Progressive Reverse Understanding Improves Text-to-SQL Generation

Xiaoxia Cheng, Bo Jiang, Qin Xu

Keywords: Reverse Understanding, text-to-sql

Abstract: Recent advances in Text-to-SQL have significantly improved natural language interfaces to databases. Despite this progress, existing approaches still exhibit limitations of dependence on high-quality prompts and costly task-specific data, underscoring the need for more efficient and adaptable solutions. In this paper, we propose Progressive Reverse Understanding (\textbf{PRU}), a novel training paradigm that incrementally enhances Text-to-SQL generation by encouraging models to reason in reverse: from SQL back to natural language. PRU decomposes the learning process into progressive stages where models first learn to construct user intents from structured SQL, and then iteratively refine the mapping between language and structured queries. This bidirectional learning enables deeper semantic alignment while alleviating reliance on prompt quality and reducing the need for additional large-scale task-specific data by leveraging the data pairs inherent in the dataset. Extensive experimental results on the Spider dataset with open-source language models demonstrate the effectiveness of our approach in enhancing Text-to-SQL performance. Moreover, ablation studies highlight the critical role of progressive reverse understanding in improving both syntactic correctness and semantic fidelity.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6986