2.1 The branching tree

On page 46 of SPR is Figure I-1, a tree diagram that shows the contrastive feature specifications of every phoneme in Russian.Footnote ² Each branch in the tree subdivides the inventory according to some distinctive feature. This tree is reproduced in our Figure 1. Since the tree is very large, we have divided it into separate subtrees for vocalic and non-vocalic segments. We have also replaced Halle’s transcription symbols by their IPA equivalents (as we have done throughout this article).

Figure 1 Contrastive hierarchy for Russian phonemes, adapted from Halle (Reference Halle1959: 46).

Halle’s numbering of features, listed in the key in Figure 1, is preserved; feature 1 in Figure 1a is [$ \pm $vocalic], which forms the top branch in his tree. Because this is the first feature, every segment in the inventory must receive a value for this feature. In Halle’s feature system, the [$ + $vocalic] class comprises vowels and liquids; non-liquid consonants, including the glide /j/, are [$ - $vocalic]. From each numbered node in the tree, the branch to the left contains phonemes with the negative value of the corresponding feature, and phonemes with the positive value are in the branch to the right.

The next feature, numbered 2, is [$ \pm $consonantal], which is contrastive in both major branches of the tree. Looking first at the [$ - $vocalic] branch of the tree in Figure 1b, only the glide /j/ is contrastively [$ - $consonantal]. Because the tree has no non-branching non-terminal nodes, no feature is assigned to any segment unless it serves to distinguish that segment from at least one other. As /j/ is now unique, no further features are assigned to it. All the other segments in Figure 1b are [$ + $consonantal], and therefore need to be distinguished from each other by additional features. On the [$ + $vocalic] side of the tree in Figure 1c, vowels are [$ - $consonantal] and liquids are [$ + $consonantal].

The third feature in the ordered list is [$ \pm $diffuse], which is distinctive only for the vowels (Hale Reference Halle1959: 126). As it does not make any contrasts in the consonants in Figure 1b, it is skipped in that tree and the next feature is 4, [$ \pm $compact]. The liquids in Figure 1c are not distinguished by any other feature until number 8, [$ \pm $continuant] (perhaps unexpectedly, /r, r^j/ are [$ - $continuant] and /l, l^j/ are [$ + $continuant]).

The division of the tree by contrastive features continues until every phoneme has been uniquely distinguished.

2.2 An argument for specification by branching trees

Before continuing, we should ask why the branching tree occurs at all: why is it necessary to specify features in this way? In SPR, Halle makes an argument on behalf of branching trees – the first such argument we have found in the literature. He argues that phonological features must be ordered into a hierarchy because this is the only way to ensure that segments are kept properly distinct, as defined in the Distinctness Condition (Halle Reference Halle1959: 32) in (1):

This formulation is designed to disallow contrasts involving a zero value of a feature. Consider, for example, the typical sub-inventory /p, b, m/ shown in (2), and suppose we characterize it in terms of two binary features, [$ \pm $voiced] and [$ \pm $nasal]. In terms of full specifications, /p/ is [$ - $voiced, $ - $nasal], /b/ is [$ + $voiced, $ - $nasal], and /m/ is [$ + $voiced, $ + $nasal].

Which of these specifications are contrastive? Many phonologists reason as follows. We observe that /p/ and /b/ are distinguished only by [$ \pm $voiced], so these specifications must be contrastive. Similarly, /b/ and /m/ are distinguished only by [$ \pm $nasal]; these specifications must also be contrastive. Specifying only these features yields the partial specification in (3).

The remaining specifications are predictable from the ones in (3). Since /p/ is the only [$ - $voiced] phoneme in this inventory, its specification for [$ \pm $nasal] is predictable, hence redundant. We can write a rule or constraint: if [$ - $voiced], then [$ - $nasal]. Similarly, /m/ is the only [$ + $nasal] phoneme, so its specification for [$ \pm $voiced] is redundant: if [$ + $nasal], then [$ + $voiced]. This is a still-popular way of thinking about contrastive specifications. Dresher (Reference Dresher2009) refers to it as ‘pairwise comparison’; we will call it the ‘Minimal Difference’ approach. By whatever name, essentially this approach is adopted explicitly by Padgett (Reference Padgett2003), Calabrese (Reference Calabrese2005), Campos Astorkiza (Reference Astorkiza and Rebeka2007), and Nevins (Reference Nevins2010), and by many others implicitly.

