Source code for ICASSP 2022 paper "MM-DFN: Multimodal Dynamic Fusion Network For Emotion Recognition in Conversations".

In this work, we focus on emotion recognition in multimodal conversations (multimodal ERC). If you are interested in textual ERC, you can refer to a related work DialogueCRN (code).

• python 3.6.10
• torch 1.4.0
• torch-geometric 1.4.3
• torch-scatter 2.0.4
• scikit-learn 0.21.2
• CUDA 10.1

Install related dependencies:

pip install -r requirements.txt

The original datasets can be found at IEMOCAP and MELD.

Following previous works (DialogueRNN, MMGCN, et al.), raw utterance-level features of textual, acoustic, and visual modality are extracted by TextCNN with Glove embedding, OpenSmile, and DenseNet, respectively. The processed features can be found by the link.

For training model on IEMOCAP and MELD datasets, you can refer to the following:

# IEMOCAP dataset
bash ./script/run_train_ie.sh
# MELD dataset
bash ./script/run_train_me.sh

Note: The optimal hyper-parameters (e.g., the number of gcn layers) are slight differences under different experimental configurations (i.e., the version of CUDA and PyTorch). To facilitate further research by interested parties, we retain the complete code including ablation and control experiments.

Results of MM-DFN on IEMOCAP dataset:

Results of MM-DFN on MELD dataset:

Evaluation matrices: Acc, macro-F1, weighted-F1, and the F1 score per class.

@inproceedings{DBLP:conf/icassp/HuHWJM22,
  author    = {Dou Hu and
               Xiaolong Hou and
               Lingwei Wei and
               Lian{-}Xin Jiang and
               Yang Mo},
  title     = {{MM-DFN:} Multimodal Dynamic Fusion Network for Emotion Recognition
               in Conversations},
  booktitle = {{ICASSP}},
  pages     = {7037--7041},
  publisher = {{IEEE}},
  year      = {2022}
}