Staged Multi-Strategy Framework With Open-Source Large Language Models for Natural Language to SQL Generation

In the field of natural language to SQL (NL2SQL), significant progress has been made with large pre-trained language models. However, these models still have deficiencies in terms of their ability to generalize, particularly in open-source Large Language Models (LLMs). Additionally, most research efforts tend to overlook the impact of key column information and data table content on the accuracy of queries during the SQL statement generation process. In this paper, we propose a staged, multi-strategy framework called Key Columns and Table Contents (KCTC). The framework is divided into two stages. Firstly, it uses fixed prompt content to extract SQL key column information from natural language questions, including selected columns and conditioned columns. It also formats the output of column information. Secondly, it combines variable prompt content to guide the model in generating SQL statements. It uses the content of the data table for constraints to reduce the impact of errors in condition values on SQL statements. We conducted experiments on the Chinese dataset TableQA using several open-source LLMs. The results demonstrate that our method significantly improved the execution accuracy of SQL statements, with an average increase of 60.29% and reaching up to 91.22% accuracy. This result validates the effectiveness of our approach. (c) 2025 Institute of Electrical Engineers of Japan and Wiley Periodicals LLC.