@inproceedings{limsuwan-rutherford-2026-chulanlp,
    title = "{C}hula{NLP} at {S}em{E}val-2026 Task 5: Regression-Calibrated {LLM} for Word-Sense Scoring",
    author = "Limsuwan, Wayu  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-1.343/",
    pages = "2723--2728",
    ISBN = "979-8-89176-414-9",
    abstract = "Word Sense Disambiguation (WSD) is typically framed as a classification task that selects one correct sense for a word. However, real language is often less clear-cut, as a homonym may support several plausible interpretations. SemEval 2026 Task 5 addresses this limitation by introducing plausibility rating, where models estimate how likely each sense is in a narrative context, aligning predictions with graded human judgments. We use GlossBERT and BEM as encoder-based baselines and show that large language models (LLMs) produce more accurate plausibility estimates. Building on this observation, we propose a regression-calibrated LLM model that applies linear regression to adjust raw LLM outputs to better match human annotation patterns. Our calibrated model achieves the highest within-standard-deviation accuracy among our evaluated systems, demonstrating that lightweight post-hoc calibration can substantially improve LLM performance on graded semantic judgment tasks."
}