According to the definition proposed by Nevins (Reference Nevins2010: 98), a segment S with specification [$ \alpha $F] is contrastive for F if there is another segment S′ in the inventory that is featurally identical to S, except that it is [$ -\alpha $F]. The condition is met by the specifications of [$ \pm $voiced] on /p/ and /b/. It is not met by [$ - $nasal] on /p/, because there is no other segment in this inventory that differs from /p/ only by this feature. For this feature to be contrastive on /p/ there would have to be a voiceless nasal /$ \underset{\circ}{\mathrm{m}} $/ in this inventory, as in (4). The same would be required by Minimal Difference to make [$ + $voiced] contrastive on /$ \underset{}{\mathrm{m}} $/. In the absence of voiceless nasal /$ \underset{\circ}{\mathrm{m}} $/, Minimal Difference gives us the specifications in (3).

According to the Distinctness Condition, however, the set of specifications in (3) is not properly contrastive. /p/ is ‘different from’ /b/, because /p/ is [$ - $voiced] and /b/ is [$ + $voiced]. Similarly, /b/ is ‘different from’ /m/, because /b/ is [$ - $nasal] and /m/ is [$ + $nasal]. But /p/ is not ‘different from’ /m/: where one has a feature, the other has no specification. Therefore, these specifications are not properly contrastive. They violate the Distinctness Condition because no feature hierarchy yields this result. If we order [$ \pm $voiced] ahead of [$ \pm $nasal] (notated [$ \pm $voiced] $ \gg $ [$ \pm $nasal]), we generate an ‘extra’ specification on /m/ (5a, 6a). If we order [$ \pm $nasal] $ \gg $ [$ \pm $voiced], we generate an ‘extra’ specification on /p/ (5b, 6b).

That is, according to SPR, the correct theory of contrast is a hierarchical one, and not Minimal Difference. To be properly contrastive, a set of feature specifications must be derivable by a branching tree. This view of contrastive specifications is reflected in the first generative phonology textbook by Harms (Reference Harms1968), but the matter became dormant after the critique of underspecification by Lightner (Reference Lightner1963) and Stanley (Reference Stanley1967) (see further Section 3.3). We believe, however, that Halle was correct in requiring contrastive specifications to be hierarchical. As demonstrated by Archangeli (Reference Archangeli1988) and Dresher (Reference Dresher2009), the Minimal Difference approach often fails to yield any intelligible set of specifications. Conceptually, the main flaw of Minimal Difference is its failure to recognize that contrastive relations in an inventory exist not just between pairs of segments, but also between groups of segments at different levels of the hierarchy.

For example, [$ + $voiced] on /m/ in (6a) is predictable from the other specifications, but it is required to distinguish /b, m/ as a group from /p/; at this point in the tree, this feature is contrastive for the entire set /p, b, m/. The same is true for [$ - $nasal] on /p/ in (6b). These contrasts involve ‘minimal differences’ at the point where they are made – in (6a) [$ \pm $voiced] makes a minimal difference between /p/ and /b, m/ and in (6b) [$ \pm $nasal] makes a minimal difference between /p, b/ and /m/. The layered character of hierarchically derived contrastive features is what results in contrastive features that are not necessarily unpredictable.Footnote ³

2.3 Rationale for feature hierarchies: Minimality

In many cases, there is more than one feature that could potentially distinguish two segments (or sets of segments), and so the ordering of the features has consequences for which of them are actually specified. In the simple example of /p, b, m/ in the previous section, we have seen that different feature orderings result in different contrastive specifications. Returning to SPR, consider, for example, the class of labial consonants /p, p^j, b, b^j, m, m^j, f, f^j, v, v^j/, which is identified within the non-vocalic inventory in Figure 1b by the features [$ + $consonantal, $ - $compact, $ + $low tonality]. Within this set, feature 6, [$ \pm $strident], distinguishes the stops and nasals /p, p^j, b, b^j, m, m^j/ from the fricatives /f, f^j, v, v^j/. This means that feature 8, [$ \pm $continuant], is not marked on any of the labials, because the distinction that it would make has already been made by [$ \pm $strident]. If the order of these two features were reversed, then [$ \pm $continuant] would mark the division, and [$ \pm $strident] would be redundant.

Although [$ + $strident] is specified on the labiodental fricatives, it is not marked on the prototypically strident postalveolar fricatives and affricate /ʃ, ʒ, ʧ/. The set of posterior consonantal segments /ʧ, ʃ, ʒ, k, k^j, ɡ, x/ identified by the features [$ + $consonantal, $ + $compact] in Figure 1b (see also Figure 2b in Section 3.1), is divided into postalveolars /ʧ, ʃ, ʒ/ and velars /k, k^j, ɡ, x/ by feature 5, [$ \pm $low tonality]. Because there are no non-strident postalveolar consonants and no strident velar ones, neither subset can be further divided by [$ \pm $strident] (feature 6). The later-ordered [$ \pm $continuant] (feature 8), on the other hand, does have a role to play here, separating the affricate /ʧ/ from the fricatives /ʃ, ʒ/ and the fricative /x/ from the stops /k, k^j, ɡ/. All these specifications could be altered if the features were ordered differently.

