An Arabic language benchmark and hybrid BERT–BiGRU–CapsNet model for optimism and pessimism detection

Accurate identification of optimism and pessimism has important applications for individual health and societal well-being. Despite the growing interest in sentiment analysis and emotion detection, limited research has focused on the detection of optimism and pessimism, particularly in the Arabic language. Existing studies primarily address general sentiment classification, leaving a significant gap in modeling optimism and pessimism detection. To the best of our knowledge, this study presents the first effort to introduce a dedicated Arabic dataset for optimism and pessimism analysis, comprehensively evaluate the dataset on seven Arabic language models, and propose a novel hybrid approach for detecting optimism and pessimism. In this paper, we introduce the Saudi optimism and pessimism detection (SOPD), the first and largest dataset for detecting optimism and pessimism in Arabic. It was constructed from posts and replies collected from the Hawamer and Kooora forums, focusing on Arabic-language content in the domains of finance and sports. The dataset comprises a total of 16,152 samples, manually annotated into three sentiment classes: optimistic, pessimistic, and neutral. The dataset was then systematically evaluated using seven Arabic language models: Arabic-BERT, CamelBERT-DA, ARBERTv2, AraBERTv02, QARiB, MARBERTv2, and SaudiBERT. Additionally, we proposed a novel hybrid deep learning architecture, BERT-BiGRU-CapsNet, which combines a Bidirectional GRU (BiGRU) layer with a Capsule Network (CapsNet) on top of a BERT encoder. The evaluation results showed that SaudiBERT achieved the highest F1-Score of 93.11%, followed by MARBERTv2 with the second-highest F1-Score of 87.25%. Nonetheless, the proposed hybrid model, BERT-BiGRU-CapsNet, outperformed all traditional Arabic models with an F1-Score of 93.34%. The results provide valuable insights into the effectiveness of current Arabic models in classifying optimism and pessimism, offering practical directions for improving context-aware sentiment analysis in Arabic NLP. Moreover, this study holds significant implications for enhancing individual well-being, predicting behavioral trends, and informing the design of targeted mental health and social interventions. It also might enable businesses and enterprises to effectively analyze customer behavior, supporting strategic decision-making, product development, and customer engagement.