Sarcasm Detection In Arabic Tweets Using Deep Learning and BERT-based Models with Data Augmentation

Abstract

Automatic sarcasm detection from text is an important classiőcation task that can help identify the actual sentiment in user-generated data, such as reviews or tweets. Despite its utility, detecting sarcasm remains a difficult problem due to the lack of vocal intonation or facial expressions in textual data. To date, the majority of solutions have relied on hand-crafted affect characteristics such as emojis and hashtags, or pre-trained models of non-contextual word embeddings, such as Word2vec. However, the inherent limitations of these models make them unsuitable for detecting sarcasm. We present in this study a deep neural network sarcasm detection application. As a starting point, we used a set of deep learning models. Moreover, we have used three main Transformer-based Models for Arabic Language Understanding.The proposed model has been evaluated using the ArSarcasm-v2 dataset, in addition to a manually collected dataset containing sarcastic AraCOVID19 tweets. Extensive experiments on different datasets demonstrate that the proposed models give a good outcome. As a result, the proposed system is capable of achieving an accuracy of 98% for the AraCOVID19 dataset and accuracy of 79% for the ArSarcasm-v2 dataset.