CHS-SQL Boosts Text-to-SQL Accuracy with Small Language Models.

Recent advancements in Text-to-SQL systems have shown that Small Language Models (SLMs) can achieve performance comparable to larger models, often requiring only a single high-end GPU and offering enhanced data security. A key challenge in these systems is Schema Linking, where redundant tables and columns are filtered to improve SQL accuracy. However, existing methods often overlook the critical precision-recall trade-off in this selection process. A new framework, CHS-SQL, addresses this by proposing an innovative Schema Linking phase. It employs a heuristic search combined with the model's internal confidence to optimize the precision-recall balance. This mechanism ensures that the most relevant schema candidates are prioritized for SQL generation while suppressing irrelevant noise, leading to more accurate queries. The same confidence-guided strategy is also applied during the SQL generation phase, helping the SLM refine candidate queries and avoid suboptimal solutions. This approach has enabled CHS-SQL to achieve state-of-the-art results for Text-to-SQL tasks using SLMs, demonstrating a significant leap in efficiency and accuracy for converting natural language into database queries.