State-of-the-art fake-news classifiers frequently report near-ceiling accuracy on widely used benchmarks such as ISOT, Misinfo, and WELFake. We argue that such results often reflect structural homogeneity and provenance-based separability rather than robust claim-level veracity inference. Anchored in the Information Disorder framework, we analyze how dataset construction operationalizes the notion of “fake” and how this shapes model behavior. We conduct systematic bidirectional cross-dataset experiments across six transfer directions and evaluate performance not only by mean accuracy, but also by variance and directional asymmetry. Results reveal substantial degradation under distribution shift and pronounced transfer asymmetries between dataset pairs. Although not always achieving the highest mean accuracy, affective augmentation combining dimensional (VAD) and categorical (Ekman) representations yields the lowest variance and smallest directional gap, indicating superior cross-domain stability. Our findings expose the disconnect between accuracy-driven benchmarking and construct-valid evaluation. We argue that progress in fake-news detection requires shifting from isolated in-domain optimization toward robustness-oriented, bidirectional, and distribution-aware assessment practices.

This paper presents the CIC system for ArGuard 2026 Subtask B2, which classifies unsafe Arabic prompts into one of seven harm domains. Our system combines four complementary components: a fine-tuned CAMeLBERT-Mix model, a modified-Huber stochastic gradient descent classifier, a character-level nearest-neighbor model, and a character class-centroid similarity model. On the development set, we achieved a Macro-F1 score of 0.916 and an accuracy of 0.923. On the test set, the proposed model achieved an accuracy of 0.766, a macro precision of 0.760, a macro recall of 0.763, and a Macro-F1 of 0.751. The results show that combining contextual transformer representations, lexical features, and retrieval-based similarity signals can improve fine-grained harmful-prompt classification.

We present our approach in this submission for ArGuard 2026 Subtask B1, which classifies Arabic prompts directed at large language models as Safe or Unsafe. Our system combines four sparse TF-IDF feature representations with dialect indicators and surface-level features. We trained logistic regression and Naive-Bayes SVM (NBSVM) classifiers and combined their predictions using late fusion. The system achieved a Macro-F1 score of 0.94 and an accuracy of 0.96 on the development set. On the test set, our submission achieved 0.79 accuracy, 0.85 macro precision, 0.71 macro recall, and 0.72 Macro-F1.

The classification task for ArGuard 2026 Track A, Subtask A1, identifies whether an Arabic meme is Hateful or Not Hateful. Our system combines three separately trained models. In the first experiment, we use word-level and character-level features with balanced linear classifiers. In the second experiment, we fine- tune MARBERT for text classification using class-balanced focal loss. In the third experi- ment, we use a frozen SigLIP image encoder with balanced logistic regression. We combine the Class-1 probabilities from the three mod- els using late fusion. On the test set, the fused model achieved a Macro-F1 score of 0.76 and an accuracy of 0.79

The classification of Arabic prompts as Safe or Unsafe is the focus of this submission to ArGuard 2026 Subtask B1. We developed a lightweight hybrid classifier that combines lexical and semantic information. The lexical branch uses normalized word-level and character-level features with a class-balanced LinearSVC classifier. The semantic branch uses frozen multilingual E5 embeddings with a class-balanced logistic regression classifier. We combined the outputs of both branches using late fusion. On the development set, LinearSVC achieved a Macro-F1 score of 0.93, while the frozen E5 branch achieved 0.88. The final hybrid system achieved a Macro-F1 score of 0.93. On the test set, the submitted system achieved 0.78 accuracy, 0.83 precision, 0.70 recall, and 0.72 Macro-F1.

This paper presents a multi-step approach for multi-label emotion classification as our system description paper for the SEMEVAL-2025 workshop Task A using machine learning and deep learning models. We test our methodology on English, Spanish, and low-resource Yoruba datasets, with each dataset labeled with five emotion categories: anger, fear, joy, sadness, and surprise. Our preprocessing involves text cleaning and feature extraction using bigrams and TF-IDF. We employ logistic regression for baseline classification and fine-tune Transformer models, such as BERT and XLM-RoBERTa, for improved performance. The Transformer-based models outperformed the logistic regression model, achieving micro-F1 scores of 0.7061, 0.7321, and 0.2825 for English, Spanish, and Yoruba, respectively. Notably, our Yoruba fine-tuned model outperformed the baseline model of the task organizers with micro-F1 score of 0.092, demonstrating the effectiveness of Transformer models in handling emotion classification tasks across diverse languages.

Immediate Textual Variation (ITV) is defined as the process of introducing changes during text transmission from one node to another. One-step variation can be useful for testing specific philological hypotheses. In this paper, we propose using Large Language Models (LLMs) as text-modifying agents. We analyze three scenarios: (1) simple variations (omissions), (2) paraphrasing, and (3) paraphrasing with bias injection (polarity). We generate simulated news items using a predefined scheme. We hypothesize that central tendency measures—such as the mean and median vectors in the feature space of sentence transformers—can effectively approximate the original text representation. Our findings indicate that the median vector is a more accurate estimator of the original vector than most alternatives. However, in cases involving substantial rephrasing, the agent that produces the least semantic drift provides the best estimation, aligning with the principles of Bédierian textual criticism.

In this paper, we share our best performing submission to the Arabic AI Tasks Evaluation Challenge (ArAIEval) at ArabicNLP 2023. Our focus was on Task 1, which involves identifying persuasion techniques in excerpts from tweets and news articles. The persuasion technique in Arabic texts was detected using a training loop with XLM-RoBERTa, a language-agnostic text representation model. This approach proved to be potent, leveraging fine-tuning of a multilingual language model. In our evaluation of the test set, we achieved a micro F1 score of 0.64 for subtask A of the competition.