LLM-Driven Knowledge Graph Construction: Intelligent Transformation from Natural Language to Formal Ontology

This article introduces a framework that uses large language models (LLMs) to automatically extract knowledge from domain texts and construct formal ontologies, aiming to address the bottlenecks of traditional manual ontology construction, such as time-consuming processes and reliance on professional knowledge. The framework integrates natural language processing and knowledge representation technologies, exploring its application prospects in scenarios like data standardization and semantic interoperability, as well as new paradigms of human-machine collaboration and future development directions.

Concepts and Roles of Ontology

An ontology is a formal specification of concepts and their relationships in a specific domain, defining classes (entity types), data properties (entity features), and object properties (relationships between entities) in the domain.

Traditional Knowledge Graph Construction Process

It includes domain analysis, concept extraction, relationship definition, formal encoding, and iterative verification. Each step requires manual participation, leading to long cycles, high costs, and difficulty in scaling.

Core Capabilities of LLMs

LLMs can identify named entities, extract relationships, infer hierarchical structures, and generate formal representations such as JSON/OWL from unstructured texts.

Framework Workflow

1. Input processing: Accept domain texts like technical documents and academic papers;
2. Automatic extraction: Identify classes, object properties, data properties;
3. Output generation: Output structured JSON, interactive HTML graphs, and Graphol files.

System Components

• Data layer: Input texts, ontology files, validation sets;
• Processing engine: Guide LLMs to perform extraction tasks based on prompt engineering;
• Visualization interface: Web application built with Streamlit, supporting process guidance, result preview, etc.

Key Innovations

• Explicit semantic constraints: Only extract explicit/implicit information from texts;
• State-driven pipeline: Ensure process integrity;
• Interactive knowledge graph: Generate interactive visualizations using Pyvis.

Data Standardization and Interoperability

Italy's National Institute of Statistics (ISTAT) uses this tool to improve the semantic consistency of official statistical data;

Rapid Domain Knowledge Modeling

Applicable to emerging fields, quickly extracting concepts from the latest literature and dynamically updating ontologies;

Enterprise Knowledge Management

Convert unstructured documents into knowledge graphs to support intelligent search and decision-making.

This framework builds a bridge between natural language and formal representation, promoting the paradigm shift of knowledge engineering from manual coding to AI-assisted. The system emphasizes human-machine collaboration (users can edit prompts and review results) and interpretability (transparent processes, traceable generation process), providing possibilities for knowledge democratization and accelerated innovation.

Current Limitations

• Output format dependency: Complex texts may have parsing errors;
• Lack of verification mechanism: No OWL/RDF semantic verification;
• Language support: Mainly for English.

Future Directions

• Support export of OWL/RDF standard formats;
• Introduce reasoning engines for semantic verification;
• Expand multilingual support;
• Add execution history tracking and ontology comparison tools.