atmahou / ludus-jluva

Code for running Joint Language Understanding Variational Autoencoder (JLUVA)

Getting Started

Setting up the environment

Clone this repository recursively:

git clone https://github.com/kaniblu/ludus-jluva --recursive && cd ludus-jluva

Download and setup GloVe:

cd datasets/glove && bash setup.sh && cd ../..

Install required packages:

pip install -r requirements.txt && pip install -r codebase/requirements.txt

(Optional) Install additional packages as specified in codebase/README.md if required.

Setup environments:

export PATH=$(pwd)/scripts:${PATH}

Create the instance directory:

mkdir instances

Creating and running an experimental instance

Create an experimental instance from the archetype:

instance-create jluva atis-jluva

Configure the dataset option in the instance:

config-set atis-jluva data dataset atis -w

Train a generative model:

instance-run atis-jluva train

Generate NLU corpus using the trained model:

instance-run atis-jluva generate --eidx 97

After above steps, the synthetic corpus (utterances + intents + slot labels) will be generated under instances/atis-jluva with the filenames word.txt, intent.txt, and label.txt. Such corpus improves LU performances when used to augment the existing corpus for training supervised LU models.