Generative artificial intelligence, and more specifically LLMs (Large Language Models), currently represents one of the major achievements of artificial intelligence research. These models have become widely used tools in everyday life thanks to their impressive natural language processing capabilities and their effectiveness in solving various problems. However, LLMs still have certain limitations, notably the phenomenon of hallucination, i.e., the generation of incorrect or fabricated answers. This problem is particularly evident in critical domains, where their knowledge remains insufficient to produce accurate and reliable answers. To address these weaknesses, several approaches have been proposed. Among them, the use of ontologies stands out as a promising solution. Ontologies constitute a structured and reliable source of knowledge, allowing LLM answers to be corrected and enriched while improving their consistency. In this work, we examine seven approaches that combine LLMs with ontologies: OG-RAG, OntoPrompt, OntoTune, TaxoLLaMA, the neuro-symbolic framework, the Recover framework, and the concrete defect diagnosis framework. For each of these methods, we describe the methodology adopted, specifying how ontological knowledge is integrated into the process, and then present their main advantages and limitations. We then offer a comparative analysis of the ontologies used and their formalisms. Finally, we conclude by outlining our research perspectives.