LLM-Empowered Knowledge Graph Construction: Paradigms, Techniques, and Future Outlook

摘要

知识图谱A structured knowledge base that represents entities and their relationships in a graph format.长期以来作为结构化知识表示与推理的基础设施。随着大语言模型Advanced AI models trained on massive text data to understand and generate human language across multiple tasks.的出现，知识图谱A structured knowledge base that represents entities and their relationships in a graph format.的构建进入了一个新的范式——从基于规则和统计的流水线转向语言驱动和生成的框架。本综述全面概述了LLM赋能的图谱构建的最新进展，系统分析了LLM如何重塑本体工程知识图谱构建的第一层流程，涉及领域概念、属性和关系的定义与组织，建立知识图谱的语义框架。、知识抽取从非结构化或半结构化数据中自动识别和提取实体、关系等知识元素的过程。和知识融合将来自不同来源的知识进行整合、消歧和统一，形成一致、完整的知识图谱。这三个经典层次。我们首先回顾传统KG方法以建立概念基础，然后从两个互补的视角审视新兴的LLM驱动方法：基于模式的范式强调结构、规范化和一致性的知识图谱构建方法，通常依赖于预定义的本体或模式。，强调结构、规范化和一致性；以及无模式范式，强调灵活性、适应性和开放发现。在每个阶段，我们综合了代表性框架，分析了其技术机制，并指出了其局限性。最后，本文概述了关键趋势和未来研究方向，包括基于KG的LLM推理、面向智能体系统的动态知识记忆以及多模态KG构建。通过这项系统性综述，我们旨在阐明LLMs与知识图谱A structured knowledge base that represents entities and their relationships in a graph format.之间不断演进的相互作用，弥合符号知识工程与神经语义理解之间的鸿沟，以推动自适应、可解释和智能知识系统的发展。

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 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. We first revisit traditional KG methodologies to establish conceptual foundations, and then review emerging LLM-driven approaches from two complementary perspectives: schema-based paradigms, which emphasize structure, normalization, and consistency; and schema-free paradigms, which highlight flexibility, adaptability, and open discovery. Across each stage, we synthesize representative frameworks, analyze their technical mechanisms, and identify their limitations. Finally, the survey outlines key trends and future research directions, including KG-based reasoning for LLMs, dynamic knowledge memory for agentic systems, and multimodal KG construction. Through this systematic review, we aim to clarify the evolving interplay between LLMs and knowledge graphs, bridging symbolic knowledge engineering and neural semantic understanding toward the development of adaptive, explainable, and intelligent knowledge systems.

引言：从符号到神经的范式迁移

Introduction: The Paradigm Shift from Symbolic to Neural

知识图谱A structured knowledge base that represents entities and their relationships in a graph format.作为结构化的语义网络，通过实体、关系及其属性来描述现实世界中的概念及其联系，已成为人工智能领域，特别是语义搜索、智能问答和推荐系统中的核心组件。传统的KG构建严重依赖专家定义的本体（模式）和复杂的自然语言处理流水线，包括命名实体识别、关系抽取和实体链接等步骤。这种方法虽然能确保高质量和一致性，但存在成本高昂、扩展性差、难以适应新领域和动态知识等固有局限。

Knowledge Graphs, as structured semantic networks that describe real-world concepts and their connections through entities, relationships, and attributes, have become a core component in the field of artificial intelligence, particularly in semantic search, intelligent question answering, and recommendation systems. Traditional KG construction heavily relies on expert-defined ontologies (schemas) and complex natural language processing pipelines, including steps such as Named Entity Recognition, Relation Extraction, and Entity Linking. While this approach ensures high quality and consistency, it suffers from inherent limitations such as high cost, poor scalability, and difficulty in adapting to new domains and dynamic knowledge.

