Lexicalization of Sound Change and Alternating Environments

Lexicalization of Sound Change and Alternating Environments

Bybee

1.Usage-based theory

Over the last twenty years a significant functionalist trend has developed in the study of morphosyntax with the aim of explaining the nature of grammar by studying how language is used in context. The basic premise of this work is that frequently used patterns become conventionalized or fossilized as grammatical patterns; that is, grammar is emergent from language use (Givon 1979; Hopper and Thompson 1980, 1984; DuBois 1985; Hopper 1987; and many more). Haiman (1994) has discussed the process by which repeated patterns become part of “grammar” in terms of ritualization, showing that the effects that repeated stimuli or repeated action has on an organism -- automatization, habituation, and emancipation—are also operative in the process of grammaticalization or the creation of new grammar (see also Boyland 1997 for a discussion of the psychological mechanisms involved).

Some comparable research in the phonological domain has begun to appear in recent years. For instance, several studies have shown that speakers’ judgments of the grammaticality of phonotactic patterns is based on the frequency of consonant and vowel combinations actually occurring in the language (Pierrehumbert 1994, Frisch 1996). In addition, speakers’ ability to access the lexicon may involve a complex interplay between the frequency of words and the number and frequency of words with similar phonological shape (Pisoni et al. 1985). Connectionism offers the possibility of formally modeling the effect of use on mental representations of language, and such models have been tested in the phonological and morphological domains (for example in Dell 1989 and Daugherty and Seidenberg 1994).

(...)

First, I present evidence that many, if not all, sound changes progress in lexical items as they are used, with more frequently used words undergoing change at a faster rate than less frequently used words. Then I examine “alternating environments” -- cases in which a sound in a particular word or morpheme is sometimes in the environment for the change to take place and sometimes not. In cases in which the targeted sound is at the edge of a word, the change can go through even where the sound is not in the appropriate phonetic environment and thus no alternation is produced. In such cases, we have evidence for the restructuring of the lexical representation of the word. However, when the alternating environment is inside of a word, the change can be retarded even in the appropriate environment, but eventually an alternation can be created, showing, again, restructuring of the lexical representation of the word. I argue that lexical representations are restructured gradually on the basis of actually occurring variants of a word and that postulating words and frequent phrases as the units of representation explains the development of word-level phonology. In addition, it will be argued that reference to the frequency with which words begin in consonants explains why a final word boundary often conditions changes as though it were a consonant.

2. The frequency effect on sound change

One of the aims of this chapter is to explore from a phonological perspective the size and nature of storage and processing units. I will present evidence that words and often longer units, such as frequent phrases, are the units of lexical storage. For the moment, however, I assume that words are the units of lexical storage. It is reasonable to assume that lexically stored words are in many ways like other mental records of a person’s experience. First, there is no reason to believe that these memorial records have details and predictable features abstracted away from them (Langacker 1987; Ohala and Ohala 1995), and second, it is reasonable to believe that new experiences are categorized, to the extent possible, in terms of the already-stored record of past experiences (see Klatzky 1980).

Each use of a word requires retrieval by the speaker and a matching of the incoming percept to stored images by the hearer (and the speaker, who is monitoring his or her own speech). My thesis in this chapter is that the act of using a word, in either production or perception, has an effect on the stored representation of the word. We already know this is true in terms of the degree of entrenchment of a word (or the resting level of activation): high-frequency words have stronger representations that make them easier to access, more resistant to change on the basis of other patterns, and more likely to serve as the basis for the creation of new forms (Bybee 1985).

In addition, certain levels of use affect the stored representation of words by actually changing their shapes. That is, along with the entrenchment effect of frequency, there is also an automation effect: words and phrases that are used a lot are reduced and compressed. This effect is very salient in grammaticizing phrases (such as 'going to' becoming 'gonna' and 'want to' becoming 'wanna') and more conventionalized contractions (such as 'won’t' and 'didn’t'), but it also occurs in a more subtle form across the lexicon when a sound change is taking place. Sound changes (phonetically motivated changes, which are usually the reduction of the magnitude of gestures or retiming of gestures; Browman and Goldstein 1992) tend to be phonetically gradual and also lexically gradual: high-frequency words undergo change at a faster rate than low-frequency words. The effects of frequency in the diffusion of a sound change through the lexicon have been shown for vowel reduction and deletion in English (Fidelholtz 1975; Hooper 1976b), for the raising of /a/ to /o/ before nasals in Old English (Phillips 1984), for various changes in Ethiopian languages (Leslau 1969), for the weakening of stops in American English and vowel change in the Cologne dialect of German (Johnson 1983), for ongoing vowel changes in San Francisco English (Moonwomon 1992), and for tensing of short a in Philadelphia (Labov 1994:506–507). In a recent paper, I have shown that there is also a frequency effect in the application of t/ d deletion in American English (Bybee 2000). Deletion occurs more in high-frequency words, including of course monomorphemic nouns and adjectives, but also regular past tense verbs, a point to which I will return later.

