Recently, there has been an increase in research within the Text-to-SQL domain utilizing Small Language Models (SLMs) for training. These approaches achieve performance close to that of large models in SQL generation, using only a single NVIDIA RTX 4090 GPU's computational power while ensuring data security.
Most existing methods filter out redundant tables and columns during Schema Linking to enhance Text-to-SQL accuracy. However, they do not consider the precision-recall trade-off when selecting candidate schema subsets. Our research indicates that both the precision and recall of Schema Linking directly affect the final SQL accuracy.
To address this, we propose a novel framework, CHS-SQL, for efficiently fine-tuning SLMs on Text-to-SQL tasks. This framework not only balances precision and recall but also improves overall performance in Text-to-SQL tasks. Its main innovation lies in the Schema Linking phase, employing a heuristic search combined with model internal confidence to achieve an optimal precision-recall trade-off.
This detailed mechanism maximizes the precision of relevant schema candidates for the generated SQL queries while suppressing irrelevant noise. The same strategy is further applied during SQL generation to refine candidate queries and assist the SLM in avoiding local optimum traps. Our method achieves state-of-the-art (SOTA) results on Text-to-SQL tasks via SLMs.
Blogger's Review: The CHS-SQL framework introduces a fresh perspective on Text-to-SQL conversion by integrating confidence-guided heuristic search, effectively balancing precision and recall. This innovation not only enhances the accuracy of SQL generation but also offers new avenues for future research, making it a noteworthy topic for further exploration.