New knowledge-aware framework boosts Text-to-SQL for low-resource domains

Text-to-SQL systems convert natural language questions into executable SQL queries, enabling non-technical users to interact with databases. However, performance in real-world scenarios is often hampered by low-resource settings where high-quality annotated question-SQL pairs are scarce—especially for domain-specific databases with opaque schema definitions, abbreviations, and implicit business logic. Existing data synthesis and prompting techniques improve coverage but fail to produce task-specific, semantically grounded examples aligned with database constraints.

To address these challenges, the researchers developed a knowledge-aware framework that builds a task-specific knowledge base capturing schema semantics, abbreviations, business logic, and query patterns. This knowledge is injected into both training and inference stages: it generates diverse, contextually grounded synthetic training data and enhances inference through targeted knowledge retrieval. Evaluated on seven benchmarks covering general and domain-specific datasets, the method significantly improves the performance of both open-source and closed-source large language models, particularly in low-resource domain-specific settings, boosting generalization, robustness, and adaptability.