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

Area of Science:

• Information Science
• Computer Science
• Artificial Intelligence

Background:

• Current bibliographic search systems lack semantic precision and struggle with complex queries.
• There is a need for advanced frameworks to improve bibliographic metadata retrieval.

Purpose of the Study:

• To develop a novel conversational search framework for the Chinese bibliographic domain.
• To enhance semantic precision and improve handling of complex queries in bibliographic searches.

Main Methods:

• Developed BibSQL, the first Chinese Text-to-SQL dataset for bibliographic metadata.
• Built a two-stage conversational system combining semantic retrieval and LLM-based SQL generation.
• Designed SoftSimMatch for supervised similarity learning and employed Program-of-Thoughts (PoT) prompting for SQL generation.

Main Results:

• The retrieval-augmented generation (RAG) approach achieved up to 96.6% execution accuracy.
• SoftSimMatch-enhanced RAG outperformed zero-shot prompting and random example selection in semantic alignment and SQL accuracy.
• PoT strategy and self-correction improved exact matching accuracy from 74.8% to 82.9% in low-resource settings.

Conclusions:

• The proposed framework demonstrates strong generalizability and practical applicability for intelligent bibliographic retrieval.
• The integration of semantic similarity learning, RAG, and PoT prompting establishes a scalable foundation for future Text-to-SQL applications.
• The study addresses limitations of current systems, paving the way for more sophisticated domain-specific search tools.