Figure 2 Subtrees resulting in underspecification of [$ \pm $voiced] in Halle (Reference Halle1959).

What, then, is the basis for the ordering? Halle (Reference Halle1959: 29–30) provides the rationale for it in his Condition (5):

Conditions (3) and (4) set out basic requirements of adequacy: Condition (3), which we discuss further in Section 3.2, requires that phonological representations contain enough information to derive the corresponding phonetic utterances without referring to other (e.g. semantic or syntactic) sources, and Condition (4) requires the phonological description to be ‘appropriately integrated into the grammar of the language’ (see further Section 4.2). Once those conditions are met, the primary desideratum in ordering the features is to minimize the number of feature specifications.

This concern for minimality stems at least in part from Halle’s interest in information theory; Cherry, Halle & Jakobson (Reference Cherry, Halle and Jakobson1953) show how binary features, parsimoniously assigned in a hierarchical order, can provide a metric for how much information phonemes can encode, and with how much redundancy. We will refer to this emphasis on efficiency of encoding as the Minimality Principle, which can be stated as in (7):

A concern with Minimality is reflected in Halle’s (Reference Halle1959: 44–45) observation that his analysis of Russian contains 43 phonemes specified by 271 feature specifications, or 6.3 distinctive feature statements per phoneme. He compares 6.3 with the lower limit of $ {\log}_243 $ ($ \approx 5.43 $) specifications, which would represent the most efficiently branching tree for 43 phonemes.Footnote ⁴ By the criterion of Minimality, this improves upon the feature hierarchy for Russian posited by Cherry et al. (Reference Cherry, Halle and Jakobson1953), which uses an average of 6.5 features per phoneme for an inventory of 42 phonemes (compared to a logical minimum of $ {\log}_242\approx 5.39 $).Footnote ⁵ Halle (Reference Halle1959: 45) cautions against reading too much into the precise numbers involved, concluding only that ‘the minimization process has achieved results of the type that might reasonably be expected’.

Minimality governs (at most) only the hierarchical ordering of features, and not the choice of features themselves; an underlying assumption for both Cherry et al. (Reference Cherry, Halle and Jakobson1953) and Halle (Reference Halle1959) is that features have phonetic content, and so the inventory can only be divided along lines of identifiable phonetic difference. ‘One might ask,’ Cherry et al. (Reference Cherry, Halle and Jakobson1953: 37–38) write, ‘Why cannot a type of feature pattern be invented which employs only’ the logical minimum average number of specifications per phoneme? Their answer is that ‘[t]his could perhaps be done; but the distinctive features used at present… serve other purposes and are intimately related to the physical production of speech’ (Cherry et al. Reference Cherry, Halle and Jakobson1953: 38). The ‘other purposes’ that distinctive features serve include not only phonetic description, but also accounting for phonological patterns; in the sections that follow, we examine the consequences of giving Minimality priority over phonological patterning in the ordering of features and show how a different approach to building feature hierarchies could have changed the agenda that SPR set for generative phonology.

3.1 The unpaired obstruents

The ordering of two features in the part of the tree we looked at in Section 2.3 had momentous consequences for the development of phonological theory. These are features 8, [$ \pm $continuant], and 9, [$ \pm $voiced]. The relevant portions of the hierarchy are repeated in Figure 2 in a more easily legible form and size.

In these subtrees, [$ \pm $continuant] is ordered above [$ \pm $voiced]. Each segment in these diagrams has a specification for [$ \pm $continuant], but the same is not the case for [$ \pm $voiced]. The three circled consonants in Figure 2, /ʦ, ʧ, x/, have no phonemic voiced counterparts; that is, there are no underlying */ʣ, ʤ, ɣ/ in Russian. These are the ‘unpaired’ voiceless obstruents that formed the focus of Halle’s (Reference Halle1957, Reference Halle1959) famous argument against a certain conception of the phoneme. The unpaired obstruents receive no specification for [$ \pm $voiced], because they have already been fully distinguished from the rest of the inventory by features higher in the hierarchy. But, as Halle (Reference Halle1959: 22–23) pointed out, these segments behave phonologically like other voiceless obstruents with respect to voicing assimilation. This fact is accounted for in SPR by rule ordering: Rule P1b, which fills in redundant [$ - $voiced] on the unpaired obstruents, precedes the assimilation Rule P3a:

