This is the official codebase of the paper
The main idea of ACQD is to alleviate the difficulty of encoding and decoding procedures in NL2SQL by decomposing a complex question into multiple simple questions, then using decoding results of simple questions to recover the result of original complex question. ACQD is consist of five modules: Intermediate Representation Completion framework (IRC), complex query decomposition, syntactic dependency modeling, grammar-based decoder and schema dependency learning.
This codebase is based on PyTorch. It supports training and inference with multiple GPUs or multiple machines.
You may install the dependencies via either conda or pip. Generally, ACQD works with Python 3.7/3.8 and PyTorch version >= 1.8.0.
Please run as follows to install all the dependencies:
pip3 install -r requirements.txt
Please go to folder IRC to see details of pre-training and evaluating of IRC.
These codes are about IRC of ACQD. More codes about ACQD are preparing, and we will upload it as soon as we can.
These codes is modified by grappa. Thanks for taoyus's contribution.