Social media platforms like X (formerly known as Twitter) are integral to modern communication, enabling the sharing of news, emotions, and ideas. However, they also facilitate the spread of harmful content, and manual moderation of these platforms is impractical. Automated moderation tools, predominantly developed for English, are insufficient for addressing online offensive language in Arabic, a language rich in dialects and informally used on social media. This gap underscores the need for dedicated, dialect-specific resources. This study introduces the Saudi Offensive Dialectal dataset (SOD), consisting of over 24,000 tweets annotated across three levels: offensive or non-offensive, with offensive tweets further categorized as general insults, hate speech, or sarcasm. A deeper analysis of hate speech identifies subtypes related to sports, religion, politics, race, and violence. A comprehensive descriptive analysis of the SOD is also provided to offer deeper insights into its composition. Using machine learning, traditional deep learning, and transformer-based deep learning models, particularly AraBERT, our research achieves a significant F1-Score of 87% in identifying offensive language. This score improves to 91% with data augmentation techniques addressing dataset imbalances. These results, which surpass many existing studies, demonstrate that a specialized dialectal dataset enhances detection efficacy compared to mixed-language datasets.

Offensive-language detection systems that perform well at a given point in time often degrade as linguistic patterns evolve, particularly in dialectal Arabic social media, where new terms emerge and familiar expressions shift in meaning. This study investigates temporal linguistic drift in Saudi-dialect offensive-language detection through a systematic evaluation of continual-learning approaches. Building on the Saudi Offensive Dialect (SOD) dataset, we designed test scenarios incorporating newly introduced offensive terms, context-shifting expressions, and varying proportions of historical data to assess both adaptation and knowledge retention. Eight continual-learning configurations—Experience Replay (ER), Elastic Weight Consolidation (EWC), Low-Rank Adaptation (LoRA), and their combinations—were evaluated across five test scenarios. Results show that models without continual-learning experience a 13.4-percentage-point decline in F1-macro on evolved patterns. In our experiments, Experience Replay achieved a relatively favorable balance, maintaining 0.812 F1-macro on historical data and 0.976 on contemporary patterns (KR = −0.035; AG = +0.264), though with increased memory and training time. EWC showed moderate retention (KR = −0.052) with comparable adaptation (AG = +0.255). On the SimuReal test set—designed with realistic class imbalance and only 5% drift terms—ER achieved 0.842 and EWC achieved 0.833, compared to the original model’s 0.817, representing modest improvements under realistic conditions. LoRA-based methods showed lower adaptation in our experiments, likely reflecting the specific LoRA configuration used in this study. Further investigation with alternative settings is warranted.