Unsupervised Classification of English Words Based on Phonological Information: Discovery of Germanic and Latinate Clusters

What it is about

The authors cluster 38,731 English lemmas from CELEX, represented as phoneme strings with no stress or morphology, using a Dirichlet-process mixture with a trigram phoneme model and variational inference. The clustering finds two main groups, "Sublex ≈Germanic" (23,354 words, 60.3%) and "Sublex ≈Latinate" (15,174 words, 39.2%), plus a small "-ity" group of 203 words. On 3,535 Germanic and 10,637 Latinate words with an etymology from Wikipedia, the V-measure is 0.198, above a permutation baseline (p < 10⁻⁵). The clusters recover the initial-stress contrast and show new cues such as word-final [i:p] and the bigram [h u:] for Germanic. For the dative alternation, the phonotactic clusters predict double-object grammaticality with accuracy 0.868, against 0.701 for true etymology.

Ideas

• I1. en/src/dict.rs: store a phonotactic class per English lemma, computed by a count-based trigram model with BTreeMap tables and a fixed number of classes. Use it as a feature for the dative choice (double object or prepositional) in 23. Complementation, and report accuracy against true etymology.
• I2. Derivation check in 17. Derivation II: suffixes: for each suffix, compare the class of its base with the class the suffix expects (for example -ion with Latinate bases). Flag mismatches as candidate lexicon errors.
• I3. Use the paper's representativeness score (a log ratio of cluster probabilities) to list the most diagnostic substrings for each class, as a table in en for the pupil dictionary. Keep only substrings with a frequency of at least ten.
• I4. Negative control: shuffle the lemmas' cluster labels before the dative test; the advantage over etymology must disappear.

For Mova

The English chapters are 11. Nouns, 19-lexicon-mwe-register.md and 23-complementation.md. The code is en/src/dict.rs and en/src/morph.rs. The paper's clustering is a Bayesian model; Mova would need a count-based version to stay deterministic. The data (CELEX) is for evaluation only, not for what Mova ships.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2504.11770

@misc{morita2025unsupervised,
  title = {{Unsupervised Classification of English Words Based on Phonological Information: Discovery of Germanic and Latinate Clusters}},
  author = {Morita, Takashi and O'Donnell, Timothy J.},
  year = {2025},
  eprint = {2504.11770},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2504.11770}
}

References

2 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Anderson, J. R. (1990). The adaptive character of thought. Studies in cognition. L. Erlbaum Associates, Hillsdale, NJ. find
2. Anshen, F., Aronoff, M., Byrd, R., and Klavans, J. (1986). The role of etymology and word-length in English word formation. M.S. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/morita-2025-unsupervised-classification-english-words-based/