A robust natural language text-to-SQL generation framework with dynamic strategies based on LLMs.

Area of Science:

• Artificial Intelligence
• Databases
• Natural Language Processing

Background:

• Natural language text-to-SQL generation (Text2SQL) aims to translate natural language questions into executable SQL queries.
• Large language models (LLMs) have advanced Text2SQL, but performance degrades with increasing question complexity.
• Current LLM methods lack dynamic adaptation to varying question complexity.

Purpose of the Study:

• To propose TriSQL, a novel three-stage framework to address the limitations of current Text2SQL methods.
• To improve the accuracy and executability of SQL generated from complex natural language questions.
• To dynamically adapt generation strategies based on question and initial SQL complexity.

Main Methods:

• Question-Guided Schema Selector: Utilizes cross-attention to identify the most relevant schema for a given question.
• Structure-Aware SQL Generator: Employs hierarchical decoding to create a syntactically valid initial SQL from the question and selected schema.
• Complexity-Aware SQL Refiner: Leverages LLMs to dynamically adjust SQL generation strategies based on question and initial SQL complexity.

Main Results:

• TriSQL achieves state-of-the-art execution accuracy on the Spider benchmark and its variants.
• The proposed method surpasses existing LLM-based Text2SQL approaches.
• TriSQL demonstrates high efficiency and strong robustness in generating accurate and executable SQL.

Conclusions:

• TriSQL offers a significant advancement in Text2SQL generation by dynamically adapting to question complexity.
• The framework provides a robust and efficient solution for translating complex natural language questions into executable SQL.
• TriSQL sets a new standard for accuracy in complex Text2SQL tasks.