@inproceedings{amal-etal-2026-ai,
    title = "{AI}@{UMS} at {S}em{E}val-2026 Task 6: Handling Long Question-Answer Pairs with Sliding Window Models for Clarity and Evasion Analysis",
    author = "Amal, Ikhlasul  and
      Zafar, Zia Ul  and
      Firdaus, Choiru  and
      Pamungkas, Endang",
    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-1.437/",
    pages = "3544--3548",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper presents the AI@UMS system for SemEval-2026 Task 6: CLARITY - Unmasking Political Question Evasions. The task requires classifying question-answer (QA) pairs from political interviews along two dimensions: clarity level (Subtask 1) and evasion technique (Subtask 2). A key challenge is that political interview transcripts often exceed the 512-token input limit of standard transformer encoder models. We address this with a sliding-window fine-tuning strategy applied to roberta-base, where each QA pair is segmented into overlapping windows of 512 tokens with a stride of 256 tokens. Per-window predictions are aggregated via softmax probability averaging across multiple windows and across an ensemble of three independently trained models with different random seeds. We further employ class-weighted focal-inspired loss and label smoothing to mitigate the pronounced class imbalance in both subtasks. Our system achieves macro F1 scores of 0.62 (Subtask 1) and 0.48 (Subtask 2) on the official evaluation set."
}