# Cyberbullying detection approaches for Arabic texts: a systematic literature review

## TL;DR

This paper reviews methods for detecting cyberbullying in Arabic texts, highlighting challenges and trends in handling linguistic and cultural nuances.

## Contribution

The study systematically analyzes 35 papers on Arabic cyberbullying detection, identifying methodological trends and gaps in cultural and linguistic adaptation.

## Key findings

- Transformer models like AraBERT outperform traditional methods in handling dialectal and orthographic variations in Arabic texts.
- Current systems are better at detecting clear cases of cyberbullying than capturing all relevant instances due to higher precision than recall.
- Evaluation metrics and annotation transparency remain significant challenges in Arabic cyberbullying detection research.

## Abstract

This study presents a comprehensive review of current methodologies, trends, and challenges in cyberbullying detection within Arabic-language contexts, with a focus on the unique linguistic and cultural factors associated with Arabic. This study reviews 35 peer-reviewed articles about the identification of cyberbullying in Arabic text. Reported accuracies across datasets and platforms range from approximately 73 to 96%, with precision frequently surpassing recall, suggesting that systems are more adept at identifying blatant bullying than at encompassing all pertinent instances. Methodologically, conventional machine learning utilizing Arabic-specific characteristics remains effective on smaller datasets, however deep neural architectures—especially CNN/BiLSTM—and transformer models like AraBERT yield superior outcomes when dialectal heterogeneity and orthographic noise are mitigated. Evaluation methodologies differ; research using a neutral class frequently indicates exaggerated accuracy, underscoring the necessity to emphasize macro-averaged F1 and per-class metrics. The evidence underscores deficiencies in dialectal representativeness, the uniformity of bullying notions compared to general abuse, and the transparency of annotation processes. Ethical and deployment considerations—privacy preservation, dialectal bias, and real-time robustness—are becoming increasingly significant. We integrate trends (models and features), standards (labeling and metrics), and future work directions, encompassing dialect-robust pretraining, cross-dataset evaluation, context-aware modeling, and human-in-the-loop frameworks. The review offers a comprehensive basis for researchers and practitioners pursuing culturally and linguistically tailored approaches to Arabic cyberbullying detection.