This is a PyTorch implementation of the PG-GSQL in our COLING 2020 paper "PG-GSQL: Pointer-Generator Network with Guide Decoding for Cross-Domain Context-Dependent Text-to-SQL Generation".

Run SParC experiment

Requirements

You can refer to requirements.txt.

Pretrained BERT

You need download the pretrained BERT from here and put them in /model/bert/data/.

Dataset

Two options are available

1. You can get dataset from https://github.com/taoyds/sparc and put them in the /data/ folder, then run the python3 preprocess_data.py --dataset sparc to preprocess the data.
2. You can use our preprocessed data and download the database from here.

Run model

Train sh ./run_sparc_pg_gsql.sh

Eval sh ./eva_att.sh

Reproduce our model

1. You need download our trained model from here and put it in /sparc_pg_gsql_paper_save/.
2. change the dir in evaluate_g.py and run sh ./eva_att.sh.
3. You can get the performance on the dev set.

	question matching	interaction matching
PG-GSQL	53.1	34.7

Reference

https://github.com/taoyds/sparc

https://github.com/ryanzhumich/editsql

https://github.com/lil-lab/atis

Our paper bibtex

@inproceedings{wang-etal-2020-pg, title = "{PG}-{GSQL}: Pointer-Generator Network with Guide Decoding for Cross-Domain Context-Dependent Text-to-{SQL} Generation", author = "Wang, Huajie and Li, Mei and Chen, Lei", booktitle = "Proceedings of the 28th International Conference on Computational Linguistics", month = dec, year = "2020", address = "Barcelona, Spain (Online)", publisher = "International Committee on Computational Linguistics", url = "https://www.aclweb.org/anthology/2020.coling-main.33", doi = "10.18653/v1/2020.coling-main.33", pages = "370--380", }