Ewan Dunbar

25
nov.
2015.
10h00
12h00

59/61, rue Pouchet, Paris; salle 159

Ewan Dunbar (LSCP)
Beyond feature economy: Geometric properties of natural inventories

Abstract - In this talk I demonstrate empirically that sound inventories show two tendencies for systematicity in their shapes that have not been previously documented. The explicit claim that sound inventories are systematic goes back at least to Martinet. Martinet's idea of "economy" was, roughly, that, sound inventories will tend to attract related sounds, and repel outlier sounds, over historical time: so {p, t, f, v, m} might perhaps attract [z] or [b] and lose [m]---anything that would allow the language to use relatively few parameters to define the largest number of sounds, predicting compact, efficient, or "economical" inventories. Economy has indeed been shown to be above chance levels in natural inventories (Clements 2003, Mackie and Mielke 2011). But, in fact, there have been claims over the years of many different kinds of systematicity in how sounds are arranged in inventories, and economy is the only one that has been rigorously investigated empirically. One property whose empirical basis is weak is "dispersion," the idea that sounds in an inventory tend to be far apart from each other to maximize contrast, so that {i, a, u} would be a better inventory that {i, ɪ, u}. Another property is "correlation": the tendency for sounds to form analogical pairs: [p]:[f]::[b]:[v]. This predicts that {p, t, f, v} would rather attract [b] than [z], even though both would make equally economical use of voicing, place, and manner. That is because we have both [p]:[f]::[b]:[v] and [p]:[b]::[f]:[v], whereas [z] only supports [v]:[z]::[p]:[t]. This talk shows in a large database of sound inventories that both of these properties are above chance in natural inventories. The tendency for correlation is particularly robust. And, despite some recent suggestions to the contrary, (Vaux and Samuels 2015), a tendency for dispersion is present in spite of the fact that both the frequency distribution of sounds and the tendency for correlation will work against it. Finally: these tendencies may have different interpretations depending on whether we think sounds are mainly related cognitively by gradient auditory or motor relations, or by classically binary distinctive features, or by privative elements, but they need to be explained regardless. Thus I sketch some possible origins for them, and I propose experiments that would help tease apart different theories, and possibly shed important new light on phonetic and phonological processing.