This page intentionally left blank chapter 23 PHONOLOGY geoff nathan 1. Basic Principles 1.1. What a Cognitive Phonology Will Look Like As with other levels of language, ‘‘doing’’ phonology within Cognitive Grammar requires a radical revision of how linguists think about their subject matter, par- ticularly as compared with the dominant worldview of Generative Grammar. How- ever, phonology is in a rather different position from other fields in that phonol- ogists have not all adopted the dominant Chomskyan paradigm, with its attendant commitments to modularity, innateness, and the independence of language structure from other cognitive processes. In fact, even generative phonologists often take a strongly functional attitude, some even arguing that all phonology is either functionally motivated or conventionalized. A second difference with syntax is that there is considerable continuity within the fields of phonology from its inception in the latter part of the twentieth century to the beginnings of the twenty-first. Many of the categories and theoretical constructs that were introduced in the early de- velopment of phonology are still considered valid by virtually all theoretical bents, despite numerous theoretical revolutions. Phoneme, syllable, consonant, vowel, feature, and even process have some status in virtually all current phonological theories, both generative and functionalist. Although the Generative Grammar tradition has evolved considerably during the past fifty years, some fundamental principles would be accepted by both generative and nongenerative phonologists, as was noted by reviewers of the proceedings volume on Formal and Functional Linguistics (Carnie and Mendoza-Denton 2003). 1.2. The Phoneme as Category Phonology can generally be defined by its practitioners’ attitude toward the idea of the phone. The phoneme originated in the nineteenth century and continued into the twentieth along a number of different, often competing, lines. Some linguists over the period emphasized the autonomy of the concept (and, indeed, of all strictly linguistic concepts) from any functional motivation or explanatory aspect rooted in language use or nonlinguistic factors such as anatomy or physiology, while in con- trast there is also a long tradition of functionalist phonology that is alive and forms a continuous thread back to the very beginnings of phonology. This is quite dif- ferent from the current state of syntactic theory, where both the dominant theory, the Minimalist Program, and most active competitors (such as Head-Driven Phrase Structure Grammar) represent radical departures from grammatical wisdom as developed over the past hundred or more years. 1 Since Cognitive Grammar is best viewed as a kind of Functional Grammar (in the broad, generic sense), we can expect to find inspiration within the functional phonology tradition that will save cog- nitive grammarians from reinventing the wheel. Over the past fifteen years, cognitive linguists have attempted to develop ways of doing phonology consistent with the assumptions of Cognitive Grammar. In part this has been a challenging task, as the theory has evolved considerably since its inception, with emphasis on the nonmodular nature of cognitive representations in the early period (and thus concern with metaphor, Idealized Cognitive Models, and Langackerian diagrams) moving to a focus on usage as a primary mechanism for acquisition and structure in recent years. Two main avenues of research have developed over that time, one continuing the emphasis on phonology as the rep- resentation of our knowledge of bodily experience (Nathan 1986, 1996, 1999, forth- coming) and the other expanding on the usage-based model (Bybee 1999, 2000, 2001). This chapter will explore the commonalities and differences between the two views and make some suggestions for future research. 1.3. The Phoneme’s Checkered Career A brief review of the history of the phoneme will set the stage for an understanding of the issues involved in developing a cognitive view of phonology. Baudouin de Courtenay ([1895] 1972), the first synchronic phonologist, argued for a psychological, embodied view, namely that phonemes were mental images of sounds that speakers systematically deformed in the ongoing process of speech according to ‘‘physiophonetic’’ principles that were universally determined by the nature of the speech production and perception apparatus. Saussure ([1916] 1974), as the originator of the structuralist autonomous view, on the other hand, argued that phonemes were arbitrary contrast points in a network of sound defined one against another. He used the analogy of a chess game, where it does not matter what each piece is made of or even whether it has the right shape, as long as it has the 612 geoff nathan value agreed upon for that piece in the game—an explicitly anti-embodiment view. Later in the twentieth century, this conflict between the phoneme as a mental con- struct and as an arbitrary linguistic pawn in the language game continued. Jones (1967: 7) argued that phonemes are a small ‘‘family of sounds, each family consist- ing of an important sound of the language together with other related sounds which, so to speak, ‘represent’ it in particular sequences or under particular conditions,’’ although he was noncommittal on the question of whether there was, in addition, a single abstract image of the basic sound (217). Trubetzkoy ([1939] 1969: 36), on the other hand, argued that a phoneme is ‘‘the sum of the phonologically relevant properties of a sound [but] actual sounds are only material symbols of the pho- neme.’’ Among American structuralists, Bloomfield (1933: 79) sided with the au- tonomous side, stating that a phoneme was a ‘‘bundle of distinctive features,’’ but Sapir ([1933] 1972: 23) agreed with Baudouin: ‘‘In the physical world the naı ¨ ve speaker and hearer actualize and are sensitive to sounds, but what they feel them- selves to be pronouncing and hearing are ‘phonemes.’’’ Pike (1947: 145) agreed with Sapir that phonemes were psychologically real entities, arguing, for example, that the Trager-Smith phonemicization of English was probably incorrect in part be- cause it was so hard to teach to native speakers of the language. Within the generative tradition there has always been a claim to the psycho- logical reality of the entities posited within the theory, but with greater and lesser degrees of seriousness. When Chomsky and Halle (1968: 259) posited the extremely abstract underlying representations for English that claimed that the underlying forms of the language had hardly changed since Chaucer’s time, other generative phonologists, such as Kiparsky, took them to task for positing entities for which there was little evidence of psychological reality as evidenced by historical and other behavior. Kiparsky showed that historical changes applied to sounds at the pho- nemic level, not at the more abstract level posited by Chomsky and Halle. He also argued that native speakers treated the classical phonemic level, and not a more abstract one, as the basis for speech production. Kiparsky (1982) is the classic locus of this debate. The term ‘‘phoneme’’ was explicitly rejected as a label for the kinds of underlying forms that Chomsky and Halle and others were positing, but Kiparsky (1982) and Schane (1971) dissented, arguing for a level equivalent to the more tra- ditional one. Since the development of lexical phonology, a similar debate has taken place, with many claiming that the output of the lexical component constitutes the only psychologically real ‘‘underlying’’ forms. Postlexical rules actually apply in speech production, unlike lexical rules, which merely represent the relationships between related sets of lexical entries. At least some proponents of lexical phonology have argued that the forms generated by the output of lexical rules are essentially the units of storage—a typical discussion can be found in Gussenhoven and Jacobs (1998: 119–24). Outside of generative phonology, two major American schools took differing tacks on this same question. Bybee, in her early work (i.e., Natural Generative Pho- nology, which was a highly constrained form of generative phonology; see Hooper phonology 613 1976), argued that only surface forms have any psychological reality, with related forms (including most allophonic alternations) being linked by networks of con- nections. Bybee has modified and expanded her views within Cognitive Grammar, arguing for a usage-based theory, as discussed in Bybee (1999, 2000, 2001) and in which the concept of the phoneme as traditionally understood is rejected entirely in favor of a ‘‘usage-based’’ approach. 2 The other major nongenerative phonological theory of the latter part of the twentieth century is Natural Phonology, which essentially adapted Baudouin’s original insight into modern linguistic theory by arguing that phonemes were men- tal sound images that speakers modified in speaking and that, in perceiving others’ speech, the same speakers sympathetically perceived the ‘‘deformed’’ output as what they would have aimed at had they said the same thing (for extensive dis- cussion, see Donegan and Stampe 1979; Donegan 1986; Stampe 1987). The defor- mations were mental adaptations caused by the inherent nature and limitations of the speech tract and perceptual system and, as such, were universal but ‘‘learned’’ (in the same sense as a child learns to control its hands or feet). 3 I have argued (Nathan 1986, 1996, 1999, forthcoming) that Natural Phonology is Cognitive Pho- nology, at least in many of its basic assumptions, and I will argue for this view below. Finally, within the past ten years a new ‘‘generative’’ paradigm has arisen, Optimality Theory (OT) (Prince and Smolensky 1993; Kager 1999). OT presents a radically different way of thinking about representations and rules in which crucial aspects of grammar are innate (for some theorists in the Chomskyan sense, but for others in the Stampean sense), but the grammar does not really consist of rules at all, but rather violable constraints which systematically compete with each other. For a particular language, a particular ordering of preferences wins out, and lan- guages differ not in which constraints are active, but rather in which ones take preference in a conflict. OT does not exactly have a view on the nature of pho- nological representations themselves—deep, surface, or otherwise. Some discus- sion has centered around whether we need to construct underlying forms or whether any input at all will produce appropriate output (this is the theory of ‘‘the richness of the base’’). On the other hand, without a notion of an input form the concept of ‘‘faithfulness’’ is not coherent, because there is nothing for the surface candidates to be faithful to. In a sense, OT cannot claim to be a theory of interest to cognitive grammarians, because no serious claims for psychological reality have been made by many of the practitioners (however, Winters and Nathan, (2006, MS), makes a contrary sug- gestion). Still, there are a subset of OT theorists who have argued that all OT con- straints must be grounded in properties of the human articulatory and perceptual system. Representative work has been done by Kirchner (1997), Boersma (1998), and Hayes (1999, 2004). To the extent that we believe phonology to be determined by the material out of which it is built, these functional OT phonologists are doing work that needs to be looked at, although their work is formed within a framework that is generally not of interest to Cognitive Grammar. 