Neural Network Based Cyberbullying Classification

Abstract

The rapid growth of social media platforms has led to a significant increase in cyberbullying, making it a major challenge for online communities. Cyberbullying includes harmful and offensive behavior targeting individuals based on factors such as age, gender, religion, and ethnicity. Automatic detection of such content remains difficult due to informal language, sarcasm, ambiguous expressions, and varying contextual meanings. Therefore, there is a need for intelligent systems capable of accurately identifying different forms of cyberbullying. This research focuses on multiclass cyberbullying detection using advanced Natural Language Processing (NLP) and Deep Learning techniques. The proposed framework classifies social media text into categories including age, gender, religion, ethnicity, other cyberbullying, and not cyberbullying. To achieve effective classification, Bidirectional Encoder Representations from Transformers (BERT) is employed to capture contextual and semantic information from text. While existing machine learning and deep learning approaches have shown promising performance, but they face challenges in handling uncertain or ambiguous content. To address this limitation, the proposed framework integrates BERT with Neutrosophic Logic. BERT generates meaningful contextual representations, while Neutrosophic Logic utilizes Truth, Indeterminacy, and Falsity measures to model uncertainty during classification. This integration enhances the reliability and interpretability of multiclass cyberbullying detection, making the framework more effective for analyzing complex social media content.

Introduction

This study focuses on the growing problem of cyberbullying on social media platforms, which can cause serious psychological harm such as anxiety, depression, and social isolation. Detecting cyberbullying is challenging because online text often includes slang, sarcasm, emojis, and context-dependent meanings that traditional methods struggle to interpret. While machine learning and deep learning models (such as SVM, CNN, LSTM, and BiLSTM) have improved detection, they still face limitations in handling ambiguity and uncertainty in real-world data.

To address this, the proposed framework combines BERT (Bidirectional Encoder Representations from Transformers) with Neutrosophic Logic for multiclass cyberbullying detection. The system classifies Twitter data into six categories: age, gender, religion, ethnicity, other cyberbullying, and non-cyberbullying. BERT is used to extract deep contextual and semantic meaning from text, while Neutrosophic Logic introduces a decision-making layer that handles uncertainty using Truth, Indeterminacy, and Falsity measures, improving reliability and interpretability.

The methodology includes data collection from a large Twitter dataset (47,692 samples), preprocessing steps such as text cleaning, normalization, label encoding, and class balancing, followed by BERT-based classification and uncertainty-aware decision refinement. An additional topic modeling module (LDA) is used to extract themes from text for better understanding.

Experimental results show that BERT effectively classifies cyberbullying content, and the addition of Neutrosophic Logic improves decision confidence in ambiguous cases. Overall, the proposed system enhances accuracy, robustness, and interpretability in detecting cyberbullying, offering a more reliable solution for safer online environments.

Conclusion

The proposed Multiclass Cyberbullying Detection System was developed to automatically identify different types of cyberbullying present in social media text using BERT and Neutrosophic Logic. The system effectively classifies textual content into categories such as Age, Ethnicity, Gender, Religion, Other Cyberbullying, and Not Cyberbullying. By leveraging BERT\'s contextual understanding capabilities, the model was able to capture meaningful linguistic patterns and semantic relationships within social media conversations, enabling accurate multiclass classification. To further enhance the reliability of predictions, Neutrosophic Logic was integrated into the framework as a decision-making layer. By incorporating Truth, Indeterminacy, and Falsity measures, the system was able to better handle uncertain and ambiguous cases, making the classification process more interpretable and trustworthy. In addition, the topic extraction module provided useful insights into the key themes discussed in the analyzed text, offering a deeper understanding of the content. The experimental results showed that the proposed framework effectively distinguished between cyberbullying and non-cyberbullying content while maintaining strong classification performance across multiple categories. The developed system demonstrates the potential of combining advanced Natural Language Processing techniques with uncertainty-aware reasoning to support the detection of harmful online behavior. Overall, the proposed approach provides a practical and intelligent solution for cyberbullying detection and can contribute to creating safer and more responsible online communication environments.

References

[1] A. F. Alqahtani and M. Ilyas, “An Ensemble-Based Multi-Classification Machine Learning Classifiers Approach to Detect Multiple Classes of Cyberbullying,” Machine Learning and Knowledge Extraction, vol. 6, no. 1, pp. 156–170, Jan. 2024, doi: 10.3390/make6010009. [2] P. G. Balaji, P. P. Katariya, S. S., and M. Venugopalan, “Cyberbullying Detection on Multiclass Data Using Machine Learning and A Hybrid CNN-BiLSTM Architecture,” in Proc. 2024 International Conference on Knowledge Engineering and Communication Systems (ICKECS), 2024, doi: 10.1109/ICKECS61492.2024.10616957. [3] M. Hzami, H. Mahersia, and T. Bejaoui, “Multi-Level Cyberbullying Detection on Social Media Using Machine and Deep Learning Models,” in Proc. 2025 5th IEEE Middle East and North Africa Communications Conference (MENACOMM), 2025, doi: 10.1109/MENACOMM62946.2025.10911024. [4] N. Islam, R. Haque, P. K. Pareek, M. B. Islam, I. H. Sajeeb, and M. H. Ratul, “Deep Learning for Multi-Labeled Cyberbully Detection: Enhancing Online Safety,” in Proc. 2023 International Conference on Data Science and Network Security (ICDSNS), 2023, doi: 10.1109/ICDSNS58469.2023.10245135. [5] M. Alinejad, “Multiclass Cyberbullying Detection Using Advanced Neural Network Architectures: A Comparative Study Amidst the COVID-19 Pandemic,” University of Central Florida STARS, Data Science and Data Mining, Jul. 2025. [Online]. Available: [6] F. Farasalsabila, E. Utami, and S. Raharjo, “Multi-Label Classification Using BERT for Cyberbullying Detection,” in Proc. 2024 4th International Conference of Science and Information Technology in Smart Administration (ICSINTESA), 2024, doi: 10.1109/ICSINTESA62455.2024.10748045. [7] A. Vadivel, N. Moogambigai, S. Tamilselvan, and P. Thangaraja, “Application of Neutrosophic Sets Based on Neutrosophic Score Function in Material Selection,” in Proc. 2022 First International Conference on Electrical, Electronics, Information and Communication Technologies (ICEEICT), 2022, doi: 10.1109/ICEEICT53079.2022.9768448. [8] E. Fan, K. Hu, and X. Li, “Review of Neutrosophic-Set-Theory-Based Multiple-Target Tracking Methods in Uncertain Situations,” in Proc. 2019 IEEE International Conference on Artificial Intelligence and Computer Applications (ICAICA), 2019, pp. 19–27. [9] Y. M. Ibrahim, R. Essameldin, and S. M. Saad, “Social Media Forensics: An Adaptive Cyberbullying-Related Hate Speech Detection Approach Based on Neural Networks With Uncertainty,” IEEE Access, vol. 12, pp. 59474–59489, 2024, doi: 10.1109/ACCESS

Copyright

Copyright © 2026 Amena Arman, Dr. P. Sammulal . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.