Printer Friendly

Similarity and contrast in segmental phonology *.


It is well-known that both similarity and contrast play an important role in language. What is much less clear, however, is why one aspect of language is sensitive to similarity while another is under the sway of contrast. This article develops a theory that allows one to predict under which circumstances similarity wins out over contrast and vice versa. Five diverse areas of language, to wit language structure, change, use, acquisition, and loss are scrutinized in an effort to erect this theory on a fairly broad empirical basis. In all of these areas, the paradigmatic axis is found to generate similarity effects whereas the syntagmatic axis gives rise to both similarity and contrast. Whether similarity, or contrast prevails on the syntagmatic dimension depends on the linear distance between the critical units. When these are very close together, contrast predominates. With increasing distance, however, contrast gives way to similarity, with a hypersimilarity effect occurring briefly in between. These results are interpreted within a model in which linguistic structure and change are understood as a response to processing strategies and biases. The representational side of the model provides a measure of determining similarity and contrast relationships between linguistic units, whereas the processing side is responsible for the strength of similarity and contrast in specific constellations. While the theory is detailed for the phonological domain, it holds' the potential of being more widely applicable.

1. Introduction

Similarity and contrast are two pervasive linguistic principles whose importance appears not to be contested by anyone. In fact, they have been shown to be operative at virtually all levels of linguistic analysis. A few examples may suffice to prove the point. A fairly general enunciation of the principle of contrast is semiotic in nature. When two signs differ in form, they also differ in meaning (e.g. Bolinger 1977; Haiman 1980; Clark 1988). The effects of contrast are most obviously felt at the lexical level. The lexicons of natural languages are structured such that violations of the principle of contrast have the status of exceptions. In English, for example, the incidence of homonymy accounts for an estimated 7.4% of the entire stock of words in general use (Berg 2001). Compared with many other languages, this figure probably puts English even at the high end of the scale. Homophony and homography are of equally limited importance. (1)

It was de Saussure (1916) who made the strongest case for contrast in the present context. According to him, this principle is in effect what enables language to function properly. A linguistic unit, be it the "signifiant" or the "signifie," cannot be defined with reference to itself, but only with reference to all other elements in the system from which it distinguishes itself. This definition is negative rather than positive in the sense that the relationship between any pair of elements is one of contrast ("X is not Y"). The attribution of properties ("the element X has the property x") does not help.

Likewise, the principle of similarity makes an impact on language. Again adopting the semiotic perspective, we may draw attention to a similarity relationship between the "signifiant" and the "signifie" which is known as iconicity. The basic insight is that an aspect of the meaning of a sign is reflected in its form and vice versa. Standard areas where iconicity plays a role include onomatopoeia, complexity (formal complexity mirrors conceptual complexity), and proximity (formal proximity mirrors conceptual proximity) (e.g. Lakoff and Johnson 1980; Haiman 1985).

A second domain where similarity can be observed is paradigms. The German definite article system is a case in point. All six allomorphs begin with the same consonant, four have the same vowel, and five end in a consonant. This is a clear instance of phonological similarity (Seiler 1967). Larger paradigms show similarity relationships within their subsets. Take the system of irregular verbs in English. A conspicuous feature of these verbs is that the members of the subclasses are close phonological neighbors (e.g. to cling, to fling, to string). Inversely, the principle of similarity leads us to expect that suppletion is uncommon, which it in fact is.

It may even be argued that paradigms thrive on the similarity relationship holding among its members. The greater the number of phonological neighbors and the greater their phonological similarity, the healthier the paradigm, that is, the less likely it is to be subject to erosion and the more likely it is even to attract new members (see Bybee and Moder 1983; Prasada and Pinker 1993; Burzio 2002). Similarity thus manifests itself as an organizing principle of paradigms. The opposite is true of dissimilarity. Comrie (1979) showed that a great phonological dissimilarity between the distinct case forms of morphological paradigms facilitates the loss of the alternation. (2)

The final example is metaphor and the principles underlying its creation. Ever since Aristotle, perceived similarity between the source and the target domain has been identified as a major motivating factor. Similarity establishes the connection between the two domains. It is a sine qua non because without it, there would be no basis for the speaker to invent a metaphor and for the listener to understand it.

It is obvious that similarity and contrast are antithetical to each other. The stronger the one, the weaker the other. To the extent that the two principles vie for the same domain, this duality creates a tension in the linguistic system which has to be released. Exactly how do these countervailing forces operate and interact? Under what circumstances does the one prevail over the other? What are the underlying mechanisms giving rise to similarity and contrast? What is universal, what is language particular about these two principles?

Surprisingly enough, these and other questions have only seldom been given the attention they deserve, let alone received satisfactory answers. To the best of my knowledge, the only two theories that explicitly address the relationship between similarity and contrast have been developed in phonetics. While one theory is production-based, the other is oriented towards perception. Attention will be focused here on the former (for the latter, see Shigeno 1991). In the production theory, similarity is the result of inertia or economy on the part of the speaker whereas contrast springs from the speaker's orientation towards the listener. The speaker's needs are epitomized by the principle of ease of articulation, the listener's needs by the principle of perceptual contrast (e.g. Liljencrants and Lindblom 1972; Lindblom et al. 1984; Janson 1986; Kawasaki-Fukumori 1992; Joanisse 1999). The speaker has to expend muscular energy to configurate her articulators appropriately and is likely to invest less rather than more energy in consonance with the principle of minimum effort. As the expenditure of energy may be defined as the degree of displacement from a resting position or a previous position, sounds are prone to resemble one another, hence the bias towards similarity. The constraints on speech perception are entirely different. The listener's task is to recover the phonological structure of an utterance. This is all the easier, the greater the identifiability of the acoustic cues and their temporal structure. Clearly, careful pronunciation enhances this identifiability and thereby implements the principle of perceptual contrast. In this theory, speaking represents a trade-off between the antagonistic needs of the speaker and the listener.

Clearly, language is more than the act of speaking. Consequently, the phonetic theory by its very nature cannot provide a general account of whether and how the linguistic territory is meted out between the principles of similarity and contrast. A clue to this issue may be taken from Jakobson (1971) who builds on de Saussure's earlier work. Examining the differing roles of the syntagmatic and paradigmatic axes of language, Jakobson (1971: 75) defines the relationship among syntagmatic units in terms of contiguity and that among paradigmatic units in terms of similarity. Let it be noted at the outset that contiguity and similarity are not opposite poles and that there is an asymmetrical relationship between them. Whereas contiguity is definitely ruled out at the paradigmatic level, there is no a priori reason why elements at the syntagmatic level cannot be both contiguous and similar. There is nothing in Jakobson's text to indicate that contiguity should co-occur with contrast at the syntagmatic level.

What led Jakobson to postulate a similarity relationship among paradigmatic units? Strangely enough, he adduced hardly any motivation for the notion of similarity. All he says in support of this notion is that synonymic and antonymic relationships characterize the members of a paradigm. However, semantic relationships other than synonymy and antonymy may be relevant in the construction of paradigms. This is not to say that Jakobson's assignment of similarity to the paradigmatic level must be spurious, only that his arguments in favor of this view are rather weak.

Stronger support for the role of similarity at the paradigmatic level can be found in de Saussure (1916). He states that paradigmatic relationships are created on the basis of shared form or meaning among lexical items. To take an example from English, he notes that teaching forms paradigms with other words such as training and education by virtue of identical formal and/or semantic features. It is obvious that similarity is a sine qua non in de Saussure's conception as paradigms stand or fall on the availability of similarity-based links among linguistic units.

One wonders what role similarity plays at the syntagmatic level. Neither de Saussure nor Jakobson appear to consider similarity a property of the syntagm (see Happ 1985 for a survey). Actually, there is good evidence for the different roles played by similarity at the paradigmatic and syntagmatic levels. This evidence comes from word association tasks in which subjects are encouraged to say the first word that springs to their mind upon presentation of a stimulus. Their reactions may be classified into syntagmatic and paradigmatic responses (e.g. acknowledge [right arrow] letter and shallow [right arrow] deep, respectively). McNeill (1966) found that the most frequent paradigmatic association was formed between items showing a minimal semantic contrast. Clearly, this result cements the link between similarity and paradigms. However, the picture is different for syntagmatic associations. To begin with, these occur far less frequently than paradigmatic ones (Clark 1970). One probable explanation for this disparity is that similarity facilitates the occurrence of associations, so the uncommonness of syntagmatic responses reflects the absence of a similarity relationship among the members of a syntagm. As a matter of fact, there is positive evidence for the importance of contrast in syntagmatic associations. Stimulus and response most usually differ in terms of their word class in syntagmatic associations, whereas word class is identical in paradigmatic associations (see examples above) (e.g. Deese 1962; Jenkins 1965; McNeill 1966). This suggests a link between contrast and syntagms.

These findings lead us to adopt the following working hypothesis. Similarity originates on the paradigmatic axis whereas contrast is generated at the syntagmatic level. The fact that the two principles jointly shape language may accordingly be put down to the conspiracy of the two linguistic dimensions under investigation. It is the objective of the present study to put this working hypothesis to the test. The analysis will be restricted to the phonological level. However, as will be suggested in the general discussion in Section 3, there is reason to expect similar patterns at other levels of description. Furthermore, the focus will be on English even though occasional reference will be made to other languages. The provisional expectation is that the relationship between similarity and contrast on the one hand and the paradigmatic and syntagmatic dimension on the other is not defined on a language-particular basis, but that English exemplifies a more general solution to this problem. This restriction to English is also imposed for practical reasons. Given the breadth of the analysis that is aimed at, only English is eligible as the language which has been relatively thoroughly investigated in many diverse areas. Note finally that the consistent focus on a single language is the only way of avoiding eclecticism.

2. Data analysis

Let us begin by defining the terms "syntagmatic" and "paradigmatic." In line with de Saussure and Jakobson, a syntagmatic relationship holds among units that appear in different places in the same planning unit. By contrast, a paradigmatic relationship holds among units that compete for the same position in an utterance. Since this work is mainly concerned with segmental phonology, similarity and contrast are amenable to a straightforward definition. Similarity is defined as a maximum, whereas contrast is defined as a minimum of phonological features shared by two phonemes. (3)

There are five areas from which pertinent empirical data will be examined. These encompass language structure, language change, language use, language acquisition, and language breakdown. Language structure includes not only the invariant phonological structure of lexical items but also the structure of phoneme systems. Phenomena such as mutation and metathesis will be treated under the rubric of language change. The use of language by competent, adult native speakers mainly concentrates on the analysis of inadvertent errors. These momentary malfunctions will be complemented by the more permanent errors made by children in the process of first language acquisition, as well as by the less predictable errors committed by various types of aphasics.

There is no doubt that these five areas are quite heterogeneous in nature. This emphasis on diversity is desirable in that a theory which is supported by many disparate data types is clearly more robust than one which relies on a single data type. In other words, heterogeneous data present a stronger test case than homogeneous ones.

Generally speaking, all five areas lend themselves to a treatment from both the syntagmatic and the paradigmatic angle. To take but one example, sound change may have a syntagmatic or a paradigmatic origin. However, some phenomena in each of these areas tend to be confined to one dimension. For instance, whereas mutation is by definition a syntagmatic effect, ablaut is of a paradigmatic nature. We will begin with the simpler story of the paradigmatic level and move on to the more complex analysis of the syntagmatic dimension thereafter.

2.1. The paradigmatic level

2.1.1. Language structure. Viewed from the paradigmatic perspective, phonological structure is tantamount to the system of phonological units, in particular phonemes. The question at the center of this investigation is, then, whether the English phoneme system is organized according to the similarity principle, as predicted by the working hypothesis. Note that this working hypothesis appears to fly in the face of Jakobson's (1944) principle of contrast on which the consonant and vowel systems of the world's languages are assumed to be built.

To investigate the role of similarity, recourse may be had to the notion of symmetry, or in structuralist terminology, pattern congruity (e.g. Hockett 1958). A characteristic trait of a symmetrical phoneme system is the maximum exploitation of a minimum number of classificatory features. The sparse use of phonological features of necessity leads to phoneme pairs differing in only one feature and to a number of phoneme pairs differing in one and the same feature.

Let us begin with consonants. Standard British English (RP) has 24 consonant phonemes which may be exhaustively classified along the three parameters of place of articulation, manner of articulation, and voice. All consonants were specified once on all three dimensions in surface-true fashion (with only positive values). For example, /f/ was coded as a voiceless, labiodental fricative and /d3/ as a voiced palato-alveolar affricate. Aside from very few exceptions, these specifications are uncontroversial. One of the problem areas is /r/ which is similar to /l/ if treated as a liquid but less similar to /l/ if treated as an approximant. The latter solution is adopted for present purposes. However, the overall pattern would not be affected if alternative solutions were chosen. With the aim of deriving an overall similarity measure, a theoretical confusion matrix was set up in which the number of shared features was calculated for all pairs of nonidentical phonemes. Taking up the above example, the number of features shared by /f/ and /d3/ is zero.

The confusion matrix contains 276 cells (24 x 24 - 24 divided by 2). The subtraction of 24 eliminates the cases of identical pairings and the division by 2 takes account of the fact that the order of the elements in the phoneme pairs is irrelevant for the determination of similarity. The 276 cells show a total of 229 shared features. Thus, the average number of features shared in all pairs of the English consonant system amounts to 0.83.

Taken by itself, this result admits of two contradictory explanations. Because it is so close to the chance mean of 1.0 in a three-value system (no features shared, one feature shared, two features shared), it may be taken to suggest the absence of any significant effect, that is, the structure of the English consonant system is sensitive neither to similarity nor to contrast. Alternatively, it may be argued that both similarity and contrast play a role, but as they are of similar strength, they cancel each other out and thereby create a situation which is not unlike that without these two principles. Further analysis is necessary before it is possible to arbitrate between these conflicting interpretations. In either case, it is clear that Jakobson's principle of contrast cannot be accepted as the major or even the only factor in the shaping of phoneme systems.

As noted above, similarity can also be understood in terms of the number of phoneme pairs being distinguished by the same feature. By their very nature, binary features serve this purpose best. Of the three parameters, voice is the binary one in English. In point of fact, voice imposes a very neat organization on the system of English obstruents, the most prototypical class of consonants. With the unique exception of /h/, all stops, fricatives, and affricates can be grouped into voiced-voiceless pairs (e.g. /b/ - /p/). This is a case of pervasive symmetry which suggests that the principle of similarity is at work.