The effect of this ordering is that the unpaired obstruents end up triggering assimilatory devoicing of obstruents to their left, despite being underlyingly underspecified. This can be seen, for example, in the derivation of [safxós] ‘state farm’ from /sovxoz/, the relevant portion of which is schematized in (9):

The underlying form is /sovxoz/, where /v/ is specified as [$ + $voiced] and /x/ has no specification for [$ \pm $voiced], indicated in (9) by $ \varnothing $. By Rule P1b this zero specification is immediately changed to [$ - $voiced]. After this point in the derivation, /x/ is like any other contrastively voiceless segment, and it acts as such to condition regressive voicing assimilation (Rule P3a) on /v/, which devoices to [f]. In sum, although ‘unpaired’ /ʦ, ʧ, x/ are not specified for [$ \pm $voiced] by the branching tree, they are assigned [$ - $voiced] early in the derivation, and subsequently behave like other voiceless segments.

With respect to contrast, then, this derivation demonstrates that the contrastive status of an obstruent with respect to voicing has no bearing on its behaviour with respect to regressive voicing assimilation. Though Halle did not focus on this point, we will see that it will have the effect of making contrastive underspecification a further, perhaps unintended, target of Halle’s argument.

3.2 The taxonomic phoneme

The analysis reviewed in the previous section formed the basis of Halle’s argument against the ‘taxonomic’ phoneme as employed in American structuralist phonology (Halle Reference Halle1957, Reference Halle1959; continued in Chomsky Reference Chomsky1964 and Chomsky & Halle Reference Chomsky and Halle1965; and see also Postal Reference Postal1964, Reference Postal1968).Footnote ⁶ From the 1930s to the 1950s, American linguists who were followers of Leonard Bloomfield attempted to make the definition of the phoneme more precise than it had been. In practice, this amounted to placing an increasing number of restrictions on the relationship between the phonemic level and the phonetic signal. One such restriction concerned ‘phonemic overlapping’ (Bloch Reference Bloch1941): one sound (a phone) may not correspond to two different phonemes in the same context. In Russian, for example, an underlying /ʦ/ can surface as [ʦ], or, when subject to regressive voicing assimilation, as [ʣ]. Because there is no underlying */ʣ/ in Russian, both [ʦ] and [ʣ] are unambiguous allophones of a single phoneme /ʦ/. This situation is permitted.

Consider now an underlying /t/: just like /ʦ/, it can surface as voiceless [t], or as [d] when voiced by regressive voicing assimilation. In this case, however, there is another source for [d] in Russian, namely underlying /d/. In some positions /d/ is in contrast with /t/, but the contrast is neutralized in assimilation environments.Footnote ⁷ Consequently, without additional information from the meaning or morphology, we would not know which phoneme any given [t] or [d] belongs to in neutralization contexts. This type of overlapping was not allowed by most American structuralist phonologists in the 1950s. The condition that rules it out came to be known as ‘bi-uniqueness’; Halle (Reference Halle1959: 21n8) cites Hockett’s (Reference Hockett1951: 340) formulation: ‘For a notation to be phonemic we require a bi-unique, one–one relation rather than a many–one relation’ between representation and utterance.

Halle (Reference Halle1959: 24) observes that bi-uniqueness did not play a role in the practice of Sapir or even of Bloomfield, but was a more recent addition to phonological theory.Footnote ⁸ As Halle (Reference Halle1959: 20–21) puts it, a relatively weak condition on phonological descriptions that is ‘accepted by all’ is Condition (3), which requires that one be able to derive an utterance from a phonological representation. To enforce bi-uniqueness, an additional condition is needed that requires that one be able to derive the (underlying) phonological representation from the utterance. Halle’s (Reference Halle1959: 21) version of this constraint is Condition (3a):

The only solution, in a theory that adheres to Condition (3a) in addition to Condition (3), is to require that every phonetic [t] must be an allophone only of the phoneme /t/, and every [d] must be an allophone only of the phoneme /d/. The fact that t alternates with d must be expressed in a mapping between the morphophonemic (lexical) level and the phonemic level; overlapping is permitted between morphophonemes. We will represent morphophonemes between double slashes, // //.

In a theory that adheres to Condition (3a) in addition to Condition (3), then, rules that change one phoneme into another must belong to the morphophonemic system, while phonemic rules that change phonemes into their various allophones in the appropriate environments apply later. Russian regressive voicing assimilation, though, does both of these: it can change some voiceless phonemes into their voiced counterparts or vice versa (e.g. mapping //t// to /d/ or //d// to /t/), but it can also change an unpaired voiceless phoneme into its voiced allophone (e.g. mapping /ʦ/ to [ʣ]). Adherence to Condition (3a) would preclude a unified account of assimilation, requiring it to be treated as two separate rules operating at two separate levels, as schematized in (10).

