@inproceedings{pedersen-2026-duluth,
    title = "{D}uluth at {S}em{E}val-2026 Task 6: {D}e{BERT}a with {LLM}-Augmented Data for Unmasking Political Question Evasions",
    author = "Pedersen, Ted",
    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-1.94/",
    pages = "650--657",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper presents the Duluth approach toSemEval-2026 Task 6 on CLARITY: Unmask-ing Political Question Evasions. We addressTask 1 (clarity-level classification) and Task 2(evasion-level classification), both of which in-volve classifying question{--}answer pairs fromU.S. presidential interviews using a two-leveltaxonomy of response clarity. Our system isbased on DeBERTa-V3-base, extended withfocal loss, layer-wise learning rate decay, andboolean discourse features. To address classimbalance in the training data, we augmentminority classes using synthetic examples gen-erated by Gemini 3 and Claude Sonnet 4.5. Ourbest configuration achieved a Macro F1 of 0.76on the Task 1 evaluation set, placing 8th outof 40 teams. The top-ranked system (TeleAI)achieved 0.89, while the mean score across par-ticipants was 0.70. Error analysis reveals thatthe dominant source of misclassification is con-fusion between Ambivalent and Clear Replyresponses, a pattern that mirrors disagreementsamong human annotators. Our findings demon-strate that LLM-based data augmentation canmeaningfully improve minority-class recall onnuanced political discourse tasks."
}