Reproduce “Exploring Database Normalization Effects on SQL Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ryosuke Kohita
An individual researcher from CyberAgent investigated how database normalization levels influence the accuracy of state-of-the-art Large Language Models when generating SQL queries. The research shows that denormalized schemas perform better for simple retrieval, while normalized schemas yield higher accuracy for aggregation queries, especially in real-world applications.
Key insights
- Optimal database schema design for NL2SQL is query-dependent: denormalized schemas are more effective for simple retrieval queries, while normalized schemas are superior for complex aggregation queries.
- Large Language Models exhibit distinct error patterns depending on schema normalization, including difficulties with join selection and base table identification in normalized zero-shot settings, and issues with duplicate counting in denormalized real-world scenarios.
- The study highlights the need for dynamic, workload-aware NL2SQL systems that can adapt schema selection based on query intent, potentially leveraging multiple schema variants.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.