DeCOCO

DEtecting emotion shift in COnversations by COmmonsense knowledge

To set up project go through the following stepd:

• Extract Features by using RoBERTa and COMET. Due to complicated steps to run RoBERTa and COMET you can download the result of step 1-6 from here
  1. You can follow step 1-3 or using tutorial of here
  2. Preprocess the data: Data should be preprocessed following the language modeling format and use GPT-2 BPE (Byte Pair Encoding) tokenizer. conv-emotion\COSMIC\feature-extraction\roberta_preprocess_iemocap.sh
  3. Train RoBERTa base: conv-emotion\COSMIC\feature-extraction\roberta_preprocess_iemocap.sh conv-emotion\COSMIC\feature-extraction\roberta_train_iemocap.sh
  4. Load your pretrained model: conv-emotion\COSMIC\feature-extraction\roberta_feature_extract_iemocap.py
  5. Clone COMET project from here
  6. Download pretrained atomic model from here to extract common-sense features.
• Keep the result of prior step in train/erc-training
• To run the project, install prerequisites
  1. python 3.8.8
  2. torch 1.8.1
  3. scikit-learn 0.24.1
  4. plotly 4.14.3
  5. pandas 1.2.3
  6. numpy 1.19.5
  7. notebook 6.3.0
  8. ipython 7.21.0
  9. tqdm 4.63.0
• Training, validation, testing, and evaluation of each dataset, have be done by the followin steps:
  • python train_iemocap.py
  • python train_meld.py
  • python train_dailydialog.py
  • python train_emorynlp.py
• You can change the default value of hyperparameter and argument. There are different hyperparameters and argumens you can change by the following stepa:
  • no-cuda: Run model on GPU or CPU
  • lr: Learning rate
  • l2: L2 regularization weight
  • recurrent_dropout: recurrent_dropout of GRU
  • dropout: Dropout rate for hidden layer
  • batch-size
  • epochs
  • class-weight: Use class weight form imbalanced dataset
  • active-listener: This is Monologue and Dialogue
  • attention: Attention type in context GRU
  • seed
  • norm: normalization strategy
• You can pass the value to each dataset like this: python train_iemocap.py --no-cuda --epochs=50 --class-weight --active-listener
• The result will be saved in train/logs