characteristics of hate
speech, assisting the research community in identifying
future research directions.

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
323 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
MacAvaney et al [15] discusses the increasing
prevalence of hate speech as online content expands,
highlighting significant challenges in automated hate
speech detection systems. Key difficulties include the
nuances of language, varying definitions of hate speech,
and the limited availability of robust datasets for training
and evaluating detection algorithms. Additionally, many
contemporary methods, particularly neural networks,
face issues with interpretability, making it challenging
for users to understand the reasoning behind their
decisions.
Yin et al [16] investigate the challenge of hate speech
detection, emphasizing that existing models often fail to
generalize to unseen data. The paper summarizes the
generalizability issues, identifies reasons for these
challenges, and reviews attempts to address them.
Finally, it suggests future research directions to enhance
the generalization capabilities of hate speech detection
systems.
Del et al [17] investigate the spread of harmful
campaigns on social networks, focusing on hate speech
in comments on public Italian Facebook pages. They
create a taxonomy of hate categories and annotate
comments, then develop classifiers using Support
Vector Machines (SVM) and Long Short Term Memory
(LSTM) networks. Their findings demonstrate the
effectiveness of these classifiers in recognizing hate
speech within the first annotated Italian Hate Speech
Corpus.
3. Methodology and Design
3.1 Data collection
The Tweepy API allows us to filter tweets by various
criteria, including date, location, language, and tweet ID
to facilitate data collection. In this study, Tweepy API
were used to collect the tweets, around 50,000 real world
tweets related to hate speech samples were collected in
both Urdu and English languages which mainly cover
the topic such as. The data was gathered from January
2023 to August 2024, focusing on race, ethnicity,
religion, gender, or sexual orientation. To ensure the
dataset's diversity and comprehensiveness, we collected
text samples. After collecting the samples were exactly
labeled as either "hate speech" or "not hate speech," in
binary classification approach. By developing a
balanced and multilingual dataset, the aim was to train
robust machine learning [22] [24] such as Support
Vector Machine (SVM) Extreme Gradient Boosting
(XGB and Logistic Regression (LR) while in deep
learning such as convolutional neural network (CNN)
with a bidirectional long short-term memory (BiLSTM),
transformer models [23] such as Bidirectional Encoder
Representations from Transformers (BERT) and
Robustly Optimized BERT Approach (RoBERTa) a and
large models capable of effectively identifying and
categorizing hate speech in its various forms across
multilingual. as seen in figure 1.
Figure 1. Work flow diagram.
3.2 Annotation.
Accurate and high-quality labeled data is essential for
training effective hate speech detection models. To
achieve this, we employed a hybrid annotation approach
that combines manual labeling with GPT-3.5 Turbo
model for efficient and scalable data annotation. We
began by manually labeling 2000 instances in binary

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
324 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
class to create a foundational dataset, which was then
used to train GPT-3.5 Turbo. The trained model was
utilized to label additional instances, and through an
iterative process of reviewing, correcting, and refining
the generated labels, we incrementally expanded the
dataset to 8700 accurately labeled instances. This
rigorous approach ensured both scalability and quality,
leveraging human expertise to enhance GPT-generated
annotations and create a reliable resource for hate speech
detection.
3.3 Annotation guidelines for initial dataset
3.3.1 Hate speech.
1. If there is racial, ethnic, religious, or gender-based
slurs that demean or dehumanize individuals or
groups.
2. Language that incites violence, promotes harm, or
threatens individuals or groups based on their
identity.
3. Terms that reduce individuals or groups to subhuman
status or animalistic characteristics, indicating a lack
of humanity or empathy.
3.3.2 Not Hate Speech
1. The sentence presents factual information without
expressing hostility or bias toward any group.
2. The statement critiques behavior or actions
respectfully without targeting individuals or groups
based on identity.
3. The sentence engages in dialogue that acknowledges
different viewpoints without demeaning any group.
3.4 Pre-Processing
Data preprocessing is an important step that ensures the
quality of the dataset for machine learning and deep
learning models. Social media data often contain
different language so the preprocess start with text
cleaning, which involves normalizing the content by
removing special characters, links, and unnecessary
whitespace, and converting all text to lowercase to
achieve uniformity. Next, tokenization is performed to
split the text into single words or tokens, allowing for
more granular analysis. Given the multilingual nature of
the dataset, language-specific tokenization techniques
may be applied to account for linguistic variations in
Urdu, and English. Additionally, TF-IDF (Term
Frequency-Inverse Document Frequency) utilized to
reflect the importance of words in the dataset as seen in
figure 2. To improve the robustness of the dataset, we
employed data augmentation on existing samples. For
this purpose we used back translation thereby increasing
the dataset's size and diversity. This comprehensive
preprocessing approach ensures that the dataset is well-
structured and ready for training models, ultimately
improving the accuracy and effectiveness of hate speech
detection efforts.
