Detecting Irony in Arabic: From Emotion AI Recognition to Cognitive AI Functionality

This paper has a twofold aim. First, to move from the development of emotion AI to cognitive AI, focusing on the potentiality for AI decision making based on the detection of such sentiment-data. We contribute to the development of linguistic/sentimental detection capabilities in AI to be utilized in translation, digital communication, and cognition. We start by detecting irony in diverse Arabic dialects by utilizing tweets from X (formerly known as Twitter). The tweets come from open data, structured based on language's semantics, and are manually annotated in the dataset. We show performance from our architecture that uses 1) Neural networks and two transformer models 2) Bidirectional Encoder Representations from Transformers (BERT) and 3) Arabic BERT, AraBERT. In our Neural network, we contribute to AI's ability of sentimental classification and cate-gorization by introducing simple preprocessing and architecture, both of which address the computational complexity problem. We show better results compared to the literature mainly from our AraBERT: 0.96 F1- score compared to 0.87 for one dataset, 0.92 versus 0.816 for a second, 0.66 F1-score versus 0.7 but a better recall 0.8 versus 0.7 for a third dataset, and our neural network achieves an F1-score of 0.82 compared to 0.73 from the literature. Second, based on the presented work, we would be able to move forward in deliberating future research to structure cognitive AI functionalities and not only irony detection. This study intends to strengthen the development of linguistic detection capabilities in AI and to move a step further in considering such mechanisms as the main building blocks towards the development of cognitive AI functions.