Greg Kochanski

Using Neologisms to Test Theories of Speech Production

Greg Kochanski and John Coleman, Oxford University Phonetics Laboratory Project funded 2008-2011 by the ESRC.

There are two commonly accepted models of speech production in linguistics: the classical view, based on Generative Phonology (e.g. Noam Chomsky and followers), and recently, exemplar models. We plan a series of experiments to test these rival models of how we process and produce speech. So far, it has not been easy to distinguish them. Experimental efforts have concentrated on differences between rare and common words and dense and sparse neighbourhoods, and they have not been decisive.

In the generative model, as we begin to say a word, we plan our tongue motion based on a set of shared mental objects and some broadly applicable rules. It assumes that the mind is parsimonious with memory, storing only what cannot be deduced and computing much of language on the fly.

Exemplar models assume that people store many different examples of words, remembering more and computing less. In this view, you might remember many variants of ``dog'', from different people, with different accents, and in different circumstances. Similarly, when speaking, you choose among stored patterns of tongue motion for each word.

Exemplar and generative descriptions of speech production make different predictions about the pronunciation of words that have never before been spoken by the experimental subject. Here, the differences are substantial and may be more decisive, because they are based on core differences between the two theories.

Exemplar models emphasize the role of individual examples of speech that the subject has heard of a particular word. Exemplar models predict that any examples you hear will become part of your collection of possible pronunciations. Thus, pronunciations you hear will affect your pronunciation. Especially with a new word that you have never heard before, the collection will be very small and even a single new example should be able to make a noticeable difference to the pronunciation of the word. And, changes should accumulate as the examples accumulate. Classical (generative) models emphasize pre-existing rules and features. They predict that a new word will be pronounced in a stable, rule-governed way. Even in extended versions of the classical models where change is possible, there is no accumulation of the effect from one example to the next. The difference between a cumulative effect or not should be detectable.

Thus, by giving subjects different examples as they pronounce a new word, we can distinguish the two types of models by a combination of experiments. The experiments involve bringing a "priming" word to the subject's attention a moment before they read the a new word they have never seen before. We then study how the pronunciation depends on the priming word. Then, in a separate experiment, whether changes accumulate or not after the new word has been pronounced a few times.

This research has the potential for substantial impact because it has a chance to disprove one or the other important model that is used to describe how the human mind produces language. Or, at least, we should be able to put substantial limits on how one or both of the models must work in order to be consistent with experimental facts.