Core Concepts
Large Language Models can be used to efficiently generate capability ontologies from natural language descriptions, reducing the manual effort required for ontology creation.
Abstract
The study investigates the use of Large Language Models (LLMs) to generate capability ontologies from natural language descriptions. Two LLMs, GPT-4 and Claude 3, were tested with three different prompting techniques (zero-shot, one-shot, and few-shot) to generate ontologies for seven capabilities of varying complexity.
The results show that even with zero-shot prompting, the generated ontologies have very few syntax errors. However, the one-shot and few-shot prompts lead to significantly better results, with the few-shot prompts generating ontologies that are almost error-free. The authors developed a semi-automated approach to test the generated ontologies for inconsistencies, hallucinations, and incompleteness using OWL reasoning and SHACL constraints.
The key findings are:
LLMs can effectively generate capability ontologies from natural language descriptions, significantly reducing the manual effort required.
The quality of the generated ontologies improves with better prompting techniques, with the few-shot prompts producing the best results.
Claude 3 outperforms GPT-4 in terms of generating ontologies with fewer contradictions and hallucinations.
The semi-automated testing approach using OWL reasoning and SHACL constraints is crucial for verifying the correctness and completeness of the generated ontologies.
Overall, the study demonstrates the potential of using LLMs to automate the creation of capability ontologies, which are essential for flexible systems and algorithms for automated planning and adaptation.
Stats
The total volume must not surpass 20.
The sum of the three input volume fractions needs to equate 1.
The current position of the input product is required to be equal to the position of the transport resource.
The transport capability guarantees that the assured position after transport is equal to the desired position to be selected.
Quotes
"LLMs can effectively generate capability ontologies from natural language descriptions, significantly reducing the manual effort required."
"The quality of the generated ontologies improves with better prompting techniques, with the few-shot prompts producing the best results."
"The semi-automated testing approach using OWL reasoning and SHACL constraints is crucial for verifying the correctness and completeness of the generated ontologies."