The similarity principle can also be shown to be operative in the English vowel system. According to this principle, diphthongs may be expected to be similar to monophthongs in that the constituent parts of the former should be phonologically identical to the latter. This is certainly true of English. All parts of diphthongs, be it the first or the second, have a corresponding element in the set of monophthongs (even though the phonetic realizations especially of the second parts are characterized by articulatory undershoot) (e.g. /e/ + /l/ [right arrow] /el/). What has been said about diphthongs can be generalized to triphthongs. These are always combinations of existing diphthongs and monophthongs (e.g. /al/+/[diff]/ [right arrow]/a[diff]/)..

A foray into other languages which have fewer vowels than English brings contrast to the fore. Languages with a minimum of three vowels almost always have /i/, /a/, and /u/ (Crothers 1978). These are the three vowels that are the furthest apart in the quadrilateral. Clearly, minimal vowel systems are shaped by the contrast principle. It is equally obvious that, given this point of departure, any enlargement of such a system can only enhance its similarity index.

In view of these distinct manifestations of both similarity and contrast, it is possible to remove the above uncertainty about the proper interpretation of the roles of contrast and similarity in the English consonant system. Since similarity and contrast can be shown to exist individually in the domain of vowels, it stands to reason that they also contribute to determining the structure of the consonant system. The tact that the similarity index is relatively low may now be construed as a compromise between the two opposing influences of similarity and contrast, which are of similar strength and therefore give rise to a similarity index which is close to chance expectation.

To conclude, the working hypothesis whereby the English phoneme system as an instance of paradigmatic choices is primarily shaped by the principle of similarity is not supported. While there is no doubt that similarity is a factor in the structuring of phoneme systems, it is counteracted by the principle of contrast, which is no less important. However, its importance has clearly been overstated by Jakobson, who views it as the major determinant of phoneme systems. Judging by this singleton analysis, one might be led to claim that the paradigmatic dimension is influenced by both similarity and contrast. Whether this combined effect is particular to the structure of phoneme systems or generally true of the paradigmatic axis will become clear after the following set of analyses.

2.1.2. Language change. Two aspects will be considered in connection with language change. Obviously, language structure is a consequence of language change. Hence, the preceding analysis of phoneme systems will be supplemented by a look at how the structure of the English phoneme system changed over time. In addition, the paradigmatic dimension reveals itself in the noncontextual sound changes that occurred in the history of the language. Let us begin with the systemic aspect.

A synchronic system which is symmetrical in nature has had to rely on diachronic forces which encourage the reduction of asymmetry and the enhancement of symmetry. The existence of such forces would constitute supplementary evidence for the role of similarity. In fact, it can be shown that asymmetrical elements are diachronically unstable whereas symmetrical elements are more stable. As mentioned before, /h/ is the only obstruent which lacks a phonological counterpart with an opposite voice specification. The prediction would accordingly be that this consonant should be more vulnerable to diachronic loss than any other. (4) This prediction is clearly borne out. The history of English <h> is an amazingly systematic "failure story." It started out as the full-fledged consonants [[x].sup.5] and [c], which began to weaken by losing their supraglottalic features and finally disappeared completely in many words, as detailed in Lutz (1988).

In other cases, symmetry was enhanced by phonemicizing allophones which had a voice specification unlike that of the corresponding phoneme. Such was the case with most fricatives. In Old English, [theta]-[eth], [s]-[z], and [f]-[v] stood in complementary distribution. By the Middle English period, the voiced allophones [eth], [z], and [v] had acquired phonemic status. It is also relevant in this connection that symmetrical phonemes, which for various reasons might be expected to be subject to obliteration, are remarkably stable. A case in point is /[eth]/, which has a low type frequency, is among the late acquisitions in child language, and is rather uncommon in the languages of the world. Despite all these potentially weakening factors, this phoneme shows no signs of being endangered in the standard language (RP). All these arguments suggest the reality of diachronic forces which serve to create symmetry in the consonant system. On the logic spelled out above, this may be attributed to the operation of the similarity principle.

Elements may not only be added to, and deleted from, the phoneme system, they may also develop into other elements. In the diachronic study of language, paradigmatic substitutions go by the name of spontaneous sound change. The prediction of the working hypothesis is that spontaneous sound change is characterized by a strong similarity between the substituting and the substituted phoneme. This prediction is unequivocally supported by the diachronic data. Almost all sound changes that occurred in the English language involve a change of only one phonological parameter. Typical cases include Middle English/y/[arrow right]/i/from the vocalic and /f/[arrow right]/v/ from the consonantal domain. While there are apparently no exceptions to this rule among consonantal substitutions, a few complex cases arise in the set of monophthongs. The largest subset encompasses vowel substitutions in which a quantitative change is accompanied by a qualitative change. A pertinent example is given in (1).

(1) O.E. wicu [arrow right] M.E. weke 'week'

The proper analysis of (1) depends on the height of/i/. If this vowel was about as high as cardinal vowel 2, the change from /i/ to /e:/ would involve both a lengthening and a fronting process; if, however, /i/ was higher than cardinal vowel 2, the change from /i/ to /e:/ would involve both a lengthening and a lowering process. While it is clear that two parameters are affected in both cases, it is equally plain that the two processes did not occur independently of each other. When the /i/ is lengthened without a concomitant qualitative change, the resultant vowel may be phonetically closer to /e:/ than to /i:/. As the lengthening process created an outcome which did not exactly match any of the established phoneme categories of the language, this outcome was assimilated to a well-established phonemic category in the vicinity, to wit /e:/. This assimilation motivated the fronting or lowering process. In a nutshell, such two-feature changes may be conceived of as a result of gaps in the vowel space and the preference for using old over new categories.

To conclude this section, unsurprisingly perhaps, there is a strong tendency for paradigmatic sound changes to be phonologically minimal, that is, to affect only one parameter. In the few cases where two parameters are involved, a case can be made for their interdependence. This provides firm support for the hypothesis that paradigmatic sound change is under the sway of the similarity principle. Substituting and substituted elements are as similar as can be, given the structure of the phoneme system. There is not a shred of evidence for the contrast principle to be at work in this area.

2.1.3. Language use. The third analysis deals with adult language use, or rather misuse. Every now and then, adult speakers make unintentional errors, so-called slips of the tongue. Like historical sound change, these may be paradigmatic in nature in that they are not facilitated by the linguistic context. A typical example is shown in (2).

(2) What do you mean, I'll loove it--I'll lose it? (from Stemberger 1985)

Example (2) documents the paradigmatic substitution of /z/ by /v/. The two phonemes at issue are minimally different. Only the place-of-articulation parameter is changed. Stemberger (1985) reports that 88.3% of the paradigmatic consonant substitutions in his large speech error collection involve a change of a single phonological feature, as in (2). A similarly high percentage is present in my corpus of German slips of the tongue.

This result motivates the claim that paradigmatic phoneme errors are strongly constrained by the principle of similarity. Contrast does not play any role in this game.

2.1.4. Language acquisition. As is well-known, language learners make systematic errors in the acquisition process. They begin to speak before they have achieved complete mastery of the phoneme system. As a result, they replace phonemes which have not yet been acquired by those that have. The exploration of similarity relationships in child language has to cope with two problems. The first is that the definition of similarity within the adult system cannot be blindly extended to children's language. Similarity can be adequately gauged only within the system which generates the particular output under investigation. Hence, the learner's current transitional system has to be taken into account in an analysis of similarity relationships.

This problem is mainly of a practical nature. Once all the relevant information is available, it disappears. However, even if only little is known about the developing systems--and this is usually so--there is no major reason for concern because young children's systems seem to be amenable to the same three-way classification that has proved useful in the examination of adult language. Stoel-Gammon's (1985) extensive study of consonant acquisition reveals that at least two features have emerged for each of the three phonological parameters in at least half of her subjects by the age of two. To illustrate, the voiced voiceless contrast was acquired. So were the oral-nasal contrast and the stop-fricative distinction, as well as several places of articulation. If the criterion of the percentage of subjects required to have mastered a certain sound is raised, the average age of mastery of the above oppositions is of course higher, though not considerably so (see Prather et al. 1975). Thus, the structure of early phoneme inventories justifies the application of the standard three-way consonant classification procedure, despite certain missing features.

The second difficulty is related to the first. It pertains to the representational status of the substituted element. To repeat, similarity is measured by comparing the feature matrices of the substituting and substituted units. This procedure presupposes that the substituted phonemes and their features are available at some representational level for the comparison to be meaningful. However, precisely this is unclear in child language. It would seem reasonable to argue that the substitutions are repairs. That is to say, they occur precisely because the target phoneme is beyond the child's phonological and/or articulatory capabilities. If this much is true, it appears hardly plausible to assume that the substituted phoneme has associated with it a fully specified feature matrix.

There is a relatively simple way of dealing with this issue. If no similarity relationship emerges between the phoneme pairs, one possible explanation would be that this is an artifact of the procedure of comparing something with nothing; if, however, the substitution patterns evince a certain degree of similarity, it cannot be presumed that the substituted element is not represented at all. Rather, it must be available at least in the form of the features on which the similarity relationship is established. Hence, the second problem vanishes with the emergence of similarity between the interacting units.

With these concerns put to rest, we may embark on the similarity analysis proper. Snow (1963) published a very large corpus of consonantal substitutions made by children from six to eight years of age. Although this age covers a relatively late period of phonological acquisition, these data are in no way atypical of the substitution patterns reported for younger children. A minor problem with this database is that no distinction is drawn between paradigmatic and syntagmatic errors. Snow used a simple picture-naming task by which single words from the children's productive repertoires were elicited. This method eliminates between-word syntagmatic substitutions. Fortunately, Snow provides all the words in which the phonological substitutions are embedded. It is possible therefore to discard all those items for which syntagmatic influences cannot be ruled out. It is a notable fact that only 0.9% of Snow's data lend themselves to a syntagmatic analysis while it is not even clear that all of these 44 cases are truly syntagmatically motivated. Thus, at least at this age level, paradigmatic substitutions are overwhelmingly more frequent than syntagmatic ones. A typical case of a paradigmatic substitution is given in (3). This error was made by a girl at a very early stage of language acquisition (1;4) but it also figures commonly in the Snow corpus. It evidences the substitution of a palato-alveolar by an alveolar fricative.

(3) [su:] for: shoe (from Ingram 1974)

The results of the similarity analysis are reported in Table 1.

It is quite evident from Table 1 that the consonantal substitutions in child language are subject to the similarity constraint. Single-feature changes make up the large majority while three-feature changes are exceedingly rare. On an average, 1.22 features are modified per substitution error. This very low average lends credence to the hypothesis that similarity constrains paradigmatic errors in child language so severely that no room is left for contrast.

2.1.5. Language breakdown. It is often observed that substitutions are the most frequent category of phonological errors in aphasia. Basically the same types of problems emerge in the analysis of pathological speech as in the analysis of child language and therefore do not need to be reiterated. The distinction between paradigmatic and syntagmatic errors is often ignored in aphasia research. Luckily, this is only a minor issue as a rather small minority of aphasic errors is of a syntagmatic origin. So even if these cases could be identified and dropped, the overall pattern would remain largely the same.

Various analyses have been performed of the role of similarity in phonological paraphasias. The classical study is Blumstein (1973). She examined the errors committed by three major aphasic groups, to wit Broca's, Wernicke's, and conduction aphasics. These syndromes were found to be quite homogeneous with respect to their sensitivity to similarity. Depending on the specific syndrome, 65% to 70% of all phonemic paraphasias involved a switch of a single feature. Unfortunately, it is impossible to calculate the average number of features changed because Blumstein did not provide separate percentages for two- and three-feature switches. Other studies, however, show a marked decrease in error rate as the number of features changed increases (e.g. Trost and Canter 1974: Martin and Rigrodsky 1974). Thus there is unambiguous support for the claim that consonantal paraphasias are sensitive to the similarity constraint.

Data on vocalic paraphasias are also available. In a careful study which distinguishes between paradigmatic and syntagmatic errors, Keller (1978) was able to replicate for vowels what had previously been found for consonants. Single-feature substitutions occurred more frequently whereas three- and more-feature errors occurred less frequently than expected by chance. There was no significant difference between observed and expected frequency in the case of two-feature substitutions.

To conclude, both vowels and consonants are not randomly misselected in pathological language but rather are replaced with phonemes which are phonologically similar to the target. It is worth adding that this finding has been confirmed for a good number of other languages, both Indo-European and other.

2.1.6. Conclusion. All sources of evidence converge on the claim that the paradigmatic dimension of language is tightly controlled by the similarity constraint. More specifically, the selection of elements from the phoneme inventory is a similarity-sensitive process. Contrast does not seem to play any role in this game. However, there is one set of recalcitrant data. Phoneme systems are not only built on the principle of similarity but also on that of contrast. There are two ways of dealing with this exception. We may either regard it as genuine and go for a theory that makes explicit provision for this inconsistency. Or we may argue that the relationship between the paradigmatic dimension and phoneme systems is not the one that has hitherto been assumed. If the structure of phoneme systems does not qualify as a piece of paradigmatic evidence, this exception would turn out to be a spurious one and a simpler model could be constructed on the basis of a consistent body of evidence. This issue will be taken up again in the general discussion in Section 3.

2.2. The syntagmatic level

While the paradigmatic dimension is unitary in nature, the syntagmatic level divides into a potentially large number of subparts. This heterogeneity is introduced by the variable distance between syntagmatically related units. This distance may vary from zero (i.e. adjacency) to as far as a single planning unit such as a sentence extends. When this unit has been fully elaborated, any distance between these two extremes is possible. The difficulty is that it is not known whether there is an interaction between the syntagmatic distance and the sensitivity to similarity and contrast. This is the question that is at the heart of the following investigation. Another complication is that the same two distances need not be equivalent. For example, a two-phoneme distance may have differing properties depending on whether the intervening phonemes are tautosyllabic or heterosyllabic, whether the syllables are stressed or unstressed, whether the interacting phonemes occupy the same syllable position, and so on. In view of this tantalizing degree of variability, it is impracticable to test the entire range of possibilities. It was decided, therefore, to focus on the most basic distinction, that between adjacent vs. nonadjacent syntagmatic units. The latter case represents an undifferentiated category which subsumes interactions of varying distances. Where possible, this variability will be taken into consideration. The two basic types of syntagmatic distance will be examined in turn. The same empirical sources will be consulted as in the analysis of the paradigmatic dimension.