Halle (Reference Halle1959) argues that rather than conforming to Condition (3a), eliminating the taxonomic phonemic level, and with it the distinction between morphophonemic and phonemic rules, allows a unified treatment of assimilation and other processes and produces a simpler phonological grammar whose rules map directly from underlying lexical representations (the ‘systematic phonemic level’ of Chomsky Reference Chomsky1964) to surface phonetic forms without any privileged intermediate level. This is schematized in (11).

3.3 The demise of underspecification and branching trees

It is widely understood that Halle’s rejection of the taxonomic phoneme played an important role in setting the course for generative phonology. But another aspect of his analysis of voicing assimilation also had far-reaching consequences that are less generally appreciated – and, we will argue, that were not necessary.

In the derivation in (9), the distinction between contrastive and redundant features is made irrelevant by the rule system: although /x/ is not assigned a value for [$ \pm $voiced] underlyingly because this feature is not needed to distinguish it from the rest of the inventory, it nonetheless receives the value [$ - $voiced] by Rule P1b in time to trigger assimilatory devoicing of /v/. More generally, any unspecified value could be filled in by rule in time to make its underlying absence phonologically inconsequential.

More than inconsequential, the derivations in (9) and (11) make it impossible to draw a line between the contrastive and non-contrastive phonology that would not incur the same problem as the taxonomic phoneme. Trying to limit the phonology to apply only to contrastively specified features – that is, reordering rules P1b and P3a in (8) – would either yield an incorrect result where regressive voicing assimilation does not apply to /v/ in (9) because /x/ has no voicing feature, or would again require regressive voicing assimilation to apply twice, first to contrastively specified features, and then again after non-contrastive features are filled in.

This made the whole notion of contrastive underspecification vulnerable to arguments such as those of Lightner (Reference Lightner1963) and Stanley (Reference Stanley1967), who pointed out that redundancy rules could be exploited to make underspecification for a given feature tantamount to a third value for that feature, increasing the expressive power of the grammar in a way that was argued to be improper (see further Section 6). Underspecification was thus abandoned altogether in SPE, along with the contrastive hierarchies that generated it.

The result was that language-particular feature contrasts did not play a role in the theory of generative grammar that developed from SPE. Rather, the underlying phonemes of a language came to be fully specified according to a universal feature theory; to a first approximation, these representations are not sensitive to language-particular phoneme inventories or contrasts. Where problems arose from this, SPE attempted to address them by introducing a theory of markedness (Chomsky & Halle Reference Chomsky and Halle1968: ch. 9; see Dresher Reference Dresher2009: 104–117 for discussion). Contrastive hierarchies like the one in Figure 1 largely disappeared from phonological theory for a generation. With a few exceptions, they did not return even with the revival of interest in theories of underspecification in the 1980s (e.g. Kiparsky Reference Kiparsky1982; Broselow Reference Broselow1984; Archangeli Reference Archangeli1984, Reference Archangeli1985, Reference Archangeli1988; Pulleyblank Reference Pulleyblank1986; Clements Reference Clements, Blevins and Carter1987; Steriade Reference Steriade1987; Hargus Reference Hargus, Gerdts and Michelson1989).

The consequences for underspecification of Halle’s (Reference Halle1959) treatment of the unpaired obstruents have gone largely unremarked in the literature, particularly in comparison to their consequences for the definition of the phoneme. A recent exception is Reiss (Reference Reiss and Samuels2017: $ \S $2), who argues – in our view rashly and erroneously – that SPR showed contrast to be irrelevant to phonology. His contention is that, because a rule like regressive voicing assimilation (in its unified form) applies in a uniform way without regard for whether its effect is neutralizing or allophonic, phonological rules in general are never affected by whether there is an underlying contrast between two segments. Reiss (Reference Reiss and Samuels2017: 26) formulates the following as a principle called ‘No contrast in phonology’:

‘In plain English,’ Reiss (Reference Reiss and Samuels2017: 26) continues, ‘the phonology doesn’t “care” about contrast, because it has no mechanism by which to do so.’ But SPR includes a mechanism by which the underlying contrasts in the lexicon can in principle affect phonological rules, namely the tree in Figure 1. If featural representations are determined by a contrastive hierarchy, then contrast can be relevant to phonological rules, because rules can refer to features. The SPR rule system happened to be rigged to make contrast irrelevant by filling in redundant features in time for them to behave identically to contrastive ones, but this was not a logically necessary property of the system – rather, it was made necessary by the fact that Halle’s (Reference Halle1959) approach to feature ordering considered underspecification primarily in broad information-theoretic terms, rather than as a potentially explanatory component of phonological theory. In the next section, we explore an alternative approach to feature ordering, and the consequences it could have had for Halle’s analysis of Russian and the course of generative phonology.