Figure 2. Pre-processing phases
3.5 Application of Models, Training and testing
phase
The application of models in this study contains a
systematic approach for both training and testing phases
to ensure effective performance in hate speech detection.
During the training phase, we utilize 80% of data on
training and remaining 20% for testing to feed our
models, for this purpose we employed various machine
learning SVM, XG, and K-NN and deep learning
BiLSTM and CNN, transformer such as BERT, Roberta
and a large language such as GPT 3.5 Turbo model
tailored for text classification. This phase involves
hyper-parameter tuning and optimization to enhance
model accuracy and generalizability. Once trained, the
models undergo a rigorous testing phase, where they are
evaluated on a separate validation dataset to assess their
performance metrics, including precision, recall, and F1-
score.
4. Results and Discussion.
In this section, we will shows the Results of our models
based on methodology, alongside a comprehensive
discussion of their implications and significance. In
experimentations process we train multiple models,
including traditional machine learning algorithms and
advanced deep learning architectures and large language
models, on a diverse dataset of annotated text.
4.1 Results for Machine learning.
Figure 3 shows the performance of machine learning
models such as SVM, LR, XGB, and K-NN in a
classification task. Each model is assessed based on four
key metrics such as precision, recall, F1-score, and
accuracy. The SVM shows superior performance with
an accuracy of 0.8, signifying a balanced ability to
correctly identify positive instances while minimizing
false positives. The LR model has slightly lower metrics,
with a precision, recall, F1-score, and accuracy all at
0.79. The XGB model matches SVM in precision, recall,
and F1-score at 0.8, while also achieving an accuracy of
0.8, indicating its robustness in making accurate
predictions. In contrast, the K-NN model lags
significantly behind the others with a precision, recall,
F1-score, and accuracy of 0.68, reflecting its limitations

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
325 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
in this context. Overall, the results indicate that SVM
outperform then all other models.
Figure 3. Performance comparisons of the machine learning models.
4.2 Deep learning models.
The Figure 4 illustrate the performance of two deep
learning models such as CNN and BiLSTM in hate
speech classification task. Both models exhibit strong
performance. In comparison, the BiLSTM model
performs better across all metrics, with a precision of
0.81, recall of 0.82, F1-score of 0.82, and accuracy of
0.82. This suggests that the BiLSTM not only accurately
classifies instances but also captures context better due
to its bidirectional nature, leading to improved recall and
F1-score.
Figure 4.performance comparisons of the deep learning models
4.3 Transformers results
The figure 5 illustrate the performance of two advanced
transformer-based models: RoBERTa and BERT. Both
models demonstrate strong capabilities, with RoBERTa
achieving a precision of 0.85, recall of 0.85, F1-score of
0.84, and accuracy of 0.85. on the other hand BERT,
exhibits superior performance across all metrics, with a
precision of 0.86, recall of 0.86, F1-score of 0.86, and
accuracy of 0.86. This suggests that BERT not only
correctly classifies instances with a higher rate but also
captures contextual information more effectively,
leading to better overall performance in hate speech
dataset.

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
326 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
Figure 5.performance comparisons of the Transfer learning models
4.4 Large Language Model
The figure 6 shows the performance of the advance
Large Language Model of OpenAi GPT-3.5 Turbo
model showcasing high scores with a precision of 0.91,
the model proves that 91% of its positive predictions are
accurate, indicating a low false positive rate. The recall,
also at 0.91, reveals that the model identifies 91% of
actual positive instances, showcasing its effectiveness in
detecting relevant cases. The F1-score, which combines
precision and recall, is also 0.91, indicating a balanced
performance between minimizing false positives and
maximizing true positives. Finally, an accuracy of 0.91
shows that the model correctly predicts 91% of all
instances, reflecting its overall effectiveness in the
classification task.
Figure 6.performance comparisons of the Transfer learning models
4.5 Error Analysis
The table 1 display the top-performing models in each
learning approaches in a classification task. The SVM)
model, achieves a precision of 0.79, a recall of 0.80, an
F1-score of 0.80, and an accuracy of 0.80. The BiLSTM
model, shows improved performance with a precision of
0.81, recall of 0.82, F1-score of 0.82, and accuracy of
0.82. The BERT model further enhances the results with
precision, recall, F1-score, and accuracy all at 0.86. The
standout performer is the our proposed model GPT-3.5
Turbo, classified as a which achieves the highest metrics
across the board with a precision of 0.91, recall of 0.91,
F1-score of 0.91, and accuracy of 0.91 and achieved the
highest performance improvement of 5.81% over the
Transformer models.

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
327 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
Models
Learning approach
Precision
Recall
F1-score
Accuracy
SVM
Machine leaning
0.79
0.8
0.8
0.8
BiLSTM
Deep learning
0.81
0.82
0.82
0.82
BERT
Transformer
0.86
0.86
0.86
0.86
GPT-3.5 Turbo
LLM
0.91
0.91
0.91
0.91
5. Conclusion
This study demonstrates the effectiveness of using
multilingual techniques and large language models for
hate speech detection across multiple languages,
specifically Urdu and English. By constructing a
comprehensive dataset and applying advanced feature
extraction and machine learning methods, we have
enhanced the accuracy and efficiency of automated
systems. Our approach not only contributes valuable
resources for future research but also advances the
capabilities of hate speech detection, particularly in
social media discourse. This work paves the way for
more robust, scalable, and context-aware systems that
can help mitigate the spread of harmful content online.
Funding: This research did not receive any funding.
Acknowledgements: This work was done with partial
support from the Mexican Government through the grant
A1-S-47854 of CONACYT, Mexico, grant 20241529 of
the Secretaría de Investigación y Posgrado of the
Instituto Politécnico Nacional, Mexico. The authors
thank the CONACYT for the computing resources
brought to them through the Plataforma de Aprendizaje
Profundo para Tecnologías del Lenguaje of the
Laboratorio de Supercómputo of the INAOE, Mexico
and acknowledge support of Microsoft through the
Microsoft Latin America PhD.
References
1. N. Djuric, J. Zhou, R. Morris, M. Grbovic, V.
Radosavljevic, and N. Bhamidipati, ‘‘Hate speech
detection with comment embeddings,’’ in Proc. 24th
Int. Conf. World Wide Web, May 2015, pp. 29–30.
2. Y. Chen, Y. Zhou, S. Zhu, and H. Xu, ‘‘Detecting
offensive language in social media to protect
adolescent online safety,’’ in Proc. Int. Conf.
Privacy, Secur., Risk Trust Int. Conf. Social
Comput., Sep. 2012, pp. 71–80.
3. B. M. Tynes, M. T. Giang, D. R. Williams, and G. N.
Thompson, ‘‘Online racial discrimination and
psychological adjustment among adolescents,’’ J.
Adolescent Health, vol. 43, no. 6, pp. 565–569, Dec.
2008.
4. M. L. Williams and P. Burnap, ‘‘Cyberhate on social
media in the aftermath of woolwich: A case study in
computational criminology and big data,’’ Brit. J.
Criminol., vol. 56, no. 2, pp. 211–238, Mar. 2016.
5. W. Warner and J. Hirschberg, ‘‘Detecting hate
speech on the World Wide Web,’’ in Proc. 2nd
Workshop Lang. Social Media, 2012, pp. 19–26.
6. H. M. Saleem, K. P. Dillon, S. Benesch, and D.
Ruths, ‘‘A web of hate: Tackling hateful speech in
online social spaces,’’ in Proc. Workshop
Programme, 2016, p. 1.
7. M. ElSherief, C. Ziems, D. Muchlinski, V. Anupindi,
J. Seybolt, M. De Choudhury, and D. Yang, ‘‘Latent
hatred: A benchmark for understanding implicit hate
speech,’’ in Proc. Conf. Empirical Methods Natural
Lang. Process., 2021, pp. 345–363.
8. V. Basile, C. Bosco, E. Fersini, D. Nozza, V. Patti,
F. M. R. Pardo, et al., "SemEval-2019 task 5:
Multilingual detection of hate speech against
immigrants and women in Twitter", Proc. 13th Int.
Workshop Semantic Eval., pp. 54-63, 2019.
