@inproceedings{fuganti-etal-2026-umusp,
    title = "{UMUSP} at {S}em{E}val-2026 Task 9: Mitigating Cross-Lingual Interference via Selective Multilingual and Multitask Specialization",
    author = "Fuganti, Julio Cesar  and
      Ferreira Leite Da Silva, Tulio  and
      Gala, Adelino  and
      S. Marcondes, Francisco  and
      Machado, Jos{\'e}  and
      Novais, Paulo",
    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-1.401/",
    pages = "3199--3208",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper proposes a selective multilingual and multitask fine-tuning strategy for online polarization detection that improves cross-lingual stability over fully joint training. Covering all three subtasks {---} polarization detection (POLARDETECT), polarization type classification (POLARTYPE), and rhetorical manifestation identification (POLARMANIFEST) {---} across all 22 languages of the shared task, the approach introduces controlled specialization, where languages and subtasks are grouped empirically and separate specialist models are fine-tuned for each subset. Restricting parameter sharing substantially improves performance even without ensemble averaging, whereas ensembling jointly trained models fails to mitigate instability. The final specialist ensemble improves Task 3 macro-F1 from 0.3330 to 0.4920 and reduces cross-lingual dispersion (CV: 0.613 {\textrightarrow} 0.321). Under the official ranking framework, the system ranks 7th among 16 submissions with complete multilingual and multitask coverage and remains within 5{\%} of the best system in 37.70{\%} of evaluation conditions."
}