Learning and learnability in phonology Adam Albright, MIT Bruce Hayes, UCLA To appear in John Goldsmith, Jason Riggle, and Alan Yu, eds. Handbook of Phonological Theory. Blackwell/Wiley. 1. Chapter content A central scientific problem in phonology is how children rapidly and accurately acquire the intricate structures and patterns seen in the phonology of their native language. The solution to this problem lies in part in the discovery of the right formal theory of phonology, but another crucial element is the development of theories of learning, often in the form of machine-implemented models that attempt to mimic human childrens’ ability. This chapter is a survey of work in this area. 2. Defining the problem Before we can develop a theory of how children learn phonological systems, we must first characterize the knowledge that is to be acquired. Traditionally, phonological analyses have focused on describing the set of attested words, developing rules or constraints that distinguish sequences that occur from those that do not. Although such analyses have proven extremely valuable in developing a set of theoretical tools for capturing phonologically relevant distinctions, it is risky to assume that human learners internalize every pattern that can be described by the theory. Indeed, it is entirely possible that there are systematic patterns that hold true of the lexicon either by sheer accident or because of a series of independent historical changes (Ohala 1981 Listener; Bybee 2001 Phonology; Blevins 2004 Evolutionary; Blevins and Garrett 2004 Metathesis; Yu 2004 Explaining). A theory of human learning should be held accountable for only that knowledge that native speakers can also be shown to have learned. Accordingly, we think it is best to begin by sticking to observables, that is, behaviors and intuitive judgments that reflect phonological knowledge that speakers demonstrably possess. We believe that one of the most powerful demonstrations of phonological knowledge is generalization of the pattern to unknown words. Using this criterion, we find support in the literature for at least three distinct types of phonological knowledge.1 Speakers possess phonotactic knowledge, meaning that they know, at least tacitly, what constitutes a legal word in their language. Halle (1978 Knowledge) gave an oft-cited example in 1 A side note: for reasons of space we will have nothing to say about a topic of great importance and relevance; i.e. phonetic learning. By this we include the induction learning of phonological categories (features, segments) from waveforms (studied by, e.g. Mielke 2005 Modeling; Lin 2005b Learning features; Maye, Werker and Gerken 2002 Infant sensitivity), language-specific patterns of phonetic realization (e.g. Keating 1985 Universal; Kingston and Diehl 1994 Phonetic), and the vast amount of free variation seen at the phonetic level (as in, for example, coarticulation; Fowler (1981 Coarticulation), Manuel and Krakow (1984 Universal), Smith (1992) Temporal). All phenomena covered here are characterizable at the level of contrasting surface entities.Albright/Hayes Learning and learnability in phonology p. 2 pointing out that brick [bɹɪk] is an existing word of English; blick [blɪk] does not exist but in principle could be a word of English, while *bnick [bnɪk] could not. Such claims can be validated not only with observations about the English lexicon, but also by observing loanword adaptation (e.g., B’nai B’rith [bəneɪ bɹɪθ], with [ə] inserted in /bn/ but not /bɹ/) and by observing experimentally elicited repetitions and ratings of nonce words. An extensive body of work has gathered phonotactic judgments on a variety of languages, documenting systematic cross-linguistic differences in structures that are deemed acceptable. Thus, the first major task of phonological learning is to determine what is phonotactically legal in the target language. Speakers of languages also have knowledge of phonological alternations. When stems and affixes are combined into words, or words into sentences, their component sounds often change in systematic ways. That speakers often internalize these patterns in their grammars is demonstrated by the substantial literature in “wug testing”, starting from Berko (1958 Child's acquisition), illustrating that speakers extend patterns of alternation to nonce stems that they learn in an experimental context. For instance the American English flapping alternation (/t/ → [ɾ] / V ___ V&) is automatically extended to novel forms; examples can be found in the wug test reported in Albright and Hayes (2003 Rules vs analogy), in which nonce verbs such as drit [ˈdɹɪt] were often pronounced with a flap in suffixed forms (dritting [ˈdɹɪɾɪŋ]). For other recent demonstrations of generalization of alternations to nonce words, see Zuraw (2000 Pattern exceptions), Albright, Andrade, and Hayes (2001 Segmental environments), Pierrehumbert (2002), Ernestus and Baayen (2003 Predicting the unpredictable), and Zhang, Lai and Turnbull-Sailor (2006 Wug-testing). Lastly, speakers possess knowledge of patterns of free variation; for example, in the idiolects of English that the authors speak, it is a predictable fact about any word containing /æ/ before /m/ or /n/ that it may be realized either as [ɛ*+ə+] or as [æ+], the latter being preferred in more formal contexts. As the research literature in sociolinguistics demonstrates, such cases could be multiplied indefinitely (for an overviews, see Hay and Drager 2007 Sociophonetics, and Pater and Coetzee, this volume). To some extent, free variation can be considered yet another form of alternation: the same word takes on different forms, but in this case alternation is conditioned by the sociolinguistic context, rather than morphological or phonological context. With this survey in mind, we can state the scientific problem at hand as follows. The goal is to “reverse-engineer” the human system, constructing a complete model that can acquire phonology exactly as people do. The model must be able learn from positive evidence, with no overt correction of its mistakes. It must learn from real-world utterances, parsing them into their component words and morphemes. It must characterize phonological well-formedness at every level (stems, words, phrases), and it must be able to synthesize novel derived, inflected, and variant forms given suitable