Enhancing Sarcasm Detection Using GAN-BERT with Multi-Task Learning
Afif Hossain Irfan, Shrabani Das, Jannatul Ferdaus, Md. Tabil Ahammed
Abstract
Sarcasm detection remains a complex task in natural language processing due to its nuanced and contextdependent nature. This paper introduces a novel sarcasm detection framework combining GAN-BERT and Multitask Learning (MTL), where sentiment and emotion serve as auxiliary supervision tasks to enhance feature representation. The proposed model uses sentiment and emotion as auxiliary tasks to improve its understanding of sarcastic cues. The Sarcasm Headlines Dataset, comprising 28,619 labeled headlines, is augmented with sentiment and emotion annotations utilizing pretrained NLP models. GAN-BERT uses 70 % of the 80 % training split as labeled data and the remaining 30 % as unlabeled, allowing semi-supervised learning. Experimental results demonstrate that the proposed model outperforms BERT, LSTM, and CNN, achieving an accuracy of 93. 88% and an F1 score of 94. 15%. These findings demonstrate the effectiveness of integrating adversarial training and multitask objectives in sarcasm detection.