An Ensemble Deep Learning Approach for Emotion Detection in Arabic Tweets

Now-a-days people use social media websites for different activities such as business, entertainment, following the news, expressing their thoughts, feelings, and much more. This initiated a great interest in analyzing and mining such user-generated content. In this paper, the problem of emotion detection (ED) in Arabic text is investigated by proposing an ensemble deep learning approach to analyze user-generated text from Twitter, in terms of the emotional insights that reflect different feelings. The proposed model is based on three state-of-the-art deep learning models. Two models are special types of Recurrent Neural Networks RNNs (Bi-LSTM and Bi-GRU), and the third model is a pre-trained language model (PLM) based on BERT and it is called MARBERT transformer. The experiments were evaluated using the SemEval-2018-Task1-Ar-Ec dataset that was published in a multilabel classification task: Emotion Classification (EC) inside the SemEval-2018 competition. MARBERT PLM is compared to one of the most famous PLM for dealing with the Arabic language (AraBERT). Experiments proved that MARBERT achieved better results with an improvement of 4%, 2.7%, 4.2%, and 3.5% regarding Jaccard accuracy, recall, F1 macro, and F1 micro scores respectively. Moreover, the proposed ensemble model showed outperformance over the individual models (Bi-LSTM, Bi-GRU, and MARBERT). It also outperforms the most recent related work with an improvement ranging from 0.2% to 4.2% in accuracy, and from 5.3% to 23.3% in macro F1 score.

@article{Mansy2022,
  title     = {An Ensemble Deep Learning Approach for Emotion Detection in Arabic Tweets},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {13},
  number    = {4},
  year      = {2022},
  publisher = {The Science and Information Organization},
  author    = {Alaa Mansy and Sherine Rady and Tarek Gharib},
  doi       = {10.14569/IJACSA.2022.01304112},
  url       = {https://doi.org/10.14569/IJACSA.2022.01304112}
}