Several studies have utilized ML and DL algorithms to detect cyberbullying, with Support Vector Machine (SVM) and Naïve Bayes (NB) being frequently applied (e.g., Haidar et al., 2017; Alakrot et al., 2018). More recently, DL methods such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have demonstrated improved performance due to their ability to capture context and semantic meanings in text (e.g., Haidar et al., 2018; Mouheb et al., 2019; Mohaouchane et al., 2019). Ensemble learning, where multiple models are combined to improve prediction accuracy, has shown promise in boosting performance. For instance, stacking, boosting, and bagging techniques have demonstrated better performance in detecting Arabic cyberbullying content (e.g., Haidar et al., 2018; Khairy et al., 2023; Table 2).

Sentiment analysis, often coupled with lexicon-based approaches, is commonly used to detect harmful content. AlHarbi et al. (2019) and Farid and El-Tazi (2020) used sentiment-based lexicons for Arabic texts, finding that pointwise mutual information (PMI) and lexicon enhancement can improve detection accuracy. Sentiment-based approaches are also utilized alongside NLP tools, such as tokenization and stemming, for feature extraction, enhancing the ability to detect cyberbullying based on emotional cues.

Dialectal Arabic presents a significant challenge, as standard ML models may not perform well on diverse dialects. Studies such as Alsubait and Alfageh (2021) and Al-Hassan and Al-Dossari (2022) indicate that datasets tailored to specific dialects (e.g., Egyptian, Levantine) enhance detection efficacy. Additionally, transformer-based models like AraBERT and multilingual BERT have emerged as effective tools for dealing with dialectal variations, as they can better capture semantic nuances across dialects (e.g., Khezzar et al., 2023; Alrashidi et al., 2023).

Arabic cyberbullying detection relies heavily on curated datasets. Studies often use platform-specific datasets from Twitter, YouTube, and Facebook, with datasets labeled for harmful or offensive language (e.g., Bashir and Bouguessa, 2021; Khairy et al., 2023). These datasets include common cyberbullying characteristics like threats, insults, and hate speech. However, the issue of dataset imbalance (more non-cyberbullying content than cyberbullying) persists, affecting model performance. Techniques like oversampling and downsampling have been employed to address this imbalance, as seen in AlFarah et al. (2022). Table 3. Shows some examples of the existing datasets addressing cyberbullying in Arabic.

The ArCybC/ArCyC corpus represents one of the few openly accessible multi-dialect Twitter datasets that makes a clear distinction between cyberbullying and general offensive content. Its development is supported by detailed documentation of the annotation pipeline and guidelines, ensuring methodological transparency (Shannag et al., 2022). The ArbCyD dataset significantly expands the available volume by including annotated Twitter posts (Aljalaoud et al., 2025).

Standards such as precision, recall, F1-score, and accuracy are commonly used to evaluate detection methods (e.g., Haidar et al., 2017; Alakrot et al., 2018). Although precision and recall are essential for accurate detection, the unique characteristics of the Arabic language and cyberbullying-specific terms often require additional metrics and customized standards. Studies such as El-Alami et al. (2022) and Amer Hamzah and Dhannoon (2023) advocate for using contextual features like sentiment polarity, emojis, and user history in cyberbullying detection. These standards help capture the nuanced characteristics of online abuse, especially within specific platforms or dialects.

Some evaluations adopt three-way labeling schemes that distinguish bullying/abusive content, non-bullying content, and neutral content. When overall accuracy is computed across all classes, the typically high prevalence of neutral instances can inflate the metric and obscure a system’s effectiveness on the bullying class, which is the primary target in safety-critical applications. For example, the Instagram-based Arabic cyberbullying benchmark provides a multi-class design with positive (bullying), negative (non-bullying), and neutral categories, together with inter-annotator agreement reporting and baseline models (Albayari and Abdallah, 2022). In such settings, macro-F1 and per-class F1 are preferable for comparing systems intended to detect bullying, whereas accuracy across all three classes can be misleading when neutral content dominates the distribution.

