1. Problem Statement

Cyberbullying has emerged as a critical challenge in the digital era, causing severe psychological harm, particularly to adolescents. While social media platforms (e.g., Facebook, Twitter, Instagram) facilitate global communication, they are increasingly exploited for harassment, hate speech, and abuse. The paper identifies several critical gaps in current detection systems:

• Modality Limitations: Most algorithms focus solely on text, failing to detect multimodal cyberbullying (images, videos, audio).
• Linguistic Bias: Research heavily favors high-resource languages (English, Chinese, French), leaving low-resource languages (Swahili, Bengali, Marathi) under-represented and leading to biased detection.
• Data Scarcity & Quality: High-quality, labeled datasets are difficult to obtain due to platform restrictions (API limitations), leading to issues with data imbalance, inconsistency, and small sample sizes.
• Evolving Tactics: Bullies use obfuscated words, misspellings, and coded languages to bypass traditional detection systems.
• Real-time & Adaptability: Existing systems struggle to adapt to rapidly changing platform features and user interfaces or to operate effectively in real-time.

2. Methodology

The authors conducted a Systematic Literature Review (SLR) following the PRISMA 2020 guidelines to ensure transparency and reproducibility.

• Search Strategy: Articles were sourced from major academic databases (IEEE Xplore, SpringerLink, ScienceDirect, ACM, MDPI) published between 2018 and November 2025.
• Selection Criteria: The review included peer-reviewed studies addressing cyberbullying from technical, policy, or ethical perspectives. Non-peer-reviewed papers, workshop papers under 4 pages, and studies lacking a cyberbullying context were excluded.
• Screening Process:
  • Initial search yielded 437 papers.
  • Title/Abstract screening reduced this to 277.
  • Full-text review resulted in 161 final studies included in the analysis.
• Taxonomy: The review categorizes detection approaches into four main technological pillars:
  1. Traditional Models (Rule-based, Statistical).
  2. Machine Learning (ML).
  3. Deep Learning (DL).
  4. Large Language Models (LLMs).

3. Key Contributions

The paper makes five primary contributions to the field:

1. Comprehensive Comparative Study: A systematic analysis of detection methods from 2018 to 2025, comparing performance across different model families.
2. Technology & Approach Analysis: A deep dive into the evolution from traditional rule-based systems to modern LLMs, highlighting the shift in feature extraction and classification capabilities.
3. Evaluation Metrics Review: An analysis of standard metrics (Accuracy, Precision, Recall, F1-score) and their applicability to imbalanced datasets.
4. Dataset Critical Analysis: A detailed examination of the source, diversity, size, and annotation quality of datasets used in the field, identifying critical flaws like class imbalance and language bias.
5. Future Roadmap: Proposing specific pathways for future research, including multimodal detection, cross-lingual approaches, and ethical frameworks.

4. Key Results & Findings

A. Performance of Detection Approaches

• Traditional Models: Rule-based and character-matching techniques are effective for small datasets and specific obfuscated words but lack semantic understanding and scalability. They suffer from high false-positive rates.
• Machine Learning (ML): Supervised methods (SVM, Naive Bayes, Random Forest) generally outperform unsupervised ones. SVM is identified as the top-performing traditional classifier, often achieving F1-scores around 0.94–0.97 when combined with TF-IDF and BoW features. However, ML models struggle with automatic feature extraction and context.
• Deep Learning (DL): Models like BiLSTM, CNN, and GRU demonstrate superior performance in capturing semantic nuances and long-term dependencies. BiLSTM is highlighted as the best-performing DL algorithm, with some studies reporting F1-scores up to 0.98. DL models excel at automatic feature extraction but face challenges with overfitting and interpretability.
• Large Language Models (LLMs): Models like BERT, RoBERTa, and GPT-3/4 achieve the highest performance (F1-scores often >0.95). They offer rich linguistic features and flexibility. However, they are computationally expensive, energy-intensive, and struggle with real-time inference and low-resource languages.

B. Dataset Analysis

• Source Dominance: Twitter is the most utilized data source due to its accessible API. Other sources include Facebook, Instagram, Reddit, and Formspring.
• Language Bias: English is the dominant language in datasets. Low-resource languages (Swahili, Bengali, Hindi) are severely underrepresented.
• Class Imbalance: A critical finding is the severe class imbalance in datasets (e.g., 10:1 to 10,000:1 ratios of non-bullying to bullying). This leads to models that are biased toward predicting the majority class (non-bullying), rendering accuracy metrics misleading. Precision, Recall, and F1-score are deemed essential for evaluation.
• Annotation Issues: Inconsistent labeling due to subjective human interpretation and a lack of expert linguists for low-resource languages compromises dataset quality.

C. Challenges Identified

• Linguistic: Code-switching, slang, emojis, and cultural context make detection difficult.
• Data: Scarcity of labeled data, high costs of annotation, and privacy restrictions.
• Technical: Lack of real-time detection capabilities, poor model interpretability (black-box nature of DL/LLMs), and high computational costs for LLMs.
• Ethical: Concerns regarding user privacy, consent, and potential bias against marginalized groups.

5. Significance and Future Directions

The paper concludes that while LLMs represent the state-of-the-art, a holistic solution requires addressing data and methodological gaps.

• Future Directions:
  • Multimodal Detection: Integrating text with images, audio, and metadata.
  • Low-Resource Languages: Developing specialized models and datasets for African and Asian languages (e.g., Swahili, Hindi).
  • Cross-Lingual Approaches: Using transfer learning to apply knowledge from high-resource to low-resource languages.
  • Automated Annotation: Utilizing Self-Supervised Learning (SSL) and Generative Adversarial Networks (GANs) to generate synthetic data and reduce manual labeling costs.
  • Ethical Frameworks: Implementing strict privacy safeguards and bias mitigation strategies in model deployment.
• Recommendations: The authors advocate for interdisciplinary research (combining CS, psychology, and sociology), the creation of multi-labeled datasets (categorizing specific types of abuse), and the development of defense mechanisms like re-authentication and community voting systems.

Overall Significance: This review serves as a foundational guide for researchers and practitioners, highlighting that while technological advancements (especially in LLMs) are promising, the field must pivot toward inclusive, ethical, and multimodal solutions to effectively combat cyberbullying across diverse global populations.