Fig 1.
A multi-hop example from the BibSQL dataset.
Table 1.
Examples of question templates and their corresponding query paths in the BibSQL dataset.
Fig 2.
Distribution of questions and queries in the BibSQL dataset.
These histograms show the length distributions for (a) natural language questions and (b) SQL queries.
Table 2.
Detailed statistics of the BibSQL dataset.
Fig 3.
End-to-end workflow of the proposed retrieval-augmented Text-to-SQL framework.
Fig 4.
The overall process of the SoftSimMatch framework.
Table 3.
Evaluation of different example selections on BibSQL-test.
Table 4.
Comparative evaluation of LLMs on BibSQL-test.
Performance is evaluated using various Retrieval-Augmented Generation (RAG) methods, where best results are in bold and second-best are underlined. These methods include using the top 1, 2, or 5 most similar questions (RAG-top1/2/5) and a random dataset sample (RAG-random1) for context.
Table 5.
Comparative evaluation of LLMs on BibSQL-test with small training set.
Table 6.
Detailed accuracy of SQL query components on BibSQL-test with small training set.
Table 7.
Error analysis of the proposed system on BibSQL-test with small training set.