هناء علي عبد مهيدي

Automated identification of political frame effects in biased news texts continues to be a difficult NLP problem, motivated by issues of label distribution imbalance and contextual richness. Despite being successful at traditional text classification tasks, identification of complex geopolitical bias effects becomes hindered by imbalanced datasets with an overrepresentation of the majority class. In this paper, we propose a binary classification pipeline for identification of geopolitical framing biases within a collection of 30,000 short news articles in the AG News dataset "World" sub-category. In order to overcome imbalances in the minority class, we used stratified partitioning and random sampling techniques for achieving class balance in a training set with 43,698 examples. We fine-tuned RoBERTa-Large (355M parameters) model for three epochs with sequence classification task using FP16 precision strategy. The resulting classifier reaches 91.08% accuracy (95% CI: [90.46%, 91.66%], Wilson score method), macro-F1 equal to 0.91 and AUC score of 0.9641 (95% CI: [0.9601, 0.9681], Hanley-McNeil approximation). Pair-wise McNemar's significance test shows that RoBERTa-Large model significantly outperforms ALBERT-base (χ2 = 5.68, p = 0.017) and DistilBERT (χ2 = 155.42, p < 0.0001) transformers which were independently fine-tuned using the same strategy. Experimental findings show that data balancing combined with transformer representations provide a statistically sound solution for bias identification.

Image Super-Resolution is one of the crucial fields in image processing and computer vision, aimed at reconstructing high-resolution (HR) images from low-resolution (LR) inputs. This method is essential in enhancing visual quality, recovering lost details, and improving the performance of various image analysis tasks in many diverse applications, such as medical imaging, satellite imaging, surveillance, and digital media. Traditionally, ISR was addressed using conventional signal processing techniques and statistical methods. However, with the introduction of deep learning techniques like Convolutional Neural Networks, Generative Adversarial Networks, and most recently, Transformer-based architectures and Diffusion Models, the field has been dramatically revolutionized. These deep learning approaches have significantly advanced the state-of-the-art and achieved unprecedented performance in terms of objective quality metrics, such as PSNR/SSIM, as well as perceptual realism. This paper presents a comprehensive review on the evolution of ISR, drawing comparisons and contrasts between traditional methods and modern deep learning approaches. It discusses their respective strengths, limitations, and challenges that remain to be conquered, such as real-world degradation modeling, the trade-off between perceptual quality and fidelity, and computational efficiency. This review also discusses emerging trends and future directions of ISR research, pointing out the continuous interplay between theoretical advances and practical applications that will lead to further improvements of visual data across diverse domains.

Abstract: Many people’s company and social connections have changed as a result of online social networking, the proliferation of these social media sites has ushered in a new age of Unfortunately extremism, where they are being used to inspire terrorist attacks, recruitment, and other forms of extremism.

For the recognition of automated modulation, algorithms based on frequency, amplitude, and signal phase were extensively used. But the algorithm is affected considerably by noise, and the performance is degraded significantly in low signal-to-noise ratio (SNR) conditions. Deep learning (DL) is an effective mechanism that has accomplished greater attainment in various applications. But the usage in transmission systems has not been well-studied. This study explores an algorithm for modulation classification and multi-signal detection that is substantial in different transmission techniques.