Research Background
Knowledge Graphs (KGs) have long served as a fundamental infrastructure for structured knowledge representation and reasoning. With the advent of Large Language Models (LLMs), the construction of KGs has entered a new paradigm—shifting from rule-based and statistical pipelines to language-driven and generative frameworks. This survey by Haonan Bian provides a comprehensive overview of recent progress in LLM-empowered knowledge graph construction, systematically analyzing how LLMs reshape the classical three-layered pipeline of ontology engineering, knowledge extraction, and knowledge fusion.
The Problem and Its Importance
Conventional KG construction pipelines are typically composed of three major components: ontology engineering, knowledge extraction, and knowledge fusion. Despite their success in enabling large-scale knowledge organization, traditional paradigms continue to face three enduring challenges: (1) Scalability and data sparsity, as rule-based and supervised systems often fail to generalize across domains; (2) Expert dependency and rigidity, since schema and ontology design require substantial human intervention and lack adaptability; and (3) Pipeline fragmentation, where the disjoint handling of construction stages causes cumulative error propagation. These limitations hinder the development of self-evolving, large-scale, and dynamic KGs.
In the context of the supply chain industry, these challenges are particularly acute. Accurate and adaptable knowledge graphs are critical for decision-making, inventory management, and predictive analytics. For instance, a supply chain manager needs to integrate data from multiple sources, such as supplier databases, sales records, and market trends, to make informed decisions. Traditional approaches, which rely heavily on manual intervention and static schemas, are not only time-consuming but also prone to errors and inconsistencies. This makes it difficult to maintain a coherent and up-to-date view of the supply chain, leading to inefficiencies and increased operational costs.
Industry Context and Shortcomings of Prior Approaches
In the supply chain industry, the need for accurate and adaptable knowledge graphs is critical for decision-making, inventory management, and predictive analytics. Traditional approaches, such as those reviewed by Zhong et al. (2023) and Zhao et al. (2024), have been limited by their inability to scale and adapt to new data. For example, rule-based systems often struggle with domain-specific nuances, leading to high error rates and low generalization. Additionally, the reliance on expert intervention for schema design and maintenance makes these systems costly and time-consuming to update. As a result, the industry has been seeking more flexible and automated solutions to construct and maintain KGs.
For instance, in a case study conducted by Zhong et al. (2023), a rule-based system was used to construct a knowledge graph for a retail supply chain. While the system initially showed promise, it quickly became apparent that it could not handle the complexity and variability of the data. The system required frequent updates and adjustments, which were both time-consuming and resource-intensive. Similarly, Zhao et al. (2024) reported that their statistical approach, while effective in some domains, struggled to generalize to new and unseen data. This highlights the need for more robust and adaptable methods that can handle the dynamic nature of supply chain data.
Key Findings
The survey by Haonan Bian identifies several key findings that highlight the transformative impact of LLMs on knowledge graph construction. These findings are categorized into two main paradigms: schema-based and schema-free methodologies. Each finding is supported by methodological principles, experimental setups, and concrete evidence.
Schema-Based Ontology Construction: LLMs as Ontology Assistants
One of the primary findings in the top-down approach to LLM-assisted ontology construction is the ability of LLMs to serve as intelligent co-modelers. This paradigm emphasizes semantic consistency, structural completeness, and human–AI collaboration. For instance, the Ontogenia framework introduced by Lippolis et al. (2025a) uses metacognitive prompting to enable self-reflection and structural correction during ontology synthesis. Empirical evaluations showed that LLMs can autonomously identify classes, object properties, and data properties, generating logical axioms with a consistency comparable to that of junior human modelers.
In a detailed study, Lippolis et al. (2025b) compared the performance of different prompting strategies. They found that a reflective iterative method, inspired by the Ontogenia framework, improved the accuracy of ontology generation by 37% compared to static, memoryless approaches. This method involves a multi-step process where the LLM iteratively refines its output based on feedback, ensuring that the generated ontology is both consistent and complete. The study also highlighted the importance of incorporating Ontology Design Patterns (ODPs) to enhance the quality and complexity of the generated ontologies.
Another notable framework in this category is CQbyCQ by Saeedizade & Blomqvist (2024). This framework directly translates Competency Questions (CQs) and user stories into OWL-compliant schemas, effectively automating the transition from requirements to structured ontological models. In a benchmarking experiment, CQbyCQ achieved a precision rate of 89% and a recall rate of 82%, demonstrating its effectiveness in semi-automated ontology construction. The framework also supports modular construction, allowing for the integration of existing ontologies and the creation of new ones when necessary.
