Bipartite Heterogeneous Graph for Emotional Reasoning

This repository contains our source code and data for the CIKM 2023 accepted paper: A Bipartite Graph is All We Need for Enhancing Emotional Reasoning with Commonsense Knowledge.

Here is an overview of the model:

Preparation

• Set up the Python 3.7 environment, and build the dependencies with the following code: pip install -r requirements.txt

• Install the torch-geometric package from this link.

• You can download our extracted training data and knowledge features for each knowledge source with the following links: COMET₂₀₁₉, COMET₂₀₂₀, Conceptnet. Due to storage limit, we don't provide extracted data for DailyDialog. You can use the following data to re-extract knowledge data for all datasets.

• Or you can build data from scratch. Firstly, set up the extraction environments for each knowledge source.

  1. For COMET₂₀₁₉, The related code is in ./comet/. First download the pre-trained COMET models from this link and put it under ./comet/pretrained_models/ directory.

  2. For COMET₂₀₂₀: Download the pre-trained model with the script: download_model.sh

  3. For Conceptnet, download our filtered english version of conceptnet from this link and put it under ./conceptnet/.

  Secondly, download the original ERC datasets from this link We also directly provide the original RECCON dataset in this repository (/RECCON_data). Put the downloaded data under the directory of each data source.

  Thirdly, run the script to extract the knowledge features:

  1. For COMET₂₀₁₉, modify and run comet_extract_origin.py. This script is adapted from the COMET source code
  2. For COMET_2020, modify and run generate_knowledge.py. This script is adapted from the source code of Zhao et al..
  3. For Conceptnet, modify and run preprocess_conceptnet.py, then run main.py.

Training:

For training on ERC datasets, we use IEMOCAP as an exmaple:

Training with COMET₂₀₂₀:

python main.py --DATASET IEMOCAP --model_checkpoint roberta-large --NUM_TRAIN_EPOCHS 10 --BATCH_SIZE 16 --model_save_dir ./model_save_dir/IEMOCAP --mode train --SEED 42 --ROOT_DIR ./bart_comet_enhanced_data/ --CONV_NAME hgt --COMET_HIDDEN_SIZE 1024 --CUDA

Training with COMET₂₀₁₉:

python main.py --DATASET IEMOCAP --model_checkpoint roberta-large --NUM_TRAIN_EPOCHS 10 --BATCH_SIZE 16 --model_save_dir ./model_save_dir/IEMOCAP --mode train --SEED 42 --ROOT_DIR ./comet_origin_enhanced_data/ --CONV_NAME multidim_hgt --COMET_HIDDEN_SIZE 768 --CUDA

Training with Conceptnet:

python main.py --DATASET IEMOCAP --model_checkpoint roberta-large --NUM_TRAIN_EPOCHS 10 --BATCH_SIZE 16 --model_save_dir ./model_save_dir/IEMOCAP --mode train --SEED 42 --ROOT_DIR ./conceptnet_enhanced_data/ --CONV_NAME multidim_hgt --COMET_HIDDEN_SIZE 768 --CUDA

Training on other ERC datasets are similar.

For training on CEE dataset RECCON:

Training with COMET₂₀₂₀:

python main.py --DATASET RECCON --model_checkpoint roberta-large --alpha 0.8 --NUM_TRAIN_EPOCHS 10 --BATCH_SIZE 1 --model_save_dir ./model_save_dir/RECCON --mode train --LR 3e-6 --SEED 42 --ROOT_DIR ./bart_comet_enhanced_data/ --CONV_NAME hgt --COMET_HIDDEN_SIZE 1024 --CUDA

Training with COMET₂₀₁₉:

python main.py --DATASET RECCON --model_checkpoint roberta-large --alpha 0.8 --NUM_TRAIN_EPOCHS 10 --BATCH_SIZE 1 --model_save_dir ./model_save_dir/RECCON --mode train --LR 3e-6 --SEED 42 --ROOT_DIR ./comet_origin_enhanced_data/ --CONV_NAME multidim_hgt --COMET_HIDDEN_SIZE 768 --CUDA

Training with Conceptnet:

python main.py --DATASET RECCON --model_checkpoint roberta-large --alpha 0.8 --NUM_TRAIN_EPOCHS 10 --BATCH_SIZE 1 --model_save_dir ./model_save_dir/RECCON --mode train --LR 3e-6 --SEED 42 --ROOT_DIR ./conceptnet_enhanced_data/ --CONV_NAME multidim_hgt --COMET_HIDDEN_SIZE 768 --CUDA# BHG

Citation

Please cite the paper as follows: