# Retrieval-augmented Chinese text-to-SQL generation for conversational bibliographic search

## TL;DR

This paper introduces a new conversational search system for Chinese bibliographic data that improves query accuracy using advanced language models and retrieval techniques.

## Contribution

The paper introduces BibSQL, a new Chinese Text-to-SQL dataset, and a two-stage conversational system enhanced with semantic retrieval and a novel prompting strategy.

## Key findings

- The proposed system achieves up to 96.6% execution accuracy using retrieval-augmented generation.
- SoftSimMatch improves semantic alignment and SQL accuracy over zero-shot prompting and random example selection.
- The PoT strategy and self-correction mechanism boost accuracy in low-resource conditions.

## Abstract

To overcome the limitations of current bibliographic search systems, such as low semantic precision and inadequate handling of complex queries, this study introduces a novel conversational search framework for the Chinese bibliographic domain. Our approach makes several contributions. We first developed BibSQL, the first Chinese Text-to-SQL dataset for bibliographic metadata. Using this dataset, we built a two-stage conversational system that combines semantic retrieval of relevant question-SQL pairs with in-context SQL generation by large language models (LLMs). To enhance retrieval, we designed SoftSimMatch, a supervised similarity learning model that improves semantic alignment. We further refined SQL generation using a Program-of-Thoughts (PoT) prompting strategy, which guides the LLM to produce more accurate output by first creating Python pseudocode. Experimental results demonstrate the framework’s effectiveness. Retrieval-augmented generation (RAG) significantly boosts performance, achieving up to 96.6% execution accuracy. Our SoftSimMatch-enhanced RAG approach surpasses zero-shot prompting and random example selection in both semantic alignment and SQL accuracy. Ablation studies confirm that the PoT strategy and self-correction mechanism are particularly beneficial under low-resource conditions, increasing one model’s exact matching accuracy from 74.8% to 82.9%. While acknowledging limitations such as potential logic errors in complex queries and reliance on domain-specific knowledge, the proposed framework shows strong generalizability and practical applicability. By uniquely integrating semantic similarity learning, RAG, and PoT prompting, this work establishes a scalable foundation for future intelligent bibliographic retrieval systems and domain-specific Text-to-SQL applications.