Haoyu Zhang

Text-to-SQL data synthesis must expand semantic coverage without losing the relational structure of observed queries. Existing SQL offers evidence of plausible join and nesting patterns, but reusing queries alone limits the range of training examples; unconstrained generation, in contrast, risks implausible structures and question–SQL mismatches. We introduce Structure-Constrained Data Synthesis (SCDS), which treats observed query structure as a prior and uses large language models (LLMs) to vary clause-level semantic content. SCDS groups source queries by coarse structural signatures and constructs semi-frozen SQL templates that retain the source FROM/JOIN context while abstracting selected clauses into semantic slots. It grounds intent generation in schema semantics and available values and represents the resulting objectives in a semi-structured intermediate representation. It then instantiates SQL, evaluates its executability diagnostically, and generates the corresponding questions and reasoning descriptions. On BIRD-dev, fine-tuning with the full set of 23,374 synthesized examples achieves a mean execution accuracy of 57.69% across three training seeds. On MaritimeSQLBench, fine-tuning with synthesized examples improves execution accuracy from 47.69% to 56.75% and reduces the execution error rate from 13.33% to 10.77%. The maritime result is measured with the gold schema elements required by each reference SQL. We publicly release MaritimeSQLBench to support further research on maritime Text-to-SQL. These results support structural priors as a useful constraint on Text-to-SQL synthesis. At a matched training size of 9428 examples, SCDS achieves performance comparable to BIRD-train; together with the maritime results, this highlights its potential for domain-specific Text-to-SQL data synthesis.