From Semantic Topics to Hierarchical Knowledge: An Evidence-Grounded LLM-Assisted Framework for Digital Transformation Risk Identification

Abstract

Digital transformation risk identification increasingly relies on heterogeneous textual evidence, yet topic discovery alone does not produce structured and traceable risk knowledge. This study proposes an evidence-grounded LLM-assisted semantic-to-structure framework that combines BERTopic evidence with schema-constrained hierarchical induction and software-enforced staged locking. Using 115 construction-sector publications, the framework identified eight semantic topics and produced five primary dimensions, ten secondary categories, and twenty-one tertiary indicators. A ten-seed benchmark against LDA, NMF, and Top2Vec, together with a descriptive TopicGPT-inspired local LLM baseline, showed method-specific trade-offs: LDA obtained the highest mean coherence, NMF the highest lexical diversity, and Top2Vec the most balanced topic sizes, while BERTopic provided contextual topic descriptors, 0.960 mean document coverage, and direct representative-document linkage. In a ten-seed hierarchy-generation ablation, the unlocked staged condition preserved $0.687~\pm ~0.174$ of the accepted upper structure, whereas software-enforced locking preserved $1.000~\pm ~0.000$ ; valid-run counts were 7/10, 10/10, and 9/10 for one-shot hierarchy generation without staged locking, unlocked staged generation, and locked staged generation, respectively. The results support structural controllability and evidence traceability, but not universal superiority or improved semantic quality from locking.