Recent research has incorporated psychological theories to enhance cyberbullying detection by analyzing underlying personality traits in text (e.g., Elzayady et al., 2023). Such frameworks align detection methods with broader behavioral standards, moving toward a more human-centered approach in identifying abusive content. Other studies, such as Boulouard et al. (2022), address multilingual standards by analyzing Arabic text in translation and leveraging cross-linguistic BERT models, thus ensuring consistency in detecting cyberbullying characteristics across languages.

Future research could focus on developing larger, dialect-specific datasets to address the significant linguistic diversity in Arabic. Datasets for Moroccan, Syrian, and Gulf dialects remain limited and would improve detection accuracy for specific regions (e.g., Rachidi et al., 2023; Ali and Kurdy, 2022). Studies also suggest expanding multilingual capabilities to improve cross-linguistic performance, with transformer models like BERT and mBERT showing potential for multilingual hate speech analysis (e.g., Alrashidi et al., 2023; Shannaq et al., 2022).

For limited-resource settings, few strategies with large language models can be grounded in complementary lines of evidence. First, in-context learning has been shown to deliver strong few-shot performance without gradient updates; GPT-3’s original study established that scaling enables task-agnostic adaptation via a handful of exemplars embedded in the prompt, a result that has shaped subsequent methodology for low-data regimes (Brown et al., 2020). Second, prompt-based and prompt-free fine-tuning methods consistently improve over naïve fine-tuning when labeled data are scarce. Pattern-Exploiting Training and its generative extension reframe supervision as cloze-style patterns to amplify supervision from very small datasets, while LM-BFF automates prompt construction and demonstration selection to yield large gains across classification and regression tasks (Schick and Schütze, 2020). Complementing these, SetFit avoids handcrafted prompts altogether by contrastively fine-tuning sentence-transformer encoders on a handful of pairs and then training a lightweight classifier on the induced embeddings, matching or surpassing larger fully fine-tuned models under strict few-shot budgets (Tunstall et al., 2022). Moreover, parameter-efficient adaptation techniques such as LoRA reduce trainable parameters by orders of magnitude while preserving or improving downstream quality, which is particularly attractive when domain transfer must be achieved under tight compute and annotation constraints (Hu et al., 2022). To mitigate the scarcity of human-written instructions, Self-Instruct bootstraps synthetic instruction–input–output triplets from the model itself and shows substantial gains over the base model, offering a practical path when labeled data are limited (Wang et al., 2022). Evidence from multilingual and domain-specific studies indicates that these approaches translate beyond English benchmarks. Cross-lingual in-context learning studies report consistent benefits for genuinely low-resource languages and highlight alignment techniques that stabilize label semantics across languages, while evaluations in biomedical and clinical NLP show that instruction-tuned LLMs can perform competitively on few-shot entity recognition, QA, and relation extraction when carefully prompted (Cahyawijaya et al., 2024).

Future research could involve advancing feature engineering, particularly through contextual embeddings, attention mechanisms, and personality inference models. These methods could enhance the interpretability of cyberbullying detection systems and better capture contextual aspects of offensive language (e.g., Mohaouchane et al., 2019; Elzayady et al., 2023). Additionally, hybrid models combining CNN, RNN, and BERT-based architectures have shown promise for handling complex language features, and future studies could explore further model fusion or ensemble methods for improved accuracy (e.g., Mohaouchane et al., 2019; Althobaiti, 2022).

Ethical standards and privacy concerns will play a growing role in future cyberbullying detection research. Privacy-preserving algorithms, especially those that anonymize or filter sensitive information, can support ethical AI use on social media platforms (e.g., Omar et al., 2021). Another area for future exploration is real-time cyberbullying detection systems that respond dynamically to harmful content as it is posted. While challenging, real-time models could be feasible with lightweight DL architectures tailored for social media monitoring (e.g., Amer Hamzah and Dhannoon, 2023; Kanan et al., 2021).

