Assessing the importance of knowledge graph formatting in improving LLM answer accuracy using data augmentation

While Large Language Models (LLMs) provide intuitive natural language interfaces, persistent hallucinations undermine their reliability, especially in specialized domains. In order to create the next generation of intelligent systems, the goal of this study is to determine which level of abstraction yields the highest question accuracy and which format would be the most scalable, as well as examine how Knowledge Graphs (KGs) can improve the accuracy of LLM responses through a neuro-symbolic approach, where structured symbolic logic forms the basis for generative artificial intelligence (GenAI). Using synthetic datasets based on the standard FOAF ontology, we modeled relational social graphs of varying complexity and tested six query attributes per entity. We evaluated performance in both in-context system prompting and supervised fine-tuning using several data formats: unstructured natural text and several structured markup languages. In total, 2400 data points were collected using data augmentation and an additional 1260 data points were collected using fine-tuning as a data integration approach. The results demonstrate a significant accuracy-complexity trade-off: utilizing KGs results in up to 29.63% higher factual precision and fewer hallucinations than using natural language alone. However, this neuro-symbolic integration requires up to a 152.42% increase in token consumption and greater overhead for data preprocessing. The presented findings provide a framework of guidelines for optimizing the data augmentation process and demonstrate that, although KGs can substantially improve LLM-based reasoning in specialized domains, they require a careful balance between symbolic representation and knowledge scalability.