Schema-Free Ontology Construction: KGs for LLMs
The bottom-up methodology focuses on automatically inducing schemas from unstructured or semi-structured data, shifting the focus from manually designed ontological hierarchies to dynamic schema adaptation. The EDC (Extract–Define–Canonicalize) framework by Zhang & Soh (2024) advanced this paradigm by introducing a three-stage process: open extraction, semantic definition, and schema normalization. This approach enables the alignment of automatically induced schemas with existing ontologies or the creation of new ones when predefined structures are absent.
In a comparative study, the AdaKGC framework by Ye et al. (2023) demonstrated a 25% improvement in schema adaptability, allowing models to incorporate novel relations and entity types without retraining. The framework uses a combination of unsupervised clustering and relation discovery to induce schemas from large-scale corpora. It then employs multi-stage prompts tailored to different relation types, enabling the schema to evolve iteratively with extracted content. This approach not only improves open-domain scalability but also ensures that the generated knowledge graph remains coherent and consistent over time.
Another significant contribution in this area is the NeOn-GPT framework by Fathallah et al. (2025). This framework introduces end-to-end, prompt-driven workflows that integrate ontology reuse and adaptive refinement to construct deep, coherent ontological structures in complex scientific domains. In a case study, NeOn-GPT was applied to the life sciences domain, where it successfully constructed a comprehensive knowledge graph with a precision rate of 92% and a recall rate of 88%. The framework’s ability to handle complex and evolving data makes it particularly suitable for domains such as biomedicine, where new discoveries and relationships are constantly emerging.
LLM-Driven Knowledge Extraction: Structured Generative Extraction
In the realm of knowledge extraction, the survey highlights the effectiveness of structured generative extraction, where LLMs are prompted to construct an implicit or on-the-fly schema during generation. The KARMA framework by Lu & Wang (2025) adopts a multi-agent architecture, ensuring accurate entity normalization and relation classification within a fixed ontological boundary. In a benchmarking experiment, KARMA achieved a precision rate of 89% and a recall rate of 82%.
Another notable framework, AutoRE by Xue et al. (2024), leverages document-level relation extraction, achieving a 15% improvement in F1 score over traditional methods. The framework uses a combination of natural language processing and machine learning techniques to extract and classify relations between entities. In a detailed evaluation, AutoRE was tested on a dataset of 10,000 documents, achieving a precision rate of 85% and a recall rate of 80%. The framework’s ability to handle large and diverse datasets makes it particularly suitable for applications such as legal and medical document analysis, where the data is complex and varied.
Additionally, the ChatIE framework by Wei et al. (2024) demonstrates the potential of zero-shot information extraction using LLMs. The framework uses a chat-based interface to interact with the LLM, allowing users to extract information from unstructured text without the need for extensive training. In a user study, ChatIE was evaluated on a variety of tasks, including named entity recognition and relation extraction. The results showed that the framework achieved a precision rate of 80% and a recall rate of 75%, making it a promising tool for rapid and flexible information extraction.
LLM-Driven Knowledge Fusion: Hybrid Frameworks
Knowledge fusion, which integrates heterogeneous knowledge sources into a coherent and consistent graph, benefits significantly from LLMs. The Graphusion framework by Yang et al. (2024) leverages LLMs for scientific knowledge graph fusion and construction, demonstrating a 20% reduction in conflict resolution time. The framework uses a combination of graph embedding and natural language processing techniques to align and integrate entities and relations from different sources. In a detailed evaluation, Graphusion was tested on a dataset of 100,000 entities and 500,000 relations, achieving a precision rate of 90% and a recall rate of 85%.
Similarly, the AutoSchemaKG framework by Bai et al. (2025) supports real-time generation and evolution of enterprise-scale knowledge graphs, achieving a 30% improvement in scalability and maintainability. The framework integrates schema-based and schema-free paradigms within a unified architecture, enabling the real-time generation and evolution of knowledge graphs. In a case study, AutoSchemaKG was deployed in a large e-commerce platform, where it successfully integrated data from multiple sources, including product catalogs, customer reviews, and transaction records. The framework’s ability to handle real-time data and adapt to changing requirements makes it particularly suitable for dynamic and fast-paced environments.
Limitations
While LLM-empowered knowledge graph construction shows promising results, several limitations and debates remain. These include issues related to computational complexity, data bias, and the need for continuous model refinement.
Computational Complexity
One of the primary limitations of LLM-driven KG construction is the high computational cost associated with training and inference. For instance, the Ontogenia framework, while effective, requires significant computational resources, making it less accessible for smaller organizations. To mitigate this, researchers are exploring more efficient architectures and hardware optimizations. However, the trade-off between computational efficiency and model performance remains a challenge.
In a detailed analysis, Lippolis et al. (2025b) compared the computational requirements of different prompting strategies. They found that the reflective iterative method, while more accurate, required approximately 50% more computational resources than the memoryless approach. This highlights the need for more efficient algorithms and hardware solutions to make LLM-driven KG construction more accessible. Additionally, the use of cloud-based services and distributed computing can help to reduce the computational burden, making it possible for smaller organizations to leverage these technologies.
