woz_extract.py
allows to parse the entire MultiWoZ dataset and extract useful statistics including the information to split the dataset as done in the paper.
adding_noise.py
allows adding targertted and random noise to a specific file combination. The same snippet can also be used to inject changephrase between synthetic mutlitask dialogues.
If you find this work useful and use it in your own research, please consider citing our [paper](link soon).
@misc{parthasarathi2020evaluate,
title={On Task-Level Dialogue Composition of Generative Transformer Model},
author={Prasanna Parthasarathi and Arvind Neelakantan and Sharan Narang},
year={2020},
booktitle = {Workshop on Insights from Negative Results at EMNLP}
}