Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LitE-SQL: A Lightweight and Efficient Text-to-SQL Framework with Vector-based Schema Linking and Execution-Guided Self-Correction

Shengmin Piao, Jieun Lee, Sanghyun Park

TL;DR

LitE-SQL presents a lightweight Text-to-SQL framework that replaces heavy LLM reliance with a vector-based Schema Retriever and a two-stage SQL Generator trained via supervised fine-tuning and execution-guided refinement. By precomputing schema embeddings and using a hard-negative filtered contrastive loss, it achieves robust schema linking with a small model footprint. The approach, evaluated on BIRD and Spider 1.0, delivers competitive execution accuracy while using 2×–30× fewer parameters than many LLM-based methods, thanks to a closed-loop, execution-guided self-correction during inference. This work offers a privacy-preserving, resource-efficient path for Text-to-SQL in data-sensitive and constrained environments, with clear avenues for reducing false positives and semantic errors in future work.

Abstract

The Text-to-SQL task translates natural language questions into SQL queries, enabling intuitive database interaction for non-experts. While recent methods leveraging Large Language Models (LLMs) achieve strong performance, their reliance on proprietary models raise concerns about deployment feasibility and data privacy. In this work, we introduce LitE-SQL, a Lightweight and Efficient framework with two components: (i) a Schema Retriever that performs efficient schema linking using a vector database of pre-computed schema embeddings, and (ii) a SQL Generator fine-tuned in two stages-supervised fine-tuning followed by execution-guided reinforcement-enabling self-correction without costly multi-candidate generation. On BIRD, LitE-SQL achieves 72.10% execution accuracy, and on Spider 1.0 it reaches 88.45%, demonstrating comparable or superior performance to LLM-based methods despite using 2x to 30x fewer parameters. Our findings demonstrate that high-quality Text-to-SQL generation is feasible with lightweight models, offering a practical solution for privacy-sensitive and resource-constrained settings.

LitE-SQL: A Lightweight and Efficient Text-to-SQL Framework with Vector-based Schema Linking and Execution-Guided Self-Correction

TL;DR

LitE-SQL presents a lightweight Text-to-SQL framework that replaces heavy LLM reliance with a vector-based Schema Retriever and a two-stage SQL Generator trained via supervised fine-tuning and execution-guided refinement. By precomputing schema embeddings and using a hard-negative filtered contrastive loss, it achieves robust schema linking with a small model footprint. The approach, evaluated on BIRD and Spider 1.0, delivers competitive execution accuracy while using 2×–30× fewer parameters than many LLM-based methods, thanks to a closed-loop, execution-guided self-correction during inference. This work offers a privacy-preserving, resource-efficient path for Text-to-SQL in data-sensitive and constrained environments, with clear avenues for reducing false positives and semantic errors in future work.

Abstract

The Text-to-SQL task translates natural language questions into SQL queries, enabling intuitive database interaction for non-experts. While recent methods leveraging Large Language Models (LLMs) achieve strong performance, their reliance on proprietary models raise concerns about deployment feasibility and data privacy. In this work, we introduce LitE-SQL, a Lightweight and Efficient framework with two components: (i) a Schema Retriever that performs efficient schema linking using a vector database of pre-computed schema embeddings, and (ii) a SQL Generator fine-tuned in two stages-supervised fine-tuning followed by execution-guided reinforcement-enabling self-correction without costly multi-candidate generation. On BIRD, LitE-SQL achieves 72.10% execution accuracy, and on Spider 1.0 it reaches 88.45%, demonstrating comparable or superior performance to LLM-based methods despite using 2x to 30x fewer parameters. Our findings demonstrate that high-quality Text-to-SQL generation is feasible with lightweight models, offering a practical solution for privacy-sensitive and resource-constrained settings.

Paper Structure

This paper contains 27 sections, 10 equations, 5 figures, 14 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Schema Linking
  4. SQL Generation
  5. Methodology
  6. Schema Retriever
  7. SQL Generator
  8. Inference
  9. Experimental Setup
  10. Dataset
  11. Baselines
  12. Metrics
  13. Implementation Details
  14. Results and Analysis
  15. Baselines Comparison
  16. ...and 12 more sections

Figures (5)

  • Figure 1: Conceptual illustration of the difference between supervised contrastive (SupCon) loss and hard-negative filtered supervised contrastive (HN-SupCon) loss.
  • Figure 2: Two-stage training strategy of SQL Generator.
  • Figure 3: The inference pipeline of LitE-SQL.
  • Figure 4: Effect of the number of self-correction iterations across different question difficulty.
  • Figure 5: The distribution of execution error types before and after self-correction for each model size. The x-axis indicates the number of errors, and the number of each bar denotes the total errors.