Categorical and gradient phonology: Theory meets methodology

James Myers, National Chung Cheng University

There has recently been an explosion of research on the phonological role of phonetically gradient representations, but much of it overlooks a central paradox. Namely, how can we distinguish gradient and categorical phonological processes if both are realized in the gradient medium of speech? A standard argument for gradient processes points to systematic differences between phonological categories along a continuous scale (e.g. incomplete neutralization). Yet since the comparisons are between categories, such evidence at best shows only that phonetic realizations may differ from informal intuitions (e.g. there may be subphonemic categories). Another approach is to note gradient phonetic shifts caused by shifts in some other gradient variable (e.g. lexical frequency). Though valid, this approach has limited usefulness, since most grammatical variables are categorical (e.g. featural, prosodic, or morphological contexts).

In this talk I discuss a different way to distinguish categorical and gradient processes, building on the assumption that the categorical/gradient distinction should correlate with other distinctions. The case study concerns variable t/d-deletion in English, which has been claimed to apply both lexically and postlexically (Guy 1991a,b). One argument for this claim is that final /t/ is more likely to delete in monomorphemic (M) words like "mist" (deletion can apply either lexically or postlexically) than in regularly inflected (R) words like "missed" (deletion can only apply postlexically, after -ed suffixation). Another argument is that word-internal context affects deletion in M words more than in R words, while word-external context affects both equally.

Phonetic research has shown that t/d-deletion involves gradiently overlapping gestures (Browman and Goldstein 1990). However, since lexical processes are generally assumed to be categorical, M words should also sometimes surface with /t/ categorically deleted, though such outputs need not neutralize with words like "miss", without underlying /t/, given the possibility of subphonemic categories (Myers 1995).

The methodological challenge is that the distributions for categorically t-less and gradiently t-reduced M tokens cannot be observed separately, since the specific path followed by any given M token is unpredictable. Fortunately, there is a recently developed statistical technique for inferring the most likely locations of overlapping distributions, and it is easily implemented in R, the free statistics package. This technique finally offers the opportunity for a conclusive test of the predictions of Myers (1995). Namely, both M and R words should share a component reflecting gradient t-reduction, but M words should show an additional distribution reflecting categorical t-deletion. Moreover, the first component should be equally sensitive to word-internal and word-external context, while the second should only be sensitive to word-internal context.

Back to the Phonology Workshop home page