4.1 Origins of contrastive feature hierarchies

Although the main criterion for ordering features in SPR is Minimality, this was not the original rationale for the use of contrastive hierarchies in phonology. The notion of contrastive specification by hierarchical ordering of features can be traced back to the Prague School in work by Trubetzkoy (Reference Trubetzkoy1929, Reference Trubetzkoy1939, Reference Trubetzkoy and Baltaxe1969) and Jakobson ([Reference Jakobson1931] 1962). Although they do not use tree diagrams of the sort found in SPR, feature hierarchies are implicit in much of their work, as discussed in some detail by Dresher (Reference Dresher2009, Reference Dresher2015, Reference Dresher2016).Footnote ⁹ A rare example of an explicit mention is Trubetzkoy (Reference Trubetzkoy and Baltaxe1969: 103), who describes the Polabian vowel inventory as involving ‘a certain hierarchy… with respect to the correlation of tongue position and the correlation of lip rounding’, which he diagrams as in (12).Footnote ¹⁰

In other words, the rounding contrast exists only within the category of front vowels. And this statement is an analytical claim, not a theory-neutral observation: given the triplets of non-low vowels /i, y, u/ and /e, ø, o/, it would be just as plausible from the inventory shape alone to say that the front–back contrast exists only within the category of rounded vowels. The difference between the two statements is the difference between the two contrastive hierarchies in (13):

The hierarchy in (13a) groups front vowels together as a natural class, whereas the one in (13b) does not; it groups rounded vowels together instead. Trubetzkoy’s (Reference Trubetzkoy and Baltaxe1969: 102) basis for choosing (13a) for Polabian is that ‘[t]he oppositions between the back and front vowels of the same degree of aperture were constant (non-neutralizable), while the oppositions between rounded and unrounded vowels of the same degree of aperture (ü-i, ö-e) were neutralizable after v and j’.Footnote ¹¹ Given the hierarchy in (13a), /y, ø/ are readily identifiable as the rounded counterparts to /i, e/; in (13b), on the other hand, where /i, e/ have no specification for [$ \pm $back], their ‘counterparts’ are all four of /y, ø, u, o/.Footnote ¹²

Underpinning Trubetzkoy’s use of features is the idea that only some properties of a segment are active, or relevant to the phonology, and these are the distinctive, or contrastive, properties. It is thus phonological activity that determines what the features are and how they are ordered, where activity can be defined as in (14), adapted from Clements (Reference Clements and Alan Hall2001: 77):

Given this definition, we can define an alternative to the Minimality Principle that gives the feature hierarchy a more important role to play in accounting for phonological patterns, and which we believe makes explicit the original Prague School motivation for ordering features:

4.2 Contrast and activity

The original intuition behind feature hierarchies is that there is a connection between phonological activity and contrast (Dresher Reference Dresher2009, Reference Dresher2015). In fact it can be shown that phonological activity still played an important role in choosing the ordering of features in SPR. Recall that Halle’s (Reference Halle1959) analysis requires 272 feature specifications. There are other orderings that use the same features as Halle uses that result in far fewer specifications; however, one has to be willing to accept some strange-looking groupings of segments. Thus, the first feature in Halle’s ordering is [$ \pm $vocalic], which divides all phonemes into [$ - $vocalic] glides and consonants and [$ + $vocalic] vowels and liquids. However, [$ \pm $vocalic] divides the inventory unequally: there are 29 [$ - $vocalic] phonemes in Figure 1, and 14 [$ + $vocalic] vowels and liquids. The next feature, [$ \pm $consonantal], also makes very unequal divisions: in the [$ - $vocalic] domain, it divides between /j/ and the other 28 consonants; in the [$ + $vocalic] sub-tree, it opposes 10 vowels against 4 liquids. Unequal divisions make for relatively inefficient trees, and increase the number of contrastive specifications.Footnote ¹³

Consider the tree in Figure 3. It has 243 specifications, 29 fewer than Halle’s. The top feature, [$ \pm $low tonality], divides the tree 25 against 18. The groupings are not very natural: only two vowels are [$ + $vocalic], and only some consonants are [$ + $consonantal]. The fact that Halle did not consider this ordering is evidence that he did not follow Minimality to the exclusion of Activity. Recall that Condition (5), Halle’s version of Minimality, is restricted to feature orders that satisfy Conditions (3) and (4). Here is Condition (4) in full (Halle Reference Halle1959: 24):

Figure 3 Contrastive hierarchy for Russian phonemes based on Minimality at the expense of Activity.

