From Fragmentation to Cohesion: An LLM-Based Iterative Approach to Ontology and Knowledge Graph Refinement
Alexandru Popa, Tiberiu-Iulian Sîrbu · 2025
Knowledge graphs and their respective ontologies often suffer from disconnected subgraphs, redundant classes, and inconsistent relationships when automatically generated from unstructured data. While recent studies involving Large Language Models (LLMs) are capable of generating knowledge graphs, they lack the capability to offer a well-connected and stable graph. In this paper, a new pipeline is proposed that repeatedly refines the ontology by identifying disconnected components of the knowledge graph and employs the LLM to find the connections between the largest connected component and all the smaller disconnected ones. Moreover, the pipeline discussed in this study aims to offer a stable ontology at the end that will create a consistent graph each time is used as input in a LLM based knowledge graph construction.