SimX-HCOME⁺: Hybrid Ontology Engineering Method

• SimX-HCOME⁺ is a hybrid ontology-engineering methodology that combines LLM automation with continuous, role-specific human oversight.
• It iteratively refines ontologies by translating natural language rules into executable SWRL constructs and validating semantic structures at every step.
• The approach addresses LLM-only limitations by improving domain coverage and precision, particularly in complex scenarios like Parkinson’s Disease monitoring.

SimX-HCOME⁺ (Simulated eXtended Human-Centered Ontology Engineering Methodology, enhanced) is an LLM-driven, hybrid ontology-engineering methodology emphasizing both automated construction and continuous human intervention at all stages. It systematically involves LLMs alongside simulated expert feedback (Knowledge Worker, Domain Expert, Knowledge Engineer) for the iterative creation, refinement, and validation of comprehensive domain ontologies. The methodology builds upon previous hybrid models (e.g., X-HCOME) and is designed for domains with complex semantic and rule-based requirements, such as Parkinson’s Disease (PD) monitoring and alerting ontologies (Bouchouras et al., 16 Dec 2025).

1. Definition and Rationale

SimX-HCOME⁺ extends human-centered ontology engineering techniques by tightly integrating LLM generative capabilities with structured, role-specific human interventions. The core objectives are:

• Automated, iterative ontology generation leveraging the semantic breadth of contemporary LLMs.
• Structured, continuous supervision, with simulated feedback provided by three expert roles at every iterative cycle.
• Immediate production of runnable ontologies after each micro-iteration, supporting early detection of semantic gaps and logical inconsistencies.
• Transformation of natural-language domain and business rules into executable Semantic Web Rule Language (SWRL) constructs, not limited to taxonomy and object properties.

This approach is motivated by empirical limitations of LLM-only ontology generation, which typically produces incomplete or unfaithful models, and by the recognition that human oversight, when structured continuously, can significantly enhance output coverage and correctness (Bouchouras et al., 16 Dec 2025).

2. Methodological Workflow and Algorithmic Structure

The SimX-HCOME⁺ methodology is formalized as an iterative, feedback-driven loop, with well-defined roles and feedback channels at each iteration. A high-level algorithmic description is provided in LaTeX pseudocode:

\begin{algorithm}[H] \caption{SimX-HCOME⁺ Ontology Engineering Loop} \begin{algorithmic}[1] \State \textbf{Input:} Aim\_Scope, CompetencyQuestions, SourceData, NL\_Rules \State \textbf{Output:} FinalOntology \State Initialize Ontology %%%%0%%%% empty \State Set iteration %%%%1%%%% \Repeat \State %%%%2%%%% \Comment{LLM Generation Step} \State %%%%3%%%% \Comment{Simulated Human Feedback} \State %%%%4%%%% \State %%%%5%%%% \State %%%%6%%%% \Comment{Integrate Feedback} \State %%%%7%%%% \Comment{Rule Conversion (NL → SWRL)} \State %%%%8%%%% \State Ontology %%%%9%%%% Ontology %%%%10%%%% \{\,%%%%11%%%%, %%%%12%%%%\} \Comment{Automated Validation} \State \textbf{validate}(\text{Ontology}) \Comment{Pellet consistency, OOPS! pitfalls} \Until{convergence or %%%%13%%%%} \State \textbf{return} Ontology \end{algorithmic} \end{algorithm}

Key characteristics of the workflow include:

• Immediate feedback incorporation: Each simulated expert role provides targeted feedback—completeness (Knowledge Worker), domain correctness (Domain Expert), and formal constraints (Knowledge Engineer)—which is directly integrated in the subsequent refinement step.
• Rule translation: Natural-language business or clinical logic is transformed into SWRL rules via dedicated LLM prompting and immediate validation.
• Automated validation: Outputs are checked for logical consistency and modeling pitfalls at each iteration via tools such as Pellet and OOPS!.

