Rethinking data in NL2SQL: a survey of what we have and what we expect

Abstract Natural Language to SQL (NL2SQL) has become a cornerstone task for enabling natural language interfaces to relational databases. With the emergence of large language models, NL2SQL systems have achieved remarkable performance gains. However, despite the focus on architectural innovations and benchmark achievements, we argue that NL2SQL is fundamentally a data-centric task — where the quality, structure, and utilization of data play a more critical role than often acknowledged. In this survey, we re-examine the NL2SQL landscape through the lens of how data are used throughout the system pipeline. Specifically, we offer a brief overview of the task challenges and evolutionary process of NL2SQL. Next, we categorize the major data types and analyze how these data sources are leveraged throughout the NL2SQL lifecycle. We then introduce the datasets and metrics used to evaluate NL2SQL systems. Finally, we highlight the remaining challenges and outline promising directions for future research. We hope that this survey can serve as a quick reference to existing work and motivate future research from a data perspective.