Social Arabic Emotion Analysis: A Comparative Study of Multiclass Classification Techniques
Hanane Elfaik, El Habib Nfaoui · 2021
The advent of social media platforms (Twitter, Facebook, etc.) has facilitated the development of user generated content, and users tend to express and share their opinions and emotions about political movements, social events, products, and services. Thus, the emotion detection task has emerged and has received significant attention from academia. It is recognized as a part of affective computing to understand and extract the human affective state expressed in a piece of writing (tweet, post, etc.). In this work, we have measured and compared the performance of different traditional machine learning models namely Linear SVC, Logistic Regression, Multinomial and Complement Naive Bayes with the accuracy produced by two of the top performing deep learning techniques namely Convolution Neural Network and Long-Short-Term Memory networks. Two types of social media datasets were utilized in this study. One is the Arabic emotions Twitter dataset (AETD) and the other one is the Iraqi Arabic emotion dataset (IAEDS). The experimental results demonstrate that Convolution Neural Network deep based technique performs better than machine learning-based algorithms on both datasets used in this study. Additionally, the results establish that proposed models outperform the current state-of-the-art Arabic emotion recognition methods, achieving a 2.4% improvement in accuracy on the AETD dataset. Furthermore, the proposed models can significantly distinguish between the different emotion labels.