@inproceedings{hoeken-etal-2025-just,
    title = "Not Just Who or What: Modeling the Interaction of Linguistic and Annotator Variation in Hateful Word Interpretation",
    author = {Hoeken, Sanne  and
      Alacam, {\"O}zge  and
      Nguyen, Dong  and
      Poesio, Massimo  and
      Zarrie{\ss}, Sina},
    editor = "Evang, Kilian  and
      Kallmeyer, Laura  and
      Pogodalla, Sylvain",
    booktitle = "Proceedings of the 16th International Conference on Computational Semantics",
    month = sep,
    year = "2025",
    address = {D{\"u}sseldorf, Germany},
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/iwcs-25-ingestion/2025.iwcs-1.7/",
    pages = "73--87",
    ISBN = "979-8-89176-316-6",
    abstract = "Interpreting whether a word is hateful in context is inherently subjective. While growing research in NLP recognizes the importance of annotation variation and moves beyond treating it as noise, most work focuses primarily on annotator-related factors, often overlooking the role of linguistic context and its interaction with individual interpretation.In this paper, we investigate the factors driving variation in hateful word meaning interpretation by extending the HateWiC dataset with linguistic and annotator-level features. Our empirical analysis shows that variation in annotations is not solely a function of \textit{who} is interpreting or \textit{what} is being interpreted, but of the interaction between the two. We evaluate how well models replicate the patterns of human variation. We find that incorporating annotator information can improve alignment with human disagreement but still underestimates it. Our findings further demonstrate that capturing interpretation variation requires modeling the interplay between annotators and linguistic content and that neither surface-level agreement nor predictive accuracy alone is sufficient for truly reflecting human variation."
}