For the Chinese track of SemEval-2026 Task 9 (Detecting Online Polarization), we address two key challenges: polarized content frequently uses implicit language (e.g., homophones and coded terms) to evade moderation, and class distributions exhibit severe long-tail imbalance. We propose a metric learning approach that frames polarization detection as semantic similarity matching, which captures implicit language patterns better than linear decision boundaries. We fine-tune an ERNIE-3.0 encoder with SoftTriple loss and apply ik/iNN retrieval for binary detection (Subtask 1). For multi-label categorization (Subtasks 2 and 3), we transfer learned representations from the detection model and fine-tune with Asymmetric Loss. A priority-based stratified cross-validation strategy ensures minority classes appear across all training folds despite extreme label skew. Evaluated on the official 1,927-sample test set using an end-to-end pipeline, our system achieved Macro-F1 scores of 0.9190 (Rank 6) on Polarization Detection, 0.8244 (Rank 5) on Type Classification, and 0.6670 (Rank 4) on Manifestation Identification.

This paper presents the KvochurHegel team’s submission to the Nakba Image Classification shared task at the Nakba-NLP 2026 Workshop. The task requires the binary classification of social media images into destruction and not_destruction categories. Given a limited and imbalanced training set of 1,400 images, we utilized a ConvNeXt-V2 Nano backbone combined with extensive data augmentation and label smoothing, prioritizing standard regularization over task-specific architectural modifications. For inference, we applied a 6-view Test-Time Augmentation (TTA) strategy using a hard-voting mechanism. The baseline system achieved a Macro F1-score of 0.8593 and an Accuracy of 0.8706 on the official private test set, ranking 6th out of 16 participating teams.

Actor-level stance detection over noisy, politically sensitive data can present challenges that standard training procedures fail to handle reliably. This paper presents KvochurHegel, our submission to the StanceNakba 2026 Shared Task, which addresses these challenges by framing stance classification as Natural Language Inference (NLI) to capture actor-level granularity. The official StanceNakba dataset contains high label noise and topic-correlated spurious features, such as texts discussing unrelated global conflicts using in-domain political vocabulary. To handle these conditions within a three-class schema, we construct templates encoding stance hypotheses for specific actors (e.g., “The author expresses support for Palestine”) and introduce a broadened neutral class designed to absorb spurious out-of-domain inputs. A DeBERTa-v3 Cross-Encoder independently evaluates the entailment between the input text and each class-specific hypothesis. Because standard cross-entropy training tends to memorize contradictory annotations under these conditions, we regularize the training procedure with R-Drop and label smoothing. This regularized setup likely contributed to robustness against distribution shifts between the competition’s evaluation phases (the public leaderboard and private test set), allowing our model to improve from a Macro-F1 of 0.9094 to 0.9384 without requiring large generative models, cross-validation, or inference-time ensembling.

This paper describes our team’s submission to AbjadMed at AbjadNLP 2026. The task involves classifying Arabic medical question-answer pairs into 82 categories, characterized by a long-tail distribution and significant semantic overlap. While domain-specific Arabic models exist, they are primarily optimized for Named Entity Recognition or span-extraction tasks rather than high-cardinality sequence classification. Consequently, our system adopts a robust optimization approach using a general-purpose encoder. We utilize ARBERTv2 as the backbone, employing Label-Distribution-Aware Margin (LDAM) loss to mitigate class imbalance and Fast Gradient Method (FGM) adversarial training to enhance generalization boundaries. Our approach achieves a Macro-F1 score of 0.4028 on the private test set, demonstrating that advanced optimization techniques can yield competitive performance on specialized taxonomies without requiring domain-specific pre-training.