KEC

The code for the paper "Neutral Utterances are Also Causes: Enhancing Conversational Causal Emotion Entailment with Social Commonsense Knowledge".

The appendix mentioned in the paper is present in here.

Some code is based on DAG-ERC, RECCON, and COMET-ATOMIC-2020.

Requirements

• Pytorch==1.8.1
• Transformers==4.3.3
• numpy=1.19.2
• nltk

Additonal Data

Edge attributes of skaig: skaig_data

Training

bash run_single.sh

Some explanations

• generate_knowledge.py is used to generate social commonsense knowledge for our method. Put it in comet-atomic-2020/models/comet_atomic2020_bart/ of COMET-ATOMIC-2020. P.S. the paths of loaded and dumped files should be modified to your own data paths. We have uploaded all the generated knowledge data in dd_data.

• knowledge_select.py is used to select sentimental related pieces of knowledge for a pair of utterances. We have uploaded all the processed data in dd_data.

• entail_construct.py is used to form the data into the entailment style with or without emotion words. The generatad files is used to train and evaluate the baseline of RECCON-DD. Furthermore, replace the train_classification.py, eval_classification.py in RECCON with RECCON_baseline/train_classification.py and RECCON_baseline/eval_classification.py in this repository. We have uploaded the entailment style data in here. Download the data and put them in data/subtask2/fold1/ in RECCON.