Optimizing Text-to-SQL Transformations: The Potential of Skeleton Decoupling in SKT-SQL

Abstract The Text-to-SQL technology faces significant challenges in converting natural language questions into SQL code, particularly in handling complexities and diversities in multi-domain and high-complexity tasks. To address these issues, this paper proposes a novel framework, SKT-SQL, which effectively decomposes the key steps in Text-to-SQL transformation, such as skeleton generation and schema linking, through multi-decoupling and skeleton prompting strategies, thereby simplifying the entire process. SKT-SQL employs a schema decoupler to filter relevant schema items linked to the natural language questions, reducing the burden on the model during the parsing process. Concurrently, the skeleton generator guides the language model in generating accurate SQL questions using an abstract SQL skeleton representation. This innovative strategy significantly enhances the accuracy of SQL skeleton generation using smaller-scale language models and optimizes the final SQL query generation results. The core contribution of this paper lies in the analysis of the potential of skeleton-based decoupling, demonstrating the advantages of SKT-SQL in improving the accuracy of Text-to-SQL generation, and providing new insights for future Text-to-SQL methods. Experimental validation on the Spider dataset shows that SKT-SQL excels in execution accuracy (EX) and exact-match accuracy (EM), showcasing its robust potential in complex query scenarios.