Xiaohui: A Text-to-SQL Application Designed for Non-technical Users

ACL ARR 2025 February Submission1863 Authors

Abstract: Text-to-SQL, a critical task in natural language processing, aims to translate natural language questions into structured SQL queries for relational databases. Despite significant advancements, existing approaches often struggle with complex queries and domain adaptation due to the inherent ambiguity and variability of natural language. In this paper, a user-centered text-to-SQL system that reduces the usage barrier for non-technical users by abstracting the intermediate language of keyword instructions is designed. The key point is to allow users to independently determine whether the predicted query and execution results are acceptable, rather than relying on the complex SQL statements of traditional methods. Specifically, our implementation, Xiaohui, first designs keyword-to-Trino SQL mapping rules based on the real needs of partner enterprises, such as growth trends, comparisons, and percentage situations commonly found in business analysis. Secondly, it utilizes a LLM-based method fine-tuned with over 9,000 text-to-keyword samples and 1,276 question-query types-keyword examples for reference. On a test dataset designed based on the actual needs of enterprises, using Qwen as the basic model, the proposed system achieves evaluation performance comparable to or even surpassing mainstream prompt-based methods like DAIL-SQL and DIN-SQL.

Paper Type: Long

Research Area: Human-Centered NLP

Research Area Keywords: user-centered design, human-centered evaluation, human-AI interaction

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources

Languages Studied: Chinese

Submission Number: 1863