The phonological grammar is probabilistic: New evidence pitting abstract representation against analogy


Abstract

Speakers and listeners extend both categorical and probabilistic regularities in the lexicon of their native language to novel forms. Ernestus and Baayen (2003); Hayes et al. (2009) demonstrate that speakers can `probability match' - producing a distribution of output forms which matches the distribution of form types found in the lexicon. As Ernestus and Baayen demonstrate, this behavior can be modeled via a probabilistic grammar with a set of abstract generalizations, or via a process such as analogy which is an epiphenomenon of the organization of the lexicon. I examine a probabilistic trend within the English stress system, showing that speakers extend it to new words, but they do not use their lexicon to do so. Speakers' knowledge of this trend is both abstract and probabilistic in nature. This supports the use of inherently probabilistic grammatical models such as Maximum Entropy (Goldwater and Johnson, 2003; Hayes and Wilson, 2008; Coetzee and Kawahara, 2013).


Documents:

Slides from the 2015 Annual Meeting on Phonology, 9 October 2015

Manuscript September 2015


Data etc:

The experiment still 'live' as of September 2015

Items: 'stress-ambiguous' prompts , fully prosodified stress choices

Experigen by Michael Becker and Jonathan Levine (2013), software used to run the experiment

WUG

    Evidence of prior linguist civilization, Fairy Glen, Skye.