FROM NATURAL LANGUAGE TO SQL: A SURVEY ON LLM-BASED TEXT-TO-SQL RESEARCH AND APPLICATIONS

Abstract

Transforming natural language queries into precise SQL queries, a process referred to as Text-to-SQL or Natural Language to SQL (NL-to-SQL), continues to be a complex and challenging problem. This is due to the fact that interpreting user intent, supporting various database schema, and producing proper SQL syntax are challenging tasks. The application of conventional methods that involve rule-based methods and neural networks has improved significantly, and Pre-trained language models (PLMs) have helped things move forward much faster. However, as the complexity of the natural language query and database schema increases, PLMs with limited parameter numbers lose their accuracy. Such constraints have led to the development of more advanced and domain-specific optimization methods, which in turn limit their generalizability.In recent years, Large language models (LLMs) have shown phenomenal ability to process natural language and offer promising new directions for improving text-to-SQL systems. This survey presents an extensive overview of methods employing LLMs for text-to-SQL translation. We first outline the history and key technical challenges in the area and then survey key datasets like Spider, WikiSQL, BIRD, and CoSQL, which have played a major role in evaluation efforts. We then survey recent advances, with focus on the role of LLMs and their influence on system effectiveness. Through the identification of recent trends and open research challenges, this paper hopes to guide and stimulate future research towards improving the development of more generalizable and reliable text-to-SQL system based on LLMs.