Boosting the Querying Accuracy of Multi-Level Occupancy Data with Ontology-Guided LLMs

As overtourism and local overcrowding are becoming increasingly critical concerns, determining and predicting occupancy levels based on real-time data and predictive models that serve as a decision-making basis for necessary countermeasures are gaining popularity. Moreover, with the rise of large language models (LLMs), approaches that automate related data access have become tempting. However, real-world databases are often inherently complex and heterogeneously structured, complicating using LLM-based text-to-SQL. Previous studies report an accuracy of only 16%, which indicates the need for better approaches. This paper investigates how ontologies can support LLMs in increasing the accuracy of querying real-world databases. Based on the need to reduce overcrowding, we propose an ontology for modeling complex, multi-level occupancy data. Our ontology, based on previous work, is theoretically well-founded and compatible with existing tourism ontologies. In a case study based on a real-world database from Outdooractive, one of the largest European outdoor tourism platforms, we compare vanilla LLM-based text-to-SQL's performance with ontology-based data access. Our results show that the ontology-based approach almost triples the querying accuracy, which illustrates the effectiveness and potential of such semantic approaches.