SD-based-on-AIEN-and-ICA

Sarcasm Detection Based on Adaptive Incongruity Extraction Network and Incongruity Cross-Attention

Requirements

• PyTorch 1.12
• Transformers

Training

The script for training is:

PYTHONENCODING=utf-8 python main.py --data_dir ./data/Ptacek \ 
--output_dir ./output/Ptacek_KL-Bert_output/ --do_train --do_test --model_select KL-Bert

where

• --data_dir can be set as ./data/SARC_politics, ./data/Ghosh, and ./data/Ptacek
• --output_dir should keep up with data_dir and model_select to be ./output/DATASETNAME_MODELNAME_output/
• --know_strategy is for different knowledge selecting strategies, which can be common_know.txt, major_sent_know.txt, contrast_sent_know.txt, and minor_sent_know.txt. Our model uses all the obtained commonsense knowledge--"common_know.txt", and the other three knowledge selection strategies are used for comparison experiments.
• --know_num is to choose how many items of knowledge are used for each sentence, which is set to '5', '4', '3', '2', '1'

The script for testing is:

PYTHONENCODING=utf-8 python run_classifier.py --data_dir ./data/Ptacek \ 
--output_dir ./output/Ptacek_KL-Bert_output/ --do_test --model_select KL-Bert