Analytic Biases for Vowel HarmonyLanguages

One of the major goals of generative linguistics is to produce a theory of language that can generate all possible languages without predicting languages that are unnatural in that they lie outside the scope of the human capacity to learn. In order to achieve this goal, the nature of language must be understood with regard to the distinction between linguistic patterns that are outside the range of possible grammars and patterns that are accidental gaps, and simply have not been documented. This difference becomes crucial when evaluating theories of generative linguistics that predict linguistic patterns that have never been observed in the set of natural languages. Specifically, grammars generated within Optimality Theory (OT) (Prince and Smolensky 1993/2004) produce a factorial typology of all linguistic patterns predicted by the given grammatical theory. All optimality theoretic grammars are predicted to be within the cognitive capacity of language users. The substantively biased theory of learning (Finley and Badecker, 2008; Wilson 2006) provides a promising means for understanding the relationship between grammars generated by linguistic theories and the characterization of patterns observed cross-linguistically. This theory of language learning hypothesizes that learning biases shape the distribution of linguistic patterns across the world’s languages. The easiest patterns to learn are consequently the most common cross-linguistically. Patterns that are phonetically grounded and/or formally concise are the easiest patterns to learn, and therefore the most likely to appear cross-linguistically, while patterns that lack these properties are avoided by the learner and are therefore cross-linguistically rare. In phonology, these learning biases are grounded in both phonetic naturalness as well as phonological naturalness. For example, learners are biased to form grammars that maximize perceptual salience and articulatory ease, but they will also be parsimonious in terms of the formalization of the grammar. These formal restrictions are characterized in terms of grammatical constructs such as natural classes and formal implementation (e.g., number of rules or constraints required to characterize the grammar); they may also include nonlinguistic factors that influence language processing such as working memory and attention. While the substantively biased theory of learning offers a means for explaining the relationship between frequently occurring and unattested linguistic patterns, there is little concrete evidence to support the notion that learning biases shape the cross-linguistic distribution of patterns in the world’s languages. Specifically, traditional methods for understanding the nature of linguistic typologies are limited to exploring attested patterns, and typically focus on frequently occurring or natural patterns. Because it is impossible to study how learners in a natural setting will cope with an unattested pattern, it is unclear why learners ultimately avoid particular unattested patterns. It may be that with exposure to the proper learning data, learners may be equally accommodating toward outwardly unnatural, unattested patterns as frequently occurring, natural patterns. Because traditional methods cannot address the ways in which learners interpret unattested patterns, this paper employs the artificial grammar learning paradigm (Finley and Badecker, in press; Reber, 1967; Wilson 2006) in order to address the question of how learners deal with data that is ambiguous between naturally occurring patterns and unattested patterns. In the artificial grammar learning paradigm, the experimenter can control the data that the learner is exposed to, making it possible to investigate the nature of learning biases towards natural versus unnatural patterns as well as