Learnability Shapes Typology: The Case of the Midpoint Pathology

Abstract

The midpoint pathology (in the sense of Kager 2012) characterizes a type of unattested stress system in which the stressable window contracts to a single word-internal syllable in some words, but not others. Kager (2012) shows that the pathology is a prediction of analyses employing contextual lapse constraints (e.g. *ExtLapseR; no 000 strings at the right edge) and argues that the only way to avoid it is to eliminate these constraints from Con. This article explores an alternative: that systems exhibiting the midpoint pathology are unattested not because the constraints that would generate them are absent from Con, but because they are difficult to learn. This study belongs to a growing body of work exploring the idea that phonological typology is shaped by considerations of learnability.* Keywords: learnability, phonology, stress, typology 1. Introduction. One of the goals of linguistic research is to construct theories that make the right typological predictions: theories that predict the existence of all and only those patterns attested in the world's languages. In constraint-based theories of phonology, such as optimality theory (OT; Prince & Smolensky 2004), the typological predictions of a constraint set can be evaluated by exploring its factorial typology. The notion of factorial typology is grounded in the classical assumption that the set of constraints (Con) is universal, but constraint rankings are language-specific. If all constraints are freely rankable, then the set of systems predicted by a given constraint set, its factorial typology, is equivalent to the set of systems generated by each possible ranking of constraints.When evaluating a factorial typology, there are at least two important questions that the analyst must ask. First, does the predicted typology undergenerate: does it fail to predict certain attested patterns? Second, does the predicted typology overgenerate: does it fail to predict only the attested patterns? Undergeneration is typically viewed as a serious problem, since we want our theories to be able to account for the full range of linguistic variation. Thus the usual response to undergeneration is to modify the contents or the structure of Con, with the goal of including all attested patterns in the predicted typology. The response to overgeneration, however, is more nuanced. Because there are multiple reasons why a proposal might overgenerate, there are multiple possible responses.One common response to overgeneration is to take a closer look at Con and propose modifications that exclude the predicted but unattested patterns from the factorial typology. These strategies can be roughly divided into two groups. Some researchers dispute the idea that all constraints are freely rankable and propose that, in order to model typological generalizations, certain constraint rankings must be universal and therefore immutable (see e.g. Prince & Smolensky 2004 on fixed rankings for peak and margin hierarchies, Steriade 2001 on fixed rankings of correspondence constraints). Others focus ...

The binary-to-ternary rhythmic continuum in stress typology: layered feet and non-intervention constraints

“…Typological underrepresentation may well be strongly related to reduced learnability. Second, the non-directional character of unattested BTU-patterns is usually only apparent in forms of considerable length, which again reduces the learnability of the non-directional pattern, given reasonable assumptions about the limited amount of relevant input to the learner in the form of long words (Stanton 2014). Consider, for example, the third BTU-pattern in category F, which is identical to the first and second ones for forms with fewer than six syllables.…”

Unbounded circumambient patterns in segmental phonology

“…discussion of word-order typology in Bickel 2007: 241), language-specific historical contingencies (Nichols 1992, Blevins 2004, Hansson 2008, Harris 2008 and differential learnability (e.g. Hayes & Wilson 2008, Culbertson 2012, Moreton & Pater 2012, Stanton 2016 17 as productive alternatives to innate constraints. We independently need explanations and theories of each of these domains, and each is a strictly simpler type of explanation of typological commonality and rarity than the non-specific, catch-all assumption of innate knowledge or constraints.…”

Using simulated learning to account for tone typology

“…Staubs [25,26] shows that learning biases explain the limited size of stress windows, obviating the need for the explicit stating of such size limits in grammar theory. Similarly, Stanton [27] shows that learning considerations can provide an alternative account for the typological finding that no language places stress on the middle positions of a stress domain. Specifically, learning such middle-stress patterns is complicated by the fact that exposure to longer words is necessary to disambiguate the pattern from edge-based patterns, and that the used learning algorithm extracts conflicting information from the various data types it is exposed to.…”