Using Machine Learning to Analyze Emotions in Arabic and Dialectical Texts

Social media is an imperative necessity in contemporary life. People can easily express their emotions and share moments on social media by writing a few words. Organizations approach Twitter as a rich data source that may be used to study emotions, but while many efforts have focused on sentiment analysis from text, emotion classification has received less attention. Emotion analysis usually provides a more in-depth assessment of the author's feelings, and in this research, we propose a dialectal Arabic text emotion classification architecture that accurately classifies the expressions into four emotions (anger, joy, fear, and sadness). Considering the improvements in natural language processing (NLP), we investigated the Bidirectional encoder representations from transformers (BERT) model. We implemented our proposed ensemble model via a majority voting technique that merges the best three versions of the pre-trained BERT models that are considered state-of-the-art in the classification field. We compared the results of our model with eight other machine learning classifiers and ten versions of the BERT model. The proposed ensemble approach accomplished around 84%, however the highest accuracy of the other investigated models was 76%. The presented experiments were examined on the Arabic tweets’ dataset for the EI-OC task provided by SemiEval, which contains 5600 tweets. KEYWORDS text classification, fine-tuning, voting technique, naive bayes, augmentation, transformers