Enhancement of Implicit Emotion Recognition in Arabic Text: Annotated Dataset and Baseline Models

Emotion recognition in textual data has emerged as a rapidly advancing task within the field of Natural Language Processing (NLP). Implicit Emotion Recognition (IER), which aims to identify emotions primarily through contextual cues rather than overt or explicit emotional expressions, remains in its early stages. Despite the significant progress in emotion recognition tasks, current research has largely ignored IER, particularly for low-resource languages such as Arabic. This study aims to comprehensively address this task for the Arabic language, including dataset construction, annotation, modeling, and evaluation. Specifically, the study 1) introduces the first annotated dataset for Arabic Implicit Emotion Recognition (AIER); 2) annotates the dataset with emotions, cues, and causes using a semi-automatic annotation tool validated by four native Arabic speakers and linguists; 3) investigates the potential of two categories of transformer-based models: masked language models, exemplified by pre-trained Bidirectional Encoder Representations from Transformers (BERT)-based architectures through a series of fine-tuning experiments, and causal-based models, such as generative Large Language Models (LLMs), through a zero-shot prompting approach; and 4) evaluates the performance of four distinct algorithmic approaches on the proposed dataset, namely classical Machine Learning (ML), Deep Learning (DL), BERT-based models, and generative LLMs. The experimental results indicate that BERT-based models outperform ML, DL, and generative LLMs. In particular, the fine-tuned MARBERTv2 model achieved the best performance among all pre-trained models, obtaining an F1-score of 79.83% on the AIER dataset.