2.2.1. Adjacency Language structure. The principal question underlying the ensuing analysis addresses the possible tendencies that govern the nature of adjacent phonemes in lexical items. Trnka (1936) was one of the first to appeal to the principle of contrast in his account of possible phoneme combinations. He claimed that an acceptable adjacent-phoneme pair /XY/ requires a minimum phonological contrast between /X/ and /Y/. Introducing some gradience into the picture, we may contend that the greater the phonological contrast between /X/and /Y/, the greater the preference for the sequence /XY/. More recently, this principle has been reformulated in terms of the "minimum sonority difference" (e.g. Steriade 1982; Clements 1990; Broselow and Finer 1991). The notion of sonority has the advantage that it applies to consonants and vowels alike. It is therefore equally appropriate for a comparison of two consonants (or vowels, for that matter) as well as of a consonant and a vowel.

Trnka's "law," as he calls it, is certainly true as a general tendency. From the typological perspective, languages tend to prefer the simple alternation of consonants and vowels to consonant clusters or hiatus. While languages vary in their tolerance for adjacent consonants (or vowels), these are never set more highly than consonant-vowel sequences. Two-consonant clusters consisting of an obstruent and a sonorant (e.g. /kr-/) occur more often than those that draw on only one of these phonological classes (e.g. /mn-/). Two-consonant clusters are more frequent than three-consonant clusters, which in turn are more frequent than four-consonant clusters (Greenberg 1965). At the same time, the immediate repetition of phonological units is banned from language structure (or at least strongly discouraged). All these diverse crosslinguistic patterns are neatly captured by the principle of contrast.

Trnka's law is also true at a more specific level. The less the sonority difference, the less preferred the phoneme combinations. Of the three phonological dimensions, manner of articulation is the main determinant of sonority. We would therefore expect the principle of contrast to have its strongest effect on adjacent phonemes with identical manners of articulation. This expectation is clearly fulfilled. Tautosyllabic pairs of stops or fricatives are licit, but highly uncommon in English (e.g. apt and sphere) while tautosyllabic pairs of nasals or glides are disallowed. This resistance to identical manners of articulation is so strong that it also extends to heterosyllabic pairs (Berg 2001), although it is of course less strong than in the tautosyllabic domain (contrast autumn vs. autumnal). Sonority may also explain many, though not all, phonotactic patterns involving identical places of articulation. Pertinent examples include the impossibility of word-initial /tl/ versus the acceptability of /tr/ in Modern English as well as the impossibility of initial /fm/ versus the acceptability of /fn/ in Old English. The former example can be accounted for by the sonority principle: the sonority difference is less between /tl/ than between /tr/. However, the Old English case is less easily accommodated by the sonority principle because there is no generally agreed-upon sonority difference between the two nasals. However, the problem disappears once standard features are had recourse to. The constituents of the illicit cluster share the same place of articulation (6) whereas those of the licit cluster do not. No matter whether the sonority- or the feature-based approach is adopted, the logic of explanation is the same. Whenever there is insufficient phonological contrast between contiguous phonemes, the sequences involved are banned from the phonotactic repertoire of the language.

It is clear then that adjacent phonemes are generally responsive to the principle of contrast. To further substantiate this claim, let us examine one more specific domain. One area where interactions between adjacent phonemes have penetrated language structure is nasal assimilation. English prohibits tautosyllabic nasal + stop sequences in which the nasal differs from the stop in terms of its place feature. Thus, we find /nt/, /nd/, /mp/, and /[??]k/ and nothing but. Because of this feature match, these cases are standardly regarded as fossilized assimilations. This view presupposes, however, that the nasal is either unspecified for place of assimilation at one level of representation or specified for one unmarked feature value (i.e. [alveolar]) which is subsequently changed into a marked one, as required by the context. The former possibility goes by the name of underspecification, the latter by the name of full specification.

The critical issue is whether similarity or contrast facilitates the occurrence of fossilized assimilation. Although relevant cases are few, there is no trace of similarity having brought about these place assimilations. On the full specification hypothesis, the clusters /nt/ and /nd/ do not count as assimilations as no feature change is involved. The two remaining clusters are made up of phonemes which are as different as can be. Not a single phonological dimension is identical. Pretty much the same picture emerges on the underspecification hypothesis, except for the fact that the comparison cannot be carried out on the place dimension. Something cannot be compared to nothing. If this dimension is left out of account, there is nonidentity on the remaining two dimensions of manner and voice. The only exception is /nd/ in which the two constituent phonemes share the voice value. On both hypotheses, then, this type of assimilation involves maximally dissimilar phonemes. (7)

Summarizing, adjacent phonemes have been shown to be highly sensitive to the principle of contrast. The greater the phonological contrast, however defined, between two given phonemic units, the more likely they are to form a sequence. No evidence was found for the similarity principle. Language change. In language change, the interaction between adjacent phonemes can be observed in three areas--assimilation, dissimilation, and metathesis. As dissimilation plays a negligible role in the history of English, attention will be focused on the other two processes.

The assimilation of neighboring units, called contact assimilation in historical linguistics, is exemplified in (4).

(4) O.E. bledsian [arrow right] bletsian 'to bless'

Here, the two-feature difference between /d/ and /s/ is reduced to a one-feature difference through voice assimilation.

All assimilations that were found in Luick (1964) and Hickey (1984) were catalogued and examined for the number of features by which the phonemes of the nonassimilated sequences differed. All cases were included irrespective of whether these sequences were tautosyllabic or heterosyllabic. Since our focus is on phonology, all sequence types were only counted once regardless of the number of words that underwent the assimilation process. The results are tabulated below.

While the absolute numbers are not very high, it emerges quite clearly from Table 2 that there is no particular skewing either way. The number of features by which two nonassimilated phonemes differ averages 2.05.

This is precisely what may be expected from a random distribution. This figure seems to suggest that neither maximum similarity nor maximum contrast is a necessary condition for assimilation to come about.

Metathesis can be divided into two major subclasses--a reversal of a consonant and a vowel, and a reversal of two consonants of which one is /s/. The two subclasses are illustrated in (5) and (6), respectively. Note that the former class appears to have affected more words than the latter in the history of the English language.

(5) O.E. purh [arrow right] Early M.E. pruh 'through'

(6) O.E. aesp--aeps 'aspen'

Example (5) shows a reordering of /u/ and /r/, (6) a reordering of /s/ and /p/. The interaction between a consonant and a vowel is incontrovertible evidence for the inoperativeness of the similarity principle. The same can be said about the /s/ class which always involves a voiceless stop. It is notable that while /s/-/p/ and /s/-/k/ interactions are well attested, /s/-/t/ interactions occur very seldom. The fact that the latter phoneme pair is more similar than the former two suggests that similarity does not play any part in metathesis. Language use. The interaction between adjacent phonemes will be approached from two angles--assimilation and language misuse as evidenced in slips of the tongue. Let us begin with the latter. The speech error patterns can be largely deduced from language structure. Given that adjacent phonemes tend towards dissimilarity, it is only to be expected that the units that are swapped in errors are phonologically dissimilar. This is so in (7).

(7) I saw that flim. for: film. (from Trevor Harley, p.c.)

Case (7) exemplifies a reversal of a consonant and a vowel, the most dissimilar units. While such errors are uncommon in speaking, they all involve dissimilar units. However, the significance of these data should not be overestimated as slips of the tongue tend to respect the phonotactic rules of the language and these rules allow the switch of consonants and vowels (as in [7]), though not the switch of adjacent consonants (e.g. /pr-/ [arrow right] */rp-/). The error patterns can therefore be taken only as ancillary evidence for the principle of contrast.

Assimilations can not only be found in language structure and change but also in language use. In fact, they originate in language use and from there make their way into language structure and instigate language change. To be best of my knowledge, there are no large-scale, token-based studies of assimilation in ordinary speech. We therefore have to confine ourselves to a type analysis of common assimilation processes as illustrated in (8).

(8) We got you to come .... [arrow right] We gotcha ...

This example testifies to one of the most frequent assimilatory processes in English--the change from /tj/ to [t[??]]. All phoneme sequences which are liable to undergo assimilation were compiled. Most of the data were taken from Roach (1991). Note that this list is not exhaustive. Unlike the fossilized and historical assimilations, these synchronic assimilations usually take place at the word or morpheme boundary, as in (8). A count of the number of feature differences between the nonassimilated phonemes is presented in Table 3. This table also includes a chance similarity count for comparative purposes. Such a count is necessary to rule out the possibility that the empirical patterns are an artifact of language structure. For example, a potential similarity effect might arise because the phonemes to be assimilated are always similar in a particular position. To eliminate this problem, the number of shared features was calculated for a sample of adjacent phonemes which are separated by a word boundary. This was done on the basis of the first two pages of the London-Lund Corpus (Svartvik and Quirk 1980). All adjacent-consonant pairs within a tone group were taken into consideration. The few cases of identical-consonant pairs were ignored.

The low number of consonant-sequence types involved in assimilations precludes any strong claims. What can, however, be argued with some confidence is that there is no skewing for a minimal or a maximal phonological difference. The average number of features by which the two phonemes differ runs to 1.86. This result is statistically indistinguishable from the chance outcome ([chi square](2) = 2.5, p > 0.2). Thus, the synchronic and diachronic assimilations behave very much alike with regard to their insensitivity to similarity. Language acquisition. The interaction of adjacent phonemes in child language takes a number of different forms, three of which were singled out for analysis. These are fusion, dissimilation, and metathesis. (8) As before, the central question is whether these processes give evidence of similarity or contrast effects.

Fusions describe a process whereby two abutting phonemes are merged into one through the recombination of their phonological features. This case of cluster reduction usually, but not always, results in a phoneme that exists in the ambient language. Stemberger (1992) states that the two most frequent cases involve the coalescence of obstruent + sonorant clusters and that of /sp/, as shown in (9).

(9) [fu:n]. for: spoon. (from Chin and Dinnsen 1992)

In this case, the manner feature of /s/ was combined with the place feature of /p/ to yield /f/. This example betrays no phonological similarity between the interactants apart from the inconsequential fact that both are voiceless. (9) The elements in all of the clusters that are liable to fusion are equally dissimilar. All of the examples reported in Bernhardt and Stemberger (1998) and Chin and Dinnsen (1992) involve clusters whose constituents differ on two or three dimensions. It is obvious then that this process is not triggered by phonological similarity. If it was, one would expect /sn/[arrow right] [[??]] fusions to be more frequent than /sin/[arrow right] [??] fusions. However, as far as can be determined from the limited body of available data, this is not so.

The process to be examined next is vocalization, which is introduced here as an instance of dissimilation. The basic observation is that syllabic consonants may turn into vowels at the end of words (Ingram 1986). A typical example is provided in (10).

(10) [babu]. for: [b[??]t[??]] bottle. (from Ingrain 1986)

The syllabic lateral [[??]] which has a velar quality, is changed to the back vowel [u]. (10) As this process occurs only after consonants, it can be viewed as a change from CC to CV, that is, as a means of increasing the contrast between the adjacent segments. In fact, any attempt to impose a CV structure on the output can be understood as being driven by the principle of maximal contrast.

Metathesis occurs not only in adult language use and change but also in first (and second) language acquisition. Children's metatheses share largely the same characteristics as those found in language history and use. Refer to (11), which happens to involve exactly the same lexical item as (7).

(11) flim. for: film. (from Smith 1973)

Like most other instances of metathesis, (11) implicates a consonant and a vowel, that is, maximally dissimilar elements. To a large extent, the swapping of adjacent consonants is prohibited by phonotactic constraints. This would seem to make the transposition of a consonant and a vowel an artifact of language structure. However, for one thing, phonotactic constraints are only gradually acquired and violable in child language. For another, even where sequence constraints are more liberal, as across syllable boundaries, the change of /[C.sub.1]-[C.sub.2]/ to [[C.sub.2]-[C.sub.1]] occurs very seldom. Thus, phonological similarity seems to play no role in bringing about metathesis.

The three processes under review (and others not mentioned here) have one property in common. They appear to be blind to the similarity principle. Some function to increase the contrast between adjacent elements. This is true not only of vocalization but also of coalescence because the latter is a type of cluster reduction, which creates maximally different phonemes. Even when a phonological process such as metathesis does not enhance contrast, it tends to involve phonemes which are no more similar, if not less similar, than would be expected by chance. Language breakdown. As far as I have been able to determine, detailed information on the interaction of adjacent phonemes in aphasia is at a premium. In particular, I am not aware of any studies dealing with assimilatory processes in aphasic speech. The only pertinent evidence I could find is metathesis, which takes the same form as in language change, use, and acquisition. The similarity between (11) and (12) is obvious enough.

(12) cloats, for: colts. (from Blumstein 1973)

Like the other cases mentioned above, this aphasiological error shows no trace of being brought about by phonological similarity. Despite the scarcity of available materials, it may therefore be suspected that the similarity principle also plays no role in these pathological language patterns. Conclusion. There is a fair level of agreement in the data. Couched in negative terms, none of the various sources of evidence show any sensitivity to the principle of similarity. This does not, however, imply that the principle of contrast is vindicated on all fronts. To be sure, it is clearly operative in language structure where more dissimilar phoneme sequences are preferred to less dissimilar ones. On the other hand, (the principle of maximum) contrast was not found to occasion processes such as assimilation. Although, trivially, this process rests on contrast, it does not require maximal contrast, that is, phonemic units which are more dissimilar than expected by chance. This process seems better described by the absence of similarity as an instigating factor. The common core of the phenomena in question thus is the absence of similarity rather than the presence of maximum contrast.

2.2.2. Nonadjacency Language structure. As explained in Section 2.2, the notion of nonadjacency is a quite heterogeneous one in that it covers any distance from one to numerous elements intervening between the potential interactants. It is certainly not clear at the outset that the number of intervening segments has no effect on the sensitivity to similarity. From among the great many phonological relationships that may be investigated within a preplanned unit, the occurrence of repeated units was singled out. It was noted in Section that the immediate repetition of phonemes is almost completely disallowed by language structure. By examining the probability of repetition in the nonadjacent domain, it is possible to gain an idea of whether linear distance interacts with sensitivity to similarity.

The first step involved constructing an adequate sample of data. To this end, the Dictionary of Contemporary English (DCE) was searched for entries containing identical phonemes. The first word on each page to meet this criterion was selected while all others on the same page were ignored. Compounds and inflected forms were disregarded because their lexical status is uncertain. The effect under investigation cannot be expected to unfold itself in the phonemic structure of morphemes which stand little chance of interacting due to their being separately represented in long-term memory. It is appropriate therefore to concentrate on those items which are likely to be lexicalized.

