Similarity-biases and pattern-coherence in phonology
For the last half century, the most influential approach to these problems in phonology (Chomsky and Halle 1967 through Prince and Smolensky 1993/2004) has accounted for categorical patterns within a language through limitations on information storage and processing imposed by an innate language faculty. Under this ‘minimal memory’ approach, the language faculty operates like a calculator, in that each time a word is produced, its pronunciation is recalculated afresh by the operation of a set of algorithms over its abstract lexical representation. This hypothesized reliance on computation at the expense of memory is the underlying explanation for pattern consistency. However, a great deal of evidence amassed over the last few decades suggests that the human language faculty operates more like Google, with a greater reliance on memory storage relative to computation. This new evidence creates a serious explanatory difficulty, for it removes the phonologist’s most compelling account for pattern consistency: if pronunciation details are largely memorized for each word, why can’t pronunciations be more phonetically idiosyncratic? My work provides a novel account for the emergence of categorical patterns within languages based on self-organization driven structure formation (Wedel 2002, 2006, 2007).
Research in many fields has shown that higher-order patterns can arise without pre-specification through feedback cycling in a process called self-organization. Systems showing this type of pattern formation are often called complex systems or dynamical systems. Language exhibits all the prerequisites for being a complex system in this sense: interacting elements exist at many levels, as do feedback loops driven by repetition and imitation within and across generations of language speakers.
How do categorical sound patterns in language arise and spread?
I have proposed a Similarity-Feedback Model resting on the interaction of two conceptually distinct sources of biases in language production and perception: similarity-biased and externally-biased variation (Wedel 2002, 2006, 2007). Considerable evidence suggests that language production and perception variants are biased toward other similar words and pronunciations, creating a steady tendency for language change to recapitulate and exaggerate existing patterns. Because errors of this type feed back upon themselves to increase similarity over time, they gradually push a system toward categorical behavior even in the absence of any intrinsic or extrinsic bias toward a particular pattern (Wedel 2006, see this simulation). A second source of bias in variation (‘externally-biased variation’) can be idealized as relatively independent of experience, such as the physiological constraints that exist on sound production or perception.
Using computational simulations as models of language use and change, I explore the interaction of these two general bias-types in the nucleation, entrenchment and spread of sound patterns. There are three major results emerging from this work.
- The Similarity-Feedback Model represents the first mechanistic account of for the development of categorical patterns across the lexicon despite highly detailed memories of word pronunciations (Wedel 2002, 2006, 2007). This account contributes to the emerging framework of Evolutionary Phonology (Blevins 2004) by providing a mechanism for several central predictions made by this model.
- In natural language, when there are multiple patterns that are cross-linguistically attested in a given phonological context, one pattern is typically found in a given language to the exclusion of the others. In Optimality Theory (a current ‘minimal-memory’ model in phonology; Prince and Smolensky 1993/2004), this outcome is guaranteed by several stipulated features of the grammatical system, in particular constraint ranking and strict constraint dominance. Using computational simulation, I have shown that the types of outcomes requiring these mechanisms in Optimality Theory emerge spontaneously in the Similarity-Feedback Model as a straightforward consequence of feedback interactions between competing patterns (Wedel 2002, 2007). In Wedel (to appear) I show that these competitive interactions are conceptually parallel to those found in instances of multi-level selection in biological systems.
- Many common phonological patterns seem to make pronunciations ‘better’ in some way, and a number of phonological theories over the last half-century have posited specific mechanisms in the language faculty to favor these type of patterns. However, this type of theory has difficulty accounting for the many patterns that appear to be entirely idiosyncratic, and the fact that even common patterns are often idiosyncratic at a finer grain of analysis. Along with a number of other phonologists, I have argued for an approach in which the factors that initiate patterns are distinct from the cognitive and social processes that drive them to become categorical (e.g. Wedel 2002, Blevins 2004, Wedel 2004, Bybee and McClelland 2004, Wedel 2006, 2007). The Similarity-Feedback Model provides a foundation for this functional separation and predicts that even common patterns will show idiosyncrasy when examined closely. These predictions support major elements of the Evolutionary Phonology framework (Blevins 2004).