@inproceedings{sriram-etal-2026-databees,
    title = "{D}ata{B}ees at {S}em{E}val-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization",
    author = "Sriram, Tanisha  and
      Shankar, Sathvika  and
      Anand, Sowmya  and
      Sivanaiah, Rajalakshmi  and
      S, Angel Deborah  and
      Thankanadar, Mirnalinee",
    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-1.268/",
    pages = "2120--2125",
    ISBN = "979-8-89176-414-9",
    abstract = "This paper describes our submission toSemEval-2026 Task 9, Subtask 1: Multilingual Text Classification Challenge {---} Polarization Detection. Our focus is on how classicaland transformer-based models compare whenapplied to multilingual polarization detection.We aim to understand where each type tendsto do well and where it breaks down, particularly once you move from high-resource tolow-resource settings. Our experimental setupevaluates classical machine learning models(TFIDF with Naive Bayes, Logistic Regression, and Linear SVM) alongside languagespecific transformer models across multiplelanguages. For Arabic, Bengali, German, Italian, and Spanish, we leveraged both multilingual and monolingual pre-trained transformers such as mBERT, XLM-R, AraBERTv2,BanglaBERT, and BETO. We compare individual classical and transformer-based modelsto identify which modeling choices work bestfor each language. Our results varied substantially across languages. We achieved our bestleaderboard rankings in Bengali (6th out of 48teams) and Italian (6th out of 43 teams), whileperformance was lower in Arabic (33rd out of44), German (41st out of 44), and Spanish (46thout of 48). The study highlights the value ofcomparing classical and transformer-based approaches for multilingual polarization detectionand identifies language-specific challenges forfuture improvement."
}