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

🔥🔥 2024.05. DEA-SQL is accepted by Findings of ACL 2024!

Set Up

Environment

# 1. Clone the repo
git clone https://github.com/FlyingFeather/DEA-SQL.git
cd DEA-SQL && mkdir data
# 2. Make a conda environment
conda create -n deasql python=3.9
conda activate deasql
# 3. Install requirements
pip install -r requirements.txt
python nltk_downloader.py

Dataset

mkdir data
unzip spider.zip -d data

.
├── argsparser.py
├── common
├── correct_sql.py
├── data
│  └── spider
│		├── ...
│		└── database
├── data_preprocess.py
├── docs
├── evaluation
├── fewshot
├── filter_characters.py
├── gen_sql.py
├── get_ner.py
├── hardness_eval.py
├── __init__.py
├── LICENSE
├── llm
├── logger.py
├── main.py
├── nltk_downloader.py
├── outputs
├── prompt
├── README.md
├── requirements.txt
└── single_eval.py

Usage

Quick Start

prediction on the Spider Dev datasets

python main.py --save_file_name "dea-sql.txt" --dataset "spider" --mode "dev" --sample "False" --few_shot_mode "masked_ques_sim" --insert_value 3 --embedding_base_model "openai" --sc_filter_nums 3 --few_shot_data "train_merge_v5"

evaluation on the Spider Dev datasets

python evaluation/test-suite-sql-eval/evaluation.py --gold "evaluation/gold_files/spider_dev_gold.sql" --pred "outputs/spider/dea-sql.txt" --db ./data/spider/database --print_file_name "outputs/spider/spider-dea-sql.txt" --table './data/spider/tables.json' --etype exec

Citing DEA-SQL

@article{xie2024decomposition,
 title={Decomposition for Enhancing Attention: Improving LLM-based Text-to-SQL through Workflow Paradigm}, 
 author={Yuanzhen Xie and Xinzhou Jin and Tao Xie and MingXiong Lin and Liang Chen and Chenyun Yu and Lei Cheng and ChengXiang Zhuo and Bo Hu and Zang Li},
 journal={arXiv preprint arXiv:2402.10671},
 year={2024}
}