EHR-SeqSQL : A sequential Text-to-SQL dataset for interactively exploring electronic health records

Abstract

Text-to-SQL parsing is a task that translates natural language into SQL, enabling users who are not database experts to retrieve information from databases using only natural language. There are several important yet under-explored objectives in this field: interactivity, compositionality, and efficiency. In this paper, we present EHR-SeqSQL, a sequential Text-to-SQL dataset for interactively exploring Electronic Health Record (EHR) databases. We demonstrate the benefits of multi-turn setting over single-turn setting with respect to compositionality, and provide a new data split and an additional test set to evaluate compositional generalization. Furthermore, we introduce unique special tokens in SQL queries to enhance execution efficiency. This study represents the first attempt in the Text-to-SQL parsing field to simultaneously consider interactivity, compositionality, and efficiency, aiming to narrow the gap between industrial demands and academic research.