Multi-turn-Dialogue-Response

Install requirements.txt

conda install --yes --file requirements.txt

Pre-processing

• Download Pretrained GloVe Embeddings and save it in /vectors.
• The preprocessed dataset is saved as /data/ED/dataset_preproc.p. If you want to create the dataset yourself or change the knowledge types generated by COMET, delete this file, download the COMET checkpoint and place it in /data/ED/Comet. The preprocessed dataset would be generated after the training script. Here, we use BART, since the GPT-2 version can not be used.

Dataset

In this study we use two different dataset. In this repository, we set the config and mapping to EmpatheticDialogues.

• EmpatheticDialogues
• DailyDialog

Data Resource

• EmpatheticDialogues
• DailyDialog

Data Structure

The data folder is organized as follows:

data
|--- ED
|    |--- Comet
|    |--- emp.pkl
|    |--- train.csv
|    |--- valid.csv
|    |--- test.csv
|    
|--- DD
|    |--- Comet
|    |--- dd.pkl
|    |--- train.csv
|    |--- valid.csv
|    |--- test.csv

Directory and File Descriptions

• data/: The root directory containing all data files.

  • ED/: This directory contains files for the "EmpatheticDialogues" dataset.

    • Comet/: A subdirectory containing the "Comet" package used to generate common-sense knowledge.

    • emp.pkl: A pickle file containing topic appearance probabilities for the "EmpatheticDialogues" dataset.

    • train.csv: The training data file for the "EmpatheticDialogues" dataset.

    • valid.csv: The validation data file for the "EmpatheticDialogues" dataset.

    • test.csv: The testing data file for the "EmpatheticDialogues" dataset.

  • DD/: This directory contains files for the "DailyDialog" dataset.

    • Comet/: A subdirectory containing the "Comet" package used to generate common-sense knowledge.

    • dd.pkl: A pickle file containing topic appearance probabilities for the "DailyDialog" dataset.

    • train.csv: The training data file for the "DailyDialog" dataset.

    • valid.csv: The validation data file for the "DailyDialog" dataset.

    • test.csv: The testing data file for the "DailyDialog" dataset.

Training

python main.py --cuda --save_path save/your_dir

Testing

python main.py --cuda --test --save_path save/your_dir --model_path save/dir_save/KITM_XXXX.XXX

Please be free to contact with us via [email protected]