Contrastive Specification in Phonology
Summary and Keywords
The fundamental idea underlying the use of distinctive features in phonology is the proposition that the same phonetic properties that distinguish one phoneme from another also play a crucial role in accounting for phonological patterns. Phonological rules and constraints apply to natural classes of segments, expressed in terms of features, and involve mechanisms, such as spreading or agreement, that copy distinctive features from one segment to another.
Contrastive specification builds on this by taking seriously the idea that phonological features are distinctive features. Many phonological patterns appear to be sensitive only to properties that crucially distinguish one phoneme from another, ignoring the same properties when they are redundant or predictable. For example, processes of voicing assimilation in many languages apply only to the class of obstruents, where voicing distinguishes phonemic pairs such as /t/ and /d/, and ignore sonorant consonants and vowels, which are predictably voiced. In theories of contrastive specification, features that do not serve to mark phonemic contrasts (such as [+voice] on sonorants) are omitted from underlying representations. Their phonological inertness thus follows straightforwardly from the fact that they are not present in the phonological system at the point at which the pattern applies, though the redundant features may subsequently be filled in either before or during phonetic implementation.
In order to implement a theory of contrastive specification, it is necessary to have a means of determining which features are contrastive (and should thus be specified) and which ones are redundant (and should thus be omitted). A traditional and intuitive method involves looking for minimal pairs of phonemes: if [±voice] is the only property that can distinguish /t/ from /d/, then it must be specified on them. This approach, however, often identifies too few contrastive features to distinguish the phonemes of an inventory, particularly when the phonetic space is sparsely populated. For example, in the common three-vowel inventory /i a u/, there is more than one property that could distinguish any two vowels: /i/ differs from /a/ in both place (front versus back or central) and height (high versus low), /a/ from /u/ in both height and rounding, and /u/ from /i/ in both rounding and place.
Because pairwise comparison cannot identify any features as contrastive in such cases, much recent work in contrastive specification is instead based on a hierarchical sequencing of features, with specifications assigned by dividing the full inventory into successively smaller subsets. For example, if the inventory /i a u/ is first divided according to height, then /a/ is fully distinguished from the other two vowels by virtue of being low, and the second feature, either place or rounding, is contrastive only on the high vowels. Unlike pairwise comparison, this approach produces specifications that fully distinguish the members of the underlying inventory, while at the same time allowing for the possibility of cross-linguistic variation in the specifications assigned to similar inventories.
Keywords: contrast, redundancy, underspecification, contrastive specification, distinctive features, contrastive hierarchy, Successive Division Algorithm, vowel harmony, voicing assimilation, consonant harmony
1. Contrast and Phonological Activity
The key motivation behind contrastive specification is the observation that features that mark phonemic contrasts often appear to be phonologically active, while features that are predictable or redundant often seem to be phonologically inert or irrelevant.
1.1 A Wholly Redundant Feature: [±low] in Turkish
For example, consider the case of Turkish vowel harmony, relevant aspects of which are discussed by Clements and Sezer (1982), Isac and Reiss (2008), Kabak (2011), and Dresher (2013), among others. The vowel inventory of Turkish is shown in (1); International Phonetic Alphabet (IPA) symbols are given first, followed by the corresponding orthographic representations in angle brackets.
The eight Turkish vowels can be distinguished from one another by the three binary features [±back], [±round], and [±high], as shown in (2).
Three binary features are the logical minimum needed to differentiate eight (= 23) phonemes. As it happens, the three specific features used in (2) cross-classify perfectly in this inventory, so that each possible combination of values corresponds to a vowel, and each vowel has a unique set of specifications. This set of features thus distinguishes the vowels of Turkish with maximum efficiency.
Phonetically, however, these three features are not the only properties in which Turkish vowels differ. As indicated in (1), the vowel /ɑ/ is phonetically lower than the other non-high vowels. (This can also be seen in the online supplement to Isac & Reiss, 2008, which illustrates the Turkish vowel inventory with audio recordings, photographs, and spectrograms at http://linguistics.concordia.ca/turkishvowels/.) If the feature [±low] were included in the representations, then /ɑ/ would be specified as [+low], and the other seven vowels would all be [−low]. This would distinguish /ɑ/ from the other vowels, but at least three other features—either the ones in (2) or some other set—would be needed to differentiate the [−low] vowels. The feature [±low] is thus redundant in Turkish, at least in the intuitive sense that its inclusion does not reduce the number of other features required to distinguish the contrasting vowels.
Would the inclusion of [±low] in the feature specifications have any phonological consequences? In principle, increasing the information contained in the representations can only increase the number of things the phonological computation can do. A system that includes [±low] in addition to the features in (2) can express all the same natural classes as a system without [±low], plus the class of non-low vowels, /i y ɨ u e ø o/.
In the case of Turkish, however, the principal effect of including [±low] would be to complicate the otherwise elegant pattern of vowel harmony. As illustrated in (3), vowels in suffixes harmonize with vowels in stems: a non-high suffix vowel assimilates to the place of the vowel before it, as in (3a), and high suffix vowels assimilate in both place and rounding, as in (3b). (The same pattern also exists within roots, though with exceptions; see Clements & Sezer, 1982; Kabak, 2011, for details.)
If Turkish vowels are specified only for the features in (2), then the alternation between /e/ and /ɑ/ in (3a) can be characterized as purely assimilatory and at least partially unified with the alternation among the high vowels in (3b). Under this view, all suffix vowels take their values for [±back] from the vowels to their left, and high suffix vowels additionally take [±round] from the same source. However, if /e/ is [−low] and /ɑ/ is [+low], then the alternation in (3a) involves something other than assimilation: back vowels in the stem, even high /ɨ/ and /u/, cause the vowel in the suffix to be low, and front stem vowels cause it to be non-low. No analogous correlation between place and height applies to the high vowels in (3b).
