@inproceedings{peerachaidecho-rutherford-2026-chulanlp,
    title = "{C}hula{NLP} at {S}em{E}val-2026 Task 6: A Hybrid {BERT}-{LLM} Framework for Political Response Clarity and Evasion Detection",
    author = "Peerachaidecho, Wisarut  and
      Rutherford, Attapol",
    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-1.236/",
    pages = "1874--1881",
    ISBN = "979-8-89176-414-9",
    abstract = "SemEval-2026 Task 6 (CLARITY: Unmasking Political Interview) focuses on detecting equivocation and evasion techniques in political interviews. While encoder-only models and Large Language Models (LLMs) individually struggle with this task, we propose a hybrid BERT{--}LLM framework to leverage their complementary strengths: the discriminative efficiency of fine-tuned encoders and the sophisticated reasoning of LLMs. We benchmarked several long-context architectures{---}DeBERTa, RooseBERT, and BigBird{---}finding that a truncated DeBERTa-large provided the most reliable candidates for the LLM. By using DeBERTa{'}s top-5 predicted labels as constrained options for LLM inference, we significantly improved evasion-level classification. This hybrid approach achieved competitive rankings in the shared task, placing 7th in Subtask 1 and 2nd in Subtask 2."
}