Seq2Seq + QCFG

The code of our ACL2023 paper "Improving Grammar-based Sequence-to-Sequence Modeling with Decomposition and Constraints".

This is the dev code containing lots of unrelated and experimental code. We will clean this repo soon.

The code heavily uses two packages:

• pytorch-lightning for most of workflows, https://www.pytorchlightning.ai/
• hydra for configuration, https://hydra.cc/

and is based on the template:

• https://github.com/ashleve/lightning-hydra-template

You can create the environment by run

bash create_env.sh

You can find data in https://github.com/yoonkim/neural-qcfg, which is also the repo I initially forked from.

See README2.md, which is the readme file of the above template, for more examples.

Example

Run experiment on ATP using the reimplemented vanilla Neural QCFG:

python train.py experiment=styleptb

Run experiment on ATP using the E model:

python train.py experiment=styleptb_d1

Run experiment on ATP using the P model:

python train.py experiment=styleptb_d7