My interpretation of the frequency effect in the diffusion of sound change (following Moonwomon 1992) is that sound change takes place in small increments in real time as words are used. The more a word is used, the more it is exposed to the reductive effect of articulatory automation. The effects that production pressures have on the word are registered in the stored representation, probably as an ever-adjusting range of variation. Thus words of higher frequency undergo more adjustments and register the effects of sound change more rapidly than low-frequency words.

The frequency with which a word is subject to the ravages of articulation is not the only factor that encourages sound change. We also have to take into account the fact that certain speech styles allow more reduction and compression than others. In particular, casual speech among familiars typically shows more reduction. Thus words that are used in casual situations will also undergo change at a faster rate (see this volume, chapter 2, and D’Introno and Sosa 1986). Of course, these words are also likely to be those that are of higher frequency overall.

Another factor affecting reduction is the status of the word within the discourse. Fowler and Housum (1987) found that the first use of a word in a spoken text was longer than in subsequent uses. This means that speakers articulate more clearly in the first use of a word, where identification by the hearer might be more difficult,
and then allow the reductive processes to apply later when identification by the listener is aided by the context and the fact that the word has already been activated. In fact, speakers may use reduction to indicate that a referent is not new but rather one that has already been accessed in the discourse. Words that are used more often within a text are produced in reduced form more often. If the produced form affects the stored form, then words that are repeated more often in a discourse will reduce at a faster rate than words that are repeated less often.

3. Exemplar-based representations

The account of phonetically gradual lexical diffusion of a sound change given in the preceding section requires a model of memory storage for linguistic units based on actual tokens of use. Each experience of a word is stored in memory with other examples of use of the same word. These memories of specific tokens are organized into clusters with more frequently ocurring exemplars, and tokens that share many properties with high-frequency exemplars are treated as more central, while less common or more deviant tokens are treated as more marginal. Thus linguistic experiences are categorized in the same way as other types of perceptual experiences. Rather than conceiving of stored representations as abstractions from the phonetic tokens, representations are considered to be the result of the categorization of phonetic tokens. This proposal, which will be referred to as “the exemplar model,” adapts proposals made by Miller (1994) for phonetic segments and Johnson (1997) for larger units. Similar arguments for phonological representations have been made by Hooper (1981) and Cole and Hualde (1998). Note that this model does not distinguish between phonetic and phonemic features in lexical representation (see Steriade 2000). Further implications of this model for sound change will be discussed in the next sections.

4. Alternating environments

Given that produced tokens affect stored representations, what would happen when a word or morpheme occurs in different environments, such that it is subject to a change in one environment but not in another? In the case of such “alternating environments” (as Timberlake [1978] calls them) two or more different surface forms map onto a stored form. How are such alternate mappings resolved? 

Here I approach this question by examining cases of sound change in progress. It is necessary to distinguish the phonetic variation that goes on while a sound change is in progress from the conventionalized alternations that can eventually arise from such sound change. By alternation I mean that a word or morpheme has two or more variants that are not phonetically continuous or variable but rather constitute discrete alternants conditioned by specific phonological, grammatical, or lexical contexts. An alternation, then, roughly corresponds to the level of variation generated by a classical phonological rule. By studying the conditions under which such alternations are conventionalized and the conditions under which they are not, we learn something about how variants of words and morphemes are organized in memory.

The study of alternating environments in cases of sound change in progress reveals that the outcome differs according to whether it is a word or a morpheme that is in the alternating environment. When the same morpheme is in an alternating environment in different words, a change is retarded even in the conditioning environment (the Timberlake Effect; see below), but an alternation can eventually arise. When the alternates are in two forms of the same word, alternations arise only under special conditions, but ordinarily only one alternate survives. I will argue that the differential behavior of morphemes and words with respect to sound change in progress provides strong evidence for the stored representation of words and frequent phrases.

(...)

Ordinarily, alternations do not develop where the conditioning is across a word boundary. This fact gives rise to the notion of “word-level phonology”—that is, the fact that most alternations occur within words. The explanation being investigated here is that ordinarily there is only one stored representation for each word. Where variation arises during change in progress, the variation is resolved in terms of one variant or the other. The exceptions to this arise only in the case of frequently used phrases, to which I will return shortly.

