With business globalization and increasing product complexity, companies often operate supply chains distributed around the world. Such globally distributed supply chains face various disruption risks, highlighting the need for procurement oficers to efectively monitor these risks. A popular paradigm is to apply information extraction technologies to open data for risk extraction. Here, we introduce an ontology-guided method for supply chain risk extraction that leverages large language models. Our method iteratively extracts a supply chain risk knowledge graph from unstructured open data, guided by a user-specified ontology. We also developed knowledge graph verification and formatting modules. Our wholistic methods enable consistent and automated identification and extraction of risk knowledge, thereby empowering procurement oficers to monitor supply chain risks.
Business background: Modern enterprise supply chains are distributed around the world. With
the broad geographical distribution, supply chains are facing various disruption risks, ranging from
natural hazards to regional conflicts [
First, previous ontology-based semantic approaches did not efectively incorporate ontology within
the risk extraction process. While these methods rely heavily on experts to build the ontology, the
resulting ontology serves as a static metadata of data schema [
Second, recent AI-based models face challenges related to low explainability, limited interpretability,
and poor adaptability to dynamic environments [
We propose an ontology-guided, iterative method that uses Large Language Models (LLMs) to extract a
supply chain risk knowledge graph (KG), addressing the limitations of previous approaches. Compared
with similar methodologies [
The process consists of three main stages: 1. Ontology Specification. An ontology should be specified to guide the LLM in the knowledge graph extraction process. While we provide templates (e.g., NaturalHazard ontology), users can create custom ontology like MilitaryConflict (e.g., involving defined labels like ConflictEvent, ConflictSideA, ConflictSideB) for monitoring conflict risks. 2. Iterative Knowledge Extraction. The extraction begins when source documents (e.g., news) are collected. Initially, the LLM is provided with a system prompt containing the complete ontology and instructions for KG extraction. Then, the extraction process iterates through each relation triple defined in the ontology (e.g., <ConflictSideA>–involvesIn–<ConflictEvent>). In each iteration, the LLM is prompted with the target relation triple, the source document, and the KG extracted so far. The LLM extracts new KG nodes and relationships in the text. An illustration is given in Figure 1. This iterative feedback loop creates an explicit chain-of-thought process, progressively assembling a comprehensive and context-aware KG. 3. Verification and Formatting. Here, we verify and format extracted nodes and relationships.
First, a dedicated judge LLM verifies nodes and relationships against the source text. The judge LLM returns a confidence score reflecting how faithfully the relation is supported by the text. Nodes and relationships with scores below a threshold are discarded, while higher-scoring ones are kept. Second, another LLM formats key entities to a standard format (e.g., dates to ISO 8601) to enhance consistency across the final KG.
The adaptability of our ontology-guided method enables diverse applications. By allowing users to specify the ontology, the resulting KG can be tailored to unique monitoring needs. We demonstrate this with the following applications: 1. Versatile Risk Monitoring: Our method allows for supply chain risk monitoring adaptable to various risk domains by changing the guiding ontology.
• Natural Hazards: Guided by a NaturalHazard ontology (see Figure 2), the LLM extracts natural hazard risk events from open data. The resulting KG, containing location information, can be linked to supplier locations, alerting POs to potential disruptions. For explainability, source URLs are included as metadata, enabling information traceability.
• Conflicts: To monitor conflict risks, users can design a Conflict ontology (e.g., with deifned labels like ConflictEvent, ConflictSideA, ConflictSideB), guiding the LLM to extract relevant knowledge from news data. 2. Impact Annotations: Our method supports analyzing the potential impact of future risks using historical data. For instance, to help domestic procurement teams contextualize overseas events (e.g., a cyclone in India), a NaturalHazardSupplyChainImpact ontology (see Figure 2) can guide the LLM to extract the semantic details of how past cyclones have afected supply chains in that region. This structured historical knowledge enables the annotation of future risk alerts with meaningful impact predictions for better-informed decision-making.
To validate our method, we introduce a brief case study of Impact Annotations. We focused on
cyclone hazards in the Indian Ocean region. We collected over 7,000 news articles via an open data API
and applied the NaturalHazardSupplyChainImpact ontology to guide an LLM in extracting a KG. The
process yielded a KG with 30,881 entities and 51,893 relationships. From this graph, we generated 9,592
specific impact annotations through Cypher queries [
We proposed an LLM-based, ontology-guided method for supply chain risk monitoring. Our ontologyguided method provides semantically rich knowledge graph extraction and achieves adaptability to diferent risk scenarios.
Future works include modeling the severity of events to make the KG not only semantically rich but also quantitatively informative. Enhancing data orchestration with active metadata is another direction. To fully realize the potential of the proposed method, developing a data platform capable of supporting active ontology metadata is necessary.
We would like to thank Qi Xiu (Hitachi, Ltd.) for providing mentorship to Shuhan during his first year of research. We also thank Tsunehiko Baba (Hitachi, Ltd.) and Mika Takata (Hitachi, Ltd.) for their supports during this research.
During the preparation of this work, the author(s) used ChatGPT in order to: Grammar and spelling check. The author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.