@inproceedings{tran-tan-dinh-2026-ttda704,
    title = "ttda704 at {S}em{E}val-2026 Task 6: Structured Chain-of-Thought Prompting for Political Evasion Detection",
    author = "Tran Tan, Tai  and
      Dinh, An",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.355/",
    pages = "2817--2829",
    ISBN = "979-8-89176-414-9",
    abstract = "We present our system for SemEval-2026 Task 6 (CLARITY: Unmasking Political Question Evasions), which addresses political evasion detection in English question-answer pairs from U.S. presidential interviews.We compare two paradigms: (1) parameter-efficient fine-tuning of Qwen3 models (4B{--}32B) using QLoRA with tiered upsampling and weighted cross-entropy loss to address severe class imbalance, and (2) structured Chain-of-Thought (CoT) prompting with reasoning-capable API models, including DeepSeek-V3.2 and Grok-4-Fast.Our best system uses Grok-4-Fast with extended reasoning and few-shot hierarchical CoT prompting, achieving Macro F1 scores of 0.5147 on Subtask 2 (9-class evasion) and 0.7979 on Subtask 1 (3-class clarity). On the official leaderboard, it ranks 8/33 on Subtask 2 and 13/41 on Subtask 1. Ablation results show that hierarchical label presentation provides a useful reasoning scaffold and that extended reasoning helps models handle subtle pragmatic distinctions, although the strongest prompt variants are not statistically distinguishable in Macro F1."
}