Introduction
In traditional linguistics, phonotactics—the rules governing sound combinations—are often viewed as categorical. A word like blork is considered a "legal" English nonsense word, while fsneep is deemed "illegal" due to its initial cluster. However, human intuition regarding these words is rarely binary; listeners often perceive a gradient of "English-likeness." This research challenges the rigid, rule-based models of Generative Phonology and Optimality Theory (OT) by proposing a stochastic (probabilistic) grammar that aligns more closely with human gradient judgments.
The Stochastic Phonological Parser
The authors developed a Stochastic Phonological Parser to model how humans evaluate the acceptability of neologisms. Unlike standard probabilistic models that assign likelihoods to individual production rules, this model utilizes Root-to-Frontier Paths. These paths represent the entire structural trajectory from the root of a word tree down to an individual phoneme.
Key Technical Features
- Positional Sensitivity: The model tracks where sounds occur (e.g., word-initial vs. word-final) using specific tags. This captures the reality that certain sounds, like /t/, are significantly more frequent in specific positions.
- Sequential Path Unification: To parse a word, the model "zips" together consistent paths. The final probability of a word is the product of the probabilities of these unified paths.
Methods and Datasets
The model was trained on the Mitton (1992) dictionary, encompassing approximately 48,580 monosyllabic and disyllabic English words, yielding nearly 99,000 unique paths. For evaluation, the authors used a dataset of 116 nonsense words from a 1996 study by Coleman, where human subjects rated words on a scale from 0 (possible) to 12 (impossible).
The study compared four scoring metrics to predict human ratings:
- Total word probability.
- Log probability of the word ($ln(p(word))$).
- Probability of the "worst" component (the single most unlikely part).
- Probability of the "best" component.
Key Insights: The Amelioration Effect
A primary finding of this research is the Amelioration Effect. The authors observed that a highly "illegal" sound at the start of a word can be offset by a very common, high-frequency ending. For instance, in a word like mrupation, the familiar suffix "-ation" helps listeners overlook the illegal initial cluster "mr-".
This integrative process suggests that human judgment is not dominated by the single worst violation—as Optimality Theory might suggest—but is instead a cumulative assessment of all parts. Phonological competence is thus better modeled as a probabilistic system rather than a set of hard constraints.
Results and Discussion
The results showed that the log probability of the whole word was the strongest predictor of human acceptability ($p < .001$). While the "worst part" was a significant predictor ($p < .01$), it was less accurate than the holistic word score. This confirms that human linguistic judgment is an integrative computational process where the likelihood of every constituent is weighed.
Limitations and Future Work
While successful for short words, the current model is limited to monosyllabic and disyllabic structures and does not yet account for the complex recursion found in longer words. Future iterations could benefit from a more granular decomposition of onsets and rhymes to better capture the interaction between morphology (prefixes/suffixes) and phonology.
Summary updated