In other words, the number of feature specifications is reduced to a minimum as long as the resulting representations make sense in terms of the grammar of the language. That is, when assigning contrastive features to segments we are guided by Minimality as long as the result makes sense with respect to Activity.

The connection between contrast and activity is made explicit by what Hall (Reference Hall2007: 20) calls the Contrastivist Hypothesis:

By identifying which features actually serve to mark contrasts in a given language, a contrastive hierarchy can limit which features can be phonologically active in that language. If, contra Halle (Reference Halle1959), redundant features are not filled in during the course of the phonological computation, then only the features identified as contrastive by the hierarchy can be active. This approach predicts that the size and shape of the underlying inventory place limits on how much information the phonological computation has access to; although it was not pursued in SPR or early generative phonology, it has been taken up in more recent work, some of whose results are discussed in Section 5.

Let us, then, revisit the Russian unpaired obstruents with (16) as our working hypothesis and Activity as the guiding principle for constructing feature hierarchies.

4.3 The unpaired obstruents redux

Regressive voicing assimilation shows that the feature [$ - $voiced] is phonologically active on Russian /ʦ, ʧ, x/: these consonants can trigger assimilatory devoicing, even though they do not contrast with underlying voiced counterparts */ʣ, ʤ, ɣ/. Under the Contrastivist Hypothesis, this is only possible if [$ - $voiced] serves to distinguish the unpaired obstruents from some other phonemes in the contrastive hierarchy. This is not the case in Halle’s (Reference Halle1959) tree, as we have seen in Figure 2, but it can be the case if [$ \pm $voiced] is given wider scope in the hierarchy. This can be accomplished with a minimal departure from Halle’s ordering of features, by making [$ \pm $voiced] take scope over [$ \pm $continuant], which yields the specifications shown in Figure 4.Footnote ¹⁴

Figure 4 Revision of Figure 2 with [$ \pm $voiced] taking scope over [$ \pm $continuant].

The feature [$ \pm $voiced] is now specified on the three circled segments, /ʦ, ʧ, x/, in each case because it marks some phonemic contrast: it distinguishes /ʦ, s, s^j/ from /z, z^j/, /ʧ, ʃ/ from /ʒ/, and /k, k^j, x/ from /ɡ/. But changing the order of features has other consequences for the specifications; a distinction that has already been marked by [$ \pm $voiced] cannot also be marked by another feature. In this case, the trade-off is that the voiced consonants /z, z^j, ʒ, ɡ/, boxed in Figure 4, are now unspecified for [$ \pm $continuant].

Is this new underspecification consistent with the facts of Russian phonology, or would the hierarchy in Figure 4 require redundancy rules to insert [$ \pm $continuant] in the same way that the hierarchy in Figure 2 requires rules to insert [$ \pm $voiced]? In fact, as discussed by Hall & Dresher (Reference Hall, Elan Dresher, Hammerly and Prickett2016), this ordering of features conforms strikingly well to Russian sound patterns. Phonetically, /ɡ/ is realized in some southern Russian dialects as a fricative [ɣ] or [ɦ]; this variability suggests that [$ - $continuant] may not be a crucial part of the phonological specification of this segment.

There is also morphophonological evidence from alternations arising from the historical First Velar Palatalization.Footnote ¹⁵ In this pattern, velar consonants ([$ + $low tonality] in Halle’s (Reference Halle1959) feature system) alternate with their postalveolar ([$ - $low tonality]) counterparts. As expected, these alternations pair /x/ with /ʃ/ and /k/ with /ʧ/; /ɡ/ alternates with /ʒ/, which would be surprising if these two segments were specified for [$ \pm $continuant], but can be elegantly characterized as part of the same pattern if they are not.Footnote ¹⁶ The alternating pairs are shown in (17) with their relevant feature specifications.

Some examples of the alternations are shown in (18); the data are from Lightner (Reference Lightner1965).

Other alternations, historically arising from the Second Velar Palatalization, pair /k/ with /ʦ/ and /ɡ/ with /z/.Footnote ¹⁷ Again, these pairings are unremarkable given the specifications in Figure 4, but would be harder to account for in a unified way if /ɡ/ and /z/ were specified for [$ \pm $continuant]. In Halle’s (Reference Halle1959) feature system, the Second Velar Palatalization alternations involve changing more features than the First Velar Palatalization ones do, but they can still readily be captured using the specifications in Figure 4. The alternating pairs are shown in (19) with their relevant features.