大语言模型Advanced AI models trained on massive text data to understand and generate human language across multiple tasks.的崛起为这一领域带来了革命性变化。LLMs在预训练过程中吸收了海量文本中的世界知识，展现出强大的语言理解、生成和上下文推理能力。这为KG构建提供了新的可能性：LLMs可以作为“通用知识处理器”，直接理解非结构化文本，并生成结构化的知识表示，从而简化甚至重构整个构建流程。这种融合标志着从符号驱动的精确工程向神经驱动的语义理解与生成的范式迁移。

The rise of Large Language Models has brought revolutionary changes to this field. LLMs, having absorbed vast amounts of world knowledge from text during pre-training, demonstrate powerful capabilities in language understanding, generation, and contextual reasoning. This opens up new possibilities for KG construction: LLMs can serve as "universal knowledge processors," directly understanding unstructured text and generating structured knowledge representations, thereby simplifying or even reconstructing the entire construction pipeline. This integration signifies a paradigm shift from symbol-driven precise engineering to neural-driven semantic understanding and generation.

核心构建范式：基于模式 vs. 无模式

Core Construction Paradigms: Schema-Based vs. Schema-Free

当前，LLM赋能的KG构建主要沿着两条技术路径演进，它们代表了在结构化控制与开放灵活性之间的不同权衡。

Currently, LLM-empowered KG construction primarily evolves along two technical paths, representing different trade-offs between structured control and open flexibility.

基于模式的范式强调结构、规范化和一致性的知识图谱构建方法，通常依赖于预定义的本体或模式。

Schema-Based Paradigm

该范式继承并增强了传统KG构建的理念，其核心是预定义的本体或模式。LLM在此框架中扮演“智能执行者”的角色，其任务是在给定模式的严格约束下，从文本中抽取或生成符合规范的知识三元组。这种方法确保了生成图谱的规范性、一致性和高质量，易于与下游的符号推理系统集成。

This paradigm inherits and enhances the philosophy of traditional KG construction, with its core being a predefined ontology or schema. Within this framework, the LLM acts as an "intelligent executor," tasked with extracting or generating knowledge triples that conform to the given schema's strict constraints from text. This approach ensures the standardization, consistency, and high quality of the generated graph, facilitating integration with downstream symbolic reasoning systems.

典型技术方法包括：

• 指令微调与约束生成：使用包含模式定义的指令对LLM进行微调，或通过解码阶段的约束（如前缀树）确保输出符合预定义的实体/关系类型。
• 思维链与分步推理：引导LLM通过“先识别实体，再判断关系”等分步推理过程，提升复杂语境下的抽取准确性。
• 检索增强生成：将外部知识库或本体作为检索源，为LLM提供上下文，辅助其做出更准确的模式对齐决策。

Typical technical approaches include:

• Instruction Fine-tuning and Constrained Generation: Fine-tuning LLMs with instructions containing schema definitions, or using constraints during decoding (such as prefix trees) to ensure outputs conform to predefined entity/relationship types.
• Chain-of-Thought and Step-by-Step Reasoning: Guiding LLMs through step-by-step reasoning processes like "first identify entities, then determine relationships" to improve extraction accuracy in complex contexts.
• Retrieval-Augmented Generation: Using external knowledge bases or ontologies as retrieval sources to provide context for the LLM, assisting it in making more accurate schema alignment decisions.

无模式的范式强调灵活性、适应性和开放发现的知识图谱构建方法，不依赖于固定的模式或本体。

Schema-Free Paradigm

该范式代表了更为激进的思路，其核心是摒弃或后置模式的定义。LLM被赋予更高的自主权，直接从文本中“涌现”出它认为重要的实体、关系及概念结构。这种方法高度灵活，能够发现预定义模式之外的潜在联系和新知识，特别适用于探索性研究或开放域知识发现。

This paradigm represents a more radical approach, with its core being the abandonment or postponement of schema definition. LLMs are granted greater autonomy to directly "emerge" entities, relationships, and conceptual structures they deem important from the text. This method is highly flexible, capable of discovering potential connections and new knowledge beyond predefined schemas, making it particularly suitable for exploratory research or open-domain knowledge discovery.

