CoMAE

Codes and data for the ACL 2021-Findings paper: CoMAE: A Multi-factor Hierarchical Framework for Empathetic Response Generation

If you have any problem or suggestion, feel free to contact me: chujiezhengchn@gmail.com

If you use our codes or your research is related to our paper, please kindly cite our paper:

@inproceedings{zheng-etal-2021-comae,
    title = "CoMAE: A Multi-factor Hierarchical Framework for Empathetic Response Generation",
    author = "Zheng, Chujie  and
      Liu, Yong  and
      Chen, Wei  and
      Leng, Yongcai  and
      Huang, Minlie",
    booktitle = "Findings of ACL 2021",
    year = "2021"
}

Data

You can download our propossed data from HuggingFace. However, the released data is a bit different from the used data in our paper.

• Data size. We found that a RoBERTa classifier may suffer from the unbalanced labels. Hence, for all the factors, we instead use BERT as classifiers. As a result, the filtered data based on CM have a larger size than that reported in our paper
• Taxonomies of DA and EM. We modified the adopted taxonomies of both DA and EM (please refer to the json files in this repo) because:
  • For DA, we found that suggestion is not categorized as a dialog act of expressed empathy (see the paper of CM). To keep consistent with the CM paper, we merged suggestion with others
  • For EM, we modified the taxonomies to reduce the overlaps between different emotions
  • Nevertheless, we think you can also modify the taxonomies as needed, and then automatically annotate the utterances

Performance of BERT-based classifiers

| Classifiers | # classes | Acc | F1-macro | | ----------- | --------- | ---- | -------- | | CM-ER | 2 | 80.5 | 76.9 | | CM-IP | 2 | 84.7 | 84.7 | | CM-EX | 2 | 96.8 | 93.6 | | DA | 8 | 91.4 | 85.9 | | EM | 9 | 65.8 | 62.8 |

Data Size

| Train | Valid | Test-Happy | Test-Offmychest | | ------ | ----- | ---------- | --------------- | | 154001 | 19940 | 13337 | 7827 |

Model Implementation

Please enter codes.