A Transformer-based Hybrid Model for Implicit Emotion Recognition in Arabic Text

Abstract

Implicit emotion recognition has emerged as an active area of research in modern Natural Language Processing (NLP). Unlike explicit emotions, which are directly expressed through emotional words, implicit emotions are inferred from the surrounding context, making their detection more challenging. While most research in Arabic NLP has focused on recognizing explicit emotions, the study of implicit emotions remains largely unexplored, primarily due to its unique linguistic and morphological characteristics. The current study addresses this gap by compiling an Arabic dataset for the implicit emotion recognition task, named Arabic Implicit Emotion Dataset (AIEmoD), which is curated from existing publicly available explicit emotion datasets. Furthermore, it proposes a novel hybrid deep learning model that integrates the Arabic transformer-based AraBERT model with a Bidirectional Gated Recurrent Units (BiGRU) network to recognize and classify implicit emotions in Arabic text. The proposed AraBERT-BiGRU model was evaluated on two widely used Arabic emotion datasets, AETD and SemEval-2018, in addition to the newly compiled AIEmoD dataset. The results show that the model achieved F1-scores of 79.87% on AETD and 70.67% on AIEmoD, significantly outperforming deep learning baseline methods. Moreover, the proposed model surpassed current state-of-the-art approaches for explicit emotion recognition, even when applied to the more challenging task of implicit emotion detection. These findings highlight the effectiveness and robustness of the proposed AraBERT-BiGRU model in recognizing implicit emotions in Arabic text.