Ethical risks arise at each stage of dataset development and deployment for Arabic cyberbullying detection, beginning with data collection. The Instagram-based benchmark demonstrates the value of reporting annotation protocols and inter-annotator agreement alongside careful corpus descriptions; however, as with Twitter- and YouTube-based datasets, the presence of user mentions and cross-post threads can inadvertently expose targets and perpetrators if not aggressively sanitized (Albayari and Abdallah, 2022; Haidar et al., 2019; Alakrot et al., 2018; Alduailaj et al., 2023; Al-Ajlan and Ykhlef, 2018; Alrougi et al., 2024). Representativeness is a second, persistent ethical and scientific concern. Arabic social media is heterogeneous across dialects, platforms, and communities; yet several widely used datasets skew toward particular dialect clusters or platform norms, such as Egyptian or Gulf Twitter, pan-Arab YouTube comments, or Instagram captions from specific demographic groups (Haidar et al., 2019). Studies that publish clear guidelines, show label distributions, and report inter-annotator agreement support more accountable modeling than those that provide only aggregate scores (Albayari and Abdallah, 2022). Curators should also protect annotator wellbeing through workload limits, content warnings, and access to support, and they should state these safeguards in their documentation. The evaluation protocol has ethical implications because metric choice shapes decision thresholds used in practice. Practical architectures therefore favor lightweight normalizers and dialect-aware tokenization before model inference, with privacy-preserving logging that stores only hashed text fingerprints or short-lived embeddings for auditing (Alakrot et al., 2018). The more explicit dataset papers are about these elements, the less likely it is that downstream models will inadvertently encode representational harms or privacy leakage.

Integrating psychological and social analysis within detection algorithms is emerging as an essential direction. Personality-based approaches could be particularly useful, helping identify users more likely to engage in or be affected by cyberbullying (e.g., Elzayady et al., 2023).

Additionally, cross-disciplinary research involving psychology, sociology, and computational linguistics could establish standards for understanding the social dynamics underlying cyberbullying, offering insights beyond linguistic patterns (e.g., Omar et al., 2021). Table 4 shows the summary of the themes related to each research question.

The results of the research emphasize the necessity of culturally sensitive detection models, sophisticated methodologies, and tailored approaches to effectively capture the distinctive characteristics of the Arabic offensive language. Arabic is an extremely diverse language, with significant variations in dialects across regions (e.g., Egyptian, Gulf, Levantine), each with its own vocabulary, syntax, and expressions. The detection of objectionable language is further complicated by this diversity, as models that have been trained in Modern Standard Arabic frequently encounter difficulties with dialectal content. These results suggest that the model’s ability to identify nuanced or implicit forms of offensive language, such as sarcasm or mockery, is improved by the inclusion of sentiment and lexicon-based features that are specifically designed for Arabic dialects and slang. Many categories of offensive language, including religious hate speech, ethnic hate, and political offence, have been classified by researchers. These types of language are particularly sensitive in Arabic-speaking societies. These categories are indicative of regional and cultural priorities, emphasizing the social and religious values that influence online discourse in Arabic contexts. The importance of accounting for these categories is underscored by research, as they pertain to highly sensitive subjects that may vary in severity and context in comparison to other languages. The results indicate that culturally aware models that identify these particular forms of objectionable language can improve the accuracy and relevance of the models.

Although numerous studies have examined cyberbullying detection methods broadly or across various languages, there is a paucity of focused analyses on Arabic-language detection, given the unique challenges presented by Arabic’s morphological intricacies and dialectal diversity (Mubarak and Darwish, 2019; AbdelHamid et al., 2022). The majority of the earlier studies predominantly analyze general patterns in cyberbullying detection, concentrating on English-language research (Alakrot et al., 2018; Bashir and Bouguessa, 2021). Although current studies recognize dataset imbalances and biases in social media-derived training data, they frequently neglect to consider privacy concerns and the ethical ramifications of automated cyberbullying detection among Arabic-speaking groups (Omar et al., 2021; Amer Hamzah and Dhannoon, 2023). This study addresses real-time detection concerns, the balance between moderation and free speech, and the necessity for privacy-preserving machine learning algorithms in social media monitoring (Kanan et al., 2021). This paper distinctly focuses on the thorough assessment of ML and DL models in detecting cyberbullying in Arabic. The prior systematic literature review by Castaño-Pulgarín et al. (2021), addressed cyberbullying detection on studies that provided exploratory data about the Internet and social media as a space for online hate speech, types of cyberhate, terrorism as an online hate trigger, online hate expressions and the most common methods to assess online hate speech. Balakrisnan and Kaity (2023) also did an SLR focusing on three main areas regarding cyberbullying detection through machine learning: the algorithms employed, the features used for detection, and the performance measures of these methods. The prior studies and reviews neglect Arabic-specific issues such as root-based word creation, tokenization complexities, and script intricacies.

