Comparison of Different Deep Learning Approaches to Arabic Sarcasm Detection

Abstract

Irony and Sarcasm Detection (ISD) is a crucial task for many NLP applications, especially sentiment and opinion mining. It is also considered a challenging task even for humans. Several studies have focused on employing Deep Learning (DL) approaches, including building Deep Neural Networks (DNN) to detect irony and sarcasm content. However, most of them concentrated on detecting sarcasm in English rather than Arabic content. Especially studies concerning deep neural networks, including convolutional neural networks (CNN) and recurrent neural network (RNN) architectures. This paper investigates several deep learning approaches, including DNNs and fine-tuned pretrained transformer-based language models, for identifying Arabic sarcastic tweets. In addition, it presents a comprehensive evaluation of the impact of data preprocessing techniques and several pretrained word embedding models on the performance of the proposed deep models. Two shared tasks' datasets on Arabic sarcasm detection are used to develop, fine-tune, and evaluate the different techniques and methods presented in this paper. Results on the first dataset showed that fine-tuned pretrained transformer-based language model outperformed the developed DNNs. The proposed DNN models obtained comparable performance on the second dataset to the fine-tuned models. Results also proved the necessity of applying preprocessing techniques with the various Deep Learning approaches for better detection performance of these models.