A Plug-and-Play Natural Language Rewriter for Natural Language to SQL

Existing Natural Language to SQL (NL2SQL) solutions have made significant advancements, yet challenges persist in interpreting and translating NL queries, primarily due to users' limited understanding of database schemas or memory biases toward specific table or column values. These challenges often result in incorrect NL2SQL translations. To address these issues, we propose REWRITER, a plug-and-play module designed to enhance NL2SQL systems by automatically rewriting ambiguous or flawed NL queries. By incorporating database knowledge and content (e.g., column values and foreign keys), REWRITER reduces errors caused by flawed NL inputs and improves SQL generation accuracy. Our REWRITER treats NL2SQL models as black boxes, ensuring compatibility with various NL2SQL methods, including agent-based and rule-based NL2SQL solutions. REWRITER comprises three key components: Checker, Reflector, and Rewriter. The Checker identifies flawed NL queries by assessing the correctness of the generated SQL, minimizing unnecessary rewriting and potential hallucinations. The Reflector analyzes and accumulates experience to identify issues in NL queries, while the Rewriter revises the queries based on Reflector's feedback. Extensive experiments on the Spider and BIRD benchmarks demonstrate that REWRITER consistently enhances downstream models, achieving average improvements of 1.6% and 2.0% in execution accuracy, respectively.