Diverse Parallel Data Synthesis for Cross-Database Adaptation of Text-to-SQL Parsers (EMNLP 2022)

[WIP] This repository provides an implementation of experiments in our EMNLP-22 paper

@article{awasthi2022diverse,
  title={Diverse Parallel Data Synthesis for Cross-Database Adaptation of Text-to-SQL Parsers},
  author={Awasthi, Abhijeet and Sathe, Ashutosh and Sarawagi, Sunita},
  journal={arXiv preprint arXiv:2210.16613},
  year={2022}
}

Requirements

This code was developed with python 3.8.8.
Create a new virtual environment and install the dependencies by running pip install -r requirements.txt

Datasets

• Spider dataset: useful files copied in data/spider
• ReFill generated datasets: data/sql-to-text/refill/jsons/spider_groups

ReFill Pipeline

• Preprocessing: Apply Masking, Convert SQLs into Pseudo-English form, Pre-compute SQL-neighbours for train and val set

  bash scripts/data/refill_postprocess.sh
  

• Train BART model for ReFill

  bash scripts/sql-to-text/train_refill.sh
  

• Train Filtering model to filter out inconsistent SQL-Text pairs

  bash scripts/sql-to-text/train_filter.sh
  

• ReFill Inference: Find SQL-neighbours of the given workload, Apply Masking and ReFilling followed by filtering

  bash scripts/sql-to-text/infer_refill.sh
  

L2S Pipeline

This pipeline makes use of relative paths. It is recommended to change directory to scripts/sql-to-text/ first before running any script

• Preprocessing + Training: Convert SQLs into L2S encoding and train a Seq2Seq model

  bash train_l2s.sh
  

• L2S Inference: Use the trained SQL-to-Text Seq2Seq model to generate text for the given workload

  bash infer_l2s.sh
  

GAZP Pipeline

This pipeline makes use of relative paths. It is recommended to change directory to scripts/sql-to-text/ first before running any script

• Preprocessing + Training: Convert SQLs into GAZP encoding and train a Seq2Seq model

  bash train_gazp.sh
  

• GAZP Inference + Filtering: Use the trained SQL-to-Text Seq2Seq model to generate text for the given workload and use a forward Text-to-SQL parser for cycle consistency based filtering

  bash infer_gazp.sh
  

SnowBall Pipeline

This pipeline makes use of relative paths. It is recommended to change directory to scripts/sql-to-text/ first before running any script

• Preprocessing + Training: Convert SQLs into SnowBall encoding and train a Seq2Seq model

  bash train_snowball.sh
  

• SnowBall Inference: Use the trained SQL-to-Text Seq2Seq model to generate text for the given workload

  bash infer_snowball.sh