The inclusion of the redundant feature [+low] in the underlying specifications of the Turkish vowel inventory would not create any actual obstacles to accounting for harmony. In a derivational model of phonology, the harmony rule might create vowels that are unattested at the surface, such as [+low, −back] [æ] or [−low, +back] [ɤ], which could subsequently be eliminated by redundancy rules making all [+back, −round, −high] vowels [+low] and all other vowels [−low]. In the surface-oriented, constraint-based framework of Optimality Theory (Prince & Smolensky, 2004), the constraints driving harmony would simply make no reference to [±low], and faithfulness to underlying specifications for [±low] would be ranked below markedness constraints that, again, would require [+back, −round, −high] vowels to be [+low] and all other vowels [−low]. In either case, the effect would be the same as if [±low] had never been specified in the first place. In other words, even if this feature is assumed to be present, the phonology of Turkish systematically behaves as if it were not. Just as [±low] is redundant in distinguishing the vowels of the Turkish inventory, it appears to play no role in the phonology.
1.2 A Partially Contrastive Feature: [±back] in Finnish
In the case of Turkish, the feature [±low] seems to have no phonological relevance at all. Finnish, on the other hand, offers an example of a case in which a feature that is contrastive and phonologically active on some phonemes is redundant and inactive on others.
The phonemic vowel inventory of Finnish is shown in (4); as before, orthographic representations are given in angle brackets after each IPA symbol.
Like Turkish, Finnish has a pattern of vowel place harmony that can be seen most clearly in the alternating forms of suffixes. (As in Turkish, harmony applies within roots as well, but not without exceptions. Loan words in particular introduce some complications not dealt with here; see Kimper & Ylitalo, 2012; Ringen & Heinämäki, 1999, for discussion.) For example, in (5) the illative case suffix -stæ/-sta and the past participle suffix -nyt/-nut agree in backness with the vowels of the stems to which they are attached (data from D’Arcy, 2004; van der Hulst & van de Weijer, 1995).
However, not all vowels participate in harmony: /i/ and /e/, though they are phonetically front, can occur with either front or back vowels. As illustrated in (6), these two neutral vowels are transparent to harmony; they do not block the propagation of [±back] from other stem vowels onto a suffix, but they themselves are unaffected (data from D’Arcy, 2004; Krämer, 2002).
The harmonizing feature [±back] is clearly contrastive on each of the vowels that participate in harmony. Every back vowel is paired with a front vowel that has the same height and (un)roundedness; [±back] (or some equivalent feature such as coronal or grave/acute) must be specified in order to distinguish /u/ from /y/, /o/ from /ø/, and /ɑ/ from /æ/.
The neutral vowels /i/ and /e/, on the other hand, do not have back counterparts. Unlike /y/, /ø/, and /æ/, they can be fully distinguished from the other members of the inventory without being specified for [−back]. Jakobson, Fant, and Halle (1952, p. 41) note the connection between harmonic transparency and the absence of contrast, writing that “those acute vowels which ceteris paribus are paired with grave vowels cannot belong to the same simple word-unit as the grave vowels […], while the plain acute vowels /e i/, which have no plain grave counterparts, are compatible with any Finnish vowel.” Kiparsky (1985, p. 115) frames the connection in terms of structure preservation, noting that the application of harmony to /i/ and /e/ would produce surface vowels *[ɨ] and *[ɤ] that are not present in the underlying inventory.
If redundant features are omitted altogether, or if, as Calabrese (1995, 2005) and Nevins (2010) propose, certain phonological processes systematically ignore them, then the transparency of /i/ and /e/ to harmony follows naturally from their position in the inventory. As illustrated in (7), vowels for which [±back] is contrastive share their values for this feature within roots, and spread it to suffixes; the two vowels on which [−back] is redundant, on the other hand, are no more involved in harmony than are consonants.
Because [±back] is demonstrably phonologically active on vowels for which it is contrastive, Finnish harmony provides a more compelling case for the special status of contrastive features than Turkish harmony can. In Turkish, it would be possible to say that [±low] is present even though it has no effect on the phonology; in Finnish, however, if non-contrastive [−back] on /i/ and /e/ were represented in the same way as the contrastive [−back] of /y ø æ/, there would be no obvious explanation for the fact that non-low unrounded vowels do not participate in place harmony.
2. Identifying Contrastive Features
To pursue any form of contrastive specification, it is necessary to have some means of determining which features are contrastive in a given system. This task, however, is not necessarily as straightforward as it may at first appear.
2.1 Pairwise Comparison
One intuitive approach to identifying contrastive features on segments in a phonological inventory is to apply the same logic used in identifying segments as contrasting phonemes. A minimal pair of words such as fan and van constitutes evidence that the distinction between /f/ and /v/ is phonemic in English, as this difference is the only means by which the two forms are distinguished. By the same token, /f/ and /v/ themselves constitute a minimal pair of segments, distinguished only by the feature [±voice]. The specification of [−voice] on /f/ and [+voice] on /v/ must therefore be contrastive.
Archangeli (1988, p. 192) presents the following procedure for identifying contrastive features and eliminating redundant ones based on pairwise comparisons of this kind:
This procedure is illustrated in (9) and (10), where it is applied to the vowel inventory of Finnish. In (9), full specifications are shown for the features [±back], [±round], [±high], and [±low]. (Features for which all the vowels would have the same value, such as [±sonorant], have been omitted.) The feature values that distinguish minimal pairs of vowels are highlighted: in cases where the members of the pair are adjacent in the table, the relevant features are enclosed together; for the non-adjacent pairs /i e/ and /y ø/, the specifications that distinguish them are separately enclosed in matching shapes.
In the case of Finnish, pairwise comparison usefully identifies [±back] as contrastive on /y u o ø æ ɑ/ and not on /i e/; this is exactly what is required for the account of place harmony sketched above in (7).
However, as Archangeli (1988) and Dresher (2009, section 2.5) have pointed out, the specifications designated contrastive by this procedure do not always adequately distinguish the members of phonemic inventories. In the Finnish example in (10), /i/ and /æ/ do not have opposing ‘contrastive’ specifications for any feature, nor do the members of the pairs /e ɑ/, /y æ/, /ø æ/, /u ɑ/, or /o ɑ/. Within each of these pairs, the two vowels have distinct representations only if the presence of a (positive or negative) specification is taken to be distinct from the absence of a specification—an effectively ternary use of binary features that has historically been controversial in phonology. (For example, both Halle, 1959, p. 32; Stanley, 1967, p. 410) define distinctness so as to require an opposition between plus and minus rather than between either and zero, and Kiparsky (1982, p. 54) likewise avoids ternarity in his approach to underspecification. On the other hand, Archangeli (1988) treats zero as distinct from plus or minus, and Inkelas (1995) specifically argues for the validity and usefulness of three-way contrasts among plus, minus, and zero.)
