@inproceedings{booth-etal-2026-ilab,
    title = "{IL}ab-{NLP} at {S}em{E}val-2026 Task 9: Comparing {XLM}-{R}o{BERT}a and {LL}a{MA}-2 for Multilingual Polarization Detection",
    author = "Booth, Declan  and
      Abercrombie, Gavin  and
      Frenda, Simona",
    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-1.189/",
    pages = "1462--1468",
    ISBN = "979-8-89176-414-9",
    abstract = "This submission describes a system for SemEval-2026 Task 9, Subtask 1, focused on binary detection of polarized versus non-polarized posts in English and Spanish. We compare two approaches: a fine-tuned multilingual encoder model (XLM-RoBERTa) and a prompted generative model (LLaMA-2 7B). Our experiments show that XLM-RoBERTa delivers stronger and more stable performance overall, while LLaMA-2 is more prone to false positives in Spanish due to a strong bias toward predicting the polarized class. In addition to headline results, we analyse model behaviour using confidence signals and SHAP, and report efficiency measurements with CodeCarbon to highlight practical tradeoffs between performance and computational cost."
}