Comparative analysis of Deep Learning and Machine Learning algorithms for emoji prediction from Arabic text

Emojis have become a crucial part of text-based communication in recent years, especially on social media and messaging services. As a result, emoji prediction has gained increasing attention as a research topic in Natural Language Processing. Emoji recommendation is a task of predicting relevant emojis based on the emotional and contextual orientation of the text. In this study, we provide a comparative analysis of several Machine Learning (ML) and Deep Learning (DL) methods for emoji prediction from Arabic text. ML models are commonly used as baselines for emoji prediction; hence, more sophisticated DL models are needed for performance enhancement. In this work, we evaluate the performance of three baseline ML models, namely Support Vector Machines (SVM), Multinomial Naive Bayes (MNB), and Random Forest (RF), as well as state-of-art DL models, namely Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Arabic Bidirectional Encoder Representations from Transformers (AraBERT), and Multilingual Bidirectional Encoder Representations from Transformers (mBERT). This research is evaluated utilizing a large corpus of Twitter dataset that is translated to Arabic and balanced to enhance the prediction performance. Throughout the experiments, the ML models achieved classification accuracies of 74%, 78.9%, and 84% for SVM, MNB, and RF, respectively. Furthermore, the DL models achieved accuracies of 91.16%, 91%, 85%, and 80% for LSTM, BiLSTM, AraBERT, and mBERT, respectively.