ExSPIN: Explicit Feedback-Based Self-Play Fine-Tuning for Text-to-SQL Parsing.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Database Management

Background:

• Self-play fine-tuning (SPIN) enhances large language models (LLMs) through self-interaction.
• Applying SPIN to text-to-SQL models faces challenges with feedback signals and schema linking.
• Existing methods struggle to capture implicit relationships between natural language and database schemas.

Purpose of the Study:

• To propose a novel self-play fine-tuning method, ExSPIN, for text-to-SQL models.
• To address limitations in feedback mechanisms and schema understanding in current SPIN frameworks.
• To improve the accuracy and efficiency of LLMs in converting natural language questions to SQL queries.

Main Methods:

• Developed ExSPIN, a self-play fine-tuning approach with explicit feedback.
• Incorporated SQL query execution results as feedback signals for parameter updates.
• Utilized in-context learning for explicit schema hints to guide LLM understanding.
• Trained and evaluated the model on two real-world text-to-SQL datasets.

Main Results:

• ExSPIN significantly improves text-to-SQL model performance.
• Explicit feedback enhances the distinction between positive and negative training samples.
• In-context schema hints improve the LLM's ability to link natural language to database structures.
• The proposed method outperforms existing state-of-the-art approaches.

Conclusions:

• ExSPIN offers a more effective approach to self-play fine-tuning for text-to-SQL tasks.
• Explicit feedback and schema guidance are crucial for robust LLM performance in database querying.
• The method demonstrates significant advancements in text-to-SQL conversion accuracy.