Torin Hopkins, PhD

AI & Music

Linguist vs. Machine

Comparing machine learning algorithms with trained linguists at uncovering morphological grammar rules.

Read the paper (PDF)DOI
Linguist vs. Machine

Can linguists writing grammars by hand beat neural networks at learning word inflection? As part of a large team, I helped develop 25+ finite-state morphological grammars and compared the effort and accuracy against sequence-to-sequence models.

Approach

Each grammar combined a lexicon-based model with a "guesser" for unseen words, built as a cascade of finite-state transducers. We also built tools to help linguists triage difficult morphophonological phenomena and hypothesize inflectional class membership.

Key findings

  • Surpassing neural models’ accuracy required significant development effort by trained linguists.
  • The paper weighs the cost/benefit of hand-written grammars, which is especially relevant for low-resource languages.

Publications

  1. Linguist vs. Machine: Rapid Development of Finite-State Morphological Grammars

    Sarah Beemer, Zak Boston, April Bukoski, Daniel Chen, Princess Dickens, Andrew Gerlach, Torin Hopkins, Parth Anand Jawale, Chris Koski, Akanksha Malhotra, Piyush Mishra, Saliha Muradoğlu, Lan Sang, Tyler Short, Sagarika Shreevastava, Elizabeth Spaulding, Tetsumichi Umada, Beilei Xiang, Changbing Yang, and Mans Hulden

    Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pp. 162–170, 2020.

    AbstractPDFDOI