In the case of Turkish vowels, the procedure in (8) cannot arrive at the specifications in (2). If [±back], [±round], and [±high] were the only features specified to begin with, then all specifications would be designated contrastive, as each vowel has a minimally different counterpart along each of these three dimensions. Step (8a), however, stipulates that the procedure begins with full specifications for all features—in this case, crucially including [±low]. As Dresher (2013) observes, while [±low] is not identified as contrastive on any Turkish vowel, its presence in the initial specifications affects the status of the other features. The values that distinguish minimal pairs are highlighted in (11):
If [±low] is initially included, then /ɑ/ does not enter into any minimal pairs. It and its closest counterparts are underspecified accordingly, as shown in (12): [−back] is not marked as contrastive on /e/, [+round] is not marked as contrastive on /o/, [+high] is not marked as contrastive on /ɨ/, and /ɑ/ itself has no ‘contrastive’ features at all.
Except for [±low] itself, the omitted values all have consequences for harmony: [±back] and [±round] are the harmonizing features, and [±high] identifies which vowels undergo rounding harmony. Excluding [±low] from the initial specifications would eliminate the problem, but pairwise comparison offers no principled basis for doing so.
The failures of the pairwise approach are even more extreme in inventories that more sparsely populate the available phonetic space. For example, if the very common five-vowel inventory /i e a o u/ is initially specified for the same four features, there are only two minimal pairs, as shown in (13):
The specifications designated as ‘contrastive’ fail to differentiate the members of the inventory, even if the absence of a value is assumed to be distinct from either a positive or a negative specification; as shown in (14), the procedure in (8) assigns identical representations to /i/ and /u/, and to /e/ and /o/:
This counterintuitive result—that the feature values identified as contrastive do not in fact suffice to identify the phonemic contrasts in the system—is not necessarily problematic for theories of phonology that accord special status to contrastive features but assume that non-contrastive features are also specified (e.g., Calabrese, 1995; Nevins, 2010). However, for any approach in which contrastive features are the only ones specified at some point in the phonological computation (e.g., Archangeli, 1988; Archangeli & Pulleyblank, 1989; Dresher, 2009; Kiparsky, 1982, 1985), pairwise comparison is clearly untenable. If putatively non-contrastive features cannot reliably be recovered from the ‘contrastive’ representations, then some of the omitted features must be contrastive, in the sense that they are not predictable.
Because it begins with fully specified representations, the pairwise comparison procedure in (8) is also incompatible with the possibility that features are emergent and cross-linguistically variable (as proposed by Mielke, 2008, among others). The procedure must be able to assign full specifications for some universal set of features a priori, including those that are not contrastive or phonologically active in the language in question. The only basis for assigning such feature values is the phonetic realization of the segments in the inventory, but that can be highly variable, especially in small inventories. For example, in Hawaiian, where there is no phonemic contrast between /k/ and /t/, the lone plosive intermediate between /p/ and /ʔ/, which is usually transcribed as /k/, can be realized phonetically as [k] or [t] (Schütz, 1995); in a set of full specifications, should it be represented as coronal, or as dorsal? (See Hall, 2011 for further discussion.) If the procedure in (8) is intended to assign contrastive specifications deterministically, its first step must be made explicit by spelling out the procedure for assigning full specifications.
The essential conceptual flaw in the minimal pairs approach is that the implication on which it rests works in only one direction. It is true that, if a phoneme x has a minimal counterpart y, differing from x only in feature F, then F must be contrastive on x (and on y), but the inverse is not necessarily true: the absence of a minimal counterpart y does not entail that F is redundant on x. All it entails is that F is predictable on x at least if all the other features of x are known. If all such features are eliminated from the representation of x, some of them may not be recoverable, as illustrated in (13) and (14).
2.2 The Successive Division Algorithm
To avoid the problems inherent in the pairwise approach, Dresher (2009) argues for an alternative approach to contrastive specification, in which features recursively divide inventories into smaller subinventories. Rather than beginning with full specification, as the pairwise method does, Dresher’s Successive Division Algorithm (SDA) starts with no feature specifications at all, adding features only if they serve to mark some phonemic distinction.
In step (15b), the wording “as many subsets as the feature allows for” indicates that the same procedure can be applied using privative, binary, or multivalent features, or any combination thereof. A privative feature or a binary one will divide a set into two subsets, either marked and unmarked or positive and negative, while a multivalent feature will produce as many subsets as it has values. The examples below all use binary features, as does much work in this framework, but Hall (2007) uses monovalent features; Dresher (2009, section 2.7.2) discusses some of the implications of feature valency for the algorithm.
The application of the SDA to particular inventories can be represented visually in branching diagrams of the sort used by Cherry, Halle, and Jakobson (1953) and Halle (1959), among others. For example, if the features [±low], [±round], [±back], and [±high] (in that order) are used to divide the vowel inventory of Finnish, the result is as shown in (16):
The initial division separates the inventory into the [−low] subset /e i ø y o u/ and the [+low] subset /æ ɑ/. Within the [−low] subinventory, [±round] divides /e i/ from /ø y o u/. In the [+low] subinventory, on the other hand, [±round] cannot make a division, and so this feature is not specified on the low vowels. Likewise, [±back] does not divide the set /e i/, and so these vowels remain unspecified for place (which is consistent with their transparency to place harmony). The sets /ø y o u/ and /æ ɑ/, however, can each be subdivided by [±back], and so their subsets /ø y/ and /æ/ are specified [−back], while /o u/ and /ɑ/ are specified [+back]. The vowels /æ/ and /ɑ/ are each now fully distinguished from all the others, and no further features can be assigned to them; within each of the other remaining subsets, [±high] makes the final division.
The specifications assigned by this sequence of divisions are shown in tabular form in (17). In (17a), the vowels and features are listed in the order corresponding to the tree in (16); in (17b), they are presented in the same order as in (10), for ease of comparison.
Because [±low] is the first feature assigned in (16), it is specified on all eight vowels, even though in some cases its value could be inferred from other features assigned later. (For example, any vowel specified as [+high] must be [−low], and in Finnish it also happens to be the case that all [+round] vowels are [−low].) Such apparent redundancies arise from the fact that the SDA relies on the notion of contrastive scope to avoid the problems of the pairwise method. In the hierarchy in (16), [±low] takes scope over [±high]; [±high] is contrastive (and thus specified) only among the [−low] vowels. If these two features were applied in the opposite order, with [±high] taking wider scope, then [±low] would be contrastive only among the [−high] vowels. Depending on the order of divisions, [+high] vowels may be specified for [−low], or [+low] vowels may be specified for [−high], but not both at the same time.
While [±low] may seem intuitively to be redundant on the Finnish high vowels, it is contrastive in two important (and related) senses. First, at the point in the algorithm at which it is assigned, it is contrastive because it separates the inventory into two contrasting subsets. Second, in the context of the complete set of specifications in (17), [±low] is the only feature for which opposite values have been assigned to the members of the pairs /i æ/, /y æ/, and /u ɑ/. If the high vowels were not specified as [−low], then the absence of a value would have to be treated as contrasting with the presence of a positive or negative value—that is, binary features would have to be interpreted as ternary. Although it includes values that seem predictable, the set of specifications in (17) is minimal in that for any feature F specified on any phoneme x, there is at least one other phoneme y such that F is the only feature for which x and y are marked with opposite values.
The Successive Division Algorithm itself does not specify what features are to be used to divide an inventory, and so it is compatible with the Emergent Feature Theory of Mielke (2008), in which features are induced by learners rather than being fully predetermined by Universal Grammar. The SDA also does not place any restrictions on the order in which the divisions should be made. The prediction implicit in this is that languages with phonetically similar inventories may divide them in different ways, resulting in different feature specifications and different phonological patterns. Some evidence for this is discussed in section 3.2. To the extent that there turns out to be cross-linguistic consistency in the ways in which features are ordered, explanations for any universals or typological tendencies may be sought in factors outside the algorithm itself, such as the principles of Robustness and Marked Feature Avoidance formulated by Clements (2009).
Unlike the pairwise method, the SDA is not deterministic; in order to produce a set of feature specifications for an inventory, it must be supplied not only with the inventory itself, but also with an ordered list of features. From the perspective of a phonologist working within the framework defined by the SDA, the necessary order of divisions in a given language is an empirical question: features that are demonstrably phonologically active must have high enough scope to be assigned to the segments that pattern as if they are specified for them, and features that are phonologically ignored on segments where they are phonetically present (such as [−back] on /i/ and /e/ in Finnish) must be low enough in the hierarchy that they are not specified on those segments. The predictive power of the SDA thus lies not in an ability to predict phonological representations from inventory shape alone, but rather in the ability to predict the range of possible representations for a given inventory, which will always involve trade-offs between potentially contrastive features. For example, as discussed above, the SDA predicts that in an inventory like that of Finnish, high vowels can be specified as [−low] or low vowels as [−high], but not both.
The SDA, in a sense, is a less complete method for determining contrastive specifications than pairwise comparison, but it is also a much more robust one. All of the features it assigns are contrastive, because it assigns features only to sets of segments in which they make phonemic distinctions. Unlike pairwise comparison, the specifications assigned by the SDA always fully differentiate the members of the inventory, because the SDA does not stop assigning features until all phonemes have distinct representations.
3. The Scope of Contrast
The crucial insight underlying the Successive Division Algorithm is that a potential feature specification can be identified as contrastive or redundant only in light of what other feature values have already been assigned. If a segment has some property that is not shared by all the other segments in the inventory, but which is not the only property that distinguishes it from any other single segment, then that property may or may not be contrastive, depending on its relative position in the hierarchy.
3.1 Intermediate Scope: [±voice] in Russian
In Russian, the feature [±voice] acts as if it is specified on some segments for which it is not minimally contrastive, but not on others. The phonemic consonant inventory of Russian, adapted from Timberlake (2002, p. 829), is shown in (18). (Timberlake’s transcription symbols have been replaced with IPA equivalents. Phonologists differ as to the phonemic status of some consonants not included here, such as long [ʃj:], historically a cluster, and the palatalized velars [kj], [ɡj], and [xj]; these questions are not crucial to the current discussion.)
As in many other languages, sonorants in Russian are consistently voiced, and have no voiceless counterparts. The obstruents mostly fall into voiced–voiceless pairs, except for the affricates /ʦ ʧj/ and the velar fricative /x/, which have no minimally different phonemic voiced counterparts.
The feature [±voice] is involved in two phonological processes: final obstruent devoicing and regressive voicing assimilation. The word-final devoicing of obstruents is illustrated in (19) (data from Hayes, 1984; Kiparsky, 1985; Padgett, 2002):
Regressive voicing assimilation applies within sequences of obstruents, including ones that cross word boundaries if the words are sufficiently close in the morphosyntactic structure. Final devoicing feeds assimilation, so that word-final obstruent clusters are voiceless. Sonorants neither trigger nor undergo voicing assimilation, nor do they devoice word-finally. (Labiodental /v vj/, historically *w, behaves like a sonorant in that it fails to trigger regressive voicing, but, patterns with other voiced obstruents in undergoing word-final and assimilatory devoicing; see, e.g., Halle, 1959; Padgett, 2002 for discussion.) Voicing assimilation is illustrated in (20) (data from Halle, 1959; Padgett, 2002; vowel reduction has been omitted from the transcriptions).
Although sonorants do not participate in assimilation or final devoicing, the unpaired obstruents /ʦ ʧj x/ are both targets and triggers of assimilation. (Because they are all underlyingly voiceless, final devoicing could apply to them only vacuously in any case.) Regressive assimilatory voicing of the unpaired obstruents produces voiced obstruents [dz ʤj ɣ] not present in the underlying inventory, as illustrated in (21) (data from Halle, 1959; Timberlake, 2002):
Regressive assimilatory devoicing triggered by unpaired obstruents is illustrated in (22), which shows their effects on the preposition /bjez/ ‘without’ (data from Calabrese, 1995; cf. also (20f) and (20g) above).
Full specification of all features would include [+voice] on sonorants, where it is phonologically inactive. Contrastive specification by pairwise comparison would leave voicing unspecified on sonorants, but also on the unpaired voiceless obstruents, which need to be able to spread [−voice] leftward in forms like those in (22). However, if contrastive specifications are assigned by the Successive Division Algorithm, then it is possible to say that [±voice] has lower scope than [±sonorant], but higher scope than the other features that might distinguish /ʦ ʧj x/ from the voiced obstruents.
The tree in (23) shows a partial contrastive hierarchy for Russian, with [±sonorant] taking scope over [±voice]:
Within the [+sonorant] subinventory, there is no contrast to be marked by [±voice], and so sonorants are unspecified for voicing. Among the [−sonorant] consonants, on the other hand, [±voice] is contrastive, and so it is assigned to all obstruents, dividing them into the two subsets shown. Even though /ʦ ʧj x/ lack minimal voiced counterparts, [−voice] is contrastive on them in that it serves to distinguish them from the whole set of voiced obstruents.
Given the hierarchy in (23), sonorants will be unable to trigger regressive assimilatory voicing, as illustrated in (24a), but the unpaired obstruents will be able to trigger regressive assimilatory devoicing, as in (24b):
In this analysis, the intermediate position of [±voice] in the contrastive hierarchy is the key to its phonological behavior. Wider scope would cause [±voice] to be specified on consonants on which it is phonologically inactive; narrower scope would cause it to be unspecified on consonants on which it is phonologically active.
3.2 Variable Scope: [±distributed] and [±sonorant] in West Nilotic
In addition to allowing for features to be contrastive on some unpaired segments but not others within a single language, the Successive Division Algorithm also permits the scope of contrasts to vary from one language to another, so that an unpaired segment that is specified for a feature in one language may be unspecified for that feature in another language with a similar inventory. Mackenzie (2009, 2011) discusses one such case involving consonant place harmony in Dholuo and Anywa, two West Nilotic languages. (See also Mackenzie, 2013, for a similar example involving laryngeal harmony in Ngizim and Hausa.)
Both Dholuo and Anywa have a phonemic contrast between dental plosives /t̪ d̪/ and alveolar /t d/. (Dholuo additionally has prenasalized /nd̪/ and /nd/; see Mackenzie, 2011, for a more detailed discussion.) Both languages also have co-occurrence restrictions requiring coronal stops within a root to agree in place: dentals can occur with dentals, and alveolars with alveolars, as in the words in (25), but no root can contain both a dental and an alveolar plosive (in either order). (All examples in this section are from Mackenzie, 2009, 2011, who cites Tucker, 1994, for the Dholuo data and Reh, 1996, for the Anywa.)
In both languages, there is a single coronal nasal /n/, with no minimally distinct dental counterpart /n̪/. However, /n/ is treated differently in the two harmony systems. In Dholuo, /n/ is ignored by dental harmony, just as non-coronal consonants are; it freely occurs with either dental or alveolar plosives, as illustrated in (26):
In Anywa, the phoneme /n/ can also occur with both alveolars and dentals, as shown in (27), but it is not invisible to place harmony; unlike Dholuo /n/, it assimilates in place to dental plosives, producing allophonic [n̪], as in (27b):
The phonemic consonant inventories of Dholuo and Anywa share the same basic asymmetry illustrated in (28): plosives minimally contrast for [±distributed], the feature that distinguishes dentals from alveolars, but nasals do not.
Assuming full specification would leave no way of relating the harmonic invisibility of Dholuo /n/ to this asymmetry; if all consonants are specified for [±distributed], then the harmony rules or co-occurrence constraints would simply have to stipulate that only coronal plosives participate. On the other hand, contrastive specification based on pairwise comparison would predict that /n/ should be ignored by harmony in Anywa as well as in Dholuo.
As Mackenzie (2009, 2011) argues, contrastive specification based on a cross-linguistically variable hierarchy of features offers a more satisfactory account of the two patterns. The crucial variable is the relative scope of the features that distinguish dentals from alveolars ([±distributed]) and nasals from plosives ([±sonorant]). The Dholuo pattern is consistent with the partial hierarchy in (29), in which [±sonorant] takes wider scope:
In (29), the specification for [+sonorant] fully distinguishes /n/ from the dental plosives /t̪ d̪/, and so it receives no specification for [±distributed]; it is accordingly ignored by dental harmony.
The Anywa pattern, on the other hand, suggests the partial hierarchy in (30), with the features in the opposite order:
If [±distributed] takes scope over [±sonorant] as in (30), then all five of the relevant consonants will be specified for the harmonizing feature. When /n/ co-occurs with /t̪ / or /d̪/, the harmony rule or co-occurrence constraint will require it to assimilate, causing its underlying [−distributed] specification to change to [+distributed] at the surface.
The variable scope of [±distributed] in West Nilotic and the intermediate scope of [±voice] in Russian both exemplify a crucial way in which specification based on a contrastive hierarchy differs from specification based on pairwise comparison. The pairwise method predicts that all and only those values that do not distinguish minimal pairs of phonemes will be unspecified (sometimes resulting in specifications that are not distinct at all). The contrastive hierarchy approach, on the other hand, predicts that only values that do not distinguish minimal pairs will be unspecified, but not all such values will necessarily be unspecified. In much the same way that Optimality Theory generates typological variation in grammars by factorial re-ranking of constraints, the Successive Division Algorithm generates typological variation in representations through different hierarchical orderings of features.
4. Predictions and Challenges
The Successive Division Algorithm is designed to guarantee that specifications designated as contrastive in any case are indeed sufficient to distinguish the phonemes of the underlying inventory, while at the same time allowing for cross-linguistic variation in the relative scope of features. Previous approaches to underspecification have largely relied on theories of markedness to determine which features to omit as redundant. For example, in the Contrastive Underspecification of Steriade (1987) and Calabrese (1988), feature values can be omitted from underlying representations if they are predictable from co-occurrence constraints, and in the Radical Underspecification of Kiparsky (1982) and Archangeli (1984, 1988), only marked values are underlyingly specified (building on chapter 9 of Chomsky & Halle, 1968).
The relative flexibility of the SDA is a response to cross-linguistic variation of the sort discussed by Mackenzie (2009, 2011, 2013), and, more generally, to the observation that substantive patterns of markedness are not always cross-linguistically consistent (Hume, 2003; Rice, 1996). This flexibility, however, potentially weakens the predictive power of the SDA relative to other approaches to contrastive specification, although the range of variation in feature scope could be constrained by combining the SDA with a theory of markedness along the lines of Clements’s (2001, 2009) feature accessibility hierarchy.
Taken by itself, though, the main predictive power of the SDA lies in its ability to limit the number of features that can be assigned to a given inventory, rather than in an ability to predict which specific features will be used. For example, if the SDA is used to assign binary features to the common three-vowel inventory /i a u/, then only two features can be assigned in total, and one segment—the one with the minority value for the first feature assigned—will be unspecified for the second feature. The logical possibilities involving the features [±low], [±back], and [±round] are exhaustively shown in (31) (and note that [±high] would make exactly the same divisions as [±low], merely reversing the positive and negative values).
The predictions of the SDA are strongest if redundant feature specifications are excluded from the phonology altogether (Hall, 2007, section 1.2.2), being filled in only at the point of phonetic implementation rather than during the phonological computation itself as posited by Archangeli (1988, p. 192). (An intermediate possibility, pursued by D’Arcy, 2003, in the spirit of Kiparsky, 1982, is that redundant features are absent during the lexical phonology, but present in the post-lexical phonology.) If a feature is observed to be phonologically active on a given segment in a given language, then the SDA predicts that it must have high enough scope to be contrastive on that segment, and that any other features that might have served to make the same distinction must therefore be redundant and phonologically inactive.
The empirical falsifiability of this approach to contrast is discussed by Nevins (2010, 2015), who claims that its predictions are in fact too strong, on the basis of what he calls ‘Oops, I Need That’ problems. For example, Nevins argues that if redundant [−back] is entirely absent on Finnish /i/ and /e/, consistent with their transparency in vowel place harmony, then there is no way of accounting for the fact that /i/ triggers assibilation of /t/ to [s] in derived environments, as in (32) (data from Kiparsky, 1973).
If [−back] must be phonologically active on /i/ in order to account for assibilation, but inactive for the purposes of vowel harmony, then, Nevins argues, some phonological patterns are sensitive only to contrastive features, but others also have access to redundant ones. However, the argument is only as strong as its premises. The assumption that [−back] is needed for assibilation in Finnish is predicated on Hall and Hamann’s (2006) phonetically grounded account of cross-linguistic assibilation patterns, but [−back] is not logically necessary for identifying /i/ as the trigger of assibilation (the combination [+high, −round] suffices), nor must it necessarily be available to spread from /i/ to /t/, as the target of assibilation changes only in manner and not in place. Still, this and other ‘Oops, I Need That’ problems raised by Nevins (2010, 2015) (some of which are taken up by Dresher, 2015) deserve careful consideration, and the question of the role and status of redundant features in general is one which any theory of contrastive specification must contend with.
This article has presented some of the basic motivations and mechanisms of contrastive specification. There is, however, much more to be said, and the following sources will provide the reader with a deeper and broader understanding of the topic. Taken together, the works described in this section offer both a diversity of approaches to contrast and a strong sense of its importance in accounting for a wide range of phonological phenomena.
Archangeli, D. (1988). Underspecification in phonology. Phonology, 5(2), 183–207. presents the original argument against specification based on pairwise comparison, arguing for an alternative which she calls Radical Underspecification. She presents key evidence for the importance of underspecification and for the criterion of predictability in identifying which features should be underspecified.Find this resource:
Calabrese, A. (1995). A constraint-based theory of phonological markedness and simplification procedures. Linguistic Inquiry, 26(3), 373–463.Find this resource:
Calabrese, A. (2005). Markedness and economy in a derivational model of phonology. Berlin, Germany: Mouton de Gruyter. Calabrese discusses the relation between contrast and markedness and argues for Visibility Theory, in which phonological patterns vary parametrically as to whether they are sensitive only to marked feature values, only to contrastive feature values, or to all feature values. Taken together, these two works provide a wealth of case studies showing the importance of contrast, but also arguing that non-contrastive features are sometimes phonologically active.Find this resource:
Dresher, B. E. (2009). The contrastive hierarchy in phonology. Cambridge, U.K.: Cambridge University Press. For a detailed discussion of the Successive Division Algorithm and a historical overview of the ways in which contrast and underspecification have been used in phonology, the best source is Dresher’s book-length treatment. Dresher traces the ways in which contrast has been treated over time, from its roots in structuralist phonology to its treatment in Optimality Theory, pointing out key assumptions about contrastiveness and redundancy that underlie work in which they have often not previously been explicitly stated. He also presents many examples of the application of the SDA to particular phonological problems.Find this resource:
Hall, D. C. (2011). Phonological contrast and its phonetic enhancement: Dispersedness without dispersion. Phonology, 28(1), 1–54. discusses the relation between phonological contrast and phonetic distinctness, arguing that the Successive Division Algorithm, combined with the notion of phonetic enhancement (, can explain not only phonological processes but also cross-linguistic generalizations about the phonetic shape of phonological inventories.Find this resource:
Lahiri, A., & Reetz, H. (2010). Distinctive features: Phonological underspecification in representation and processing. Journal of Phonetics, 38(1), 44–59. Lahiri and Reetz argue for the Featurally Underspecified Lexicon (FUL), which posits privative features and a combination of universal and contingent underspecification. Some features, in particular , are universally unmarked, while others, such as , are specified if, and if, they serve to mark a contrast in the inventory. The authors provide evidence for this approach both from phonological patterns and from experimental data on perception.Find this resource:
Mackenzie, S. (2011). Contrast and the evaluation of similarity: Evidence from consonant harmony. Lingua, 121(8), 1401–1423.Find this resource:
Mackenzie, S. (2013). Laryngeal co-occurrence restrictions in Aymara: Contrastive representations and constraint interaction. Phonology, 30(2), 297–345. Mackenzie uses contrastive specification based on the Successive Division Algorithm to account for a range of different consonant harmony patterns (including the two briefly sketched in section 3.2). She argues that the notion of similarity, which is widely posited to underlie such patterns, is best understood in terms of formal phonological representations based on contrastive feature specifications, rather than on surface phonetic resemblances.Find this resource:
Morén, B. (2006a). Consonant–vowel interactions in Serbian: Features, representations and constraint interactions. Lingua, 116(8), 1198–1244.Find this resource:
Morén, B. (2006b). The division of labor between segment-internal structure and violable constraints. In Blaho, S., Bye, P., & Krämer, M (Eds.), Freedom of analysis? (pp. 313–344). The Hague, Netherlands: Mouton. Morén’s Parallel Structures Model uses privative features in minimal representations based on contrast. As in other contrast-based approaches to feature specification, the Parallel Structures Model includes more complex representations only when they contrast with simpler ones; unlike most other feature-based approaches, it enforces minimality by requiring that, for each feature used in an inventory, there must be one segment specified only with that feature (analogous to the Autonomous Interpretation Hypothesis in Element Theory, on which see ).Find this resource:
Nevins, A. I. (2010). Locality in vowel harmony. Cambridge, MA: MIT Press. Further applications of Calabrese’s approach to contrast are presented by , who uses Visibility Theory in conjunction with a model of searching and feature valuation inspired by the mechanisms of minimalist syntax (. Nevins provides several case studies in vowel harmony and illustrates the role of contrast in relativized minimality.Find this resource:
Nevins, A. I. (2015). Triumphs and limits of the contrastivity-only hypothesis. Linguistic Variation, 15(1), 41–68. More recently, Nevins proposes a synthesis that retains the parametric availability of redundant features from Visibility Theory, but adopts the Successive Division Algorithm as the means for identifying contrastive features.Find this resource:
Oxford, W. (2015). Patterns of contrast in phonological change: Evidence from Algonquian vowel systems. Language, 91(2), 308–357. Oxford explores the relevance of contrast for historical phonology in a detailed treatment of diachronic developments in the vowel systems of Algonquian languages. He shows that contrastive hierarchies offer a way of understanding the structure of phonological systems that yields new insights into how those systems change over time, through mergers, splits, shifts, and reanalyses.Find this resource:
Steriade, D. (1995). Underspecification and markedness. In Goldsmith, J. A. (Ed.), The handbook of phonological theory (pp. 114–174). Oxford: Blackwell. Steriade provides an extensive comparison of Radical Underspecification and Contrastive Underspecification, ultimately proposing that both of these approaches, rooted in derivational assumptions, may be superseded in a declarative model of phonology by licensing constraints on features. This view has been explicitly or implicitly adopted in much subsequent work, particularly within the framework of Optimality Theory.Find this resource:
Trubetzkoy, N. S. (1939). Grundzüge der phonologie. Travaux du cercle linguistique de Prague. This foundational work by Trubetzkoy (also available in English translation as ), contains one of the earliest and most lasting and influential discussions of the role of contrast in phonology. While Trubetzkoy’s work predates the mechanisms of modern generative phonological theory, his insights into the relation between phonemic contrast and phonological patterning remain relevant today.Find this resource:
Archangeli, D. (1984). Underspecification in Yawelmani phonology and morphology. (Doctoral dissertation). Massachusetts Institute of Technology.Find this resource:
Archangeli, D. (1988). Underspecification in phonology. Phonology 5(2), 183–207.Find this resource:
Archangeli, D., & Pulleyblank, D. (1989). Yoruba vowel harmony. Linguistic Inquiry 20(2), 173–218.Find this resource:
Calabrese, A. (1988). Towards a theory of phonological alphabets. (Doctoral dissertation). Massachusetts Institute of Technology.Find this resource:
Calabrese, A. (1995). A constraint-based theory of phonological markedness and simplification procedures. Linguistic Inquiry 26(3), 373–463.Find this resource:
Calabrese, A. (2005). Markedness and economy in a derivational model of phonology. Berlin, Germany: Mouton de Gruyter.Find this resource:
Cherry, E. C., Halle, M., & Jakobson, R. (1953). Toward the logical description of languages in their phonemic aspect. Language 29(1), 34–46.Find this resource:
Chomsky, N. (1995). The minimalist program. Cambridge, MA: MIT Press.Find this resource:
Chomsky, N. (2000). Minimalist inquiries: The framework. In R. Martin, D. Michaels, & J. Uriagereka (Eds.), Step by step (pp. 89–155). Cambridge: MIT Press.Find this resource:
Chomsky, N., & Halle, M. (1968). The sound pattern of English. New York, NY: Harper and Row.Find this resource:
Clements, G. N. (2001). Representational economy in constraint-based phonology. In T. A. Hall (Ed.), Distinctive feature theory (pp. 71–146). Berlin, Germany: Mouton de Gruyter.Find this resource:
Clements, G. N. (2009). The role of features in phonological inventories. In E. Raimy, & C. E. Cairns (Eds.), Contemporary views on architecture and representations in phonology (pp. 19–68). Cambridge, MA: MIT Press.Find this resource:
Clements, G. N., & Sezer, E. (1982). Vowel and consonant disharmony in Turkish. In H. van der Hulst, & N. Smith (Eds.), The structure of phonological representations (pp. 213–255). Dordrecht, Netherlands: Foris.Find this resource:
D’Arcy, A. (2003). Yowlumne reexamined: A challenge for contrastive specification. Toronto Working Papers in Linguistics 20, 21–46.Find this resource:
D’Arcy, A. (2004). Unconditional neutrality: Vowel harmony in a two-place model. Toronto Working Papers in Linguistics, 23(2), 1–46.Find this resource:
Dresher, B. E. (2009). The contrastive hierarchy in phonology. Cambridge, U.K.: Cambridge University Press.Find this resource:
Dresher, B. E. (2013). Contrastive features and microvariation in vowel harmony. In S. Keine, & S. Sloggett (Eds.), NELS 42: Proceedings of the Forty-Second Annual Meeting of the North East Linguistic Society (pp. 141–153). Amherst, MA: GLSA.Find this resource:
Dresher, B. E. (2015). The motivation for contrastive feature hierarchies in phonology. Linguistic Variation, 15(1), 1–40.Find this resource:
Hall, D. C. (2007). The role and representation of contrast in phonological theory (Doctoral dissertation). University of Toronto.Find this resource:
Hall, D. C. (2011). Phonological contrast and its phonetic enhancement: Dispersedness without dispersion. Phonology, 28(1), 1–54.Find this resource:
Hall, T. A., & Hamann, S. (2006). Towards a typology of stop assibilation. Linguistics 44(6), 1195–1236.Find this resource:
Halle, M. (1959). The sound pattern of Russian: A linguistic and acoustical investigation. The Hague: Mouton.Find this resource:
Harris, J., & Lindsey, G. (1995). The elements of phonological representation. In J. Durand, & F. Katamba (Eds.), Frontiers of phonology (pp. 34–79). London: Longman.Find this resource:
Hayes, B. (1984). The phonetics and phonology of Russian voicing assimilation. In M. Aronoff, & R. T. Oehrle (Eds.), Language sound structure (pp.318–328). Cambridge, MA: MIT Press.Find this resource:
Hume, E. (2003). Language specific markedness: The case of place of articulation. Studies in Phonetics, Phonology and Morphology 9(2), 295–310.Find this resource:
Inkelas, S. (1995). The consequences of optimization for underspecification. In J. N. Beckman (Ed.), Proceedings of the North East Linguistic Society 25 (pp. 287–302). Amherst, MA: GLSA.Find this resource:
Isac, D., & Reiss, C. (2008). I-language: An introduction to linguistics as cognitive science. Oxford: Oxford University Press.Find this resource:
Jakobson, R., Fant, C. G. M., & Halle, M. (1952). Preliminaries to speech analysis. Technical Report 13, MIT Acoustics Laboratory.Find this resource:
Kabak, B. (2011). Turkish vowel harmony. In M. van Oostendorp, C. J. Ewen, E. Hume, & K. Rice (Eds.), The Blackwell companion to phonology, Vol. 5 (pp. 2831–2854). Oxford: Wiley-Blackwell.Find this resource:
Kimper, W., & Ylitalo, R. (2012). Variability and trigger competition in Finnish disharmonic loanwords. In N. Arnett, & R. Bennett (Eds.), Proceedings of the 30th West Coast Conference on Formal Linguistics (pp. 195–204). Somerville, MA: Cascadilla Proceedings Project.Find this resource:
Kiparsky, P. (1973). Abstractness, opacity and global rules. In O. Fujimura (Ed.), Three dimensions of linguistic theory (pp. 57–86). Tokyo, Japan: Tokyo Institute for Advanced Studies of Language.Find this resource:
Kiparsky, P. (1982). Lexical morphology and phonology. In The Linguistic Society of Korea (Ed.), Linguistics in the morning calm: Selected papers from SICOL-1981 (pp. 3–91). Seoul, South Korea: Hanshin.Find this resource:
Kiparsky, P. (1985). Some consequences of lexical phonology. Phonology Yearbook 2, 85–138.Find this resource:
Krämer, M. (2002). Local constraint conjunction and neutral vowels in Finnish harmony. Belfast Working Papers in Language and Linguistics 15, 38–64.Find this resource:
Lahiri, A., & Reetz, H. (2010). Distinctive features: Phonological underspecification in representation and processing. Journal of Phonetics, 38(1), 44–59.Find this resource:
Mackenzie, S. (2009). Contrast and similarity in consonant harmony processes. (Doctoral dissertation). University of Toronto.Find this resource:
Mackenzie, S. (2011). Contrast and the evaluation of similarity: Evidence from consonant harmony. Lingua, 121(8), 1401–1423.Find this resource:
Mackenzie, S. (2013). Laryngeal co-occurrence restrictions in Aymara: Contrastive representations and constraint interaction. Phonology, 30(2), 297–345.Find this resource:
Mielke, J. (2008). The emergence of distinctive features. Oxford: Oxford University Press.Find this resource:
Morén, B. (2006a). Consonant–vowel interactions in Serbian: Features, representations and constraint interactions. Lingua, 116(8), 1198–1244.Find this resource:
Morén, B. (2006b). The division of labor between segment-internal structure and violable constraints. In S. Blaho, P. Bye, & M. Krämer (Eds.), Freedom of analysis? (pp. 313–344). The Hague: Mouton.Find this resource:
Nevins, A. I. (2010). Locality in vowel harmony. Cambridge, MA: MIT Press.Find this resource:
Nevins, A. I. (2015). Triumphs and limits of the contrastivity-only hypothesis. Linguistic Variation, 15(1), 41–68.Find this resource:
Oxford, W. (2015). Patterns of contrast in phonological change: Evidence from Algonquian vowel systems. Language, 91(2), 308–357.Find this resource:
Padgett, J. (2002). Russian voicing assimilation, final devoicing, and the problem of [v] (or the mouse that squeaked). Manuscript, University of California, Santa Cruz. ROA #528. Available online at http://roa.rutgers.edu/article/view/538Find this resource:
Prince, A., & Smolensky, P. (2004). Optimality theory: Constraint interaction in generative grammar. Oxford: Blackwell.Find this resource:
Reh, M. (1996). Anywa language: Description and internal reconstruction. Cologne: Ruddiger Koppe.Find this resource:
Rice, K. (1996). Default variability: The coronal–velar relationship. Natural Language & Linguistic Theory, 14(3), 493–543.Find this resource:
Ringen, C. O., & Heinämäki, O. (1999). Variation in Finnish vowel harmony: An OT account. Natural Language & Linguistic Theory, 17(2), 303–337.Find this resource:
Schütz, A. J. (1995). The voices of Eden: A history of Hawaiian language studies. Honolulu: University of Hawai‘i Press.Find this resource:
Stanley, R. (1967). Redundancy rules in phonology. Language, 43(2), 393–436.Find this resource:
Steriade, D. (1987). Redundant values. Chicago Linguistic Society, 23(2), 339–362.Find this resource:
Steriade, D. (1995). Underspecification and markedness. In J. A. Goldsmith (Ed.), The handbook of phonological theory (pp. 114–174). Oxford: Blackwell.Find this resource:
Stevens, K. N., & Keyser, S. J. (1989). Primary features and their enhancement in consonants. Language, I(1), 81–106.Find this resource:
Timberlake, A. (2002). Russian. In B. Comrie, & G. G. Corbett (Eds.), The Slavonic languages (pp. 827–886). London: Routledge.Find this resource:
Trubetzkoy, N. S. (1939). Grundzüge der phonologie. Travaux du cercle linguistique de Prague, 7.Find this resource:
Trubetzkoy, N. S. (1969). Principles of phonology (C. A. M. Baltaxe, Trans.). Berkeley and Los Angeles: University of California Press.Find this resource:
Tucker, A. (1994). A Grammar of Kenya Luo (Dholuo). Cologne: Ruddiger Koppe.Find this resource:
van der Hulst, H., & van de Weijer, J. (1995). Vowel harmony. In J. A. Goldsmith (Ed.), The handbook of phonological theory (pp. 495–534). Oxford: Blackwell.Find this resource: