Evaluating Large Language Models for Text-to-SQL Tasks: Comparing First Normal Form (1NF) and Star Schema

Large Language Models (LLMs) have shown promise in generating SQL queries from natural language, but their adaptability to Business Intelligence (BI) schemas, such as star schemas, is underexplored. This study evaluates six LLMs on SQL query generation tasks using a sales dataset structured in normalized and star schemas. Results reveal high semantic accuracy across models. However, variability in execution time and performance among models highlights areas for improvement. This research underscores the importance of schema design in optimizing LLM-driven text-to-SQL systems for practical BI applications. We further report statistical significance for schema effects (two-proportion tests for validity and exact-match; paired t-tests for per-LLM runtime means), and situate the contribution within Cognitive Infocommunications as an introductory bridge between natural-language intent and relational semantics for BI practice.