Intelligent fusion of heterogeneous data for armed group recognition and attack attribution using the Global Terrorism Database

Abstract Attributing terrorism-related incidents to responsible armed groups remains challenging because incident narratives contain ambiguous descriptions, inconsistent organization naming, heterogeneous contextual information, and substantial class imbalance. This study proposes a unified transformer-based framework for two complementary tasks: armed-group named entity recognition (NER) and attack attribution using the Global Terrorism Database (GTD). A multi-modal preprocessing pipeline transforms 209,706 incidents with 135 heterogeneous variables into an 80-feature analysis-ready representation and supports structured-information supplementation when relevant incident information is absent from the narrative. This supplementation provides limited benefit for information-rich narratives but improves attribution accuracy by up to 13.43 percentage points under severe narrative sparsity, demonstrating its value as an information-completion mechanism. For armed-group extraction, a RoBERTa–BiLSTM–CRF architecture augmented with gazetteer features is developed and evaluated on a silver-standard NER corpus of 67,386 incidents, achieving 99.23% F1. For attack attribution, BERT, RoBERTa, DistilBERT, and ConfliBERT are evaluated across 10-, 20-, 50-, and 100-group classification settings, achieving up to 99.73% accuracy at 10 groups and 98.52% at 100 groups. Fine-tuned RoBERTa also consistently exceeds the best observed 7B in-context-learning configurations on matched evaluation subsets, with accuracy differences of 7.56–12.00 percentage points across the classification scales and a 27.41-point entity-level F1 difference in NER. The fine-tuned encoder models further exhibit substantially lower inference latency under the evaluated hardware configuration. Additional robustness analyses show that attribution accuracy remains 93.08% under closed-set temporal evaluation. The entity-disjoint NER evaluation further demonstrates that the models retain transferable recognition capability even when target organizations are excluded from task-specific supervised training, while also highlighting the greater difficulty of this stricter generalization setting. Overall, the results show that task-specific transformer models provide strong predictive performance, computational efficiency, and robust generalization across multiple evaluation conditions for terrorism-related information extraction and attribution.