Research

I am, very broadly speaking, interested in the interrelated questions of how languages are learned, how they change, and why they are so similar to each other (except when they're not). Somewhat more specifically, the goal of my research is to develop a cognitive theory of the emergence of phonological language universals. I shift the burden away from an innately and substantively specified grammar and onto a relatively powerful and flexible learner. This approach is structured by the following working hypotheses:

    1)  Phonological patterns arise from a substrate of phonetic and acoustic variation (a basic tenet of Evolutionary Phonology).
    2)  Detail at this level (and others) is retained in the grammar or in representations that are accessible to the grammar in the mind of the speaker/listener (a view consistent with exemplar theory).
    3)  Abstract grammatical representations are not always induced from every observable pattern, but when they are
    4)  they are the result of a learning process operating on an input which is the product of a noisy transmission process, filtered through the additional bottlenecks of memory limitations and the idiosyncrasies of language-specific lexical inventories.
    5)  The learning function itself is inherently capable of producing a richer set of possible languages than that which is attested
    6)  and the commonly occurring patterns observed in the world's languages can be attributed in large part to the properties of the data typically available to this learner.

My research program involves investigating, via TYPOLOGICAL, EXPERIMENTAL, FORMAL & COMPUTATIONAL METHODS, the above components of human language learning which I posit to give rise to phonological grammars.




Phonological Theory & Typology

Standard Optimality Theory (OT) makes several strong predictions about phonological patterns that should not exist. These predictions are based on assumptions about the make-up of the universal constraint set, and certain fixed relative rankings within that constraint set. While there are not, to my knowledge, any formal rules about what may or may not constitute a legal constraint, a pair of constraints which are direct inverses of one another seem implicitly disallowed. Thus, for example, the set {Onset, NoCoda} accounts for the typology in which the CV syllable is the optimal structure, and in which CVC and VC may be tolerated, but not preferred. {NoOnset, Coda}, allowing for systems which show the reverse preferences, is not part of the posited universal constraint set.

The typology, for the most part, appears to support asymmetries like this. However, there is a certain amount of evidence that systems that are in disagreement with the conventional OT constraints and rankings, 'anti-markedness' systems, can exist. For example, non-harmonically complete consonant inventories (/k/ but no /t/) are found in a number of languages (such as Hawaiian, Fort Chipewyan and Luangiua), and assimilation systems in which /k/ and /t/ pattern together, but differently from /p/ (harmonically gapped) can be found (for example, in Harar Oromo). Epenthesis of /g/ (rather than /t/) appears to be an active process in Buryat. And final neutralization to aspirated (rather than plain) stops seems to occur in more than a few languages (such as Eastern Pomo, Klamath and Koyukon).

Minimally, evidence from languages like these and others provides us with reason to doubt the most extreme claims of what I will call synchronic determinism: namely, inviolable restrictions on allowable grammars. Placing the universal commonalities to the world's languages upstream from the grammar itself allows {Onset, NoCoda}, for example, to arise under most circumstances, but {Coda, NoOnset} to remain a possible outcome, in certain, particular instances (perhaps as in the Australian language Arrernte). If all constraints, even the most common ones, are learned, and most conceivable systems are in fact possible, then we are left with two inter-related questions; how does that learning proceed, and why do certain grammars seem to be (sometimes vastly) more widespread than others?




Experimental Work

I have been working with an artificial grammar learning paradigm in a number of different phonological domains in order to investigate questions related to the learning process: the relationship between phonetics and phonology, and the circumstances under which learners will generalize, or regularize over their input. These experiments involve training participants on made up languages composed of naturally produced speech tokens - often just lists of lexical items paired with pictures, or mini-paradigms (such as a singular/plural alternations). While listeners are explicitly taught one thing (such as the meaning or inflectional endings for words) they are implicitly presented with other information, in some cases at the sub-phonemic level.

  • Nasal Assimilation:

    Listeners can pick up on very small acoustic cues (amount of nasalization on vowel) and learn correlations between this phonetic information and abstract morphological structure (word boundaries).


    • Click on the pictures to hear one of the singular/plural pairs participants hear during training in this experiment:

    Condition A

               

    During test, participants have to decide which of two choices is the right one for pictures like this they've never seen before (can you hear the difference?):

                      Choice 1 Choice 2


    Participants are more likely to pick Choice 2 because it has the same amount of nasalization on the last vowel (the vowel right before the plural suffix).



  • Epenthesis:

     Learners prefer to assign a single sound pattern to a single meaning (that is, avoid allomorphs for suffixes). If their phonetic expectations are met (input is phonetically natural), then they are highly likely to produce only a single suffix form. This seems to be true even when there is a large amount of acoustic variation in that input. If their input is only phonologically natural and predictable, on the other hand, they typically produce multiple allomorphs.

    • Click on the pictures to hear some of the singular/plural pairs participants hear during training (all singular words end in a vowel):

    Condition A

                

                

    During test, participants hear new singular words, like this one:         

    But they have to produce the plural themselves.
    Click the picture to hear one participant's reponse:         





Simulations & Models

I also investigate the properties of the learner from a formal, computational perspective. I assume a powerful algorithm (such as a Bayesian learner) that is capable of producing both the widely observed language systems of the world, but also the less common ('impossible') patterns. The premise is that the universal properties of cognition (here broadly construed as learning mechanisms) will produce the same types of output grammars, over and over, cross-linguistically, when operating on the same types of data (the natural products of articulation and perception). But when the learner is confronted, for whatever reason, with a different set of language data, the outcome of learning may be different. With the universals situated in the mechanisms rather than the substantive structures of grammars, the idiosyncrasies of individual lexicons and inventories may, on occasion, produce the circumstances from which an unusual language system arises. The question is how often, and under what circumstances this could happen. Answering this question is not as simple as it might first appear, and involves first answering quite a number of other questions, such as

  • What can sound change consist of, how do changes propagate, and how are they initiated?

  • What are the likely shape of lexicons (how uniformly sampled is the space of possible words)?

  • How is language acquired as a function of the words that are learned (what statistics do learners use)?

  • What is the outcome of learning (what do phonological grammars look like)?
Even if we don't know the answers to all (or any) of these questions, we can create models to test how different answers create different outcomes.
  • Sonority-To-Stress
    • Simulations of a naive Bayesian learner inducing a stress grammar over surface forms after the application of a uniform context-free vowel shift

    • Sonority-To-Stress Grammar:
      Stress high-sonority vowels (a) in preference to low-sonority (schwa)

    • Complete Sound Change:
          a > schwa


    • 1000 Randomly Generated Lexicons:
      Most generous estimate: 11.5% of cases produce anti-markedness grammar (Reversed-Sonority-To-Stress)

    • Conclusion:
      Insufficient evidence for necessity of substantive UG constraints

      •    
  • Velar Palatalization
    •  Simulations of the emerge of a new phoneme category from an existing one via probabilistic sound change

    • Probabilistic Sound Changes:


    • 1000 Randomly Generated Lexicons:

      Unnatural Phonotactics: Random Associations exceed predictability of Natural Associations in 36% of cases:
      E.g., [C] likely to occur two segments after /b/

      Some Natural Association with high, front vowel preserved in 97% of cases

      Anti-Natural Tendencies: Gradient Violations of Implicational Hierarchy: 22% of cases
         

      Anti-Natural Phonotactics: Reversal (Trend) of Implicational Hierarchy: 0% of cases