SHU at AdabEval 2026: Category-Aware Fine-Tuning of MARBERT for Arabic Politeness and Pragmatic Function Classification

The 7th Workshop on Open-Source Arabic Corpora and Processing Tools (OSACT7) with 5 Shared Tasks

Abstract

This paper describes our submission to the AdabEval 2026 shared task, addressing Subtask A (politeness classification) and Subtask B (multi-label pragmatic category prediction). For Subtask A, we fine-tuned MARBERT using weighted cross-entropy to mitigate class imbalance. For Subtask B, we apply BCEWithLogitsloss with inverse-frequency positive weighting to address the minority categories, and we introduce a category merging strategy to reduce categories’ sparsity and annotation variation. Finally, we propose a stacked architecture where predicted pragmatic categories are injected as auxiliary features into the politeness classifier. Our results demonstrate that dialect-aware modelling, class-imbalance handling, and category-aware stacking improve Macro-F1 across both subtasks, achieving 0.85 for Subtask A and 0.55 for Subtask B on the test set.