First let us consider how variation at the word level is represented and how cases of sound change in an alternating environment would eventually be resolved. In the exemplar model described earlier, the representation of a word is a cluster of actually occurring tokens, with more frequent tokens accumulating greater weight or strength. Thus each word has its own range of variation dependent upon its frequency and the contexts in which it is used. When little or no sound change is affecting a word, the range of variation in the tokens may be small and relatively stable. During change, however, the range of variation increases and the center of the cluster gradually shifts.

When the same word occurs in both an environment that conditions a change and a non-conditioning environment, as in the Spanish s-aspiration case, the cluster for a word may divide into two (or more) subclusters, each one with a strong center of high-frequency tokens. In this case, each subcluster is associated with one environment -- the word-final [s] tokens with the environment before a vowel and the word-final [h] tokens with the environment before a C. It appears that such a situation is unstable when the environment is not also part of the representation, because it tends to be resolved in favor of one variant for all environments. That is, the most frequent variant, the weakened consonant, [h], wins out and tends to be chosen even in contexts before a vowel.

In contrast, when the environment is part of the stored unit, an alternation can be established, in the sense that the [s] can remain before a vowel. This happens in frequent phrases.

(...)

Thus my hypothesis is that words and frequent phrases are storage units and that ordinarily there is only one representation per word, so that variations in the form of a word are normally reconciled to a single form and no alternation is created through sound change. Exceptions to this occur when a word is used in high-frequency phrases and/or phrases involving grammatical morphemes, such as pronouns and articles. This hypothesis makes strong predictions about the conditions under which sandhi phenomena will develop. It predicts that sandhi processes will only occur in phrases of high frequency and most commonly in those involving grammatical morphemes or other high-frequency words. This prediction is borne out by the most famous cases of sandhi, such as French liaison (Tranel 1981).1 The hypothesis also predicts that cases of reduction restricted to certain “syntactic” environments, such as English auxiliary contraction and the reduction of don’t, will occur only in the most frequent contexts in which the form appears. This prediction is also borne out where it has been tested (Krug 1999; Bybee and Scheibman 1998; see also Bybee 1998).

4.2. Sound change inside of words

Of course, alternations do develop inside of words. A morpheme inside a word may undergo phonological change, producing a new allomorph and thus an alternation. This fact follows from the hypotheses presented earlier: if sound change permanently affects stored units and words—even morphologically complex ones—are the units of storage, then the same morpheme in different words will take on different phonological shapes.

Further evidence for the hypotheses developed here is the fact that the effect of an alternating environment inside a word is very different from the effect across word boundaries. Inside a word, a variable process never applies outside of its phonetic environment (as, say, the aspiration of /s/ in Spanish occurs even ____##V). Instead the effect is the reverse: there is evidence that a change can be retarded even in its phonetic environment if it occurs in a morpheme that also has alternates that appear outside of the phonetic environment. Timberlake (1978) has made this point by presenting examples of changes that progress faster in uniform environments and are retarded in alternating ones.

(...)

These examples show that the implementation of a sound change in particular words (that is, its lexical diffusion) depends heavily on the contexts in which the sound is used. Since I have been arguing that the unit that serves as the context for a sound as it undergoes change is the word, then we must now consider how to account for the fact that the environment of a morpheme in one word affects the rate of change for the same morpheme in a different word. To understand this issue, we must understand the nature of morphological relationships, a matter to which I now turn.

5. A network model

In various works (Bybee 1985, 1988, 1995) I have proposed that the lexicon is organized into a complex set of relations among words and phrases by connections drawn among phonologically and semantically similar items. Parallel phonological and semantic connections constitute morphological relations if they are repeated across multiple pairs of items. As Dell (2000) points out, morphological relatedness is the joint effect of the organization of words into phonological and semantic neighborhoods.

In this model, the relations between base and past-tense forms of English verbs are diagrammed as in figure 10.1,3 where semantic relations are not explicitly shown and where relations of similarity (rather than identity) between segments are shown with broken lines. Affixes are not explicitly listed in storage but emerge from sets of connections made among stored words and phrases. The very high type frequency of the regular English past tense strengthens its representation in memory and makes it highly productive. It can then apply to verbs whose past-tense forms are not accessible because they have never been encountered or are of such low frequency as to not be easily accessible.

Past tense constitutes a category, but not one that can be accessed independently of a particular verb, because it is a category to which verb forms may belong or not belong. How then do individual tokens of the past-tense suffix relate to one another? This is, of course, an empirical question, and here is the evidence we
have so far.

