Recently, the use of social media platforms has increased with ease of use and fast accessibility, making such platforms a place of rumor proliferation owing to a lack of posting constraints and content authentication. Therefore, there is a need to leverage Artificial intelligence techniques to detect rumors on social media platforms to prevent their adverse effects on society and individuals. Most existing works that detect rumors in Arabic only target the textual features of the tweet content. Nevertheless, tweets contain different types of content such as (text, images, videos, and URLs), and the visual features of tweets play an essential role in rumor diffusion. This study proposes an Arabic rumor detection model to detect rumors on Twitter from textual and visual image features through two types of multimodal fusion: early and late fusion. In addition, we leveraged the transfer learning of the pre-trained language and vision models. Different experiments were conducted to select the best textual and visual feature extractors for building a multimodal model. MARBERTv2 was used as a textual feature extractor, whereas the ensemble of VGG-19 and ResNet50 was used as a visual feature extractor to build a multimodal model. Subsequently, the language and vision models of the single models were used as a baseline to compare their results with those of the multimodal models. Finally, the experimental results demonstrate the effectiveness of textual features in rumor detection tasks compared with multimodal models.

The rapid adoption of multimodal learning systems across high-stakes applications has intensified concerns about fairness and social bias, yet existing research remains scattered across modalities, tasks, protected attributes, mitigation strategies, and evaluation methodologies. This survey provides an overview of fairness and social bias in multimodal artificial intelligence. By reviewing 44 papers and 16 datasets, we synthesize the current landscape of fairness research in multimodal systems and identify emerging challenges and research opportunities. Our primary contribution is twofold. First, we introduce a multi-dimensional taxonomy that organizes the literature across nine operational dimensions, providing a unified perspective on multimodal fairness research. Second, we propose an end-to-end operational lifecycle framework that illustrates how biases emerge within shared latent spaces and identifies targeted pre-processing, training-time, and post-training intervention strategies. Overall, this survey discusses key performance-fairness trade-offs and provides a structured roadmap for evaluating and mitigating bias, ultimately supporting the development of more equitable multimodal AI systems.

Background and objective: The increasing use of online healthcare platforms has generated largevolumes of Arabic medical consultations, creating a need for automated methods to organizeand route medical questions to appropriate specialties. However, Arabic medical specialtyclassification remains relatively underexplored, particularly on large publicly available datasets.This study investigates the effect of medical label-space organization on Arabic medical specialtyclassification using the Arabic Healthcare Dataset (AHD).Methods: We evaluated two transformer-based models, AraBERT and XLM-RoBERTa,for medical specialty classification. To address semantic overlap and excessive granularity inthe original specialty labels, we applied a semantic label-space refinement pipeline based oncontextual representations and Hierarchical Density-Based Spatial Clustering of Applicationswith Noise (HDBSCAN). The refined label space was then evaluated using both transformermodels and compared with classification performance obtained using the original labels.Results:Label-space refinement substantially improved classification performance for both models.AraBERT’s macro-F1 increased from 24.16% to 69.16%, while its weighted-F1 increased from65.51% to 79.94%. XLM-RoBERTa showed a similar improvement, with macro-F1 increasingfrom 22.40% to 68.50% and weighted-F1 from 64.60% to 79.70%. Conclusions: The resultsdemonstrate that the organization and granularity of medical specialty labels can stronglyinfluence classification performance. Improving the label space before model training providesan effective data-centric strategy for Arabic medical specialty classification without modifyingthe underlying transformer architectures. These findings highlight label-space design as animportant consideration in biomedical text classification systems.