Wolf-pack intelligence with multi-leadership Wolf search hybrid model for text-based deception-risk assessment in executive background checks

Text-based deception-risk assessment in executive background checks is a challenging task because potentially misleading statements must be identified from linguistically complex and highly imbalanced data. This paper proposes a Multi-Leadership Wolf Search Algorithm (ML-WSA)-based framework for automated deception-risk assessment and statement prioritization. The core contribution lies in a dynamic multi-leadership search mechanism that enables parallel exploration of multiple regions in high-dimensional feature space, effectively alleviating premature convergence in traditional Grey Wolf Optimization. ML-WSA integrates a nonlinear convergence strategy and Cauchy mutation to balance global exploration and local exploitation. To capture deceptive linguistic cues, a dual-channel feature representation combining BERT-based semantic embeddings and linguistic statistical features is constructed, and a semantic-constrained oversampling method is introduced to address minority-class scarcity. Experimental results show that the complete ML-WSA-based framework achieves an F1 score of 0.8733 (95% CI: 0.861–0.886) and an AUC of 0.9012, significantly outperforming standard GWO and BERT fine-tuning baselines (adjusted p < 0.001). Cross-dataset evaluations on three heterogeneous public deception benchmarks provide evidence of robustness under domain shift, with a macro-average F1 of 0.8067. These results support the potential of the integrated framework for human-supervised text-based deception-risk screening under high-stakes conditions.