614 geoff nathan In attempting to formulate a theory of phonology within the general worldview entailed by Cognitive Grammar, I have argued (Nathan 1986, 1996, 1999) that the original insight of the earliest phonologists was not mistaken and that people re- ally do perceive phonemically. That is, they hear their language as a string of basic sounds, traditionally called phonemes. They are not normally aware of the varia- tions in those basic sounds that are induced by their position in the word or larger prosodic unit, only becoming aware of those variants if they show up in the ‘‘wrong’’ place (note that I use the scare quotes so as to include not only nonnative speakers producing nonnative patterns, but also the ability to recognize other dialects of one’s own language—although normally without the ability to say exactly what is ‘‘wrong’’). Furthermore, production processing errors that displace those basic sounds normally result in the sounds being produced with the variant appropriate to the new environment (as virtually all research on so-called Spoonerisms has found; see Fromkin 1973, 1980, 1988). Evidence from children’s acquisition and perception of speech also indicates that speakers are actively constructing phonetic patterns in the process of speaking and that this processing may well produce forms that speakers (especially children) may never have produced or even heard before. One fundamental insight that Cognitive Grammar can bring to phonology is that the identification of phonemes is simply a matter of categorization; that is, phonemes are cognitive/mental categories. As such, the principles of categorization that Cognitive Grammar crucially relies on will apply in phonology as much as in other areas of linguistic behavior. This means that all of the apparatuses that Cognitive Grammar has developed to understand the structure of categories applies also to phonemes. 1.4. Radial Sets and Processes I will begin by establishing that the basic building block of phonological structure, the phoneme, is a psychologically real entity, the existence of which linguists need to account for, and that therefore there is some place for phonological theory within Cognitive Grammar. This is an important point, because one of the fundamental tenets of Cognitive Grammar is the ‘‘content requirement’’ (Langacker 1987; Langacker 2000: 8), which holds that the only real linguistic units are semantic, phonological, or symbolic structures (i.e., there are no intermediate level units, such as D-structures, that are not either sounds or senses). In order to postulate an apparently abstract unit such as the phoneme, which in some sense is neither a sound nor a meaning, we need to justify its existence by showing that speakers and hearers behave as if they speak and hear in phonemic, not merely phonetic, surface terms. Although much more could be said on this point (and much of this is discussed in Nathan, forthcoming), I will merely point out that the vast majority of languages of the world have writing systems where the basic symbols are virtually identical phonology 615 to phonemes—it seems to be very easy to learn such a writing system, especially compared to the more complex morphologically based ones such as Chinese and Japanese. In order to understand how we can classify sounds into categories despite the fact that they constitute categories not associated with semantics, we need to begin by noting that individual phonemes are not semantic units at all—sounds asso- ciated with semantics are, of course, morphemes. Phonemes are abstract categories of sounds qua sounds. There is, of course, no semantics associated with sounds such as thunder or the wind blowing in the leaves (aside from the placing of the particular sound in some category), but this does not make these sounds in any way less of a category, nor does this make the assignment of individual instances to the category in any way problematic. On the other hand, it is possible that fragments of words may acquire some semantic associations—for some discussion of the possibility of semantics inherent in sounds or groups of sounds themselves, see Palmer’s (1996: 279–89) discussion on sound symbolism. The work of Lakoff (1987) established the importance of radial categories as a fundamental linguistic organizing principle. He shows how members of fairly dis- parate categories, such as the senses of the word over or classifier systems, could be unified once we give up the idea that all categories are Aristotelian in the sense that all members have to share common identifying features. This leads quite naturally to a view of phonemes that allows insightful clarification of many of the problems that had confronted classical and generative phonological theory. In the first major published work dedicated solely to the issue (Nathan 1986), I presented an analy- sis of the problematic (and nonproblematic) aspects of the American English pho- neme /t/. The problem with American /t/ is that the instantiations of that particular phoneme are wildly diverse and, taken as a whole, share no common point or manner of articulation. The facts are as follows: There are at least five different variants of the /t/ phoneme in American En- glish, as shown in (1): (1)[t h ] tall [t] stall [?] button (as in [b¼?nµ]) [t’] What! [Q] water 4 Notice that there is no single feature that all of these sounds share, not even [–voice] (flap is voiced); yet there is no doubt that native speakers categorize all of them as ‘‘kinds of /t/.’’ In fact, it normally takes several months in a phonetics class before students can begin to be aware of the fact that these variants exist at all. The fact that naive speakers do not normally notice these differences constitutes what phonologists and psychologists have referred to as ‘‘phonemic perception.’’ If speakers are asked to make judgments about sounds as sounds, they may well notice that things are off, or odd, but such things normally pass unremarked, unlike, say, 616 geoff nathan a wrong ending or article choice. But if we assume that all of these sounds con- stitute ‘‘the phoneme /t/,’’ we will be unable to adequately describe even these basic facts unless the category that we set up is non-Aristotelian (and hence, un- like any standard structuralist model). In Nathan (1986) (see also Taylor 1995: 222–34), I argued that this sound cate- gory could be understood as a classic example of a radial set, with a prototypical central member and the other members of the phoneme radiating outwards ac- cording to well-defined phonetic principles, analogous to the extensions described by Lakoff involving such principles as metaphor, metonymy, and image schema transformation. Evidence from a number of different sources suggests that the voiceless unaspirated form [t] is the central member. For example, according to Maddieson (1984: 32), 99.7% of languages have either dental or alveolar stops, and children appear not only to acquire them early, but in most cases to acquire them by replacing their native language voiced and aspirated stops with voiceless unaspirated ones, whatever the phonemic system of the target language. Hurch (1988) argues that aspirated stops are nonprototypical compared to unaspirated ones. From the prototypical voiceless aspirated stop it is possible to adjust the target toward alternative forms, such as the aspirated variety, the form with simultaneous glottal closure (all voiceless stops in English are produced with simultaneous glottal closure when syllable-final), and, by extension, the glottal stop. The idea of sounds having prototype structure is not unique to Cognitive Grammar; it has been suggested by various researchers within the phonetics com- munity. Representative work can be found in papers by Samuel (1982) and Kuhl and Iverson (1995). A clear discussion of one aspect can be seen in Lotto (2000:194), which looked at perception of vowels and found that ‘‘the best /i/ exemplars were judged to be those furthest from the /i/ distribution (i.e., low F 1 and high F 2 ). Interestingly, these exemplars of /i/ would be very rare in natural speech because vowels are often reduced (moved away from the extremes of the F 1 Â F 2 space) in normal speaking contexts’’ 5 (194). Each of these extensions is licensed by what Stampe called a ‘‘natural process.’’ According to Stampe (1979: 1), a natural process is ‘‘a mental operation that applies in speech to substitute, for a class of sounds or sound sequences presenting a specific common difficulty to the speech capacity of the individual, an alternative class identical but lacking the difficult property.’’ In Cognitive Grammar terms, this means that allophones are image schema transformations of prototypical sounds in ways that allow them to fit the particular environments (see Nathan 1996 for more extensive discussion). Let me make an analogy here to Lakoff’s (1987) discussion of the word over. We prototypically think of this word as referring to a trajector lo- cated ‘above’ the landmark (i.e., oriented vertically with respect to the force of gravity), but we can certainly put some wallpaper ‘‘over’’ a hole in the ceiling. Just as we see the hole in the ceiling ‘‘upside down’’ without noticing that we have made any kind of change or extension, so we produce an aspirated /t/ without being aware that we have made it sound any different than the original target, which, as I argued above, is unaspirated. phonology 617 The idea that we come pre-equipped with natural responses to motor difficulties that we need to unlearn is not mysterious, nor is it at all foreign to the basic principles of Cognitive Grammar. For example, in teaching the skills involved in karate I have observed that students are resistant to moves that require the arms to do two different activities at the same time. Even after numerous repetitions of some move, students revert (often under conditions of cognitive overload, such as standing in an unusual position or stepping backwards rather than forwards) to hand movements that I have not taught them, that they have never seen before but are more ‘‘natural’’ in the sense that the human body has a natural preference for limb movements to follow a pattern of bilateral symmetry. Similar ‘‘spontaneous’’ errors occur in learning to dance when an asymmetrical move is required. The emergence of naturally motivated patterns is a part of all human skilled motor learning, and speech production is unlikely to be exempt. Natural Phonology’s ‘‘natural processes’’ are nothing more than the articulatory and perceptual in- stantiations of such ‘‘errors,’’ and they clearly have a role in a Cognitive Phonology. Why do I argue that these changes are made ‘‘online’’ in the context of speech production? There are a number of reasons for making this argument—many of them assembled in Stampe (1968), a famous unpublished conference paper titled ‘‘Yes, Virginia ’’ 6 The following reasons strike me as completely persuasive, al- though recently Bybee and others have challenged this view. Their views will be discussed below. We first begin by noting that speaker behavior strongly supports the idea that