Transferring Styles between Sarcastic and Unsarcastic Text using SHAP, GPT-2 and PPLM

In order to realize the text style transfer in sarcasm, improve the performance of style transfer in the text. We use News Headlines Dataset for Sarcasm Detection as the dataset and GPT-2 as the pre-training model to train a discriminator that can accurately identify sarcasm. Through the combination of shap and sarcasm detection model, the deletion of text attributes is more intuitive and precise. We fine-tuned the GPT-2 model on the sarcasm dataset, and used the trained sarcasm discriminator on the PPLM as the generation part. Compare the generation effect of fine-tuned GPT-2, PPLM, Combination of PPLM and fine-tuned GPT-2. The results show that the generation effect of using PPLM is better. The new method we propose, in the deletion and generation part of the text conversion, realizes the style transfer of the text between sarcastic and unsarcastic.