9. N. Ousidhoum, Z. Lin, H. Zhang, Y. Song and D.-Y.
Yeung, "Multilingual and multi-aspect hate speech
analysis" in arXiv: 1908.11049, 2019.
10. Alfina, R. Mulia, M. I. Fanany and Y. Ekanata, "Hate
speech detection in the Indonesian language: A
dataset and preliminary study", Proc. Int. Conf. Adv.
Comput. Sci. Inf. Syst. (ICACSIS), pp. 233-238, Oct.
2017.
11. P. Burnap and M. L. Williams, "Cyber hate speech
on Twitter: An application of machine classification
and statistical modeling for policy and decision
making", Policy Internet, vol. 7, no. 2, pp. 223-242,
2015.
12. Yaosheng, Z., Tiegang, Z., Tingjun, Y., & Li, H.
(2024). Domain-enhanced Prompt Learning for
Chinese Implicit Hate Speech Detection. IEEE
Access.
13. Wang, C. C., Day, M. Y., & Wu, C. L. (2022).
Political hate speech detection and lexicon building:
A study in Taiwan. IEEE Access, 10, 44337-44346.
14. Alkomah, F., & Ma, X. (2022). A literature review
of textual hate speech detection methods and
datasets. Information, 13(6), 273.
15. MacAvaney, S., Yao, H. R., Yang, E., Russell, K.,
Goharian, N., & Frieder, O. (2019). Hate speech
detection: Challenges and solutions. PloS one, 14(8),
e0221152.
16. Yin, W., & Zubiaga, A. (2021). Towards
generalisable hate speech detection: a review on
obstacles and solutions. PeerJ Computer Science, 7,
e598.
17. Del Vigna12, F., Cimino23, A., Dell’Orletta, F.,
Petrocchi, M., & Tesconi, M. (2017, January). Hate
me, hate me not: Hate speech detection on facebook.
In Proceedings of the first Italian conference on
cybersecurity (ITASEC17) (pp. 86-95).
18. Mubarak, H., Darwish, K., & Abdelali, A. (2017).
Abusive Language Detection on Arabic Social
Media. Proceedings of the First Workshop on
Abusive Language Online.
19. Mishra, A., Yannakoudakis, H., & Shutova, E.
(2018). Neural Character-based Composition
Models for Abusive Language Detection.
Proceedings of the 2018 Conference on Empirical
Methods in Natural Language Processing (EMNLP
2018).
20. Ahmad, Muhammad, et al. "Elegante: A Machine
Learning-Based Threads Configuration Tool for

Hate Speech Detection Using Social Media Discourse: A Multilingual Approach With Large Language Model
328 Afr. J. Biomed. Res. Vol. 28, No.2s (February) 2025 Usman Sardar, Carlos Aguilar-Ibañez et al.
SpMV Computations on Shared Memory
Architecture." Information 15.11 (2024): 685.
21. Ahmad, Muhammad, et al. "Cotton Leaf Disease
Detection Using Vision Transformers: A Deep
Learning Approach." crops 1 (2024): 3.
22. Ahmed, M., Usman, S., Shah, N. A., Ashraf, M. U.,
Alghamdi, A. M., Bahadded, A. A., & Almarhabi, K.
A. (2022). AAQAL: A machine learning-based tool
for performance optimization of parallel SPMV
computations using block CSR. Applied
Sciences, 12(14), 7073.
23. Ullah, F., Ahmed, M., Zamir, M. T., Arif, M., Felipe-
Riverón, E., & Gelbukh, A. (2024, March). Optimal
Scheduling for the Performance Optimization of
SpMV Computation using Machine Learning
Techniques. In 2024 7th International Conference
on Information and Computer Technologies
(ICICT) (pp. 99-104). IEEE.
24. Ullah, F., Zamir, M., Arif, M., Ahmad, M., Felipe-
Riveron, E., & Gelbukh, A. (2024, March). Fida@
DravidianLangTech 2024: A Novel Approach to
Hate Speech Detection Using Distilbert-base-
multilingual-cased. In Proceedings of the Fourth
Workshop on Speech, Vision, and Language
Technologies for Dravidian Languages (pp. 85-90).
25. Ahmad, M., Sardar, U., Humaira, F., Iqra, A.,
Muzzamil, M., Hmaza, A., ... & Batyrshin, I. (2024).
Hope Speech Detection Using Social Media
Discourse (Posi-Vox-2024): A Transfer Learning
Approach. Journal of Language and
Education, 10(4), 31-43.