Multi-task learning for multi-dialect Arabic sentiment classification and sarcasm detection
摘要
Sentiment analysis (SA) is a widely recognized and increasing field of research in the science of natural language processing (NLP). A wide range of methods exist by which individuals express their sentiments and emotions. Sarcasm is occasionally employed with sentiments, particularly when expressing intense emotions. Sarcasm is characterized by using positive language to express a negative intention. In current research, these two aspects are often treated as separate tasks. However, recent advancements in deep learning algorithms have greatly improved the efficiency of standalone classifiers for both sentiment and sarcasm tasks. Despite these improvements, a major challenge remains: correctly classifying sarcastic sentences as negative. Furthermore, there has been an important increase in the number of research efforts focused on Arabic dialects. In this research paper, we explore both Sentiment and Sarcasm within multi-dialect Arabic language corpora to set up a highly accurate sentiment classification and sarcasm detection. To be more specific, we develop a system of classification that employs a Multi-Task Learning (MTL) algorithm using a pre-trained Arabic language model to accurately determine sentiment classification and sarcasm detection. Considering this, we claim that having the ability to identify sarcasm will improve the accuracy of sentiment classification. The performance of our approach showed notable results, surpassing the performance of previously developed models described in the literature on all of the three datasets, for sentiment classification with up to an F1-score of 73.96% on ArSarcasm