Emotion Detection in Arabic Short Text: A Deep-Learning Approach

Abstract In modern society, writing down thoughts, ideas, and pleasant experiences has become a widespread way to express feelings. X, formerly known as Twitter, is a rich data source that companies can use to study individuals’ ideas, thoughts, and emotions for a range of useful applications. The analysis of positive and negative feelings is a key focus of NLP research literature, while emotion detection receives relatively little attention. Very few studies to date have examined the classification of emotions in text, particularly Arabic written content. The new study uses deep learning approaches to solve this difficulty and close these existing gaps in the literature. A number of different deep learning models are available and each one has been developed based on a unique feature engineering approach to classify the emotions conveyed in the SemEval-2018 dataset into four groups, namely joy, fear, anger and sadness. The results show that the CNN model that employs Word2vec outperforms the other models, with an accuracy of 80%. Furthermore, it has been found to perform better in Arabic than the most current comparable model, with improvements varying between 12–54.96% for F1 scores to 5–26% for accuracy.