First, instances of the suffix attached to a verb are affected by the token frequency of the whole verb form: the rate of deletion for a final [t] or [d] on a pasttense verb is affected by the frequency of the form, as shown in Bybee (2000). In that study, past-tense forms with a frequency in Francis and Kucera (1982) of 36 or greater are considered high frequency and those with a frequency of less than 36 as low frequency, following a suggestion by Stemberger and MacWhinney (1988), who establish that the mean frequency of inflected verbs in Francis and Kucera is 35. Using this cutoff point, I (Bybee 2000) find that there is a significant difference between high- and low-frequency verbs in the position for deletion. See table 10.4. 

Second, in the data I cited earlier, the overall trend for past-tense [t] and [d] is that they delete less often than [t] or [d] in monomorphemic words.

A related third point is that Losiewicz (1992) has shown that not only are monomorphemic [t] and [d] shorter than past-tense [t] and [d], but high-frequency past-tense [t] and [d] are shorter than low-frequency [t] and [d]. Losiewicz proposes to account for her data by a dual-access model in which high-frequency morphologically complex forms are stored and retrieved as wholes, while low-frequency forms are composed by adding the suffix to a base form using a schema. But this model would predict the same rate of deletion in high-frequency regular past tense forms as in monomorphemic forms of comparable frequency, and this prediction is not borne out by the data. Rather, in the data used in Bybee (2000), the rate of deletion for all words with frequencies of 36–403 was 54.4%, while the rate for past-tense forms of the same frequency was 39.6%. Thus we must posit that the past tense in low-frequency verbs, which is longer and phonetically fuller, can have some effect on the past-tense suffix on high-frequency verbs, which is shorter and more prone to deletion but not as short and prone to deletion as the [t] and [d] of monomorphemic forms. Thus the fuller form of the suffix on low-frequency verbs has some impact on the suffix on other verbs.

(...)

This account of the Timberlake Effect makes predictions about the circumstances under which the effect will be the strongest. A change will be retarded most noticeably in an alternating environment when the alternates that are not in the environment to undergo the change are the most frequent—either there are more conditions in the paradigm in which the change does not take place or the environments that do not condition a change occur in the unmarked or most frequent categories. Furthermore, it is less likely that the Timberlake Effect will be observed in high-frequency paradigms in which individual forms have a greater lexical strength (accessibility) and weaker connections with related words (see Bybee 1985).

Now compare again these word-internal cases to those where change is occurring at a word boundary. The word may for a time have multiple variants, suggesting either a range of variation in the changing segment or even multiple representations for a single word. However, in this case the tendency is to resolve the variation in favor of a single form for each word, except in the case of high-frequency or grammatical words. It appears that the cases in which distinct alternates become established are just those cases in which the conditioning environment is registered in storage with the alternating item. Thus in the case of progu, progi each variant can be registered because we are dealing with two different (although related) words, one of which consistently has the palatalizing environment and one that does not. Similarly in frequent phrases, such as muchos año(s), the [s] preceding the vowel may be preserved (as though it were word-internal) because the conditioning vowel occurs with it in storage and processing. Other instances of the same word may occur without the [s], as [mucoh] or [muco]. Such variation would not necessarily exist indefinitely. Unless one variant is in a highly entrenched phrase, the variation is likely to be eventually leveled out.

Thus by registering words in the lexicon and establishing connections among them we are able to account for the two different effects on sound change of alternating environments inside of words and across word boundaries.

6. Lexical phonology

Some of the effects of variable processes that I have discussed earlier have been addressed by Guy (1991a, 1991b) in the context of Lexical Phonology. This proposal is relevant here, even though I will argue that it does not work for all the cases at hand, because it incorporates the notion argued for here, that some words behave as if variable processes have applied to them more than once. Guy proposes that variable rules may apply cyclically and at all levels of a Lexical Phonology and offers an account for the variation in t/d deletion that is conditioned by the morphological structure of the word. The facts are as follows: on average, the highest rate of deletion of /t/ or /d/ takes place in monomorphemic words, the next highest rate in pasts with vowel changes (such as slept, left, told), and the lowest rate in regular past-tense forms.

The advantage of the Lexical Phonology approach is that it does recognize that fairly low-level variable phonology is deeply entwined with the lexicon and morphology. It also suggests that the greater progress of a sound change can be attributed to more applications of the “rule.” The problem with it is that it makes incorrect predictions in some cases and it cannot deal at all with frequency effects across lexical items.

Consider first another case of an alternating environment that Timberlake describes. Timberlake brings up this case to show that it is the alternating environment itself, and not the morpheme boundary, that causes the retardation of change.

(...)