The pronunciation norm followed in the identification of identical elements was Standard British English (RP). Diphthongs were regarded as holistic units. To illustrate, the word bailiff was not taken to have repeated /1/'s even though it contains /e1/ and /1/. Because triphthongs spread across two different syllables, they were treated as diphthongs + /diff/. Consequently, the /diff/ was considered repeated in a word like diamond [da1[diff]m[diff]nd]. As a general rule, the pronunciation of slow, careful speech was targeted. In particular, schwas were counted also in those cases where they could have been dropped, as in garden. A distinction was also made between long and short vowels such that /i:/ and /1/ were not classified as identical. On the other hand, minor qualitative differences such as between [i] and [1] as in foliage [f[diff][u]li1dz] were left out of account. This was for the reason that [i] and [1] are not distinct phonemes in English (assuming that [i] is an allophone of /1/ rather than /i:/) and that the level of representation to be probed into is the phonological one.

For reasons to be explained below, the DCE was perused twice, once for identical consonants and once again for identical vowels. This strategy unearthed 1031 words with identical consonants and 988 with identical vowels. It is worth stressing that this sample is almost certainly representative, given the size of the DCE and that a great many words, especially monosyllabic ones, do not contain identical phonemes at all.

Some words do not only have two identical phonemes but even three, as in consistency [k[diff]ns1st[diff]ns1]. These cases pose an analytical problem. Should the double repetition of phonemes be treated as a separate phenomenon or broken down into two singleton repetitions? I opted for the latter alternative and counted words with three identical phonemes twice. This procedure yielded a total of 1110 consonantal and 1055 vocalic repetitions. Words may also contain two different pairs of identical consonants, as in the above example in which repeated /s/'s occur alongside repeated /n/'s. Of course, both types of repetition were taken into consideration.

The next step was to calculate the length of the words as well as the distance between the identical units. The easiest way would have been simply to count the number of phonemes, but this would have blurred the distinction between long and short vowels as well as that between (short) monophthongs and diphthongs. It was decided therefore to have recourse to the skeleton tier and count short vowels as one (timing) unit and long vowels and diphthongs as two (timing) units. At the consonantal level, affricates were treated like simple consonants as both their constituents are associated with a single consonantal slot. Obviously, this procedure was applied to the calculation of word length and linear distance alike. The decision to refer to the CV tier has the great advantage that it is phonetically more faithful without compromising the phonological nature of the analysis. As is well-known, phonologically long vowels generally are of longer duration than phonologically short ones.

The final step required the calculation of the null hypothesis. This was done by building two lexicons of artificial words. Each comprised 9000 items which divided into nine even subsets of varying length (from two to ten phonemes). The one lexicon was designed to derive the null hypothesis for consonant repetition, the other served to determine the chance level for vowel repetition. Separate lexicons were constructed as a response to the fact that the chance probability of phoneme repetition is not identical for consonants and vowels. This is because of the differing number of consonantal and vocalic phonemes in English. Evidently, the higher the number of elements per subset, the lower the probability of repetition.

The artificial lexicons were created by running what comes close in spirit to a Monte Carlo simulation. Words were formed by having a computer program randomly select a succession of units (from two to ten) from a pool of 24 consonants (or 20 vowels, for that matter). These selections were independent of one another such that the same element could in principle be selected immediately again. One thousand runs were performed per subset to obtain a large enough sample of pseudo-words.

An examination of the real-word data revealed that word length had little impact on the frequency of repetitions as a function of linear distance. For example, the highest number of repeated consonants occurred at distance 2 irrespective of whether the lexical items consisted of anything from four to ten phonemes. This relative homogeneity in the data provided the justification for collapsing the different word lengths. This was done by averaging across all word lengths from three to ten phonemes. The results are diagrammatically presented in Figure 1 for consonants and Figure 2 for vowels.


We will begin the discussion of Figures 1 and 2 by focusing on their commonalities. The most basic observation to make about both consonants and vowels is that phoneme repetitions are significantly less likely than chance at distance 0, but significantly more likely than chance at slightly greater distances while the attested and the chance outcome approximate to each other at considerably greater distances. This finding largely replicates for English what MacKay (1970) observed for Hawaiian and Serbo-Croatian on the basis of a very different method of chance estimation. We are thus led to distinguish three temporal phases in the analysis of phoneme repetitions--an initial phase of underrepresentation, a subsequent phase of overrepresentation, and a final phase of chance representation. A first conclusion is then that the probability of phoneme repetition interacts heavily with the distance between the repeated units.

Looking at the two phonological-segment types separately, we notice that consonantal repetitions occur significantly less frequently than expected by chance at gap lengths 0 and 1 (p < 0.001 for both gap lengths). From distances 2 to 5, repetitions occur above chance frequency even though the [chi square] values decrease with increasing distance. From distance 6 on, actual and chance outcome are statistically indistinguishable.

Vocalic repetitions are significantly underrepresented at distance 0 (p < 0.001), significantly overrepresented at distance 1 (p < 0.001) and again significantly underrepresented at distance 2 (p < 0.001). At distances 3 and 4, the rate of vowel repetition is again above chance (p < 0.001). Finally, from gap length 5 on, actual and chance outcome are almost identical.

There are two major differences between consonantal and vocalic repetitions in English. For one thing, the vowel curve peaks twice whereas the consonant curve peaks only once. For another, the vowel curve reaches its (first) peak faster than the consonant curve. By the time the consonant curve peaks, the vowel curve has already reached its first trough (at distance 2). In addition, the vowel rate reaches a higher peak than the consonant rate. These observations may be summarized by the claim that the vowel graph is more extreme and changes faster than the consonant graph. We will not go into the possible reasons for the disparate behavior of consonants and vowels here. Suffice it to point out that the stress-timed nature of English is propitious to the occurrence of unstressed vowels in adjacent syllables. As there are so few of them that occur frequently, the repetition of unstressed vowels, in particular schwa, is difficult to avoid. The main point in the present connection is that there is agreement between consonants and vowels in that we find a below-chance occurrence of repeated phonemes at the shortest gap lengths but an above-chance occurrence of repeated phonemes immediately after.

The issue of unit repetition may not only be tackled at the phoneme but also at the feature level. A sufficiently restrictive domain for the syntagmatic analysis of phonological features is the syllable. As syllables by definition do not accommodate more than one vocalic peak, the investigation will be confined to consonant features. The guiding question is therefore whether identical features from tautosyllabic onset and coda consonants exhibit a higher or lower than chance occurrence in English. The several studies that have addressed this issue are unanimous that identical features are generally discouraged within the syllable.

Locke (1983) discovered an interaction between the phonological dimension and the probability of feature repetition. For both place and manner of articulation, feature repetition occurred less often than would be expected by chance. By contrast, onsets and codas more frequently agreed than disagreed in terms of voicing. Unfortunately, Locke did not report having run tests of statistical significance on these data, so we have to take them at their face value. It is clear, however, that voice repetition is less discouraged than place and manner repetition.

Focusing on CVC syllables, Kessler and Treiman (1997) confirmed that nonidentical features are preferred to identical ones. They compared the actual to the chance frequency of onset-coda pairings and found that some pairings occurred significantly more often and others significantly less often than would be expected on the basis of the combined frequency of the individual consonants. Those that occurred less often included identical (e.g. /l/-/l/) or similar (e.g. /b/-/p/) onset and coda consonants whereas those that occurred more often included more dissimilar (e.g. /f/-/n/) onset and coda consonants.

By examining all monosyllabic words, Berkley (1994) was able to introduce the parameter of distance, which is of particular relevance in the present context. Concentrating on place of articulation, she uncovered a below-chance occurrence of homorganic consonants. Interestingly, this underrepresentation of identical place features diminished with increasing distance. To be specific, the antihomorganicity constraint was strongest for distances 1 and 2, weaker for distance 3, and absent for distance 4. Although Berkley does not break down her data according to the nature of the intervening elements, the overall [chi square] values she reports suggest that the avoidance of homorganic consonants is stronger in syllables with short monophthongs than in syllables with long monophthongs and diphthongs. This provides further justification for referring to the skeleton rather than the melody tier in the calculation of linear distance.

In brief summary, a structural pattern of identity avoidance has been established which operates both at the phoneme and at the feature level. The repetition of phonemes and features is dispreferred at very short lags (roughly within the same syllable). (11) At slightly longer lags, however, this avoidance of repetition turns into a predilection for repetition at the phoneme level. Eventually, this effect dissipates at even longer lags at which observed and chance frequencies converge. Language change. In the history of English, one phenomenon stands out which is clearly syntagmatically motivated and which typically does not involve adjacent phonemes. This is the process of mutation by which the vowel of the prior syllable is made more like the vowel of the subsequent syllable in the same word. It is thus an instance of assimilation at a distance. There are two types of change termed [i/j] and backmutation. The former involves the raising (and fronting) of vowels through the action of the palatal vowel/i/or glide/j/, the latter turns monophthongs into diphthongs, the second constituent of which has a velar quality. This latter process is induced by the back vowels /a/, /o/, and /u/. (12) One example of each type of change follows.

(13) O.E. * musi [right arrow] mys 'mice'

(14) Proto-Germanic * hessan [right arrow] O.E. heofon 'heaven'

In (13), the /i/ fronts the /u:/ to /y:/, and in (14) the vowel /e/ diphthongized into /eo/ under the influence of /a/. This change made the stressed vowel more velar.

The critical question is whether mutation is sensitive to phonological similarity. More precisely, is it more likely when the potential interactants are similar than when they are dissimilar? We think that this is not the case. [i/j] mutation affected most vowels, irrespective of the similarity between target and trigger. The nonaffected vowels were /i/, /y/, /e:/, and /ae:/. This comes as no surprise in the case of /i/ and/y/ as these two vowels can hardly be made more palatal. The case of /e:/ may be explained away on the assumption that it was raised to /i:/ in pre-Old English time. The resistance of /ae:/ to mutation is not compatible with the similarity hypothesis as /ae:/ is more palatal than the velar vowels that underwent this process. It seems that similarity does not play a major role here. Luick (1964: 185) even argues for contrast as a factor in this process. The greater the dissimilarity between target and trigger, he claims, the earlier and the more consistently the change took place.

The analysis of backmutation also provides no support for the similarity hypothesis. The three vowel qualities that were immune to change were /o/, /u/, and /y/ in both their long and short counterparts. The cases of /o/ and /u/ are without any significance as they are identical with the triggering vowels. The only remarkable case is /y/, which could have been backed to /u/. This would be the expected outcome under the similarity hypothesis because /y/ shares its rounding feature with the back vowels. However, it is the unrounded front vowels--which are so clearly distinguished from the rounded back vowels--that were umlauted. It transpires therefore that mutation is insensitive to phonological similarity.

We will now give more emphasis to the syntagmatic axis and calculate the linear distance between the to-be-mutated vowel and its trigger. Unfortunately, work in historical linguistics has not been conducted in such a way as to allow us to examine the interaction between distance and the incidence of mutation. The standard approach confines itself to documenting and analyzing mutated forms. A comparison of mutated and unmutated forms has never been carried out. While it seems to be implicitly assumed that all the words which meet the phonological description undergo mutation, no empirical study has shown this to be true. On the contrary, it has been known for quite some time that sound change is not independent of the lexicon (e.g. Wang and Cheng 1977). It thus remains a real possibility that mutation is (also) lexically determined. If this hypothesis turned out to be correct, it would be conceivable that varying linear distances between the mutated vowel and the trigger may exhibit a varying sensitivity to similarity or contrast.

In view of this lack of detailed information, we have to content ourselves with examining the linear distance in the words that actually underwent mutation. On this basis, it may at least be argued that the insensitivity to the similarity principle applies to the distance that typically separates the to-be-mutated vowel and its trigger. To this end, all mutated items that appear in Luick (1964) and Lehnert (1973) were culled and categorized according to the number of consonants intervening between the critical vowels. A total of 147 words were taken into account. The results of this analysis are summarized in Table 4.

It is readily apparent from Table 4 that the to-be-changed vowel and its trigger are usually separated by a singleton consonant. The average distance between the two is 1.29 consonants. (13) Extrapolating from these data, we may conclude that the absence of a similarity relationship between the mutated vowel and its trigger is a short-distance effect. This absence holds for the one-phoneme gap and perhaps also for the two-phoneme gap. As larger distances are not covered in mutated forms, the historical data remain silent on the role of similarity and contrast in wider spans. This limitation will be overcome by the data from language use to which we now turn. Language use. It was shown above that phonological similarity is not a causal agent in adjacent-phoneme reversals in adults' slips of the tongue. The next step is to enquire into the role of similarity in interactions of nonadjacent phonemes. In actual fact, there is a huge difference between adjacent and nonadjacent slips. Whereas the former often involve the reversal of a consonant and a vowel, the latter always involve either two consonants or two vowels. Differently put, nonadjacent consonant-vowel interactions do not occur. This is a first hint that similarity kicks in once we leave the adjacent for the nonadjacent domain.

This hypothesis receives some support from a detailed look at consonant errors. It has become customary to divide phoneme slips into within-word and between-word cases. As a rule, the distance that the interactants span is larger in the latter than in the former set, witness (15) and (16).

(15) He'd be as pick as a sarrot. for: as sick as a parrot. (from Trevor Harley, p.c.)

(16) A research gropramme. for: programme (from Trevor Harley, p.c.)

As can be seen, five phonemes are skipped over in (15) but only three in (16) (with the diphthong being counted as two). The great majority of within-word slips are between-syllable slips, as in (16), so the number of within-syllable slips is too low to determine with certainty whether they differ from between-syllable within-word slips in their sensitivity to the similarity constraint. The two subsets will therefore be collapsed. The analysis of all within-word slips leaves little doubt that they are constrained by the similarity principle. Stemberger (1985) notes that 51.5% of these slips in his corpus involve phonemes that are maximally similar. Thus, nonadjacent interactions may be claimed to be sensitive to phonological similarity.

We may go a step further and compare the similarity index of within-word and between-word sequencing errors. Stemberger (1985) reports that 70.4% of the between-word errors differ by a single feature from their targets. This makes between-word errors significantly different from within-word errors (p < 0.001). The conclusion is that the sensitivity of tongue slips to phonological similarity increases with the distance between the interactants (with distance being defined as the crossing or otherwise of word boundaries). The sensitivity to similarity may therefore be regarded as a general characteristic of nonadjacent between-syllable speech errors.

