MuCDN

• Code for COLING 2022 accepted paper titled "MuCDN: Mutual Conversational Detachment Network for Emotion Recognition in Multi-Party Conversations"
• Weixiang Zhao, Yanyan Zhao, Bing Qin.

Requirements

• Python 3.7
• PyTorch 1.8.2
• Transformers 4.12.3
• CUDA 11.1

Preparation

Discourse Parsing

We use a powerful dialogue discourse parser to obtain the structure for the explicit detachment.

Details for training the parser could be found in https://github.com/shizhouxing/DialogueDiscourseParsing. We have already processed the ERC data in the same formmat with the input of the paser, which is in the folder of erc_data.

With the trained parser, run main_inference.py to get the parsed ERC data.

You can also direcly use the preprocessed features in the following.

Preprocessed Features

You can download the preprocessed features including dataset, extracted utterance feature and dialogue discourse structure we used from: https://pan.baidu.com/s/1gMIyK4mXVSvis1f1_DSQsQ 提取码:j84f

or from: https://drive.google.com/drive/folders/1-VzZjNmdZMO9rzzQD-JOt46w53fsBwJc?usp=sharing

and place them into the corresponding folds like emorynlp and meld

Training

You can train the models with the following codes:

For EmoryNLP: python train_emorynlp.py --hidden_dim 300 --pos

For MELD: python train_meld.py --hidden_dim 300 --pos --norm