CAB: Empathetic Dialogue Generation with Cognition, Affection and Behavior

This is the official implementation for paper CAB: Empathetic Dialogue Generation with Cognition, Affection and Behavior.

Model Architecture

Setup

• Check the packages needed or simply run the command:

pip install -r requirements.txt

• Download GloVe vectors from here (glove.840B.300d.txt) and put it into /data/.

• Download the completely processed dataset data from Google Drive and place it into /data/ for experiments.

• If you want to see the annotated dataset with dialogue act label and both interlector's emotion lables, then you could:

  • Download the processed EmpatheticDialogues dataset with dialogue act label and both interlector's emotion lables from here and place processed dataset train.json, valid.json and test.json into /data/ed_data/.

• If you want to reconstruct knowledge paths, then you could:

  • Download the processed ConceptNet data and place processed data ConceptNet_ranked_dict.json into /data/knowledge_data/, meanwhile, download dataset_preproc.p and place it into /data/.

• For reproducibility purposes, we place the model checkpoints. You could download and move it under /save/final/.

• To skip training, please check folder /result/CAB/output.txt/.

Training

CAB (Our)

python main.py \
--cuda \
--batch_size 16 \
--lr 1e-4 \
--hidden_dim 300 \
--emo_dim 300 \
--act_dim 300 \
--latent_dim 200\
--hop 1 \
--heads 2 \
--pretrain_emb \
--model CAB \
--multi_hop 5 \
--K_num 5 \
--k_num 3 \
--path_num 15 \
--pointer_gen \
--emb_file data/glove.840B.300d.txt

Testing

Add --test into above commands.

You can directly run /evaluate_result.py script to evaluate the model predictions.

Citation

If you find our work useful, please cite our paper as follows:

@article{CAB2023,
      title={CAB: Empathetic Dialogue Generation with Cognition, Affection and Behavior}, 
      author={Pan Gao, Donghong Han, Rui Zhou, Xuejiao Zhang, Zikun Wang},
      journal={arXiv preprint arXiv:2302.01935},
      year={2023},
}