The second part of this subsection is devoted to vowel harmony. The consideration of vowel harmony requires justification at two levels. Why is it included in this subsection and why at all in this article? One argument against its inclusion might he its nonoccurrence in English. However, the primary aim of this study is the analysis of similarity and contrast in language rather than the analysis of English (even though a focus on one language was deemed preferable to a mishmash of languages). It is quite consonant with this approach to select an area which does not exist in the focal language and thereby widen the empirical range. Furthermore, the comparison of (synchronic) vowel harmony and (diachronic) vowel mutation may shed more light on the factors that are responsible for the prevalence of similarity or contrast.

Vowel harmony is more aptly treated as an instance of language use than of language structure. This decision does justice to the claim that it is a dynamic aspect which is computed online during language production. While it cannot be entirely ruled out that harmonized forms, especially high-frequency items, are lexically represented, its historical development and synchronic productivity suggest that it plays an active role in language use. This view is compatible with the autosegmental, process-oriented approach to vowel harmony which dominates the phonological literature.

The ensuing analysis is focused on vowel harmony in Hungarian, in particular backness harmony. The general principle is that a vowel adopts the backness specification of another vowel in the same word. The process is unidirectional in that the vowel of the suffix assimilates to the vowel of the stem, not vice versa. Two simple examples follow, one illustrating front harmony and the other back harmony. The underspecified affix is the dative marker -nVk, which is realized as -nek following a front-vowel stem in (17), and as -nak following a back-vowel stem in (18). The examples are given in standard orthography.

(17) orom + DAT [right arrow] oromnek 'joy'

(18) haz + DAT [right arrow] haznak 'house'

In the following, the various constellations of vowels in stems and suffixes as categorized in Hulst (1985) will be investigated from the perspective of similarity relationships among them. The presentation is largely based on Hare's (1990) insightful analysis. Generally speaking, if similarity is a factor in determining the shape of the suffix, we would expect a stem vowel which is similar to the underspecified suffix vowel to have a greater assimilating effect than a stem vowel which is dissimilar to the suffix vowel. However, not only the similarity between stem und suffix vowel but also the similarity between two stem vowels in polysyllabic stems may play a part. The higher their similarity, the stronger their potential influence on the suffix vowel. These predictions will now be tested.

The least interesting case from the perspective of similarity is polysyllabic stems which are internally harmonic with respect to the front-back dimension, as in (17). There is no conflict here and hence no room for similarity to make itself felt, apart from the obvious observation that the realization of the suffix vowel increases the similarity of the vowels in the word.

More interesting is the case of disyllabic, disharmonic stems in which a front vowel cooccurs with a back vowel. The backness value of the suffix vowel is determined by the nature of the second vowel of the stem, as shown in (19) and (20).

(19) zsonglor + DAT [right arrow] zsonglornek 'juggler'

(20) buro + DAT [right arrow] buronak 'bureau'

There are two possible reasons for this pattern. Proximity to the suffix vowel is a determining factor and/or the second vowel of the stem is generally more similar to the suffix vowel than the first and therefore induces harmony. Hare (1990) makes a case for the latter option. In (20), the first vowel of the stem is maximally dissimilar from the suffix vowel (on the phonological dimensions that are specified). By contrast, the second vowel of the stem is more similar to the suffix vowel than the first and is therefore in a better position to impart its backness feature to the underspecified vowel.

For (19), the same argument would be expected to hold. In order to be able to assign its frontness to the suffix vowel, the second stem vowel should be more similar to the suffix vowel than the first. However, this is not the case because the two stem vowels agree on all features except backness. Thus there is no difference in numbers of specified differences in similarity between the first stem vowel and the suffix vowel on the one hand and the second stem vowel and the suffix vowel on the other. For these cases, Hare invokes the similarity relationship between the two stem vowels. She claims that when these are highly similar, the second wins out. This happens not only as a result of the proximity principle but also because of a perseveratory effect. When the prior element is similar to the subsequent element, it increases the strength of the latter and thereby its availability as a trigger. Consequently, the suffix vowel takes on the front quality of /[theta]:/ and surfaces as [[epsilon]] in zsonglornek.

The next general class to be dealt with is mixed stems. These differ from disharmonic stems in the presence of so-called transparent vowels (i.e. long and short /i/ and /e/), which are ignored in the harmony process. Mixed stems come in four subclasses, two of which will be discussed here. The first is called "vacillating" because these stems accept a front or a back vowel, as shown in (21). The front vowel of the stem is always /e/ or /e:/ and the back vowel frequently /a/ or /a:/. The back vowel precedes the front one.


When transparency is taken for granted, it is easy to explain (21b). However, the so-called transparent vowel is not only close but also similar to the suffix vowel and therefore likely to make an impact, as in (21a). Thus there are good arguments for both options and the competition ends in a draw. The net result of this is the observed vacillation.

The second subclass is called "mixed neutral." The words of this class usually combine a high front vowel with a back vowel in variable order. Unlike the preceding case, the suffix vowel is always back. This constellation is exemplified below.

(22) taxi + DAT [right arrow] taxinak 'taxi'

What has to be explained is the difference in behavior of vacillating and neutral stems. Specifically, why do non-high front vowels generally induce vacillation whereas high front vowels do not? Again, Hare argues that similarity holds the key. Because high front vowels are more dissimilar from the suffix vowel than are non-high front vowels, the former have no effect on the harmony process. Therefore, cases like (22) for example always adopt the back specification. This view implies that the transparency of certain vowels may actually be fictitious. The true reason that some vowels are blind to harmony may be their dissimilarity to the suffix vowel. This hypothesis would account for the fact that trisyllabic stems with one back vowel and two /i/'s most usually prefer a front vowel suffix. If high front vowels were transparent, it should not matter whether they occur once or twice. However, it does. On a similarity-based approach, it even makes sense that it does. The repetition of /i/ strengthens the front feature and therefore outstrips its back contender for the specification of the suffix vowel.

To conclude, while it is doubtful that Hare's model can handle all of the Hungarian data, it seems to be the case that a weak version of the similarity approach can be upheld. Similarity appears to be one relevant factor in determining the nature of suffix vowels. It interacts with others such as proximity and, perhaps, underspecification and may be masked by idiosyncratic lexical choices. The main conclusion is then that vowel harmony is sensitive to similarity. This sensitivity appears to be restricted neither to Hungarian nor to backness harmony. It has also been shown to hold for other harmony types such as rounding harmony in Khalkha Mongolian (Goldsmith 1985) and Altaic languages more generally (Walker 2001), as well as in Tiv, a Niger-Congo language (Pulley-blank 1988). Language acquisition. Harmony occurs not only in adult but also in child language even though there is some complementariness involved. While vowel harmony is rather frequently and consonant harmony rather infrequently found in the world's languages, the opposite is true of children's language. Thus, the same questions that were raised for vowel harmony in Hungarian will now be raised for consonant harmony in the acquisition of English. This will be done on the basis of a single case study conducted by Smith (1973). This work is ideal for our present purposes because Smith's subject was a frequent harmonizer at the onset of his language acquisition process and because Smith published the entire diary notes on which his investigation is based. This goldmine allows one to address a further issue that played a major part in the analysis of language structure (Section, namely the linear distance between the two critical units. We will therefore not only examine whether consonant harmony is similarity-sensitive, but also whether the assumed similarity interacts with linear distance.

Consonant harmony is a process whereby one consonant is assimilated to another consonant, which typically occurs in the same word. The assimilation is either partial, as in (23), or complete, as in (24).

(23) wip. for: zip. (from Smith 1973)

(24) [ri:br[diff]]. for: zebra. (from Smith 1973)

In both cases, the child struggles with [z], which is replaced by [w] under the influence of [p] in (23) ("bilabial harmony") but by an [r], which also appears in the second syllable in (24) ("rhotic or post-alveolar harmony"). The target and trigger consonants differ by three features in (23) and by two in (24). Note also that the distance between the interactants is 1 in (23) and 3 in (24).

The complete appendix in Smith (1973) was searched for harmonic forms. Each item was counted only once even if it appeared at more than one stage in the child's development. Inflectional variants were also only counted once. Compounds were discarded provided the harmony process occurred inside one of their constituents and this constituent was separately listed in the appendix. When one word gave rise to two different harmony types, both were included in the analysis. Altogether 230 items met these selection criteria.

The distance between the interactants was gauged on the basis of the child's rendition rather than the adult model. This decision followed from the intention to measure the number of units that were demonstrably skipped. Phonetically unrealized segments might also be skipped, but there is no way of knowing for sure. In any event, the number of words with segments that appear in the adult though not in the child's form was so low that their inclusion would not have dramatically altered the picture. By contrast, the analysis of the featural differences was based on the target word because the factors that give rise to harmony are under scrutiny. This calls for an examination of the phonological distance of the critical consonants in the nonharmonized (adult) forms.

Table 5 presents the results of the similarity analysis as a function of linear distance.

Several observations are readily apparent from Table 5. The interacting consonants stand fairly closely together. Almost all harmonic forms have one or two units between their interactants. The average distance between these is 1.36 segments. Since we are dealing with consonant harmony, it is justified to say that the great majority of interacting consonants skip a single vowel (be it long or short, a monophthong or a diphthong). The reason for this closeness has to do with the early age at which children introduce harmonic processes into their language. Children at this stage use words with only a relatively simple phonological structure. Consequently, the distances to be covered are relatively short. When more complex words enter the child's lexicon, the harmony process has already disappeared.

On an average, the nonharmonized consonants differ by 2.13 features. This result shows quite clearly that consonant harmony is not facilitated by phonological similarity (contra Bernhardt and Stemberger 1998). If anything, there is a slight tendency towards the involvement of dissimilar units. This would make sense in the light of the structure of the English syllable as well as the fact that some harmonies are of the within-syllable type (see [23]). As discussed in Section, the relationship between onsets and (tautosyllabic) codas tends to be one of dissimilarity. It might therefore seem obvious that this dissimilarity is reflected in the target words that serve as input to the child. However, it must be emphasized that the lack of similarity between the critical consonants is not artifactual, that is, a necessary consequence of linguistic structure. It would be perfectly possible for similar consonants to be propitious to harmony and for dissimilar ones to be resistant to it. This, however, is not the case, given that the above analysis is based on a relatively comprehensive corpus.

Unfortunately, the heavy concentration of the data in the distances 1 and 2 provides little opportunity for investigating the interaction between similarity and distance. However, some small pieces of evidence are worthy of note. The three cases of zero distance all involve maximally dissimilar consonants whereas the two cases of maximum distance involve maximally similar ones. While the numbers are low, it does not seem unreasonable to suggest that the sensitivity to similarity increases with distance. Contrast prevails in adjacent interactions, wanes at short distances and turns into similarity at longer distances. Language breakdown. Some aphasics are liable to produce phonological paraphasias which bear a superficial resemblance to harmony patterns in child language. A pertinent case is (25) in which the velar stop is assimilated to the alveolar one.

(25) tat. for: cat. (from Hatfield and Walton 1975)

The chief difference between (25) and the child language patterns is that the latter are largely predictable at a certain stage in the learner's development whereas the former is not. Children produce harmonic forms because their immature processing system does not allow them to generate a disharmonic output. Aphasics, by contrast, suffer from disrupted, not from underdeveloped processing. Thus, their output is more variable and may also include disharmonic forms, as in (26), which do not normally occur alongside harmonic forms in child language. This variability likens paraphasias to the slips of the tongue of normal adult speakers.

(26) tick. for: kick (from Hatfield and Walton 1975)

Hatfield and Walton (1975) tested an aphasic patient on a repetition task to elicit errors. Unfortunately, all their CVC stimulus syllables contained no other consonants but voiceless stops. Thus, the similarity of the interactants is part of the design of the study and therefore without any significance. Equally unrewarding is a paper by Kohn et al. (1995), who probe into the susceptibility of individual feature dimensions to harmonize in a word. Alas, their data are not presented in a way which would permit a similarity analysis of the nonharmonized consonants. Kohn and Smith (1990: 150) note that the "harmonies" made by a conduction aphasic involved dissimilar consonants. As no quantitative analysis was provided and as the data were not broken down by linear distance, their finding is difficult to evaluate in the present context.

The other type of nonadjacent syntagmatic error is reversals which occur only seldom in aphasic speech. One such example is given below.

(27) ephelant, for: elephant. (from Blumstein 1973)

Blumstein (1973) reports that the interacting consonants tend to be close to each other in aphasic reversals. In almost two-thirds of cases, they are separated by a single phoneme. It is unfortunate that Blumstein did not carry out a similarity analysis of the interacting consonants. Judging from the few examples that she gives and others published elsewhere, there is no hint that short-distance reversals such as (27) predominantly involve similar consonants. No claims can be made about long-distance reversals as these hardly ever occur in aphasia (Kohn and Smith 1990). Conclusion. This section has yielded rather heterogeneous results. Some areas (e.g. vowel harmony) were found to be sensitive to the similarity principle whereas others (e.g. mutation) were not. Still others produced a variable sensitivity. Particularly illuminating is the analysis of the repetition effect which revealed a resistance to phoneme repetition at very short lags but a predilection for phoneme repetition at somewhat longer lags. This result brings the length of the lag as the critical variable into the picture. Similarly, normals' slips of the tongue were found to be all the more sensitive to similarity, the greater the distance between the interactants. The low degree of similarity in aphasic speech errors, although seemingly contradictory, meshes nicely with the data from normal subjects. In all probability, it is a simple spin-off of the fact that the distance between the interactants in phonological paraphasias is so small.

3. General discussion

We started out from the working assumption that the paradigmatic dimension is the locus of similarity and the syntagmatic dimension the locus of contrast. As the empirical section has shown, this hypothesis cannot be upheld. There is no simple one-to-one correspondence between the paradigmatic and the syntagmatic dimension on the one hand and similarity and contrast on the other. It is equally untrue to say that both dimensions host similarity and contrast alike. Rather, the empirical results may be summarized by the claim that the syntagmatic level generates both similarity and contrast, while the paradigmatic level gives rise only to similarity effects. The hypothesis that the syntagmatic level accommodates two antagonistic principles calls for a parameter which regulates when the one principle prevails over the other. The data analysis suggests that linear distance is the crucial variable. Smaller distances, including the extreme case of adjacency, bring about contrast whereas greater distances bring about similarity. Thus there is a turning point, or turning span, at which contrast develops into similarity.

