A review of knowledge graph construction using large language models in transportation: Problems, methods, and challenges

In an era of unprecedented data availability and increasingly complex transportation systems, there is a pressing need for computational paradigms that can unify cross-disciplinary knowledge and systematically deduce new hypotheses. Knowledge graphs (KGs) provides a powerful approach in organizing and connecting fragmented evidence from multiple disciplines into a single, holistic analysis framework for deep scientific discoveries. The challenge is to automate the KG construction process by integrating diverse data sources and, importantly, harmonizing fragmented, incomplete, or even contradictory evidence that arises from multiple domains. Large language models (LLMs), trained in extensive corpora from multidisciplinary data, serve as a vast knowledge repository with advanced cognitive and reasoning capabilities. LLMs lends great opportunity to automate the KGs’ construction and expansion with transdisciplinary data integration and harmonization capabilities. To facilitate the quick adoption of KGs and LLMs in transportation, this paper presents a comprehensive review of LLMs for the construction of KGs with a particular focus on methodological development, including classifications, definitions, and challenges of KG construction tasks, and methodological pipelines and techniques using LLMs for these tasks. Building on these, we propose a LLM-driven pipeline for ontological transportation KG construction harmonizing un-/structured data across disciplines and generic purpose KGs. The graph evolves iteratively through an adaptive graph-refinement process, enabling updates with new findings and data while ensuring logical consistency and theoretical coherence. The transportation ontology system provides the structural backbone for the process, ensuring knowledge alignment to maintain semantic consistency across domains. Finally, we summarize the challenges from aspects of data quality, model capability, and computational costs, and outline future research directions. The study advances the use of LLMs for KG-based knowledge representation, facilitating automated discoveries and innovations in transportation.