Multi-Agent LLMs for Automatic Ontology Generation: A New Paradigm in Knowledge Engineering

From Unstructured Text to Formal Ontologies: A Long-Standing Challenge

Ontologies, as core building blocks of knowledge engineering, bear the critical responsibility of formally representing domain concepts, relationships, and rules. However, automatically generating high-quality formal ontologies from massive volumes of unstructured natural language text has long been one of the central challenges in knowledge engineering. Traditional approaches rely on manual construction by domain experts — a process that is time-consuming, labor-intensive, and difficult to scale. In recent years, the rise of large language models (LLMs) has brought new hope to this problem, but which architectural design choices truly drive improvements in generation quality? And why do existing methods frequently fail? These critical questions have lacked systematic answers.

A paper recently published on arXiv (arXiv:2604.23090v1) directly addresses these questions. The research team proposed an automatic ontology generation method based on multi-agent LLMs and conducted rigorous controlled experiments using the highly specialized domain of insurance contracts as the experimental scenario.

Core Method: The Architectural Leap from Single-Agent to Multi-Agent

The core idea of this research is to decompose the ontology generation task into multiple subtasks, completed collaboratively by different LLM agents with distinct roles. The research team first established a single-agent LLM baseline system, then progressively introduced multi-agent architectures to systematically examine the impact of different design choices on final ontology quality.

Specifically, different agents in the multi-agent framework may be responsible for the following tasks:

• Concept Extraction Agent: Identifies core domain concepts and entities from unstructured text
• Relationship Discovery Agent: Uncovers hierarchical relationships and semantic associations between concepts
• Axiom Generation Agent: Constructs formal logical constraints and inference rules
• Consistency Verification Agent: Checks the logical consistency and completeness of the generated ontology

This "divide and conquer" strategy allows each agent to focus on a specific stage of the ontology construction pipeline, thereby achieving higher precision and reliability in their respective subtasks.

Experimental Design: Insurance Contracts as the Litmus Test

The research team chose insurance contracts as the experimental domain — a deliberate and meaningful choice. Insurance contract texts feature highly specialized terminology systems, complex clause-level logical relationships, and strict legal constraints, imposing extremely high demands on the ontology generation system's concept comprehension, relational reasoning, and formal expression capabilities.

Through controlled experiments, the research team compared the performance differences between single-agent and multi-agent architectures across the following dimensions:

• Concept coverage of the generated ontology
• Accuracy of relationship extraction
• Logical consistency of ontology structure
• Degree of alignment with ontologies manually constructed by domain experts

Experimental results showed that the multi-agent architecture significantly outperformed the single-agent baseline across multiple key metrics, particularly excelling in handling complex concept hierarchies and cross-clause semantic associations.

In-Depth Analysis: Why Is the Multi-Agent Approach Superior?

From a technical perspective, the advantages of the multi-agent architecture stem primarily from three aspects:

First, task decomposition reduces cognitive load. Ontology generation is a complex task involving multi-level reasoning. A single LLM struggles to simultaneously balance concept extraction, relationship modeling, and logical verification within a single inference pass. The multi-agent approach, through task decomposition, allows each agent to focus on a single objective, effectively reducing reasoning difficulty.

Second, inter-agent collaboration introduces a "review and revision" mechanism. Interactions between different agents can form a quality control process similar to peer review, where one agent's output can be examined and refined by another, thereby reducing hallucinations and logical errors.

Third, modular design enhances interpretability and debuggability. When issues arise in the generated results, researchers can precisely pinpoint which specific stage of which agent went wrong, facilitating targeted improvements.

This research also revealed typical failure modes of current LLMs in ontology generation tasks — such as inconsistent concept granularity, incorrect relationship directionality, and omission of implicit constraints — providing important diagnostic references for subsequent research.

Industry Impact and Future Outlook

This research carries significant methodological implications for the knowledge engineering field. As enterprise digital transformation deepens, specialized domains such as finance, healthcare, and law face increasingly urgent demands for domain ontologies, while traditional manual construction methods can no longer meet scalability and timeliness requirements. The multi-agent LLM approach offers a viable technical pathway for automatic ontology generation.

Looking ahead, several directions in this research area warrant further exploration: How can this method be extended to additional domains? How can human-AI collaboration mechanisms be introduced to further improve ontology quality? How can dynamic ontology evolution and version management be handled?

Notably, the multi-agent collaboration paradigm itself is becoming a major trend in LLM applications. From code generation to scientific discovery, from data analysis to knowledge engineering, the idea of "having multiple AI agents collaborate with division of labor" is demonstrating potential to surpass single models in an increasing number of complex tasks. This research provides solid empirical support for the implementation of this trend in the knowledge engineering domain.