This paper presents a lexicon-augmentedRoBERTa system for the SemEval-2026 Task2 valence–arousal regression challenge. Themodel integrates deep contextual embeddingswith a 6-dimensional feature vector derivedfrom the NRC VAD lexicon, achieving a hightoken coverage rate of 72.05%. Under officialuser-aware evaluation, the system reached acompetitive average composite correlation of0.547, significantly outperforming the ridgeregressionbaseline. The system demonstratedparticular robustness in valence (r = 0.656)and achieved strong generalization to unseenusers (rarousal = 0.519). These findings indicatethat lightweight lexicon-based statisticsprovide valuable complementary cues for longitudinalemotion modeling in modern transformerarchitectures.

This paper presents the OZemi team’s submission to SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization.We propose a unified multilingual approach that addresses multiple languages and subtasks efficiently. Our system combines multilingual models with data-level techniques and a class-weighted cross-entropy loss to mitigate data imbalance across languages, subtasks, and categories. Results show consistent performance across languages, achieving macro F1 scores above 70% in most languages for Subtask 1 achieving our highest rank in subtask 1 for Persian (1 out of 44). These results suggest that the proposed framework provides a flexible foundation for multilingual and multi-task polarization analysis.