A Provenance-Aware Seven-Class Dataset and Attention-Enhanced Distilbert–Bilstm–Bigru Framework For Violence Detection in Digital Text

Abstract

In the history of democratic governance, the year 2024 was unprecedented as in this year more than sixty naAutomatic identification of violence-related language in digital text remains challenging because harmful content may appear as explicit threats, harassment, coercive behaviour, personal disclosures, identity-directed hostility, and context-dependent narratives. Existing harmful-language datasets frequently employ heterogeneous labels such as hate speech, toxicity, abuse, offensive language, cyberbullying, and domestic-violence relevance, making direct dataset integration semantically unreliable. This study presents a provenance-aware dataset-construction framework together with an attention-enhanced DistilBERT–BiLSTM–BiGRU architecture for seven-class violence detection: physical violence, sexual violence, emotional violence, economic violence, harmful practices, hate speech, and non-violence. Source-supported labels were harmonised while preserving dataset provenance, mapping decisions, and duplicate-conflict controls. A rapid experimental benchmark comprising 1,13,040 unique records was used for model evaluation. The proposed DistilBERT–BiLSTM–BiGRU–Attention configuration achieved an accuracy of 0.9315, balanced accuracy of 0.9527, macro F1-score of 0.9231, weighted F1-score of 0.9317, MCC of 0.9149, and Cohen’s kappa of 0.9145 in the ablation experiment. Ablation analysis demonstrated progressive improvement from DistilBERT alone (macro F1 = 0.8638) to DistilBERT–BiLSTM (0.8931) and DistilBERT–BiLSTM–BiGRU (0.9113), with attention producing the strongest hybrid performance. The TF-IDF + Linear SVM baseline achieved an accuracy of 0.9477 and macro F1-score of 0.9599, outperforming the rapid frozen-transformer configuration on aggregate classification metrics. These findings demonstrate both the importance of provenance-aware label harmonisation and the contribution of recurrent and attention mechanisms, while also highlighting the competitiveness of conventional sparse-text baselines for structured violence-language classification.Keywords—violence detection; digital text; BERT; BiLSTM; BiGRU; attention; provenance; gender-based violence; source-supported labels; harmful language; multi-class classification; dataset harmonisation; explainable NLP.