CoE-SQL

This is the project containing the source code for the NAACL2024 paper CoE-SQL: In-Context Learning for Multi-Turn Text-to-SQL with Chain-of-Editions in NAACL 2024 main conference. If you find it useful, please cite our work.

@misc{zhang2024coesql,
      title={CoE-SQL: In-Context Learning for Multi-Turn Text-to-SQL with Chain-of-Editions}, 
      author={Hanchong Zhang and Ruisheng Cao and Hongshen Xu and Lu Chen and Kai Yu},
      year={2024},
      eprint={2405.02712},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Run CoE-SQL

1. Create the data directory and move the downloaded datasets into this directory. Here is the example of the directory structure.

data
├── cosql
│   ├── database (directory)
│   ├── database-testsuite (directory)
│   ├── dev.json
│   ├── tables.json
│   └── train.json
└── sparc
    ├── database (directory)
    ├── database-testsuite (directory)
    ├── dev.json
    ├── tables.json
    └── train.json

2. Run edit.py to automatically generate the chain-of-editions for all examples in the train set. Here are two examples.

python edit.py --dataset sparc --max_len 4
python edit.py --dataset cosql --max_len 3

3. Run main.py to run CoE-SQL on the dev set. Here are two examples.

python main.py --dataset sparc --coe
python main.py --dataset cosql --coe