@inproceedings{costa-pado-2026-comparison,
    title = "A Comparison of Methods to Bias Translation Toward {P}ortuguese Variants",
    author = "Costa, Catarina  and
      Pad{\'o}, Sebastian",
    editor = "Souza, Marlo  and
      de-Dios-Flores, Iria  and
      Santos, Diana  and
      Freitas, Larissa  and
      Souza, Jackson Wilke da Cruz  and
      Ribeiro, Eug{\'e}nio",
    booktitle = "Proceedings of the 17th International Conference on Computational Processing of {P}ortuguese ({PROPOR} 2026) - Vol. 1",
    month = apr,
    year = "2026",
    address = "Salvador, Brazil",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-dnd/2026.propor-1.8/",
    pages = "78--92",
    ISBN = "979-8-89176-387-6",
    abstract = "Portuguese serves as the official language of multiple countries across four continents. It is classified into two primary variants (European Portuguese and Brazilian Portuguese), but there is limited research on and resources for European Portuguese compared to the Brazilian variant.In this paper, we consider the task of Machine Translation (MT) into Portuguese. Given the resource imbalance, standard MT systems produce translations that are typically closer to the Brazilian standard. We compare four methods available to bias the translation toward the minority European Portuguese variant that target different places in the MT lifecycle: (1) reranking $n$-best MT outputs according to a variant classifier; (2) biasing hypothesis generation at inference time toward the target variant; (3) fine-tuning for the target variants; (4) moving completely to an LLM-based approach. We find that all methods can bias translation outputs to an extent. The LLM-based approach yields numerically the highest results, but the impact of memorisation remains unclear."
}