SKT-SQL: multi-decoupling and skeleton prompt framework for text-to-SQL on small-scale LMs

SKT-SQL: multi-decoupling and skeleton prompt framework for text-to-SQL on small-scale LMs

연구 분야: Databases

학회: Journal of Intelligent Information Systems

To address the strong coupling between schema linking and structural parsing in text-to-SQL tasks for small-scale language models, as well as the neglect of SQL skeleton guidance in existing decoupling methods, this paper proposes SKT-SQL, a multi-stage decoupling framework. The framework redesigns the generation process through a three-stage decoupling mechanism: (1) leveraging a schema decoupler to eliminate irrelevant schema items and reduce semantic noise; (2) predicting query hardness to generate an abstract SQL skeleton, forming a structured template with operator logic; (3) utilizing the skeleton as dynamic prompts to guide a transformer-based seq2seq T5 model in precisely filling specific schema items, followed by execution-guided beam search to derive the final SQL query. This "structure-first, entity-later" paradigm eliminates the need to simultaneously resolve syntactic complexity and schema correlations, significantly reducing cognitive load. Experiments on the Spider 1.0 benchmark show that SKT-SQL-base achieves 80.8% execution accuracy (EX), surpassing the 12 larger T5-3B model by 6.4% and outperforming T5-base by 22.9%. Compared to existing decoupling-based methods, this framework still demonstrates prominent application value, with skeleton prompting notably enhancing small-scale model performance. This study proves that explicit modeling of SQL syntactic skeletons can break through the structured semantic parsing bottleneck of small-scale language models, offering a novel paradigm for database interaction tasks in low-resource scenarios. Our code is available at https://github.com/JarvenYi/SKT-SQL.