The results of this study underscore the necessity of creating extensive, dialect-specific datasets and enhancing NLP models to address syntactic and lexical discrepancies among Arabic dialects. Deep learning architectures such as CNNs and BiLSTMs generally surpass classical baselines once training sets exceed the low-thousands and when preprocessed to handle orthographic variation, elongation, and code-mixing. Transformer models fine-tuned on Arabic corpora—especially variants trained with substantial dialectal coverage—consistently lead when the label definitions align with the pretraining distribution and when macro-averaged F1 rather than accuracy guides optimization. A recurring empirical pattern is precision outpacing recall, reflecting systems that confidently flag explicit bullying but struggle with implicit attacks, sarcasm, and context-dependent harassment. Performance differences are driven first by data composition. Dialectal diversity, platform genre, and class design are the most decisive factors. Models trained on tweets in Egyptian or Gulf dialects tend to degrade on Levantine, Maghrebi, or code-mixed content because lexical cues and morphological patterns shift, and subword tokenizers learned on Modern Standard Arabic under-segment dialectal forms. Domain shift between platforms—short, slang-heavy tweets versus longer Instagram captions or YouTube comments—likewise reduces transfer, as does the prevalence of emojis, creative spellings, and Arabizi. Class definitions also vary: some corpora equate cyberbullying with general abuse or toxicity, whereas others require intent, repetition, or power imbalance. The broader the “bullying” label, the higher the apparent scores, but the weaker the comparability across studies. Evaluation choices amplify these effects. Where annotation guidelines were explicit and inter-annotator agreement documented, models learned more stable decision boundaries; where guidelines were minimal or borrowed from sentiment analysis, models overfit to superficial polarity and miss community-specific bullying norms. Pretraining and representation learning explain the remaining variance. Yet, when fine-tuning data are severely imbalanced, even strong encoders prioritize surface toxicity over nuanced bullying constructs. In contrast, classical models augmented with curated lexicons and character-level features sometimes outperform deep baselines on noisy, low-resource dialects because they are less sensitive to tokenization errors and require fewer examples to generalize.

The most promising methodological direction is dialect- and domain-robust modeling anchored in standardized evaluation. Progress depends on a benchmark suite that harmonizes label schemas for cyberbullying versus general abuse, publishes class priors, and mandates macro-F1 and per-class F1 with clear treatment of the neutral class. Cross-dataset testing should be routine, with models trained on one corpus evaluated zero-shot on another to measure real-world robustness. Data and supervision strategies also offer leverage. Active learning and disagreement-focused annotation can densify minority bullying phenomena such as threats, doxxing, or body-shaming. Weak supervision that combines lexicon rules, community guidelines, and pattern matchers can cheaply label large pools for pretraining, followed by human verification on hard examples. Span-level rationales and multi-label tags for bullying types improve transparency and enable error analysis beyond single-label outcomes, while adversarial training with paraphrases and sarcasm transformations hardens models against implicit aggression. Context modeling is a further frontier. Many failures stem from sentence-level isolation. Incorporating conversation threads, author–target history, and lightweight social signals can disambiguate teasing from harassment and detect repetition, a hallmark of bullying. Graph-based representations of interactions, when coupled with privacy-preserving design and strict ethical safeguards, can capture power asymmetries and coordinated attacks without storing sensitive personal attributes.

Finally, instruction-tuned large language models adapted to Arabic show potential as few-shot labelers, error analyzers, and data generators, but their deployment must be paired with rigorous calibration, bias auditing across dialects and demographics, and conservative thresholding in safety-critical pipelines. Taken together, the evidence suggests that the field is moving from accuracy on single, homogeneous datasets toward robust, dialect-inclusive systems evaluated under standardized, recall-sensitive protocols, with the integration of context and improved supervision likely to yield the next substantive gains.