Source code to reproduce results of our paper "DIET: Lightweight Language Understanding for Dialogue Systems".

In order to reproduce the experiments results, execute the following steps:

(1) We used Rasa for running the experiments. You first need to clone the repository, checkout the branch diet-paper and install Rasa.

# clone the repository and checkout 'diet-paper' branch
git clone https://github.com/RasaHQ/rasa
cd rasa
git checkout diet-paper
# installation
curl -sSL https://raw.githubusercontent.com/python-poetry/poetry/master/get-poetry.py | python3
make install-full

(2) Execute the run.sh script to train and test models on ATIS, SNIPS, and NLU Evaluation Data. You need to specify what results you want to reproduce by specifying the table from the paper. Make sure you are at the root directory of this repository before executing the script. Consider executing the experiments on a machine with a GPU to speed up the experiments.

./run.sh [table-1|table-2|table-3|table-4|table-5]

(3) The experiments results can be found in the folder experiments.

To test other featurization or hyperparameters update the configuration files in the configs folder. Available featurization components and a list of available hyperparameters can be found here.