@inproceedings{li-heinz-2026-matters,
    title = "What Matters in Tonotactic Learning",
    author = "Li, Han  and
      Heinz, Jeffrey",
    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.18/",
    pages = "180--190",
    ISBN = "979-8-89176-412-5",
    abstract = "This paper investigates whether tonotactic learning differs across representations and learning models. We conduct an experiment using the same dataset encoded in three representations: segments, features, and autosegmental representations (ARs). To the extent possible, two learning models are evaluated, the Maximum Entropy (MaxEnt) model and the Bottom-Up Factor Inference Algorithm (BUFIA), to examine how learning outcomes interact with both model type and representations. A follow-up experiment further explores the roles of frequency and complexity thresholds. The results show that (1) AR-based learning gives the strongest overall performance; (2) there is no consistent advantage between segmental and featural representations across learning models; (3) MaxEnt performance improves substantially when frequency information is introduced and lastly (4) the effects of complexity bounds interact with representation type and frequency information. These findings suggest that tonotactic learning benefits from structurally explicit representations. Overall this work highlights the importance of using linguistically meaningful representations into learning."
}