AraSenti-MARBERT-DynGCN: An Advanced Framework for Sarcasm Detection in Arabic Text

Abstract

Sarcasm detection in Arabic texts is challenging because of the complexity in Arabic morphology, the multitudes of dialects used by native speakers, and its strong reliance on context. The lack of standardized orthography in most written forms of Arabic, combined with the frequent use of colloquialisms and metaphorical expressions, makes the task all the more difficult when trying to detect sarcasm. In this paper, we present a hybrid approach for Arabic sarcasm detection by integrating sentiment analysis and contextual embeddings with Dynamic Graph Convolutional Networks (DynGCNs). Our model uses pre-trained language model MARBERT for building rich contextualized word embeddings, AraSenti for accurate sentiment polarity classification, and adopts DynGCNs for capturing syntactic dependencies in the text and updating them dynamically. We also applied state-of-the-art preprocessing to handle informal characteristics such as emoticons and punctuation, which are very necessary in recognizing sarcasm in informal Arabisc speech. The proposed approach was empirically evaluated on two well-known Arabic Sarcasm detection repositories: iSarcasmEval and ArSarcasm-v2 through extensive experiments. As a result, on the iSarcasmEval data split it reached an accuracy of 92.8% and F1-score reaching up to 78.5%, significantly outperforming its counterparts such as AraBERT, QARiB, and MARBERT based models. Its performance on the ArSarcasm-v2 dataset gave an accuracy of 86.5% while the F1-score stood at 71.7%, hence making it a robust method across different datasets.