Fine-Grained Emotion Recognition from Arabic and English Social Media Texts

Abstract

Emotion recognition in social media using (NLP) is a critical step toward understanding human affect and mental states in digital communications. The present study aims to address fine-grained emotion recognition in English and Arabic usergenerated contents, involving multi-class and multi-label emotion classification. This study introduces a comprehensive comparative analysis of multiple approaches based on (DL), Transfer Learning (TL), Transformers, and attention mechanism. Whereby, various combinations of transformer-based models with BiLSTM/BiGRU layers and self-attention mechanism are rigorously trained and tested to assess their performance on fine-grained emotion classification using 3 publicly available datasets. These include English-specific transformers such as BERT and RoBERTa and Arabic-specific transformers such as AraBERT, MarBERT, and CamelBERT. All of the proposed models are evaluated based on macro-average and per-class scores of accuracy, recall, precision, AUC, and F1-score metrics using ISEAR, CBET, and SemEval2018 datasets. Experimental results demonstrated the high effectiveness of fine-tuning hybrid transformerbased models for fine-grained emotion classification across Arabic and English social media texts by achieving 0.7308, 0.6482, and 0.6251 macro-average f1-scores on ISEAR, CBET, and SemEval2018 datasets, respectively