Abstract
Background: Ontologies and knowledge graphs have become critical for structuring data into machine-interpretable knowledge, especially in dynamic domains like IT service management (ITSM). Traditional ontology engineering relies heavily on domain experts, making it costly and slow. This study investigates whether a domain-specific ontology can be extended from a top-level ontology without expert involvement, using the IT service management ontology (ITSMO) and the descriptive ontology for linguistic and cognitive engineering (DOLCE-lite) as a test case used in this study.
Methodology: We propose an automated mapping approach integrating lexical approaches, embeddings, graph neural networks (GNN), and large language models (LLMs). Two primary mapping methods were developed: (1) embedding-based matching, computing cosine similarity between class embeddings from DOLCE and ITSMO; and (2) LLM-based matching, prompting a language model (GPT-4o) to evaluate class compatibility on a numeric scale. We also experiment with GraphSAGE GNN to enrich embeddings with ontology structure. Z-score clustering is applied to similarity scores to select top candidate mappings while filtering out outliers from the top cluster. The methodology operates with no annotated data and was validated using three-steps approach: GPT-4o as a surrogate expert for baseline class matching evaluation, expert spot-check, and OWL reasoner (Pellet and HermiT) to prove logical consistency (Glimm et al., 2014; Sirin et al., 2007).
Results: The automated method successfully mapped ITSMO classes under DOLCE, yielding an integrated ontology (80 classes) that extends DOLCE into the ITIL domain with minimal expert intervention (expert consolidated suggestions into a result ontology). The LLM-based approach (GPT-4o) achieved the best performance with 73.5% accuracy for top-1 mappings and 82.4% for top-3 (cluster) inclusion. Transformer-based embeddings (e.g., DeBERTa) also performed well (up to 39.3% top-1, outperform random matching with 27.6% accuracy), but classical graph embeddings (RDF2Vec/Node2Vec) failed due to the small ontology size. Incorporating a GNN provided smoother embedding distributions and increased correct mappings within top-3 clusters, but it slightly reduced top-1 precision in this small-graph setting. These findings underscore the effectiveness of LLMs in zero-shot ontology alignment and the limitations of purely structural methods on limited data.
Conclusions: This work demonstrates, as a proof-of-concept, that an upper-level ontology can be extended to a domain ontology automatically, with no or minimal expert involvement, by leveraging AI-based mapping techniques. The resulting new ontology integrates ITSMO into DOLCE, providing a consistent semantic foundation for IT domain knowledge graphs. The approach is immediately applicable to ITSM and suggests a generalizable framework for ontology expansion in other domains. Future work will focus on scaling the method to larger ontologies, automatically discovering new classes/relations from text, and evaluating the approach’s practical impact on IT service management processes.