Structural health monitoring (SHM) domain continuously generates large volumes of textual data whose knowledge remains largely implicit. Most existing natural language processing (NLP) applications in this field emphasize syntactic or semantic informa...
Structural health monitoring (SHM) domain continuously generates large volumes of textual data whose knowledge remains largely implicit. Most existing natural language processing (NLP) applications in this field emphasize syntactic or semantic information extraction, providing limited support for pragmatic-level understanding. This study proposes a pragmatic NLP framework that uncovers latent domain knowledge by integrating an NER component based on a BERT–BiLSTM–CRF architecture, a domain-specific knowledge graph (KG), and a hypothesis generation (HG) module driven by contextual reasoning. A corpus of SHM-related literature was manually annotated into five entity categories and used to fine-tune the NER model, achieving 0.8998 accuracy and 0.8705 F1-score. To analyze complex knowledge within the SHM domain, a novel dataset was constructed. The extracted entities were represented as nodes, and a top-k relationship assignment strategy was used to define weighted edges, yielding a KG comprising 3,509 nodes and 5,660,655 edges that aggregate information from various structural and SHM studies. Within this KG, path-finding algorithms identified candidate hypothesis paths, and additional intermediate nodes were introduced to form richer, multi-step relational patterns that approximate scientific reasoning chains. To evaluate their plausibility, each hypothesis path was converted into a textual statement and compared with sentences in the source corpus using a Sentence-BERT model, which quantified semantic alignment. The generated hypotheses closely aligned with relevant literature, indicating that the framework can uncover meaningful, previously unexplored connections across studies and highlight underexplored combinations of mechanisms, conditions, and responses. Overall, the proposed approach demonstrates the potential of pragmatic NLP to support automated hypothesis reasoning and latent knowledge discovery in SHM research, while providing a scalable foundation for future integration with interactive decision-support and design tools.