Skip to content

The code for "Past, Present, and Future: Conversational Emotion Recognition through Structural Modeling of Psychological Commonsense Knowledge" plus the code of models in "A Hierarchical Transformer with Speaker Modeling for Emotion Recognition in Conversations"

Notifications You must be signed in to change notification settings

LeqsNaN/SKAIG-ERC

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

50 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Requirements

  • Pytorch==1.8.1
  • Transformers==4.3.3
  • Torch-geometric==1.7.0

Additional Data

Edge attributes (extract files and put them into bert_data/DATASET_NAME/):
IEMOCAP
DailyDialog
EmoryNLP
MELD

We additionally provide the utterance representations and edge attributes of MELD to perform the training style like COSMIC.
The data is in here meld. Extract the files and put them in cosmic_data/meld/

Training

Training SKAIG models on datasets can use:
IEMOCAP: python gnn_train.py -index 1 -lr 1e-5 -choice 'cn' -pretrain 'roberta-base' -cn_num_layer 5 -sent_dim 300 -cn_nhead 6 -edge_dim 300 -cn_ff_dim 600 -cn_dropout 0.1 -hip 7 -seed 7 -residual_type 'none' -dataset_name 'IEMOCAP'

DailyDialog: python gnn_train.py -index 1 -lr 1e-5 -choice 'cn' -pretrain 'roberta-large' -cn_num_layer 5 -sent_dim 300 -cn_nhead 6 -edge_dim 300 -cn_ff_dim 600 -cn_dropout 0.1 -hip 2 -seed 7 -residual_type 'none' -dataset_name 'DailyDialog'

EmoryNLP: python gnn_train.py -index 1 -lr 1e-5 -choice 'cn' -pretrain 'roberta-large' -cn_num_layer 5 -sent_dim 300 -cn_nhead 6 -edge_dim 300 -cn_ff_dim 600 -cn_dropout 0.1 -hip 4 -seed 7 -residual_type 'none' -dataset_name 'EmoryNLP'

MELD: python gnn_train.py -index 1 -lr 8e-6 -choice 'cn' -pretrain 'roberta-large' -cn_num_layer 2 -sent_dim 200 -cn_nhead 4 -edge_dim 200 -cn_ff_dim 200 -cn_dropout 0.1 -hip 2 -seed 7 -residual_type 'none' -dataset_name 'MELD'

For COSMIC style training, the command is:
python cosmic_style_train.py -index 1 --edge_mapping --shuffle --beta --root_weight --use_gpu -dataset_name 'meld' -batch_size 8 -sent_dim 200 -ff_dim 200 -heads 4 -dropout 0.5 -num_layers 2 -edge_dim 200 -hop 3 -seed 7 -lr 1e-4 -l2 3e-4

TRMSM

The model is inserted into MentalModel, which can be activated by -choice 'tr' in gnn_train.py. Additionally, the hyper-params for it is started with 'tr_'.
The commands for training TRMSM: Pending.

About

The code for "Past, Present, and Future: Conversational Emotion Recognition through Structural Modeling of Psychological Commonsense Knowledge" plus the code of models in "A Hierarchical Transformer with Speaker Modeling for Emotion Recognition in Conversations"

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages