# Decomposition for Enhancing Attention: Improving LLM-based Text-to-SQL through Workflow Paradigm

**Authors:** Yuanzhen Xie, Xinzhou Jin, Tao Xie, MingXiong Lin, Liang Chen, Chenyun Yu, Lei Cheng, ChengXiang Zhuo, Bo Hu, Zang Li  
**arXiv:** [2402.10671](https://arxiv.org/abs/2402.10671)  
**Date:** 2024-02-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.10671v3/

## Key points

- Authors propose a novel approach titled "Decomposition for Enhancing Attention" to address limitations faced by large-language models (LLMs) in complex tasks like text-to-SQL.
- The approach introduces a workflow paradigm method that aims to improve attention and problem-solving scope through decomposition.
- Method includes an information determination module to eliminate redundant information and a new prompt structure based on problem classification to enhance the model's attention.
- Inclusion of self-correction and active learning modules significantly expands the problem-solving capabilities of LLMs.
- Extensive experiments show that the approach outperforms existing methods by achieving about 2-3 percentage point improvements compared to baseline results, setting new state-of-the-art results on the Spider Test dataset.
- The refined approach not only enhances attention and problem-solving scope but also pushes the upper limit of LLM-based approaches in text-to-SQL tasks.

## Summary

Authors Yuanzhen Xie, Xinzhou Jin, Tao Xie, MingXiong Lin, Liang Chen, Chenyun Yu,
Lei Cheng, ChengXiang Zhuo, Bo Hu and Zang Li have proposed a novel approach titled "Decomposition for Enhancing Attention: Improving LLM-based Text-to-SQL through Workflow Paradigm" to address the limitations faced by large-language models (LLMs) in complex tasks like text-to-SQL. The existing single-step chain-of-thought prompting approach often encounters challenges such as attention diffusion and inadequate performance. To enhance the contextual learning capabilities of LLMs in text-to-SQL tasks, the authors introduce a workflow paradigm method that aims to improve attention and problem-solving scope through decomposition. The proposed method includes an information determination module designed to eliminate redundant information and a new prompt structure based on problem classification to enhance the model's attention. Additionally, the inclusion of self-correction and active learning modules significantly expands the problem-solving capabilities of LLMs. Through extensive experiments conducted on three datasets (Spider Dev, Spider-Realistic, Bird Dev), the authors demonstrate that their approach outperforms existing methods by achieving about 2-3 percentage point improvements compared to baseline results. Moreover, their method sets new state-of-the-art results on the Spider Test dataset. This refined approach not only enhances the attention and problem-solving scope of LLMs but also pushes the upper limit of LLM-based approaches in text-to-SQL tasks. The code for this research is available on GitHub at https://github.com/FlyingFeather/DEA-SQL.