EmpSOA

The official implementation for the Findings of ACL 2023 paper Don't Lose Yourself! Empathetic Response Generation via Explicit Self-Other Awareness.

Requirements

Python 3.7
PyTorch 1.8.2
Transformers 4.12.3
CUDA 11.1

Preparation

Download Pretrained GloVe Embeddings and save it in /vectors.

The processed dataset and generated commonsense knowledge can be downloaded from (https://drive.google.com/drive/folders/1zr6xlD9ryCmVx-P5REfO1Run4aT069aT?usp=share_link). Then move it to /data.

We have tried two ways to use the commonsense knowledge: hidden features and decoded texts. And the former one can achieve better results.
If you want to generate the hidden features:

cd comet-atomic-2020/models/comet_atomic2020_bart

Download the pretrained COMET model in download_model.sh.

Then run generate_knowledge.py to get commonsense knowledge features.

python generate_knowledge.py

or you can use the decoded commonsense texts following the dataset processing step. And the preprocessed dataset and decoded commonsense texts would be generated after the training script.

Training

python main.py --cuda --save_decode [--wo_sog] [--wo_som] [--wo_sod] [--only_user] [--only_agent] [--wo_dis_sel_oth]

The extra flags can be used for ablation studies.

Evaluation

Create a folder results and move the obtained results.txt to this folder:

python src/scripts/evaluate.py

Citation

If you find our work useful for your research, please kindly cite our paper as follows:

@article{zhao2022don,
  title={Don't Lose Yourself! Empathetic Response Generation via Explicit Self-Other Awareness},
  author={Zhao, Weixiang and Zhao, Yanyan and Lu, Xin and Qin, Bing},
  journal={arXiv preprint arXiv:2210.03884},
  year={2022}
}

Credits

The code of this repository partly relies on CEM and I would like to show my sincere gratitude to authors of it.

circle-hit/EmpSOA