SemEval-2026: 1st Place Winner (Task 13)

In the SemEval-2026 competition (Task 13), our team achieved the 1st Place global ranking in Subtask C. We developed a sophisticated ensemble architecture that pushed the boundaries of current NLP evaluation techniques.

Technical Approach & Contributions:

• EDA & Feature Engineering: Conducted extensive Exploratory Data Analysis to identify critical semantic nuances, followed by developing handcrafted linguistic features that significantly improved the performance of our team’s base classifiers (CatBoost, XGBoost).
• MIL-based Learning: Implemented Multiple Instance Learning (MIL) strategies to handle weak supervision and improve model robustness across diverse evaluation tasks.
• Ensemble Integration: Collaborated on the integration of high-precision classifiers into a larger stacking architecture (leveraging ModernBERT and UniXcoder), ensuring optimal feature synergy and predictive performance.

Results:

• Rank #1: Achieved the highest score in Subtask C among international participants.
• Research Impact: The methodology is currently being documented for a research publication, focusing on the effectiveness of hierarchical ensembles in low-resource semantic tasks.