Sarcasm Detection in Arabic Short Text using Deep Learning

Recently, a growing interest among researchers has emerged in discovering ambiguous information in short sarcastic texts in Arabic. Nevertheless, sarcasm is a particularly challenging problem for sentiment analysis algorithms due to its considerable impact on emotions. A short text evaluation can provide important information about a product or service. However, due to the currently small number of sarcastic datasets and their unbalanced nature, it is also important to preprocess data before classification, especially those with dialects. Furthermore, to detect sarcasm, language models must be capable of capturing complicated relationships and ambiguous semantic meanings. In this paper, we attempt to detect sarcasm in Arabic text using a large pre-trained language model (BERT). Therefore, a new dataset for the "iSarcasmEval" shared task is examined in this paper. Preprocessing of the dataset is performed first. Moreover, the data is balanced by applying both Random Swap and Random Deletion, which are both data augmentation techniques. Following that, two transformer-based models, MARBERT and AraBERT, were used to analyze the data. Experimental results reveal that the MARBERT model outperforms the AraBERT model in this dataset.