典型技术方法包括：

• 开放信息抽取：提示LLM直接从文本中生成（主语，谓语，宾语）形式的三元组，不预设关系类别。
• 概念聚类与抽象：利用LLM的语义表示能力，对抽取出的实体和关系进行聚类、归纳，自底向上地形成概念层次和模式。
• 对话式交互构建：通过多轮人机对话，逐步引导LLM澄清、修正和扩展知识图谱A structured knowledge base that represents entities and their relationships in a graph format.，实现协同构建。

Typical technical approaches include:

• Open Information Extraction: Prompting LLMs to directly generate triples in the form of (subject, predicate, object) from text, without preset relationship categories.
• Concept Clustering and Abstraction: Leveraging the semantic representation capabilities of LLMs to cluster and summarize extracted entities and relationships, forming conceptual hierarchies and schemas in a bottom-up manner.
• Conversational Interactive Construction: Through multi-turn human-machine dialogue, gradually guiding the LLM to clarify, correct, and expand the knowledge graph, achieving collaborative construction.

技术评估与对比分析

Technical Evaluation and Comparative Analysis

为了清晰展示两种核心范式在不同构建阶段的特点与适用场景，我们将其关键维度对比如下：

To clearly illustrate the characteristics and applicable scenarios of the two core paradigms across different construction stages, we compare their key dimensions as follows:

Comparison Dimension	Schema-Based Paradigm	Schema-Free Paradigm
Core Driver	Predefined Ontology/Schema	Data-Driven and Model Emergence
LLM Role	Executor under Schema Constraints	Autonomous Knowledge Discoverer
Knowledge Quality	High Consistency, Low Hallucination	High Flexibility, but Consistency Hard to Control
Domain Adaptability	Requires Domain Schema, High Migration Cost	Strong Zero-shot/Few-shot Adaptation
Main Advantages	Standardized Output, Easy Integration, High Reliability	Open Discovery, Adapts to the Unknown, Agile Construction
Main Challenges	Schema Design Cost, Rigidity in Information Extraction	Uncontrollable Results, Hallucination, Complex Post-Integration
Typical Application Scenarios	Vertical Domain Graphs (Finance, Healthcare), Enterprise Knowledge Bases	Scientific Frontier Exploration, Open-Domain QA, Creative Assistance

未来研究方向

Future Research Directions

LLM与KG的融合方兴未艾，以下几个方向有望成为未来研究的重点：

The integration of LLMs and KGs is still in its early stages, and the following directions are poised to become focal points for future research:

1. KG增强的LLM推理：研究如何将动态检索到的、精确的结构化知识（KG三元组）作为“外挂内存”或“推理依据”注入LLM的推理过程，以克服其事实性幻觉和缺乏可追溯推理链的问题。
2. 面向智能体的动态知识记忆：在智能体系统中，构建能够实时更新、存储和利用交互经验的知识图谱A structured knowledge base that represents entities and their relationships in a graph format.，作为智能体的长期记忆，支持其持续学习和复杂规划。
3. 多模态知识图谱A structured knowledge base that represents entities and their relationships in a graph format.构建：突破纯文本限制，利用多模态大模型从图像、视频、音频中联合抽取实体与关系，构建包含丰富模态特征的知识图谱A structured knowledge base that represents entities and their relationships in a graph format.。
4. 可信与可解释性：开发能够评估和解释LLM所生成知识可信度的方法，并提供清晰的溯源，增强整个构建流程的透明度。

1. KG-Augmented LLM Reasoning: Investigating how to inject dynamically retrieved, precise structured knowledge (KG triples) as "external memory" or "reasoning grounds" into the LLM's reasoning process to overcome issues of factual hallucination and lack of traceable reasoning chains.
2. Dynamic Knowledge Memory for Agents: In agent systems, constructing knowledge graphs capable of real-time updating, storing, and utilizing interactive experiences as the agent's long-term memory to support continuous learning and complex planning.
3. Multimodal Knowledge Graph Construction: Breaking through the limitations of pure text, leveraging multimodal large models to jointly extract entities and relationships from images, videos, and audio, constructing knowledge graphs rich with multimodal features.
4. Trustworthiness and Explainability: Developing methods to evaluate and explain the credibility of knowledge generated by LLMs, providing clear provenance to enhance the transparency of the entire construction pipeline.

