This paper addresses the problem of actor-level stance detection in English social media posts concerning the Palestinian issue, a subtask of the StanceNakba-2026 Shared Task. The objective is to classify posts into one of three categories: Pro-Palestine, Pro-Israel, or Neutral, which is more challenging than the traditional favor/against/neutral formulations. This study uses a dataset comprising 1,401 posts, collected from X (formerly Twitter) after October 7, 2023, and annotated with one of the three stance labels. As Yafa’s Team, we tried to solve this problem using BERT-based models, which have proven their superiority in similar tasks. Several BERT-based models were fine-tuned and compared, including ARBERT, MARBERT, and PoliBERTweet, among others. Our winning model is the “MARBERT-Y”, where the “Y” comes from Yafa, a MARBERT-based model that has achieved a macro-F1 score of 95% on the test set. We argue this to two main factors: the structured and harsh preprocessing steps applied and the fine-tuning process employed. This indicates that domain-adapted transformer models, i.e., those pretrained on large-scale Twitter data are highly effective for politically stance detection tasks.

Cross-domain Arabic Named Entity Recognition (NER) remains challenging because models trained on one domain can lose performance when applied to unseen domains with different vocabulary, writing styles, and entity distributions. This paper addresses this challenge, and presents Yafa’s Team submission (ranked 7^th) to Subtask 1 of the KnowledgeGraphEval 2026 Shared Task. Our system leverages an AraBERTv2-based model for nested Arabic NER to approach it. The proposed approach combines task-adaptive (TAPT) pretraining on unlabeled Konooz data, R-Drop fine-tuning on Wojood, strong data augmentation, teacher-student knowledge distillation, span-level confidence and entropy filtering, and an entity-specific confidence-aware weighted-logit ensemble. Three teacher models generate soft labels for Konooz, and three distilled student models are combined in the final ensemble to produce the final predictions. Wojood contains approximately 448K tokens and 72K entity mentions across 21 entity types, while Konooz contains data from 10 domains for cross-domain evaluation. The final system achieves a Micro F1 of 71.33% on the official Konooz test set, improving over the non-adapted AraBERTv02 baseline of 61.00% by 10.33 points. The results show that combining domain adaptation, filtered knowledge transfer, and entity-specific ensembling can improve cross-domain Arabic NER performance.

Relation extraction (RE) identifies semantic relations between entities, converting unstructured text into structured facts for downstream applications such as knowledge graph construction. Arabic RE remains challenging due to rich morphology, complex syntax, dialectal variation, and the scarcity of high-quality annotated resources. We present Yafa’s Team submission (ranked 7th) to Subtask 2 of the KnowledgeGraphEval 2026 Shared Task, which frames Arabic RE as multiclass classification over 40 relation types from the dataset. We propose an ArBERTv2-based entity-marker classifier that fuses global sentence context ([CLS]) with entity-specific representations (h_e1, h_e2) into a joint representation for relation prediction. Building on this baseline, we exploit a transductive self-training step that pseudo-labels the released test inputs at a softmax confidence threshold of 0.95 and retrains on them, raising Micro-F1 from 70.20% to 91.92%. We further conduct a systematic analysis of cross-backbone ensembling with MARBERTv2 and AraBERTv2, finding that ensembling does not outperform the single transductive model. Our results highlight transductive self-training as a simple yet effective strategy for improving Arabic RE robustness.

Measuring semantic similarity and analyzing authorial style are fundamental tasks in Natural Language Processing (NLP), with applications in text classification, cultural analysis, and literary studies. This paper investigates the semantic similarity and stylistic features of Nakba short stories, a key component of Palestinian literature, using transformer-based models, AraBERT, BERT, and RoBERTa. The models effectively capture nuanced linguistic structures, cultural contexts, and stylistic variations in Arabic narratives, outperforming the traditional TF-IDF baseline. By comparing stories of similar length, we minimize biases and ensure a fair evaluation of both semantic and stylistic relationships. Experimental results indicate that RoBERTa achieves slightly higher performance, highlighting its ability to distinguish subtle stylistic patterns. This study demonstrates the potential of AI-driven tools to provide more in-depth insights into Arabic literature, and contributes to the systematic analysis of both semantic and stylistic elements in Nakba narratives.

In this paper, we enrich Arabic Natural Language Processing (NLP) resources by introducing the “Nakba Topic Classification Corpus (NTCC),” a novel annotated Arabic corpus derived from narratives about the Nakba. The NTCC comprises approximately 470 sentences extracted from eight short stories and captures the thematic depth of the Nakba narratives, providing insights into both historical and personal dimensions. The corpus was annotated in a two-step process. One third of the dataset was manually annotated, achieving an IAA of 87% (later resolved to 100%), while the rest was annotated using a rule-based system based on thematic patterns. This approach ensures consistency and reproducibility, enhancing the corpus’s reliability for NLP research. The NTCC contributes to the preservation of the Palestinian cultural heritage while addressing key challenges in Arabic NLP, such as data scarcity and linguistic complexity. By like topic modeling and classification tasks, the NTCC offers a valuable resource for advancing Arabic NLP research and fostering a deeper understanding of the Nakba narratives

In this paper, we present a high-performing model for Arabic stance detection on the STANCEEVAL2024 shared task part ofARABICNLP2024. Our model leverages ARABERTV1; a pre-trained Arabic language model, within a single-task learning framework. We fine-tuned the model on stance detection data for three specific topics: COVID19 vaccine, digital transformation, and women empowerment, extracted from the MAWQIF corpus. In terms of performance, our model achieves 73.30 macro-F1 score for women empowerment, 70.51 for digital transformation, and 64.55 for COVID-19 vaccine detection.