The official repository for ACL-IJCNLP 2021 paper "DialogueCRN: Contextual Reasoning Networks for Emotion Recognition in Conversations". In this work, we propose a cognitive-inspired network that uses multi-turn reasoning modules to capture implicit emotional clues in conversations.

There are some recent works for the ERC task:

• MM-DFN: We propose a graph-based method named MM-DFN that combines multimodal knowledge and contextual features for the ERC task. The paper can be found here and the code is available on here. This work has been accepted by [ICASSP 2022].
• SACL-LSTM: We introduce a new sequence-based method named SACL-LSTM, which achieves better performance for the ERC task. This method combines the contextual perception module of DialogueCRN with an adversarial contrastive learning technique. The paper can be accessed here and the code is available on here. This work has been accepted by [ACL 2023].

• 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, SEMAINE and MELD.

Following previous works (bc-LSTM, DialogueRNN, DialogueGCN, et al.), raw utterance-level features of textual modality are extracted by TextCNN with Glove embedding. The pre-extracted features can be found in bc-LSTM. If you want to train the features by yourself, you can refer to DialogueGCN.

Besides, another alternative is to use BERT/RoBERTa to process text features, which will achieve better performance in most cases. You also can find the code and processed features with RoBERTa embedding in COSMIC.

DialogueCRN with Glove features (paper)

For training model on IEMOCAP dataset, you can refer to the following:

WORK_DIR="$(pwd)" # your work path
DATA_DIR="${WORK_DIR}/data/iemocap/IEMOCAP_features.pkl" # your data path

EXP_NO="dialoguecrn_base"
DATASET="iemocap"
OUT_DIR="${WORK_DIR}/outputs/${DATASET}/${EXP_NO}"

python -u ${WORK_DIR}/code/run_train_ie.py   \
    --status train  --feature_type text --data_dir ${DATA_DIR} --output_dir ${OUT_DIR}  \
    --gamma 0 --step_s 3  --step_p 4  --lr 0.0001 --l2 0.0002  --dropout 0.2 --base_layer 2

For training model on MELD dataset, you can refer to the following:

WORK_DIR="/DialogueCRN" # your work path
DATA_DIR="/DialogueCRN/data/meld/MELD_features_raw.pkl" # your data path

EXP_NO="dialoguecrn_base"
DATASET="meld"
OUT_DIR="${WORK_DIR}/outputs/${DATASET}/${EXP_NO}"

python -u ${WORK_DIR}/code/run_train_me.py   \
    --status train  --feature_type text --data_dir ${DATA_DIR} --output_dir ${OUT_DIR}  \
    --gamma 1.0 --step_s 2  --step_p 0  --lr 0.0005 --l2 0.0002  --dropout 0.2 --base_layer 1 --valid_rate 0.1

Run examples:

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

DialogueCRN with RoBERTa features

For training model on IEMOCAP dataset, you can refer to:

WORK_DIR="/DialogueCRN" # your work path
DATA_DIR="/DialogueCRN/data/iemocap/iemocap_features_roberta.pkl" # your data path

EXP_NO="dialoguecrn_bert_base"
DATASET="iemocap"
OUT_DIR="${WORK_DIR}/outputs/${DATASET}/${EXP_NO}"

python -u ${WORK_DIR}/code/run_train_bert_ie.py   \
    --status train  --feature_type text  --data_dir ${DATA_DIR}  --output_dir ${OUT_DIR}    \
    --gamma 0  --step_s 3  --step_p 0  --lr 0.0001  --l2 0.0002  --dropout 0.2  --base_layer 2 --valid_rate 0.1

For training model on MELD dataset, you can refer to:

WORK_DIR="/DialogueCRN" # your work path
DATA_DIR="/DialogueCRN/data/meld/meld_features_roberta.pkl" # your data path

EXP_NO="dialoguecrn_bert_base"
DATASET="meld"
OUT_DIR="${WORK_DIR}/outputs/${DATASET}/${EXP_NO}"

python -u ${WORK_DIR}/code/run_train_bert_me.py   \
    --status train  --feature_type text  --data_dir ${DATA_DIR}  --output_dir ${OUT_DIR}    \
    --gamma 1 --step_s 0  --step_p 1  --lr 0.0001 --l2 0.0002  --dropout 0.2 --base_layer 1 --use_valid_flag

Run examples:

# IEMOCAP dataset
bash ./script/run_train_bert_ie.sh 
# MELD dataset
bash ./script/run_train_bert_me.sh 

Note: The optimal hyper-parameters (e.g., the number of turns in Reasoning Modules) are selected according to the performance of validation set, with slight differences under different experimental configurations (i.e., the version of CUDA and PyTorch).

IEMOCAP dataset:

MELD dataset:

EmoryNLP dataset:

@inproceedings{DBLP:conf/acl/HuWH20,
  author    = {Dou Hu and
               Lingwei Wei and
               Xiaoyong Huai},
  title     = {DialogueCRN: Contextual Reasoning Networks for Emotion Recognition
               in Conversations},
  booktitle = {{ACL/IJCNLP} {(1)}},
  pages     = {7042--7052},
  publisher = {Association for Computational Linguistics},
  year      = {2021}
}