Prior to presenting an account of the empirical results, it is necessary to ask whether a unified account is in principle possible. It might after all be that the data types are too heterogeneous to be fitted gracefully into a single model. Indeed, the data types drawn on in this article are quite diverse. Slips of the tongue and paraphasias are processing data, (some types of) assimilations and mutation are historical data, and phonotactic constraints and phoneme repetitions are structural data. Undoubtedly, language structure, change, and processing are all different from one another. To give but one illustration, language change is something eminently social in that an incipient innovation has to be accepted by a linguistic community. By contrast, nothing social inheres in slips of the tongue. In view of this diversity, it may be unwise to expect a single theory to cover everything. However, if a single theory can be shown to successfully account for so many diverse data types, it is all the more robust for it.

What should be the foundation of such a general theory? The answer to this question depends on the assumed relationship between language structure, change, and processing. Several scholars have argued that linguistic structure and change can be fruitfully explained through language processing (e.g. Hawkins 1994; Hare and Elman 1995; Berg 2001). That is, processing is given a certain priority as an explanation of linguistic structure and change. From this angle, linguistic structure is understood as a set of fossilized processing strategies while linguistic change is seen as a response to processing biases (Berg 2001). Adopting this general perspective, we will turn to a processing model which is widely accepted in the psycholinguistic literature and which has been devised to deal with production data. It thus has been developed fully independently of the synchronic and diachronic data to which it will be applied below.

The following processing model belongs in the general class of network-based, activation spreading, local connectionist frameworks. Major proponents of this theory include Stemberger (1985), Dell (1986), and MacKay (1987). There are only two representational primitives--nodes and connections. The nodes represent linguistic units, the connections the relationships among them. This theory is mainly a theory of retrieval, that is, it specifies the mechanisms by which a linguistic unit is selected at a particular point in time. Three phases may be distinguished in retrieval--preselection, selection, and postselection. The selection process is prepared by the mechanisms of activation spreading whereby higher-level nodes activate lower-level nodes which in turn spread activation to all other nodes connected to them. After a period of information exchange among competing nodes, the system settles into a stable state, with one node at each level having accumulated a larger amount of activation than its competitors. According to a given time schedule, the most highly activated node is selected for production. Subsequent to its selection, this node undergoes a period of self-inhibition during which its activation level is zero. This unavailability is assumed to last for approximately 100 ms. (MacKay 1987). This state of inertia is followed by a rebound phase in which the activation level of the just-selected node rises again as a consequence of its being connected to other nodes that remained active as they did not themselves undergo selection. However, a node's activation level during postselection is lower than that during preselection. Therefore, the processing system naturally generates a right-to-left bias that is stronger than a left-to-right bias. After this rebound phase, the activation level tapers off until the node reaches its resting level.

As self-inhibition is a critical factor in the model to be proposed below, it is useful to dwell on it for a moment. Although it is modelled on the refractory phase that neurons undergo subsequent to firing, it is a genuinely psychological concept which serves a well-defined function in a psychological model. Its purpose is to prevent a unit which is fully activated, and therefore ready to be selected, from being selected over and over again even though it is only once needed. Self-inhibition thus ensures that units are not more frequently selected than they are actually needed. It is the major claim of this article that this processing model exactly predicts the pattern of results reported in the previous section. (14)

3.1. Accounting for the paradigmatic data

Let us begin with the paradigmatic level and its causation of similarity. From the psycholinguistic point of view, the paradigmatic dimension translates into the noncontextual selection of linguistic units. Under what constraints does noncontextual selection operate? As explained above, the network structure of, and the principle of activation spread in, the processing system make sure that a number of nodes are simultaneously active prior to selection. Crucially, the degree of co-activation depends on the similarity of the competitors and the target. This is a direct consequence of the architecture and the processing characteristics of the network. Take as a simple example the activation of the target phoneme/p/. During its preparation for output, activation accumulates on all nodes that are linked to it. For instance, both /k/ and /g/ are secondarily activated in the course of target processing because both share the [stop] node with /p/. However, /k/ and/g/are not activated to the same degree. The voiceless velar stop is more strongly activated than its voiced counterpart since the former receives additional activation from the [voiceless] node, which in turn was activated by the target. Hence, the greater the similarity between target and competitor, the higher the activation level of the latter.

This principle has an important implication. If the target unit is weak, (15) a similar competitor profits from the target's weakness more than a dissimilar one. As the similar element is more highly activated than the dissimilar one, the former is more likely to oust the target than the latter, given the design of the processing system. Thus, a paradigmatic substitution is prone to involve maximally similar interactants.

This is precisely what we find in slips of the tongue, one of the most faithful reflections of the processing constraints in the production system (Section 2.1.3). The similarity principle also operates in underdeveloped and disrupted networks (Sections 2.1.4 and 2.1.5). From the perspective of the model, this comes as no surprise. Let us suppose a developing system lacks nodes that exist in a fully developed system. Such an underdeveloped network generates the similarity effect in much the same way as a full-fledged network. The similarity effect emerges independently of network size because the processing principles are independent of network size. Alternatively, a developing system may be construed as supporting a reduced activation spread (relative to an adult system). Even in this case, the similarity effect persists. This is because the reduced transmission efficiency not only lowers the activation level of the target, but also that of the competitors. The ranking on the activation scale, with most activation for the target, less activation for similar and least activation for dissimilar competitors, remains unchanged.

The robustness of the similarity effect can also be seen in pathological language. If language breakdown is conceptualized as an inefficient transmission of activation between nodes, the same comments apply that were just made in regard to child language. Alternatively, the processing system of aphasics may be argued to be subject to excessive noise levels. However, random noise would also not abolish the similarity effect because, by definition, it affects all nodes with roughly the same probability. It therefore tends to preserve the difference in activation between similar and dissimilar competitors in the majority of cases. (16)

The pattern observed in spontaneous sound change (Section 2.1.2) fits very neatly into this picture. It may be inferred from this that sound change starts out under the same processing constraints as inadvertent speech errors, and that the dissemination of the innovation throughout a linguistic community does not affect the similarity relationship between the old and the new form. In fact, it stands to reason that this similarity is propitious to the spread of the innovation, as a small phonetic difference between the old and the new form diminishes listeners' awareness of, and hence their resistance to, the change (see Cole [1973] for relevant experimental evidence). In this type of language change, then, speakers and listeners reinforce one another in promoting similarity.

3.2. Accounting for the syntagmatic data

We now turn to the syntagmatic level. Psycholinguistically speaking, it captures the process of serialization, which is the selection from a pool of concurrently active units in a particular order. This definition shows that serialization is more complex than the isolated selection process discussed above. The syntagmatic level involves all phases of selection, in particular postselection self-inhibition as a prerequisite for the selection of the next unit in line. This self-inhibitory process is claimed to be at the heart of the dissimilarity effect found in adjacent and some nonadjacent units to be serialized. The logic is relatively straightforward. Recall that self-inhibition entails the momentary unavailability of the selected unit and that a unit's phonological specification is distributed, that is, shared with other units at the same analytical level. So after, let us say, the phoneme /b/ has been selected, not only the /b/ node but also the feature nodes for [bilabial], [stop], and [voiced] undergo self-inhibition. (17) This means that for a short period of time, these four nodes are less available than the other nodes. The implication of this processing principle is that /b/ will be most likely to be followed by a phoneme that shares none of its features with it, less likely to be followed by a phoneme that shares some features with it and least likely to be followed by a phoneme which shares all of its features, that is, which is identical to it.

This is exactly what we find in language structure (Section The largest difference between phonemes is that between consonants and vowels, and so vowels and consonants are most likely to alternate. This is true of most, if not all, languages. The second best choice is the succession of two consonants that are maximally different. The less this difference, the greater the phonological overlap and, hence, the greater the necessity of accessing nodes which are difficult to activate. This processing problem is elegantly and permanently solved by shaping language accordingly, that is, by avoiding the succession of similar elements. The worst choice is the immediate repetition of units, which appears to be universally dispreferred.

There are many other alternations which might be put down to the avoidance of identity, such as the alternation of stressed and unstressed syllables especially in stress-timed languages. Moreover, complementary quantity patterns bear mention in the present connection. In Swedish rimes, for example, a long vowel is followed by a short consonant while a short vowel is followed by a long consonant (or consonant cluster). The avoidance of adjacent phonemes of identical length may join other factors such as structural ones in bringing about this phonotactic constraint.

It is important to note that language structure prefers, but does not slavishly adhere to, the principle of maximum contrast. That is, it allows the principle of maximum contrast to be violated and contents itself with respecting the principle of minimum contrast when the need for greater information density arises. To give an example, a consonant may be added to a CV structure by creating a cluster or another syllable. Both options have their assets and their liabilities. The drawback of the former is the violation of the principle of maximum contrast, while the disadvantage of the latter is the doubling of word length. Which option is preferred is decided on a language-particular basis. However, the main point in the present connection is that the principle of maximum contrast has to compete with other principles such as shortness. Through this interaction it may be weakened into the principle of minimum contrast.

The child language data (Section lend themselves well to being explained in terms of self-inhibition. Developmental processes that increase the dissimilarity between adjacent phonemes may be understood as an attempt at avoiding access problems created by the self-inhibitory mechanism. Fewer nodes have to be selected in quick succession when adjacent phonemes are made dissimilar. It also follows naturally from the theory espoused here that children resort more to the principle of contrast than do adults. It may be argued that this is because children have more difficulty with overcoming the processing limitations imposed by self-inhibition. They therefore respect it more faithfully and produce an output that is close to what may, by the standards of the self-inhibitory mechanism, be regarded as an ideal output.

As mentioned above, phonological nodes recover rather quickly from their refractory phase, enter a phase of hyperexcitability, and finally return to their original state. We thus have three temporal phases which may be linked to three distinct empirical results. Firstly, the refractory phase is characterized by the absence of similarity and/or the presence of dissimilarity effects. Secondly, the hyperexcitability phase predictably leads to an excessive similarity. Thirdly, the return to resting level signals a phase in which the node regains its normal sensitivity to the similarity principle. In point of fact, evidence for all three behaviors has been gathered in Section 2.2.2. Let us consider each phase in turn.

The refractory phase seems to be responsible for quite a few patterns reported in Section 2.2.2. These include phonotactic restrictions within the syllable (Section, vowel mutation (Section, consonant harmony in child language (Section, and perhaps also aphasics' phoneme reversals (Section Notably, a trend towards phonological contrast emerges in all of these data types. While this tendency is usually not strong, the absence of a similarity effect can be established beyond doubt. Why the tendency towards dissimilarity is not more pronounced may be explained along the following lines. The essence of mutation is the anticipation of a front or back feature. If this anticipatory process is focused only on the backness dimension and therefore blind to the other features constituting the to-be-mutated vowel, similar and dissimilar vowels will be affected likewise.

A slightly different explanation may account for the child language data. It is generally recognized that children resort to harmony in an attempt to overcome the difficulty involved in producing disparately specified phonemes in the same planning unit. Thus, the trigger of harmony is the avoidance of contrast. However, not all phonological contrasts are equally objectionable. Children tend to experience most difficulty with disparate place specifications. Given this weighing, language learners are likely to harmonize consonants with disparate place specifications, irrespective of whether these consonants are otherwise similar or dissimilar. As in the historical data, the blindness of the process precludes the restriction of harmony to dissimilar consonants.

Unlike the diachronic and developmental data, the phonotactic evidence produces quite a strong dissimilarity effect. Similar (tautosyllabic) onset and coda consonants tend to occur less frequently, and dissimilar onset and coda consonants more frequently than would be predicted by chance. While it is clear that onset-coda restrictions are governed by several factors (such as the type of phonological contrast and the nature of the vowel), it is almost certain that similarity is not among them. It may be concluded from this that the effects of self-inhibition may extend throughout a CVVC syllable.

The only piece of evidence for the hyperexcitability phase comes from the above-chance occurrence of repeated phonemes in language structure (Section I adopt MacKay's (1970) explanation to the effect that the former is the cause of the latter. This is then another example of processing constraints penetrating the structure of language. The phoneme repetition effect is a hyperexcitability effect in the sense that processing facilitates the recurrence of identity rather than similarity. The claim would consequently be that the repetition effect applies at the phoneme level though not at the feature level. MacKay (1970) hints that the refractory period for feature nodes is shorter than that for phoneme nodes. Obviously, features are needed more often than phonemes. If MacKay's hypothesis is correct, the probability of feature repetition should peak earlier than that of phoneme repetition. As the linear distance between phonemes repeated at above-chance level is fairly small, the distance between features repeated at above-chance level should be even smaller, possibly involving adjacent phonemes or at most nonadjacent tautosyllabic phonemes. However, as shown above, there is no such feature repetition effect within the syllable. As it is unlikely to occur across syllable boundaries, it seems that it does not exist at all.

The two areas exhibiting a sensitivity to similarity are vowel harmony and slips of the tongue (Section According to the theory sketched out above, this is because these processes occur subsequent to the recovery phase. In other words, they are supported by a processing system that carries out the requisite task largely in ignorance of previous tasks. Therefore, its natural tendencies may unfold themselves without restraint and impose the similarity bias.

The sensitivity to similarity is stronger in slips of the tongue than in vowel harmony. Tongue slips are mainly brought about by similarity and noise, so it is hardly surprising that similarity has such an overwhelming effect. However, the origin of vowel harmony is different. As in consonant harmony in child language, its major function is to create phonemes which are identical on one phonological dimension. What matters in Hungarian, for example, is that all vowels have the same backness value, not that all vowels as such are identical. This aim may be reached by harmonizing both similar and dissimilar vowels. Hence, there is no compelling reason why dissimilar vowels cannot be involved in the harmony process, and so, relatively speaking, dissimilarity plays a larger role in vowel harmony than in slips of the tongue.

The preceding analysis has related the various data types to different temporal phases in the process of selecting linguistic units. On the basis of the claims that vowel harmony is sensitive to similarity whereas mutation is not, and that linear distance is the critical relevant variable, mutation may be predicted to be a short-distance and vowel harmony a long-distance effect (in relative terms). Indeed, this seems to be generally true. Vowel harmony in Hungarian affects all vowels in a word (excepting the so-called neutral ones), irrespective of the number of syllables in the word and the number of consonants intervening between the vowels. Vowel harmony may therefore be considered a word prosody and as such phonologically of a global nature.

By contrast, mutation in the history of English evinces different properties. Almost all cases discussed in the literature involve disyllabic words. What about trisyllabic words (with initial stress)? When the trigger occurred in the middle syllable, the stressed vowel was mutated whereas the vowel in the final syllable was not. This asymmetry reveals that the mutation process spread from right to left, never the other way around. Such a restriction is not usually found in vowel harmony.

However, when the trigger occurred in the final syllable, both the first and the second vowel could undergo mutation, as exemplified in

(28) West Germ. * apuling [right arrow] *apyling [right arrow] *apiling [right arrow] O.E. aepeling 'noble man' (from Lehnert 1973)

Case (28) may be interpreted as showing that the trigger in the final syllable fronted <a> to <ae> and <u> to <y>. As noted in Section, these are the two phonetic processes involved in mutation. The sequence of reconstructed forms shown in (28) allows us to infer that the influence of the trigger on the two target vowels proceeded in two steps. At first, the trigger changed the adjacent vowel and only later was the nonadjacent (i.e. initial) vowel changed. If this interpretation is correct, the adjacency condition on mutation can be maintained. Mutation spread only to adjacent syllables, not throughout the word.

A further indication of the local nature of mutation is that its domain is largely the single morpheme. The mutation process did not extend to prefixes, as can be seen in O.E. asettan 'to set up' in which the <e> of the stem represents mutated <a> (compare Gothic satjan) and in which the <a> of the prefix was not raised. Similarly, an &lt;i&gt; in the suffix often did not occasion mutation in the stem, for example, usic 'us' (Campbell 1959). Also in compounds, the operation of mutation was restricted. All these pieces of evidence show that mutation should not be described as a word prosody.

It seems safe to conclude that mutation in the history of English is a more local process than vowel harmony in Hungarian. As shown in Section, the linear distance between trigger and target is very small in the great majority of cases of mutation. By contrast, vowel harmony in Hungarian is not so restricted. It applies throughout the word and thus is less subject to locality conditions than mutation. This difference motivates the distinction between mutation as a similarity-insensitive and vowel harmony as a similarity-sensitive process. All that has to be assumed is that mutation spans a distance that is too short for similarity to unfold itself, whereas the distance between target and trigger in vowel harmony is large enough to allow for similarity effects to emerge.

The remaining data fall neatly into place. The greatest distance is covered by slips of the tongue, which more frequently than not cross word boundaries. Little wonder then, that they display the greatest sensitivity to similarity. The fact that between-word slips have a higher similarity index than within-word slips may be argued to ensue from the same principle. As noted before, the assumed absence of a similarity effect in aphasics' phoneme reversals probably follows from the minimal linear distance between the interacting units.

It would be nice if the exact location of the switch point from dissimilarity to similarity could be pinpointed. From the data analysis it can be gathered that it may lie between distance 1 (i.e. one intervening unit between the interactants) and distance 3. However, it is difficult to be more precise because the switch point is not in fact a point, but more probably a continuum in which (viewed from the perspective of increasing distance) dissimilarity gradually gives way to similarity. This continuum serves to explain why the difference between certain phenomena such as mutation and vowel harmony is gradual rather than absolute. For instance, locality plays a larger role in mutation than in harmony but neither process does without it completely.

It also has to be borne in mind that linear distance is a simplistic measure which blatantly ignores two critical properties of language and speech. One is that language is not a linear concatenation of units but hierarchically organized. Thus, linear distance has to be supplemented with hierarchical chunking. For example, a tautosyllabic distance 2 (e.g. .CVVC.) is almost certainly not the same as a heterosyllabic distance 2 (e.g. CV.VC) (with dots representing syllable boundaries). As the precise influence of hierarchical structure on the refractory phase is not known, this issue cannot unfortunately be pursued here.

The other critical property is speech rate. It is well known that speech rate varies from speaker to speaker and from situation to situation and that it is modality-specific. Theoretically, the duration of the refractory phase may be constant or depend on production rate. Again, there is no evidence one way or another, so the possibility of a further source of variation remains.

The temporal aspects of the processing theory allow one to make an intriguing prediction regarding the likelihood of repetition. Let us assume with MacKay (1987) a modality-neutral network of mental nodes. In view of the fact that writing is generally three to four times slower than speaking, the recovery cycle will be completed earlier in the production of written than in the production of spoken words. More specifically, given the hypothesis that the recovery cycle is completed after two or three spoken units, it may be completed even before the next written unit is selected. This implies that the written representation should allow closer distances between repeated units than the spoken medium. In particular, zero distances should be found more frequently in written than in spoken language. This is almost certainly so. As shown in Section, the immediate repetition of phonemes hardly ever occurs in English. In contrast, the immediate repetition of graphemes is not only possible, but even commonplace. Witness the letter doubling in soot and rabble as two examples for a host of others. This difference is argued to emanate from the differing temporal constraints to which speaking and writing are subject.

One empirical problem remains to be addressed. In Section 2.2.1, the structure of the English phoneme system was found to be shaped by both similarity and contrast although all other paradigmatic pieces of evidence exhibited a sensitivity to contrast only. It appears reasonable to argue that phoneme systems are not exclusively paradigmatic in nature but also implicate the syntagmatic dimension. Differently put, they are responsive not only to selectional but also to sequential constraints. This is a highly plausible assumption to make. Phonemes have to be selected not as isolated events but as parts of larger structures, such as words, for which members of the same class are required. For example, two vowels are generally needed for a disyllabic word. The listener's task is facilitated if these two vowels are dissimilar rather than similar. This brings us back to the principle of contrast discussed in the introductory section. As will be recalled, similarity is in the interest of the speaker whereas contrast is in the interest of the listener. Take vowel harmony as one line of indirect evidence for the operation of contrast. If there was no principle of contrast, vowel harmony as a similarity effect would unfold in unrestrained fashion. However, it is a minority pattern in the world's languages. This observation finds a ready explanation in the hypothesis that the speaker-oriented bias is counteracted by the listener's desire to maintain a fair amount of contrast among the vowels of a word. Of course, this is only possible if the vowel system is also constructed on the principle of contrast. Hence, it may be concluded that phoneme systems are sensitive to both similarity and contrast because they are shaped by both paradigmatic and syntagmatic influences to ensure efficient but not too effortful communication.

3.3. Going beyond phonology

The preceding analysis of similarity and contrast has focused exclusively on phonology. However, the model that has been developed to account for the phonological patterns has a wider applicability. In fact, in the light of the general processing principles invoked, the theory predicts basically the same patterns at all linguistic levels. This is because units at all levels are selected in essentially the same manner: they undergo self-inhibition and a subsequent rebound phase before returning to resting level. We would therefore expect adjacency relations to be characterized by contrast and a short later phase that is characterized by identity. Notwithstanding this overall parallelism in the processing system, it should be reiterated that the temporal constraints differ from level to level. To be specific, the recovery cycle is longer, the higher up in the linguistic hierarchy the level is located. A more serious problem is that the further we move up the linguistic hierarchy, the greater the impact of semantics. This means that the predicted processing effects on language structure may be less apparent at the higher levels because these are more immediately governed by the semantics and because an efficient coding system such as language must grant its contents a good deal of independence from the forms into which these contents are cast.

Let us briefly consider the morphology of English. Adjacent morphological categories (18) are indeed dissimilar, as predicted by the theory. A stem is flanked by an affix rather than another stem. Compounding is of course possible, but it occurs less frequently than derivation or inflection. Similarly, affixes prefer a stem rather than another affix by their side. Two successive prefixes are almost completely ruled out. Although two successive suffixes are perfectly possible, they are preferentially of different subtypes (i.e. derivational followed by inflectional) rather than of the same subtype.

The rebound phase predicts that the same category has a certain probability of coining up again after self-inhibition has been overcome. Given the morphological structure of English, there are not many possibilities where such an effect might manifest itself. However, one suitable testing ground is prefix-stem-suffix combinations. According to the rebound hypothesis, the likelihood of a suffix would be increased by the presence of a prefix. While this prediction has not yet been seriously examined, it is worth noting that Scalise (1988) made a relevant claim for these morphological complexes in Italian. He treats the so-called parasynthetic words, which are defined by the fact that the suffix can only be added to the prefixed form, not to the stem alone. Scalise comments that a prefix may favor the creation of a suffix. For example, the verb centrare 'to center' cannot be nominalized by adding the suffix--mento to give *centramento. However, suffixing is perfectly possible in the presence of a prefix such as de-, as in decentramento 'decentralization'. Parasynthetic words also occur in English. For example, embankment and to embolden are clearly three-morphemic items even though their two-morphemic

parts do not exist. This, then, might be a piece of morphological evidence for the rebound phase.

At the lexical level, we find strong evidence for contrast between adjacent categories. Items from the same word class do not normally occur next to each other. Of course, it is possible for two adjacent adjectives to modify the same noun, but such cases are rather infrequent. Significantly, when two nouns occur side by side, they tend to undergo morphologization, namely, compounding. From the perspective of the psycholinguistic model, this process may be seen as a means of avoiding repetition. By integrating two independent nouns into one lexical unit, the sequence of two identical lexical classes disappears even though a sequence of two identical morphological classes (i.e. stems) is thereby created. It does not seem way off the mark to claim that identity of category at the lower morphological level is less offensive than at the higher lexical level, given that the phonological level rebuffs identical adjacent categories even less (witness the incidence of consonant clusters). (19)

Needless to say, these considerations are only spotlights which serve to illustrate the range of application of the theory proposed above. Clearly, detailed studies are required to validate it in these domains. Its success in the phonological domain makes this endeavour seem a promising one, even though it is likely that the processing principles at issue may be weakened by semantic effects.

4. Conclusion: on the status of similarity and contrast in phonological theory

What are the implications of the foregoing analysis for phonological theory? Bernhardt and Stemberger (1998) complain that current phonological theory has no way of accommodating similarity although it plays an important role in phonological patterns. The data examined in this article suggest that this complaint be widened to encompass the notion of contrast, which also lacks an appropriate place in phonological theory. In addition to these notions, it will be necessary to take the distance between critical elements into account.

One objection to these suggested extensions might be that they are not required by a competence theory of language because the data, including the child language data discussed by Bernhardt and Stemberger, belong in the realm of performance. However, this objection cannot be sustained. While it is true that language acquisition and breakdown are aspects of language performance, language change and (even more so) language structure are generally recognized as providing data which have a clear bearing on issues of linguistic competence. As the so-called competence and performance areas treated in this study yielded highly similar results, there is really no basis for emphasizing this distinction in the present context.

Basically, there are two ways of providing a theoretical account of similarity and contrast. Either these notions are incorporated into a model of theoretical phonology, or the latter is supplemented by a processing model which is then made the locus of similarity and contrast effects. Let us briefly in turn examine the two alternatives.

To accommodate the two notions in question in an exclusively linguistic model, the application of a rule such as vowel harmony must be made contingent upon the number of nodes that the potential interactants share in the phonological representation. In this view, contrast is implemented by requiring a minimum number of shared nodes while similarity is implemented by requiring a maximum number of shared nodes for a rule to apply. Similar devices are not unknown in theoretical phonology. For example, syllabification rules are claimed to be sensitive to the sonority difference between adjacent consonants (e.g. Vennemann 1988), a mechanism which relies on similar computational principles. However, the real challenge is to account for the fact that the sensitivity to the number of shared nodes varies with the distance between the critical units. It is unclear how this interaction can be worked into the phonological representation in a non-ad-hoc manner.

Because of this unclarity, it may be preferable to propose a compound theory which accommodates representational and processing aspects alike. In such a model, the phonological representation provides the basis for defining relationships of similarity and contrast as well as those of adjacency and nonadjacency. How these relationships are instantiated is, however, left to the processing component. The processing mechanisms by which similarity and contrast are brought about were detailed in the preceding section and need not be recapitulated. The overall conclusion is, then, that the whole gamut of synchronic and diachronic data can be most adequately handled by a comprehensive model that views language as both a linguistic and a psychological phenomenon.
Table 1. Frequency of paradigmatic consonant substitution errors
in child language as a function of phonological similarity (based
on data presented in Snow 1963) (N = 4895)

 One-feature Two-feature Three-feature
 change change change

Number 3855 1023 17

Percentage (%) 78.8 20.9 0.3

Table 2. Frequency of contact assimilation in language change as
a function of similarity of the nonassimilated phonemes (N = 43)

 One-feature Two-feature Three-feature
 difference difference difference

Numbers 12 17 14

Percentage (%) 27.9 39.5 32.6

Table 3. Phonological similarity of adjacent consonants from adjacent
words in synchronic assimilations (N = 21) and in ordinary language
use (N = 129)

 One-feature Two-feature Three-feature
 difference difference difference

Assimilation 7 (33.3%) 10 (47.6%) 4 (19.0%)

 language 30 (23.3%) 53 (41.1%) 46 (35.7%)

Table 4. Frequency of nonmutated vowel pairs as a function of
linear distance (N = 147)

Distance Zero- One- Two- Three-
 phoneme phoneme phoneme phoneme

Number 1 104 41 1

Percentage (%) 0.7 70.7 27.9 0.7

Table 5. Similarity of critical consonants in one child's
harmonic forms as a function of linear distance (N = 230)

Distance Zero- One- Two- Three-
 phoneme phoneme phoneme phoneme


One-feature - 28 12 1
Two-feature - 71 41 2
Three-feature 3 50 20 -
Total 3 149 73 3

Distance Four- Total


One-feature 2 43
Two-feature - 114
Three-feature - 73
Total 2 230


* I am immensely grateful to Stefan Gries for calculating the chance probability of phoneme repetition. Without his able assistance, Section would have materialized as a shadow of itself. Beata Zaide has also had her share in the empirical work that went into this part of the story. A further debt of gratitude is due to Julia Schluter and two anonymous reviewers for their comments on an earlier version. This report is based on a talk I was delighted to give at the Scuola Normale Superiore, Pisa, in April 2003. Correspondence address: Department of English, University of Hamburg, VonMelle-Park 6, 20146 Hamburg, Germany. E-mail:

(1.) The present article concerns itself with the analysis of similarity and contrast in form rather than function (for a recent analysis of functional contrast in phonology, see Anderson 2001).

(2.) On the role of similarity in the acquisition of paradigms, see Rispoli (1994).

(3.) There are more complex, and probably more accurate, similarity metrics, such as the one developed by Frisch (2000), but the simple one used here will be sufficient for the present purposes.

(4.) Also, it would be consonant with the working hypothesis to predict that a voiced glottal fricative was added to the system (or rather elevated to acquire phonemic status). However, languages only seldom develop a phonemic contrast between voiceless and voiced glottal fricatives (see Maddieson 1984), and so this was not a real alternative.

