Contrast and Generation Make BART a Good Dialogue Emotion Recognizer

• torch==1.7.1
• transformers==4.11.0
• numpy
• pickle
• tqdm
• sklearn
• fitlog

Model runs on one GPU by default, and we didn't try it on CPU.

We recommend using GPU with memory more than 24G , otherwise you may need to adjust the hyperparameters and the results may vary significantly.

To run the model on test sets of four datasets,

1. Download the pre-trained models:

   For MELD, download the checkpoint , unzip it to ./save/MELD

   For EmoryNLP, download the checkpoint , unzip it to ./save/EmoryNLP

   For IEMOCAP, download the checkpoint , unzip it to ./save/IEMOCAP

   For DailyDialog, download the checkpoint , unzip it to ./save/DailyDialog

2. Execute the following command in terminal:

   For MELD: bash eval.sh MELD ./save/MELD/best_model_762985

   For EmoryNLP: bash eval.sh EmoryNLP ./save/EmoryNLP/best_model_468186

   For IEMOCAP: bash eval.sh IEMOCAP ./save/IEMOCAP/best_model_981929

   For DailyDialog: bash eval.sh DailyDialog ./save/DailyDialog/best_model_738497

For MELD: bash train.sh MELD

For EmoryNLP: bash train.sh EmoryNLP

For IEMOCAP: bash train.sh IEMOCAP

For DailyDialog: bash train.sh DailyDialog

It should be noticed that performance is greatly affected by random seed. So we recommended some seeds in the script for reproduction.

For MELD: bash eval.sh MELD save/MELD/best_model_939239

For EmoryNLP: bash eval.sh EmoryNLP save/EmoryNLP/best_model_552848

For IEMOCAP: bash eval.sh IEMOCAP save/IEMOCAP/best_model_625968

For DailyDialog: bash eval.sh DailyDialog save/DailyDialog/best_model_269130

If you find this work useful, please cite our work:

@misc{li2021contrast,
    title={Contrast and Generation Make BART a Good Dialogue Emotion Recognizer}, 
    author={Shimin Li and Hang Yan and Xipeng Qiu},
    year={2021},
    eprint={2112.11202},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Some code of this project are referenced from TodKat and DialogXL. We thank their open source materials for contribution to this task.