结论

Conclusion

大语言模型Advanced AI models trained on massive text data to understand and generate human language across multiple tasks.正在深刻重塑知识图谱A structured knowledge base that represents entities and their relationships in a graph format.构建的理论与实践。基于模式的范式强调结构、规范化和一致性的知识图谱构建方法，通常依赖于预定义的本体或模式。与无模式的范式强调灵活性、适应性和开放发现的知识图谱构建方法，不依赖于固定的模式或本体。并非相互取代，而是构成了一个从“严格规范”到“完全开放”的连续光谱。未来的智能知识系统 likely 将是混合架构，能够根据任务需求灵活选择或融合不同范式。通过将LLM的神经语义能力与KG的符号结构化优势相结合，我们正朝着构建更加自适应、可解释、且真正智能的知识基础设施稳步迈进。这一交叉领域的持续探索，不仅将推动KG技术本身的发展，也将为下一代人工智能系统奠定坚实的认知基础。

Large Language Models are profoundly reshaping the theory and practice of knowledge graph construction. The schema-based and schema-free paradigms do not replace each other but rather form a continuous spectrum from "strictly normative" to "completely open." Future intelligent knowledge systems will likely be hybrid architectures, capable of flexibly selecting or integrating different paradigms based on task requirements. By combining the neural semantic capabilities of LLMs with the symbolic, structural advantages of KGs, we are steadily advancing towards the construction of more adaptive, explainable, and truly intelligent knowledge infrastructure. The ongoing exploration in this interdisciplinary field will not only propel the development of KG technology itself but also lay a solid cognitive foundation for the next generation of artificial intelligence systems.

---

论文信息

• 标题: 大语言模型Advanced AI models trained on massive text data to understand and generate human language across multiple tasks.驱动的知识图谱A structured knowledge base that represents entities and their relationships in a graph format.构建：范式、技术与未来展望 (LLM-Empowered Knowledge Graph Construction: Paradigms, Techniques, and Future Outlook)
• 作者: Haonan Bian 等
• arXiv链接: https://arxiv.org/abs/2510.20345
• 领域: 人工智能 (cs.AI)

Paper Information

• Title: LLM-Empowered Knowledge Graph Construction: Paradigms, Techniques, and Future Outlook
• Author: Haonan Bian et al.
• arXiv Link: https://arxiv.org/abs/2510.20345
• Field: Artificial Intelligence (cs.AI)

常见问题（FAQ）

LLM如何改变传统的知识图谱A structured knowledge base that represents entities and their relationships in a graph format.构建流程？

LLM将传统基于规则和统计的流水线转变为语言驱动和生成的框架，重塑了本体工程知识图谱构建的第一层流程，涉及领域概念、属性和关系的定义与组织，建立知识图谱的语义框架。、知识抽取从非结构化或半结构化数据中自动识别和提取实体、关系等知识元素的过程。和知识融合将来自不同来源的知识进行整合、消歧和统一，形成一致、完整的知识图谱。这三个经典层次。

基于模式和无模式的LLM知识图谱A structured knowledge base that represents entities and their relationships in a graph format.构建有什么区别？

基于模式强调结构、规范化和一致性；无模式则注重灵活性、适应性和开放发现，两者代表了互补的技术范式。

LLM知识图谱A structured knowledge base that represents entities and their relationships in a graph format.的未来研究方向有哪些？

包括基于KG的LLM推理、面向智能体系统的动态知识记忆以及多模态KG构建，旨在推动自适应、可解释的智能知识系统发展。