Modeling Distinctive Feature Emergence

While these two claims are not contradictory, this paper provides an account of the abundance of natural classes without recourse to innate features. Evidence is provided in Section 3 from a simulation of class emergence which is based on readily-observable aspects of phonetic similarity, provided in turn by a phonetic similarity metric detailed in Section 2. One way to examine the claim that common classes are the ones which are statable in terms of distinctive features is to conduct a survey of classes that are involved in sound patterns, and see how many of these can be accounted for in terms of features in different theories.1 Mielke (2004) reports on phonologically active classes from grammars of 561 languages, about 17,000 sound patterns. A phonologically active class is defined as any group of sounds which, to the exclusion of all other sounds in a given inventory, (a) undergo a phonological pattern, or (b) trigger a phonological pattern. The database contains 6077 distinct classes fitting this description. All of these classes are naturally occurring, and the terms “natural” and “unnatural” will only be used in reference to a specific feature theory. A class of sounds is natural with respect to a particular theory if the class is statable as a conjunction of features in that theory. A class of sounds is unnatural with respect to a particular theory if it is not statable as a conjunction of features, but rather requires special treatment, such as disjunction or subtraction of natural classes, or is unstatable. The results show that unnatural classes are not particularly rare. 3640 classes (59.90%) are natural according to the feature system of Preliminaries to Speech Analysis (Jakobson, Fant, and Halle 1954), while 4313 classes (70.97%) are natural according to the feature system of The Sound Pattern of English (Chomsky and Halle 1968), and 3872 classes (63.72%) are natural according to Unified Feature Theory (Clements and Hume 1995). 1496 classes (24.65%) are unnatural according to all three of these feature theories. Figure 1 shows the distribution of natural and unnatural classes in terms of the Unified Feature Theory feature system, which in this case is fairly representative of the three systems examined. In this chart, unique feature specifications are arranged along the x-axis in order of decreasing frequency. Light bars represent classes which are natural in Unified Feature Theory, and dark bars represent