SyntaxMind at SemEval-2026 Task 6: Exploring Transformers and LLMs for Unmasking Political Question Evasions

AbstractThis paper describes our approach to Subtask 1: Clarity-level Classification in SemEval-2026 Task 6. The task focuses on determining the clarity of political responses with respect to their corresponding questions. To enhance model performance, we introduced a direct answer generation strategy as an additional input feature and applied Task-Adaptive Pre-Training (TAPT) to enhance encoder-only Transformer models with the task domain. We further explored both cross-entropy and focal loss to address potential class imbalance. Experimental results show that TAPT enhanced encoder models, particularly DeBERTa-V3-base, achieved the strongest performance, while generative small language models fine-tuned via parameter-efficient methods exhibited comparatively lower results. Our system obtained a macro-F1 score of 0.72 on the official evaluation set, ranking 24th out of 40 teams.