(5.) [x] had the voiced counterpart [??] in Old English but the latter is generally treated as an allophone of/g/.

(6.) The phonemes/f/and/m/may both be assumed to be coded as [labial] at an abstract phonological level (Berg 1989).

(7.) Of course, this is not to deny the trivial observation that assimilation enhances the similarity between the interacting units. However, this is not the point here.

(8.) Assimilation will not be considered here because this term is customarily used to refer to the interaction of nonadjacent units in the developmental literature.

(9.) This fact is inconsequential because the identity in voicing is enforced by the phonotactic rules of English.

(10.) This is therefore an instance of paradigmatic similarity, as discussed in Section 2.1.4.

(11.) If there is identity nonetheless, the words tend to have a certain expressiveness (Fudge 1970).

(12.) The discussion of backmutation in the standard handbooks includes vowel change under the influence of an immediately preceding labio-velar glide (/w/), termed "graded" backmutation in Luick (1964). Since this is a case of interaction between adjacent phonemes, it will not be given any attention here.

(13.) It is worthwhile to point out that this distance is significantly smaller for backmutation than for [i/j] mutation ([chi.square](1)= 5.1, p < 0.025), ignoring the two cases of zero and three-phoneme distance. This difference may be taken as one piece of evidence in favor of the lexical conditioning of mutation. If the phonological structure of words was irrelevant, such a difference would not be expected.

(14.) See also Schluter's (2003) recent paper, which nicely illustrates the explanatory power of this model.

(15.) This weakness may have several reasons ranging from random noise in a quasi-neurological system to secondary thoughts in the speaker's mind.

(16.) Although noise may not abolish the similarity effect, it may weaken it, depending on how much noise interferes with normal processing strategies.

(17.) The latter part of this claim has an interesting implication at the phonetic level provided the high-level phonological constraints are powerful enough to reach the stage of articulation. The postselection self-inhibition hypothesis predicts that coarticulation effects are more of an anticipatory than a perseveratory nature. This prediction appears to accord quite well with the phonetic facts (Laver 1994).

(18.) We are dealing here with categories rather than individual linguistic items. It is just not realistic to expect the same morpheme or word to pop up again unless there is a semantic motivation for it.

(19.) Note in parentheses that diphthongization may be viewed in a similar light. Through this process, identical adjacent categories in phonology (i.e. vowels) are merged into one complex category (i.e. diphthongs). As a consequence, the undesired identity of adjacent categories (hiatus) disappears.


Anderson, John (2001). A major restructuring in the Old English consonant system: the delinearization of [h] and the de-consonantalization of [w] and [j]. English Language and Linguistics 5, 199-212.

Berg, Thomas (1989). How phonetic is a phonological feature representation? The case of labiodental fricatives. Speech Communication 8, 329-346.--(2001). Linguistic Structure and Change: An Explanation from Language Processing. Oxford: Oxford University Press.

Berkley, Deborah Milam (1994). Variability in obligatory contour principle effects. In Papers from the 30th Regional Meeting of the Chicago Linguistic Society. Vol 2: The Parasession on Variation in Linguistic Theory, K. Beals et al. (eds.), 1-12. Chicago: Chicago Linguistic Society.

Bernhardt, Barbara H.; and Stemberger, Joseph P. (1998). Handbook of Phonological Development. San Diego: Academic Press.

Blumstein, Sheila E. (1973). A Phonological Investigation of Aphasic Speech. The Hague: Mouton.

Bolinger, Dwight (1977). Meaning and Form. London: Longman.

Broselow, Ellen; and Finer, Daniel (1991). Parameter setting in second language phonology and syntax. Second Language Research 7, 35-59.

Burzio, Luigi (2002). Missing players: phonology and the past-tense debate. Lingua 112, 157-199.

Bybee, Joan L.; and Moder, Carol Lynn (1983). Morphological classes as natural categories. Language 59, 251-270.

Campbell, Alistair (1959). Old English Grammar. Oxford: Oxford University Press.

Chin, Steven B.; and Dinnsen, Daniel A. (1992). Consonant clusters in disordered speech: constraints and correspondence patterns. Journal of Child Language 19, 259-285.

Clark, Eve V. (1988). On the logic of contrast. Journal of Child Language 15, 317-335.

Clark, Herbert H. (1970). Word association and linguistic theory. In New Horizons in Linguistics, John Lyons (ed.), 271-286. Harmondsworth: Penguin.

Clements, George N. (1990). The role of the sonority cycle in core syllabification. In Papers in Laboratory Phonology L John Kingston and Mary E. Beckman (eds.), 283-333. Cambridge: Cambridge University Press.

Cole, Ronald A. (1973). Listening for mispronunciations: a measure of what we hear during speech. Perception and Psychophysies 13, 153-156.

Comrie, Bernard (1979). Morphophonemic exceptions and phonetic distance. Linguistics 17, 51-60.

Crothers, John (1978). Typology and universals of vowel systems. In Universals of Human Language, vol. 2, Joseph H. Greenberg (ed.), 93-152. Stanford: Stanford University Press.

de Saussure, Ferdinand (1916). Cours de linguistique generale. Paris: Payot.

Deese, James (1962). Form class and the determinants of association. Journal of Verbal Learning and Verbal Behavior 1, 79-84.

Dell, Gary S. (1986). A spreading-activation theory of retrieval in sentence production. Psychological Review 93,283-321.

Frisch, Stefan (2000). Temporally organized lexical representations as phonological units. In Papers in Laboratory Phonology V, Michael B. Broe and Janet B. Pierrehumbert (eds.), 283-298. Cambridge: Cambridge University Press.

Fudge, Erik (1970). Phonological structure and "expressiveness." Journal of Linguistics 6, 161-188.

Goldsmith, John (1985). Vowel harmony in Khalkha Mongolian, Yaka, Finnish, and Hungarian. Phonology Yearbook 2, 253-275.

Greenberg, Joseph H. (1965). Some generalizations concerning initial and final consonant sequences. Linguistics 18, 5-34.

Haiman, John (1980). The iconicity of grammar: isomorphism and motivation. Language 56, 515-540.

--(1985). Natural Syntax: Iconicity and Erosion. Cambridge: Cambridge University Press.

Happ, Heinz (1985). "Paradigmatisch"--"Syntagmatisch." Heidelberg: Winter.

Hare, Mary (1990). The role of similarity in Hungarian vowel harmony: a connectionist account. Connection Science 2, 123-150.--; and Elman, Jeffrey L. (1995). Learning and morphological change. Cognition 56, 61-98.

Hatfield, F. M.; and Walton, K. (1975). Phonological patterns in a case of aphasia. Language and Speech 18, 341-357.

Hawkins, John A. (1994). A Performance Theory of Order and Constituency. Cambridge: Cambridge University Press.

Hickey, Raymond (1984). Remarks on assimilation in Old English. Folia Linguistica Historica 5, 279-303.

Hockett, Charles F. (1958). A Course in Modern Linguistics. New York: Macmillan.

Hulst, Harry van der (1985). Vowel harmony in Hungarian: a comparison of segmental and autosegmental analyses. In Advances in Nonlinear Phonology, Harry van der Hulst and Norval Smith (eds.), 267-303. Dordrecht: Foris.

Ingram, David (1974). Phonological rules in young children. Journal of Child Language 1, 49-64.

--(1986). Phonological development: production. In Language Acquisition, Paul Fletcher and Michael Garman (eds.), 223-239. Cambridge: Cambridge University Press.

Jakobson, Roman (1944). Kindersprache, Aphasie und allgemeine Lautgesetze. Uppsala: Almqvist & Wiksell.

--(1971). Two aspects of language and two types of aphasic disturbances. In Fundamentals of Language, Roman Jakobson and Morris Halle (eds.), 67--96. The Hague: Mouton.

Janson, Tore (1986). Cross-linguistic trends in the frequency of CV sequences. Phonology Yearbook 3, 179-195.

Jenkins, James J. (1965). Transitional organization: association techniques. In Psycholinguistics. A Survey of Theory and Research Problems, Charles E. Osgood and Thomas A. Sebeok (eds.), 112-116. Bloomington: Indiana University Press.

Joanisse, Marc F. (1999). Exploring syllable structure in connectionist networks. In Proceedings of the 14th International Congress of Phonetic Sciences, San Francisco, John J. Ohala et al. (eds.), 731-734.

Kawasaki-Fukumori, Haruko (1992). An acoustical basis for universal phonotactic constraints. Language and Speech 35, 73-86.

Keller, Eric (1978). Parameters for vowel substitutions in Broca's aphasia. Brain and Language 5, 265-285.

Kessler, Brett; and Treiman, Rebecca (1997). Syllable structure and the distribution of phonemes in English syllables. Journal of Memory and Language 37, 295-311.

Kohn, Susan E.; Melvold, Janis; and Smith, Katherine L. (1995). Consonant harmony as a compensatory mechanism in fluent aphasic speech. Cortex 31,747-756.

--; and Smith, Katherine (1990). Between-word speech errors in conduction aphasia. Cognitive Neuropsychology 7, 133-156.

Lakoff, George; and Johnson, Mark (1980). Metaphors We Live By. Chicago: University of Chicago Press.

Laver, John (1994). Principles of Phonetics. Cambridge: Cambridge University Press.

Lehnert, Martin (1973). Altenglisches Elementarbuch. Berlin: Walter de Gruyter.

Liljencrants, Johan; and Lindblom, Bjorn (1972). Numerical simulation of vowel quality systems: the role of perceptual contrast. Language 48, 839-862.

Lindblom, Bjorn; MacNeilage, Peter; and Studdert-Kennedy, Michael (1984). Self-organizing processes and the explanation of phonological universals. In Explanations for Language Universals, Brian Butterworth, Bernard Comrie and Osten Dahl (eds.), 181-203. Berlin and New York: Mouton de Gruyter.

Locke, John L. (1983). Phonological Acquisition and Change. New York: Academic Press.

Luick, Karl (1964). Historische Grammatik der englischen Sprache. Stuttgart: Bernhard Tauchnitz.

Lutz, Angelika (1988). On the historical phonotactics of English. In Luick Revisited, Dieter Kastovsky and Gero Bauer (eds.), 221-239. Tubingen: Narr.

MacKay, Donald G. (1970). Phoneme repetition in the structure of languages. Language and Speech 13, 199-213.

--(1987). The Organization of Perception and Action. A Theory for Language and Other Cognitive Skills. New York: Springer.

Maddieson, Ian (1984). Patterns of Sounds. Cambridge: Cambridge University Press.

Martin, A. Damien; and Rigrodsky, Seymour (1974). An investigation of phonological impairment in aphasia, Part 2: Distinctive feature analysis of phonemic commutation errors in aphasia. Cortex 10, 329-346.

McNeill, David (1966). A study of word association. Journal of Verbal Learning and Verbal Behavior 5, 548-557.

Prasada, Sandeep; and Pinker, Steven (1993). Generalization of regular and irregular morphological patterns. Language and Cognitive Processes 8, 1-56.

Prather, Elizabeth M.; Hedrick, Dona Lee; and Kern, Carolyn A. (1975). Articulation development in children aged two to four years. Journal of Speech and Hearing Disorders 40, 179-191.

Pulleyblank, Douglas (1988). Underspecification, the feature hierarchy, and Tiv vowels. Phonology 5, 299-326.

Rispoli, Matthew (1994). Pronoun case overextensions and paradigm building. Journal of Child Language 21, 157-172.

Roach, Peter (1991). English Phonetics and Phonology. Cambridge: Cambridge University Press.

Scalise, Sergio (1988). Inflection and derivation. Linguistics 26, 561-581.

Schluter, Julia (2003). Chomsky's worst possible case: phonological determinants of grammatical variation. In Determinants of Grammatical Variation in English, Gunter Rohdenburg and Britta Mondorf (eds.), 71-118. Berlin and New York: Mouton de Gruyter.

Seiler, Hansjakob (1967). On paradigmatic and syntagmatic similarity. Lingua 18, 35-79. Shigeno, Sumi (1991). Assimilation and contrast in the phonetic perception of vowels. Journal of the Acoustical Society of America 90, 103-111.

Smith, Neilson V. (1973). The Acquisition of Phonology. A Case Study. Cambridge: Cambridge University Press.

Snow, Katherine (1963). A detailed analysis of articulation responses of "normal" first grade children. Journal of Speech and Hearing Research 6, 277-290.

Stemberger, Joseph P. (1985). The Lexicon in a Model of Language Production. New York: Garland.

--(1992). A connectionist view of child language. In Phonological Development. Models, Research, Implications, Charles A. Ferguson, Lise Menn and Carol Stoel-Gammon (eds.), 165-189. Timonium, MD: York Press.

Steriade, Donca (1982). Greek prosodies and the nature of syllabification. Unpublished doctoral dissertation, MIT, Cambridge, MA.

Stoel-Gammon, Carol (1985). Phonetic inventories, 15-24 months: a longitudinal study. Journal of Speech and Hearing Research 28, 505-512.

Svartvik, Jan; and Quirk, Randolph (1980). A Corpus of English Conversation. Lund: CWK Gleerup.

Trnka, B. (1936). General laws of phonemic combinations. Travaux du Cercle Linguistique de Prague 6, 57-62.

Trost, Judith E.; and Canter, Gerald J. (1974). Apraxia of speech in patients with Broca's aphasia: a study of phoneme production accuracy and error patterns. Brain and Language 1, 63-79.

Vennemann, Theo (1988). Preference Laws Jor Syllable Structure and the Explanation of Sound Change. Berlin and New York: Mouton de Gruyter.

Walker, Rachel (2001). Round licensing, harmony and bisyllabic triggers in Altaic. Natural Language and Linguistic Theory 19, 827-878.

Wang, William S.-Y.; and Cheng, Chin-Chuan (1977). Implementation of phonological change: the Shuang-Feng Chinese case. In The Lexicon in Phonological Change, William S.-Y. Wang (ed.), 148-158. The Hague: Mouton.

Received 28 May 2002 Revised version received 7 May 2003

University of Hamburg
COPYRIGHT 2004 Walter de Gruyter GmbH & Co. KG
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2004 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Berg, Thomas
Publication:Linguistics: an interdisciplinary journal of the language sciences
Date:Nov 1, 2004
Previous Article:Third person effects on binding *.
Next Article:Hybrid agreement in English *.

Terms of use | Privacy policy | Copyright © 2020 Farlex, Inc. | Feedback | For webmasters