This example shows that the Lexical Phonology approach is fundamentally the wrong approach, for it is a fact of usage, not structure, that is accelerating the change: since the /d/ in /ado/ is in the context for reduction and deletion no matter what verb it is added to and since many verbs with this suffix are of very high frequency, there is a frequency effect to accelerate the change and nothing to impede it. 

Finally, Lexical Phonology, as a theory of structure and not a theory of usage, cannot account for the frequency effects demonstrable in the lexical diffusion of sound change. In Bybee (2000) I have shown that t/d deletion occurs more often in words of higher frequency. This is true of all of the 2,000 tokens studied; this relation also holds when nouns and adjectives, semi-weak past-tense verbs, and regular past-tense verbs are considered as well. Since the semi-weak past-tense verbs are all of high frequency, frequency of use alone can account for their higher rate of deletion over the regular past verbs. I conclude, then, that variable rates of phonological change are the product of usage, not of structure.

7. Conclusions

The evidence discussed in this essay bears on two issues regarding the nature of stored memory for linguistic forms. First, the minimal unit of independent storage is the word, which is also the minimal unit of production, since smaller units cannot be used in isolation. I hasten to add, however, that that does not mean that other much longer sequences are not stored and processed as wholes. Here we have seen evidence that frequently used phrases behave like single processing units (just as words typically do) in that they preserve segments that might otherwise be lost at word edges. In various papers I have argued for a highly redundant storage mechanism that includes specific instances of phrases and clauses as well as more generalized constructions as storage and processing units (Bybee and Scheibman 1999; Bybee 2000).

The view of sound change as affecting sounds in words according to their context of use allows us to understand why most phonological alternations occur at the word level: alternations can only be established in cases in which the conditioning environment is present in the storage and processing unit. Words or other units that occur in alternating environments that are not part of the stored unit will not have variants but rather will resolve any variation in favor of one form or the other. This proposal also allows us to make interesting predictions about the development of liaison or sandhi phenomena. Conventionalized alternations across traditional word boundaries indicate that at least one alternate is part of a larger stored unit. Thus such liaison alternations can be used to study the nature and size of storage units. Finally, the view of sound change as affecting sounds in words provides an account of the different effects of alternating environments inside of words and across word boundaries.

The second major aspect of the model presented in this essay is that sound change has an immediate and permanent effect on stored representations. This view contrasts with the generative and structural view that underlying representations remain fixed and sound change is “rule addition”—nothing more than a change in the phonological component. The evidence that sound change has an immediate effect on the lexicon is that words change gradually and at different rates according to their token frequency, even while a “rule” is still “variable.” The evidence that such change is permanent is the fact that old underlying forms never resurface, even when the “phonological rule” becomes unproductive (see Cole and Hualde 1998 for more evidence on this point). Instead the progress of a change is inexorably unidirectional in both a phonetic and a morpho-lexical sense. In the phonetic sense, we see the unidirectionality in chains of reduction and assimilation changes, such as those shown
in (8), where one change builds on the other and continues its direction:

(8)

t→d→ð→Ø
s→h→Ø
k’ → kj → c

If stored items are changed gradually and the motivation for increased automation remains fairly constant, then the continuous nature and strict directionality of such changes is predicted. If sound change were “rule addition,” there would be no explanation for why, for example, after “adding the rule” d → ð / V_V, a language would
go on to “add the rule” d → Ø / V_V.

Inexorable unidirectionality is also apparent in the morphologization and lexicalization of the results of sound change. While no one disputes that morphologization eventually takes place, I have shown here and elsewhere that involvement with the lexicon and grammar occurs very early (Hooper 1976a, 1981; Bybee 2000). Examples given earlier are the word frequency effect in lexical diffusion, the lower rate of deletion of morphemic /t/ and /d/ in American English, the appearance of aspiration for earlier /s/ before a vowel at the end of a word in Cuban Spanish, and the examples described as the Timberlake Effect. Once involved with the lexicon and morphology, alternations become more and more entrenched and can only be undone by the strong pattern pressures we know as analogical leveling.

The two hypotheses of words as storage units and the immediate and permanent effect of sound change on words explain why most phonological alternations occur at word level, that is, why word boundaries block phonological “rules” and morpheme boundaries do not. I have also shown here that the tendency of final word boundaries to act like consonants follows from these hypotheses and from the fact that the segment most frequently following a final word boundary is a consonant.

The larger theoretical message is that use impacts representation, a point often made in studies of the discourse origins of syntax and a point that is also being made by connectionist modelers of language. As I have argued here, many cases of what was earlier postulated to be structural turn out to be derivable from the way language is used. I also see many instances where a careful look at use brings to light new data that was ignored before. I suspect that a usage-based perspective will be very productive in generating new questions and new answers in phonology.