they do not store allophonic variants. Speakers cannot, under normal circum- stances, even hear differences among allophones. It was Sapir’s famous paper mentioned above that first pointed this out when he noted that his informant did not perceive the intervocalic voicing and spirantization in his language ‘‘in terms of the actual sounds, but in terms of an etymological reconstruction [we would now say, underlyingly]’’ ([1933] 1972: 24). Similarly, all of the early ‘‘contrastive’’ second- language literature pointed out that the phonemic filter of a first language pre- vented speakers from hearing their own allophonic variation as an instance of a phoneme in some other language (this was first discussed extensively in the classic work by Weinreich 1970). Not only can speakers not hear allophonic variation, virtually all writing systems discount it, and rhyme systems do not count as rhyme sounds that are identical if they are members of different phonemes, but do count sounds as identical if they are members of the same phoneme, even if they are dif- ferent (see Stampe 1987 for discussion). Second, there is the question of how many variants speakers store. Since each phoneme may have a number of distinct allophones, each lexical item could be pronounced differently each time it is used, depending on which words precede and follow it. This, for example, includes such allophonic variation as that dis- cussed above, so that, for example, hat will come out differently as follows: (2) [hæQ] Put your hat on your head. [hæ?] I can’t find my hat now. 618 geoff nathan However, not only does every obvious allophone used as an example in in- troductory textbooks need to be dealt with, there is in fact far more variation than those standard textbooks normally discuss. For example, all vowels are affected by the height and backness of the preceding and following vowel, even over a word boundary. Thus, each vowel will have to be sensitive to the initial vowel in the following word, as well as each final vowel in the preceding word. This fact seems to require that we attribute phonological behavior to active, online computation, since otherwise every monosyllabic word in the language would have to have a separate entry for every vowel in the language and every multisyllabic word would have to have two entries for every vowel in the language, not including the sepa- rate entries required for the initial and final consonants and variations induced by rate and formality. There would thus be an exponential increase in the size of the lexicon unless variations could be computed in the process of speech. While it is true that much more of speech is stored than is normally believed (and certainly irregular, and even some regular, morphological alternations are probably stored), it would seem likely that the computational load required to adjust the target toward the appropriate allophone for each environment would be much less than that required to select the appropriate form of each entire word relative to the surrounding words. Additional reason to believe that allophonic variation is computed during speech comes from speech errors. Numerous researchers over the years have found that speech errors normally occur at the phonemic level and that the sounds moved to their new environments are always adjusted to their new positions. If words were stored in their purely surface form, we would expect displaced voiceless stops to be aspirated after /s/ and unaspirated initially, but neither change occurs. Furthermore, speech errors produce otherwise nonexistent forms that cannot possibly have been produced on the basis of existing stored forms unless those forms are ‘‘spelled’’ in strings of separate segments. For example, while aiming at same time zone I caught myself saying [t h eimzain], producing something which does not correspond to an existing word. Unless the segments are stored separately, in some sense, they should be unavailable to be moved in a production error such as this one. Finally, child language acquisition indicates that children are doing online processing, reconstructing abstract phonemic representations rather than simply recording the ambient words and later reconstructing more abstract schemas from them. For example, children often produce words in ways that they have never heard them. One classic example concerns an American child who systematically referred to mittens as [mit h @nz] despite the fact that both parents always and exclusively said [mi?nµz]. Numerous parallel examples can be found in such sources as Smith (1973). One striking example involves a child’s pronunciation of dog as [gO] (David Stampe, p.c.). The initial velar is attributable to a commonly reported process of velar harmony, but we cannot say this child is simply repeating what he or she has heard, but rather that the child has stored the target form correctly, is processing it, with the child’s production not even including the trigger of the harmony process that caused the deformation in the process. It is impossible for phonology 619 . (this is the theory of ‘ the richness of the base’’). On the other hand, without a notion of an input form the concept of ‘‘faithfulness’’ is not coherent, because there is nothing for the surface candidates. considerable continuity within the fields of phonology from its inception in the latter part of the twentieth century to the beginnings of the twenty-first. Many of the categories and theoretical constructs. research. 1.3. The Phoneme’s Checkered Career A brief review of the history of the phoneme will set the stage for an understanding of the issues involved in developing a cognitive view of phonology. Baudouin