3. Prompt Engineering and LLM-Human Interaction

SimX-HCOME⁺ operationalizes a multi-stage, conversational prompt framework inspired by X-HCOME but refined for continuous oversight. The interaction sequence follows:

• Ontology Sketch: The LLM receives the aim, scope, and competency questions, then produces a basic skeleton (classes/properties) in a formal language (e.g., Turtle).
• Detail Enrichment: The ontology is iteratively detailed using source data, with focus areas (e.g., movement patterns, patient states) explicitly provided via prompt.
• SWRL Rule Translation: Natural-language rules (e.g., clinical scenarios) are converted into formal SWRL atom representations.
• Micro-Chain-of-Thought (micro-CoT): Each stage passes its output as the next prompt's context, enforcing local coherence and enabling ongoing expert supervision.

This “micro-CoT” mechanism differs from one-shot or static Chain-of-Thought prompting by imposing tight stage-wise control and immediate expert correction, resulting in incremental, contextually anchored ontology evolution (Bouchouras et al., 16 Dec 2025).

4. Evaluation Metrics and Comparative Results

Quantitative evaluation focuses on class-level alignment with a gold standard reference ontology (41 classes for PD domain) via standard information retrieval metrics:

• Precision: $$\displaystyle\frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP}}$$
• Recall: $$\displaystyle\frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}}$$
• F1 score: $$2 \cdot \frac{\mathrm{Precision} \times \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}}$$

Empirical results:

Method	#Classes	TP	FP	FN	Precision	Recall	F1
ChatGPT-4	17	9	8	32	52%	21%	31%
ChatGPT-3.5	21	14	7	27	66%	34%	45%
Gemini	22	15	7	26	68%	36%	48%
Claude	24	12	12	29	50%	29%	37%

• SimX-HCOME⁺ showed clear improvements over one-shot and traditional Chain-of-Thought prompts (F1 ≈ 17–20%) and performed comparably or better than X-HCOME (F1 = 42%), with best results (F1 = 48%) for Gemini.
• SWRL rule translation, measured by logical atom match (F1ₗₒ𝚌), remained challenging, with best performance at F1ₗₒ𝚌 = 20% (Claude).

5. Qualitative Observations and Methodological Limitations

Comprehensive evaluation indicated:

• Produced ontologies were consistently syntactically correct and passed standard validation tools (Pellet, OOPS!), except for minor Gemini errors.
• LLMs, within SimX-HCOME⁺, discovered domain-relevant classes sometimes missed by human experts, thus achieving higher recall on certain sub-domains.
• Object and data property modeling persisted as the primary point of weakness (F1 typically <15%).
• Conversion of natural-language rules into executable SWRL was difficult, with incomplete logical structure capture, suggesting a requirement for domain-specific prompt tuning or LLM adaptation.

Key limitations include:

• Residual hallucinations and biases, not fully eliminated by structured feedback.
• Emphasis on taxonomic and object property engineering, with less coverage of data properties, complex axioms, or constraint patterns.
• Results rely on simulated, not real, human supervision; actual team deployment may face different feedback latencies and cognitive constraints (Bouchouras et al., 16 Dec 2025).

6. Implications and Future Directions

Emerging research directions for SimX-HCOME⁺ include:

• Training or fine-tuning "Ontology-GPT" models on curated domain tasks and SWRL examples for improved logical translation.
• Explicit integration of data property and logical axiom generation into the core iterative loop.
• Exploration of genuinely multi-expert, asynchronous feedback processes to simulate real-world collaboration more closely.
• Application of SimX-HCOME⁺ to broader domains (beyond healthcare) with rigorous tracking of cost and supervision-efficiency trade-offs.

The methodology demonstrates that tightly coupled human-LLM iterative collaboration—formalized through structured, role-based cycles—yields significantly more comprehensive ontologies than fixed, one-shot, or even static hybrid approaches, and it provides a framework for addressing unresolved challenges in automated and semi-automated ontology engineering (Bouchouras et al., 16 Dec 2025).