@inproceedings{cho-kwon-2026-language,
    title = "Can language models process linguistic deference?",
    author = "Cho, Youngdong  and
      Kwon, Chloe Dokyung",
    editor = "Voigt, Rob  and
      Warstadt, Alex  and
      Feldman, Naomi  and
      Linzen, Tal",
    booktitle = "Proceedings of the Society for Computation in Linguistics 2026",
    month = jul,
    year = "2026",
    address = "San Diego, CA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.scil-main.35/",
    pages = "369--378",
    ISBN = "979-8-89176-412-5",
    abstract = "Honorifics are linguistic forms that encode respect toward a socially valued individual or entity. This paper investigates how language models process Korean subject honorifics, which signal the social status of the subject through specific morphological markers. We evaluate a set of language models to determine whether they process honorifics in a human-like way by capturing the socio-pragmatic constraints governing their use, rather than merely relying on surface co-occurrence patterns. Our results indicate a systematic dissociation: models generally succeeded in detecting surface morphosyntactic mismatches, successfully treating unacceptable honorific constructions as less expected. However, models consistently favored overt honorific marking regardless of the subject{'}s social status, suggesting reliance on surface heuristics over genuine pragmatic knowledge. These findings suggest that language models have not fully acquired the socio-pragmatic constraints underlying honorific use, even when extensively trained on Korean text."
}