北大 & 作业帮团队提出 Text-to-SQL 新框架 Interactive-T2S，攻克宽表处理与低资源对齐难题

Core Insights - The article discusses the development of the Interactive-T2S framework, which transforms large language models (LLMs) into intelligent query agents capable of multi-turn interactions with databases, addressing inefficiencies in handling complex, wide tables [2][5][6]. Text-to-SQL Technology - Text-to-SQL serves as a bridge between natural language and databases, allowing users to convert natural language queries into executable SQL without needing SQL syntax knowledge, which is valuable in various sectors like enterprise data analysis and public services [4]. Challenges in Current LLM-based Text-to-SQL Methods - Existing methods face three main challenges: inefficiency in processing wide tables, poor adaptability in low-resource scenarios, and lack of interpretability in the interaction process [5][8]. Interactive-T2S Framework - The Interactive-T2S framework views LLMs as intelligent query agents and databases as data environments, utilizing a multi-turn interaction logic to generate and validate SQL queries step-by-step, requiring only two annotated examples for few-shot learning [6][10]. Core Tools of Interactive-T2S - The framework includes four core tools designed to reduce the reasoning burden on LLMs: - SearchColumn for semantic column identification - SearchValue for fuzzy value searching - FindShortestPath for table association - ExecuteSQL for real-time execution and validation of SQL queries [7][12]. Experimental Validation - The research team conducted experiments on various datasets, demonstrating that Interactive-T2S outperforms existing methods in execution accuracy and efficiency, particularly in complex and noisy data environments [11][14][15]. Application Value and Future Directions - Interactive-T2S has potential applications in smart education, enterprise data analysis, and public service queries, simplifying data retrieval processes for users [18]. Future enhancements will focus on optimizing tool efficiency and exploring capabilities in multimodal data queries [19].