SD-CVD corpus: Towards robust detection of fine-grained cyber-violence across Saudi dialects in online platforms
Abstract
This paper introduces SD-CVD, a large, balanced Saudi-dialect corpus for fine-grained detection of cyber-violence on social media. The dataset comprises 88,687 tweets spanning Najdi, Hijazi, Eastern, and Qassimi dialects and is annotated hierarchically across three levels: (i) benign vs. cyber-violent, (ii) cyberbullying vs. online hate speech (with benign split into positive/negative/neutral), and (iii) seven hate-speech subtypes (incitement to violence, gender, national, social-class, tribal, religious, and regional discrimination). Annotation followed detailed guidelines with double review and adjudication, yielding Fleiss’ κ > 0.89. We benchmark classical machine learning (TF–IDF n-grams) and deep learning (embedding-based) models. SVM achieved the best overall performance (Accuracy = 0.854, F1 = 0.853), followed closely by Logistic Regression (F1 = 0.849), while MLP and LSTM were the strongest deep models (F1 ≈ 0.83); CNN underperformed on context-dependent categories. Error analysis shows most confusions within benign subclasses (neutral vs. negative), whereas hate-subtypes such as regional, tribal, and social-class discrimination are reliably identified. The results confirm that carefully engineered lexical features remain highly competitive for Arabic dialectal text, and they establish robust baselines for future hybrid and transformer-based approaches. SD-CVD directly supports national digital-safety efforts by enabling more precise detection and monitoring of cyber-violence in Saudi online discourse.