TLDR: This survey paper by Haonan Bian provides a comprehensive overview of how Large Language Models (LLMs) are transforming the construction of Knowledge Graphs (KGs). It details the shift from traditional rule-based methods to language-driven, generative frameworks across ontology engineering, knowledge extraction, and knowledge fusion. The paper explores schema-based and schema-free paradigms, highlighting how LLMs enhance adaptability, scalability, and semantic understanding in KG creation, and outlines future directions like KG-based reasoning for LLMs, dynamic memory for agents, and multimodal KGs.
Knowledge Graphs (KGs) have long been essential for organizing and understanding structured information, forming the backbone of many intelligent applications like search engines and question-answering systems. Traditionally, building these graphs involved complex, multi-step processes that often required significant human expertise and struggled with scalability and adaptability. However, a new era has dawned with the rise of Large Language Models (LLMs), fundamentally changing how KGs are constructed.
This comprehensive survey, authored by Haonan Bian, explores how LLMs are transforming the entire pipeline of knowledge graph construction. It highlights a shift from rigid, rule-based systems to more flexible, language-driven, and generative frameworks. The paper delves into how LLMs are reshaping ontology engineering (defining concepts and relationships), knowledge extraction (pulling information from text), and knowledge fusion (integrating diverse knowledge sources).
The Traditional Approach to Knowledge Graph Construction
Before LLMs, KGs were built through a three-layered pipeline: ontology engineering, knowledge extraction, and knowledge fusion. Ontology engineering involved experts manually defining concepts and their relationships, a process that was precise but often slow and difficult to scale. Knowledge extraction relied on handcrafted rules or statistical methods to identify entities and relations from text, often struggling with new domains or sparse data. Finally, knowledge fusion aimed to combine information from different sources, a challenging task due to semantic differences and potential conflicts.
These traditional methods faced several challenges: they were difficult to scale across different domains, heavily dependent on human experts, and prone to error propagation because each stage was handled separately. These limitations hindered the creation of KGs that could evolve and adapt dynamically.
LLMs as Game Changers
LLMs bring a transformative approach by offering generative knowledge modeling, semantic unification, and instruction-driven orchestration. This means they can directly create structured representations from unstructured text, integrate various knowledge sources through natural language understanding, and manage complex construction workflows using simple prompts. Essentially, LLMs are moving beyond simple text processing to become “cognitive engines” that bridge the gap between human language and structured knowledge.
Rethinking Ontology Construction
The paper discusses two main ways LLMs are enhancing ontology construction. The “top-down” approach uses LLMs as intelligent assistants to help human experts define formal ontologies from natural language descriptions or specific questions. For example, systems can now translate competency questions (like “What are the key concepts in this domain?”) directly into formal ontology schemas. This significantly speeds up the process and makes it more consistent.
Conversely, the “bottom-up” approach focuses on automatically creating schemas from raw data, especially useful for systems like Retrieval-Augmented Generation (RAG). Here, the KG acts as a dynamic memory for LLMs, providing factual grounding. This involves generating instance-level graphs from text and then abstracting concepts and relations through clustering and generalization, allowing schemas to adapt and evolve continuously.
Innovations in Knowledge Extraction
LLM-driven knowledge extraction has also evolved into two main paradigms: schema-based and schema-free. Schema-based methods still use an explicit knowledge schema for guidance, but now the schema can be dynamic and adaptive, rather than static. This means LLMs can use parts of an ontology relevant to a specific context, making extraction more flexible while maintaining precision.
Schema-free methods, on the other hand, aim to extract structured knowledge directly from text without any predefined ontology. LLMs are prompted to create an “on-the-fly” schema during generation, using advanced reasoning patterns. This includes techniques like Chain-of-Thought prompting and open information extraction, where LLMs discover all possible entity-relation-object triples from text, prioritizing broad coverage and discovery.
Advancements in Knowledge Fusion
Knowledge fusion, the process of integrating heterogeneous knowledge sources, is also being revolutionized by LLMs. This involves unifying the structural backbone (schema-level fusion) and aligning specific knowledge instances (instance-level fusion). LLMs are moving beyond simple matchers to become adaptive reasoning agents that can integrate contextual, structural, and retrieved information for scalable and self-correcting fusion. Hybrid frameworks are emerging that combine both schema and instance-level fusion into unified, prompt-driven workflows, leading to more autonomous and self-evolving knowledge graphs.
Also Read:
- Structured Answers: Interpretable QA Using Knowledge Graphs
- Simplifying LLM Interaction: A New Approach to Adaptive Prompt Generation
Future Directions
The survey concludes by outlining exciting future applications. KGs are expected to be further integrated into LLM reasoning mechanisms, enhancing logical consistency, causal inference, and interpretability. They are also envisioned as dynamic memory systems for LLM-powered agents, allowing for continuous learning and multi-agent collaboration. Furthermore, multimodal knowledge graph construction aims to integrate diverse data types like text, images, and audio into unified representations. Beyond their role in RAG systems, KGs are becoming a “cognitive middle layer” for LLMs, providing structured support for querying, planning, and decision-making, leading to more explainable and grounded AI systems.
This survey clarifies the evolving relationship between LLMs and knowledge graphs, bridging symbolic knowledge engineering with neural semantic understanding. It paves the way for the development of adaptive, explainable, and intelligent knowledge systems. For more details, you can read the full paper here.
