How many grammars am I holding up? Discovering phonological differences between word classes
Adam Albright, Massachusetts Institute of Technology
It has been widely observed that different subsets of the vocabulary can show different phonological patterns. For example, in many languages, different parts of speech show different stress or accentual patterns (English: Kelly 1992; Lenakel: Hammond 1984; Fukuoka Japanese: Smith 1999), different prosodic shapes (English: Cassidy and Kelly 1991; Classical Arabic: McCarthy 2005), and even different segmental composition (English: Sereno and Jongman 1990, Kelly 1992). Phonological differences can also be found between words in different syntactic classes ("dative shift" verbs: Grimshaw and Prince 1986; verb+particle combinations: Jackendoff 1997), different inflectional classes (German pluralization classes: Köpcke 1993; Italian verb classes: Albright 2002; English regular verbs, Albright and Hayes 2003), or even between arbitrary sets of words that are not otherwise distinguished by morphological or syntactic properties (Ito and Mester 1995; Pater, to appear). In some cases, the phonological differences make sense when we take into account the different contexts that words occur in (e.g., McCarthy 2005), but in other cases, they appear to lack any synchronic motivation.
The possibility of arbitrary phonological differences between word classes raises numerous questions about how such differences are learned. Do learners start by assuming different grammars for each subset of words, unifying them only later (if at all)? Or do learners posit fractured grammars only on the basis of overt evidence that different rules are actually required? Current trends in statistical and inductive learning might lead one to suppose that learners always start by learning detailed rules about local subsets of the data. Under this view, learning correlations between word class and phonology is simply a necessary and automatic by-product of how linguistic generalizations are discovered. In this talk, I contrast two examples that show that this simple account cannot be right. The first example (from Spanish) is a case in which speakers fail to generalize across different inflectional classes, even though for at least a subset of the structures involved, there is no overt evidence that the classes behave differently. The second example (from English) is a case in which speakers apparently ignore segmental differences between nouns and verbs, in spite of the fact that statistical -- and even categorical -- differences can be observed in the lexicon. The challenge, then, is to understand mismatches between what is found in the lexicon (as diagnosed by statistical corpus analysis) and what speakers actually learn (as diagnosed by psycholinguistic experimentation).
I propose that the answer lies in a constraint on possible grammars: the choice of word class-specific rules is a global decision, and cannot be specified on a property-by-property basis. Under this view, the decision to learn distinct sub-grammars for each class of words must be made based on the data of the entire language. The choice involves an estimate of the trade-off between achieving a more "tailored" fit with parochial rules vs. the redundancy of having to specify invariant properties separately for each word class. As a result, differences among word classes will be encoded only if there is a payoff for doing so in the language as a whole; one or two small differences, even if they are categorical, are not likely to motivate a bifurcation of the entire grammar. For English nouns vs. verbs, the overall parallelism between segmental phonotactics of the two classes precludes encoding the small differences that can be observed. Conversely, once the decision to encode separate subgrammars has been made globally, the learner is "trapped"; phonological patterns must be stated separately for each class, even if there is no evidence that the classes differ. This correctly describes the Spanish case, in which differences between some inflectional classes requires distinct encoding for every inflectional class. I sketch a Hierarchical Bayesian approach to deciding whether to adopt unified or subdivided grammars that captures this restriction.