CGML

ML-TranX and ML-TranX'

These two models are based on a general-purpose Transition-based abstract syntaX parser ACL '18 paper and
TranX' is a variant of TranX which decodes in a breadth-first manner. By adopting a mutual learning based model training framework, both models can fully absorb the knowledge from each other and thus could be improved simultaneously.

System Architecture

The System Architecture is the same as TranX ACL '18 paper. Please refer to https://github.com/pcyin/tranX .

Usage

cd tranX
conda env create -f config/env/tranx-py2.yml
./scripts/django/train-mutual-share-embedding.sh 0  # train on django code generation dataset  with random seed 0

conda env create -f config/env/tranx.yml  # create conda Python environment. 
./scripts/atis/train-mutual-share-embedding.sh 0  # train on ATIS semantic parsing dataset
./scripts/geo/train-mutual-share-embedding.sh 0  # train on GEO dataset
./scripts/ifttt/train-mutual-share-embedding.sh 0  # train on IFTTT dataset

./scripts/django/test_used.sh    # modify the configuration to test the model on django code generation dataset
./scripts/atis/test_used.sh     # modify the configuration to test the model on atis code generation dataset
./scripts/geo/test_used.sh      # modify the configuration to test the model on geo code generation dataset
./scripts/ifttt/test_used.sh     # modify the configuration to test the model on ifttt code generation dataset