Knowledge graph-enhanced pre-trained language model for organizational resilience assessment and dynamic capability reconstruction path mining

Abstract Organizational resilience now shapes which firms survive prolonged uncertainty, yet most assessment instruments remain fragmented across indicators and keep structured domain knowledge apart from semantic understanding. This study develops a domain-specific framework in which knowledge graph reasoning and pre-trained language model encoding are coupled through cross-attention fusion. We build an organizational resilience knowledge graph covering enterprises, capabilities, resources and risk factors, inject it into the middle transformer layers, and score absorptive, adaptive and recovery capacities separately through attention-weighted aggregation; a reinforcement learning path miner then searches the graph for dynamic capability reconstruction trajectories. Ground-truth labels for 856 listed enterprises (2018–2023) combine rule-based scoring on post-disruption recovery indicators, financial as well as non-financial, with three-rater expert annotation, for which we report an intraclass correlation coefficient of 0.81 alongside a Cohen’s kappa of 0.78. Because the labelling rubric and the feature space partly overlap, we deliberately test the framework beyond the rubric: an isolation ablation strips out rule-aligned features, and predictions are validated against outcomes the rubric never observes, including post-shock financial distress and the time needed to regain pre-shock operating levels. Data are partitioned by enterprise-year under strict temporal stratification, and the graph is built only from training-period documents. Under this protocol the framework reaches 87.92% accuracy and an F1-score of 86.51%, ahead of six baselines and of three open-source large language models tuned through an explicit LoRA hyperparameter search; path mining attains 56.82% Hit@1, 77.24% Hit@5 and a mean reciprocal rank of 0.672. Results hold across alternative normalisation, weighting and discretisation schemes, and degrade predictably as graph noise or edge sparsity increases. Roughly two fifths of top-ranked paths remain incorrect, most often through misjudged resource constraints, so we position the recommendations as decision aids rather than prescriptions. The study contributes to computational resilience assessment while mapping, rather than concealing, the boundaries within which its outputs can be trusted.