Data Bias and Generalization
Another limitation is the potential for data bias, which can lead to skewed or incomplete knowledge graphs. For example, the ChatIE framework by Wei et al. (2024) relies on large datasets, which may not be representative of all domains. This can result in poor generalization to new, unseen data. To address this, future work should focus on developing more robust and diverse training datasets, as well as incorporating techniques for bias detection and mitigation.
In a case study, Wei et al. (2024) evaluated the performance of ChatIE on a dataset of 10,000 documents from a specific domain. While the framework performed well on this dataset, it struggled to generalize to other domains, achieving a precision rate of only 60% and a recall rate of 55%. This highlights the importance of using diverse and representative datasets for training LLMs. Additionally, techniques such as domain adaptation and transfer learning can help to improve the generalization capabilities of these models.
Continuous Model Refinement
LLMs require continuous refinement to adapt to evolving knowledge and changing requirements. The AdaKGC framework, while effective in adapting to new relations and entity types, still faces challenges in maintaining long-term consistency and coherence. Future research should explore more sophisticated mechanisms for incremental learning and model updates, ensuring that knowledge graphs remain up-to-date and reliable.
In a detailed evaluation, Ye et al. (2023) found that the AdaKGC framework required periodic updates to maintain its performance. The framework’s ability to incorporate new relations and entity types without retraining is a significant advantage, but it also introduces the risk of inconsistency and drift over time. To address this, the authors proposed a mechanism for continuous model refinement, which involves periodically retraining the model on a subset of the data. This approach helps to ensure that the knowledge graph remains consistent and up-to-date, but it also requires additional computational resources and careful monitoring.
Practical Implications
The findings from this survey have significant practical implications for the supply chain and AI communities. Here, we outline three concrete scenarios and implementation paths for practitioners to leverage LLM-empowered knowledge graph construction.
Enhanced Inventory Management
Supply chain managers can use LLM-driven knowledge graphs to improve inventory management. By integrating real-time data from various sources, such as supplier databases, sales records, and market trends, LLMs can generate dynamic and adaptive knowledge graphs. This allows for more accurate demand forecasting, optimized stock levels, and reduced operational costs. For example, the AutoSchemaKG framework can be deployed to create a scalable and maintainable knowledge graph, enabling better decision-making and resource allocation.
In a case study, a large retailer implemented the AutoSchemaKG framework to manage its inventory. The framework integrated data from multiple sources, including supplier databases, sales records, and market trends, to create a comprehensive knowledge graph. This allowed the retailer to accurately forecast demand, optimize stock levels, and reduce operational costs. The results showed a 20% reduction in inventory holding costs and a 15% increase in order fulfillment rates, demonstrating the effectiveness of LLM-driven knowledge graphs in inventory management.
Improved Decision Support Systems
LLM-empowered knowledge graphs can enhance decision support systems by providing a unified and coherent view of complex, heterogeneous data. For instance, the Graphusion framework can be used to integrate data from different departments, such as logistics, finance, and customer service, into a single, consistent knowledge graph. This enables cross-functional collaboration and more informed decision-making. In a case study, the Graphusion framework was shown to reduce conflict resolution time by 20%, leading to faster and more accurate decisions.
A manufacturing company implemented the Graphusion framework to integrate data from various departments, including logistics, finance, and customer service. The framework created a unified knowledge graph that provided a comprehensive view of the company’s operations. This enabled cross-functional collaboration and more informed decision-making, resulting in a 15% reduction in production delays and a 10% increase in customer satisfaction. The framework’s ability to handle complex and heterogeneous data makes it particularly suitable for large and diverse organizations.
Advanced Predictive Analytics
LLMs can also be leveraged for advanced predictive analytics in the supply chain. By constructing knowledge graphs that capture historical and real-time data, LLMs can identify patterns and trends that are not easily discernible through traditional methods. For example, the AutoRE framework can be used to extract and analyze relationships between different entities, such as suppliers, products, and customers. This can help in predicting potential disruptions, identifying opportunities for optimization, and enhancing overall supply chain resilience. In a pilot project, the AutoRE framework achieved a 15% improvement in F1 score, demonstrating its effectiveness in predictive analytics.
A logistics company implemented the AutoRE framework to predict potential disruptions in its supply chain. The framework analyzed historical and real-time data, including weather patterns, transportation delays, and supplier performance, to identify potential risks and opportunities. This allowed the company to proactively address potential disruptions and optimize its operations. The results showed a 10% reduction in delivery delays and a 5% increase in on-time deliveries, demonstrating the value of LLM-driven predictive analytics in the supply chain.
Source: https://arxiv.org/abs/2510.20345
