MASEO: A Multi-Agent System for Explainable Ontology Generation

This repository provides the artifact for MASEO, a research-oriented multi-agent system that automated generate ontologies from competency questions, with a built-in focus on explainability. It aims to make the process of ontology generation more transparent, modular, and intelligent by distributing tasks among specialized agents. Each specialized agent is designed to keep track the logic behind each entity in generated ontology.

MASOE Structural Overview

The pipeline consists of four sequential stages:

Agent	Responsibility	External Tools
Ontology Generation Agent	Generates the initial OWL ontology from CQs	None
Syntax Repair Agent	Fixes RDF/XML syntax errors reported by the parser	rdflib
Logical Consistency Agent	Repairs logical inconsistencies reported by HermiT	HermiT Reasoner
Pitfall Resolution Agent	Resolves ontology modeling pitfalls reported by OOPS!	OOPS!

The illustration of the MASOE framework:

Features

• End-to-end automation — from a list of CQs to a validated ontology
• Role-based agents — each stage is handled by a dedicated LLM agent with a specific instruction and responsibility
• Provenance tracking — every ontology entity carries an append-only vaem:rationale log attributed to the agent that made each change, and a dc:source log linking each change back to the CQ, pitfall, or error that motivated it

External tools requirements

Tool	Purpose	Setup
HermiT Reasoner	Logical consistency checking	Download HermiT.jar and update the path in reason_ontology()
OOPS! REST API	Ontology pitfall detection	No local setup required — uses the public REST endpoint
Java (JRE 8+)	Required to run HermiT	sudo apt install default-jre

Execution

MASEO support execution over single set of competency questions with sepcific LLM (CLI Execution) as well as batch run over a selection of models and sets of competency questions over various domains (Batch Execution).

CLI Execution

python -u cli.py \
    --config       ./config.yaml \
    --cqs_file     ./dataset/cqs/wine_cqs.json \
    --save_file    ./wine.owl \
    --agent_method true

Argument	Required	Description
--config	Yes	Path to config.yaml. Defaults to ./config.yaml.
--cqs_file	Yes	JSON file with competency questions: [{"id": "CQ1", "value": "..."}, ...].
--save_file	Yes	Where to write the produced OWL ontology.
--agent_method	No	true (default) runs the full multi-agent pipeline; false runs single-pass generation only.

Batch Execution

To sweep multiple models and competency-question files in one command, use run_batch.py:

python -u run_batch.py --batch ./batch.yaml

batch.yaml only contains the list of models that you wish to run.

models:
  - provider: openrouter
    id: qwen/qwen3.6-flash
  - provider: deepseek
    id: deepseek-chat
  - provider: ollama
    id: qwen3:32b
  ...

Place your competency-question files in ./dataset/cqs/. For every (model, cqs_file) pair the runner invokes MASEO generation (--agent_method true) and normal agent generation (--agent_method false). All generated ontology and log file will be saved independently.

Documentation

Additional documentation of the project is available at readthedocs

• The full document for configuration can be found at: Configuration
• The full document for Input file structure can be found at Input
• The full document for Output file structure can be found at Output

Acknowledgements

This work was supported by the grant SOEL: Supporting Ontology Engineering with Large Language Models PID2023-152703NA-I00 funded by MCIN/AEI/10.13039/501100011033 and by ERDF/UE. The authors would also like to thank the EDINT (Espacios de Datos para las Infraestructuras Urbanas Inteligentes) ontology development team for sharing the project resources for evaluation purposes.