A neural local coherence model trained in a multi-task learning scenario where dialogue act prediction is used as an auxilary task. Special thanks to my student Sebastian Bücker for implementing many parts of my ideas for this project. We appreciate your interest to this project. Please cite this paper if you use the code in this repository.
Also don't forget to give the repo a Github star (on top right). Thanks.

• platform: linux-64
• anaconda: 2019.07
• conda: 4.7.10
• Python: 3.6.8
• GCC 7.3.0

You can install all required packages as follows:

conda create --name dicoh --file conda_env.txt

conda activate dicoh

pip install -r pip_spec.txt

We conduct our experiments on two English dialogue corpora:

• DailyDialog

The dataset can be downloaded from here or from here

However, we do it in the exec_dataset_creation.sh script.

• SwitchBoard

You can get the data and scripts for processing SwitchBoard from here. The exec_dataset_creation.sh script will get them automatically.

bash exec_dataset_creation.sh

python mtl_coherency.py --logdir logs --seed $RANDOM --datadir data/daily_dialog --task up --do_train --do_eval --model model-3 --loss mtl --cuda 0

This project is licensed under the terms of the MIT license.

Mohsen Mesgar, Sebastian Bücker, and Iryna Gurevych. ACL 2020.