Phonotactic Probability and the Maori Passive: A Computational Approach

Oiwi Parker JonesSigmorphon 2008

TL;DR

This paper challenges the long-held assumption that thematic consonants in Māori passive verbs are unpredictable. By employing a neural network trained on phonotactic probabilities of active verbs, the authors demonstrate that these "unpredictable" consonants can be predicted from the phonological context of the verb stem, offering a computational alternative to traditional linguistic analyses.

What Problem Does This Paper Address?

The Māori passive verb construction has been a long-standing puzzle in Polynesian linguistics. Two main analytical frameworks, "morphological" and "phonological," have attempted to explain how passive forms are derived from active verbs. A key feature of these analyses is the assumption that the thematic consonants (e.g., /k/, /m/, /ŋ/) that appear in some passive forms are unpredictable and must be "lexicalized" – meaning they are stored as part of the verb's inherent properties. This paper questions this assumption of unpredictability, proposing that phonotactic probabilities, which describe the likelihood of sound sequences in a language, might offer a predictive mechanism for these thematic consonants.

Methods

The core of this research is a computational approach using an artificial neural network. Specifically, a 3-layer feed-forward network was designed to learn the mapping from active verb representations to their corresponding passive verb suffixes. The network architecture included 199 input units, 100 hidden units, and 10 output units. The input layer was designed to represent verb stems using various coding schemes: segmental, and two syllabic approaches. The network was trained using back-propagation, a standard algorithm for training neural networks. The authors also explored "degraded" inputs (e.g., removing all consonants or vowels) to probe which aspects of the input representation were most crucial for prediction.

Datasets

The primary dataset for this study was drawn from the Māori-English section of The Revised Dictionary of Modern Māori by P.M. Ryan (1989). This provided 476 passive patterns, with 12 identified as English borrowings. A subset of 464 patterns was used for experiments, excluding the loanwords. For gerund analysis, a small number of gerunds from Ryan's dictionary were supplemented by a larger set of gerund-like tokens from the Māori Broadcast Corpus (MBC), totaling 139 gerund-like types.

Key Insights and Ideas

  • Challenging Lexicalization: The most significant insight is that thematic consonants, previously considered unpredictable and lexicalized, can be predicted based on the phonotactic probabilities of the active verb stem. This suggests that linguistic rules or patterns might be more systematic than previously assumed.
  • Power of Phonotactics: The study highlights the predictive power of phonotactic constraints in language, demonstrating their relevance beyond simple sound sequencing to morphological processes like passive formation.
  • Synergistic Linguistic and Computational Analysis: The paper exemplifies how computational methods, particularly neural networks, can be powerful tools for testing linguistic hypotheses and uncovering patterns that might be difficult to discern through manual analysis alone.
  • Syllabic Representation Superiority: The experiments indicated that representing verb stems using syllabic units, rather than individual segments, led to better prediction accuracy, suggesting that the syllable is a more relevant unit for this task. The "de Lacy analysis" (a specific syllabic coding) performed best.

Key Results

  • Representation Experiment: The best-performing input representation was the "Syllable 2" (de Lacy analysis) coding scheme, achieving 93.91% accuracy in predicting passive forms, significantly outperforming the segmental approach (90.43%) and "Syllable 1" (91.74%).
  • Thematic Consonant Prediction: For the subset of test cases involving thematic consonants, the network achieved 46.67% accuracy, a substantial improvement over chance (23.44%) and a "majority choice" strategy.
  • General Passive Prediction: The network achieved 93.91% accuracy in predicting passive forms overall, a massive improvement over the null hypothesis (4.49%).
  • Gerund Prediction: The network demonstrated strong performance on gerund prediction, achieving 100% accuracy on the two gerunds from Ryan's dictionary and an average of 90.36% on the larger MBC dataset.
  • Loanword Adaptation: The network successfully predicted passive forms for 12 loanwords with 100% accuracy, suggesting it can generalize learned patterns to new vocabulary. It also outperformed a proposed "default analysis" for loanwords.

Limitations and Future Work

  • Preliminary Nature: The authors acknowledge that the work is preliminary and that further investigation is needed.
  • "Black Box" Nature: While the network performs well, understanding why it succeeds (i.e., which specific input features or internal representations are driving the predictions) remains a challenge. The paper begins to address this by examining weight matrices and suggesting experiments with "degraded" inputs.
  • Further Investigation of Degraded Inputs: The study mentions a preliminary finding that vowel-only input outperformed consonant-only input, but calls for further investigation into the role of different phonetic features and representations.
  • Addressing "Systematic Holes": While the network outperformed the default analysis on loanwords, the authors note that additional work is needed to fully understand how the network handles cases where the default analysis fails.
  • Broader Implications: The authors suggest that this approach could be extended to explore allomorphy in other Polynesian languages, which has not been extensively studied using phonotactic probabilities.

Summary updated

BibTeX
@inproceedings{jones2008phonotactic,
  author = {Oiwi Parker Jones},
  title = {{Phonotactic Probability and the Maori Passive: A Computational Approach}},
  booktitle = {Sigmorphon},
  year = {2008},
}
Figure