In sum, reordering [$ \pm $voiced] and [$ \pm $continuant] in the SPR feature hierarchy improves the empirical status of both features: in the former case by extending [$ - $voiced] to the ‘unpaired’ /ʦ, ʧ, x/, and in the latter case by leaving /z, z^j, ʒ, ɡ/ unspecified for [$ \pm $continuant].Footnote ¹⁸

4.4 Consequences of the revised analysis

This analysis suggests a picture of phonological levels that is different both from the post-Bloomfieldian (10) and from the one in (11) advocated by Halle (Reference Halle1959) and Chomsky & Halle (Reference Chomsky and Halle1968). This model is shown in (21). Underlying lexical representations correspond to the systematic phonemic level of Chomsky and Halle, but with the proviso that they are made up only of contrastive features. These representations serve as input to the contrastive phonology, which, subject to the Contrastivist Hypothesis, computes only contrastive features. The representations which we designate here between straight brackets $ \mid \quad \mid $ are the output of the contrastive phonology.Footnote ¹⁹ These representations serve as the input to additional levels not governed by the Contrastivist Hypothesis, which are the domain of enhancement and other phonetic processes that add or modify contrastive as well as non-contrastive features.Footnote ²⁰

The Halle–Chomsky arguments against the taxonomic phoneme, defined by a series of conditions as discussed in Chomsky (Reference Chomsky1964), still go through, but the duplication problem raised by regressive voicing assimilation disappears. This is because the rule applies one time to segments bearing a contrastive specification of [$ \pm $voiced], whether the result is an already existing phoneme or a new allophone of a phoneme. While the phonemic/allophonic distinction does not mark out a special level, the difference between contrastive and non-contrastive phonology does. Unlike the taxonomic phoneme, however, there are no constraints against overlapping or any other constraints on the relation between this level and phonetic representations. Therefore, all the arguments against the taxonomic phoneme would have still been the same.

What is different is that between underlying lexical representations (now simply called phonemic) and the surface phonetic level (which has never been well defined) there is an intermediate level. This level is not the old taxonomic phoneme, but is the output of the contrastive phonology, where only contrastive features are in play, and the input to enhancement and other processes that add non-contrastive information (Dyck Reference Dyck1995; Hall Reference Hall2011).Footnote ²¹

5.1 Cross-linguistic variation

Because the ordering of features is not universal, different languages with similar inventories, or different stages of the same language, may assign different featural representations to phonetically similar phonemes. Here we will briefly review two such cases: laryngeal harmony in two Chadic languages, and labial harmony in a Tungusic and a Mongolic language. In each case there are subtle differences as to which segments participate in the relevant harmonies that can be accounted for by positing a difference in the feature hierarchies of segment inventories that are very similar in the relevant respects.

5.1.1 Laryngeal harmony in Chadic

Mackenzie (Reference Mackenzie2009, Reference Mackenzie2013) discusses the case of laryngeal harmony in Ngizim and Hausa, two Chadic languages. Each language has a three-way phonemic laryngeal contrast among obstruent stops, distinguishing plain voiced stops, plain voiceless stops, and implosives. Using the features [$ \pm $voice] and [$ \pm $constricted glottis], there are two possible contrastive hierarchies that can be applied to such an inventory. If [$ \pm $constricted glottis] takes wider scope, as in (22a), then the implosives will be unspecified for [$ \pm $voice]; if [$ \pm $voice] takes wider scope, as in (22b), then the voiceless stops will have no specification for [$ \pm $constricted glottis].

Mackenzie (Reference Mackenzie2013) argues that (22a) is the correct hierarchy for Ngizim, which requires pulmonic obstruents to agree in voicing, but allows implosives to co-occur with either voiced or voiceless obstruents, as illustrated in (23):

In Hausa, on the other hand, the hierarchy in (22b) appears to be the correct one. Hausa disallows a mismatch in [$ \pm $constricted glottis] on homorganic voiced obstruents, as in (24a), but a voiceless pulmonic consonant can occur with an implosive, as in (24b).

As Mackenzie shows, the two different patterns of underspecification in (22) predict exactly the two different patterns of harmony and transparency found in Ngizim and Hausa. In Ngizim, with the hierarchy in (22a), implosives are outside the scope of the harmonizing feature [$ \pm $voice], while in Hausa voiceless stops are outside the scope of harmonizing [$ \pm $constricted glottis].

5.1.2 Labial harmony in Tungusic and Mongolic

The Ngizim and Hausa obstruent stop systems are a kind of ‘minimal pair’: they have very similar phonemes and make use of similar features, but differ in how they pattern. In this case, the difference can be attributed to a difference in how the features are ordered. Van der Hulst & Smith (Reference Van der Hulst, Smith, Coopmans and Hulk1988) called attention to another such ‘minimal pair’: Tungusic and Mongolic languages have similar vowel inventories and similar types of labial (rounding) harmony; however, in Tungusic /i/ blocks labial harmony, and in Mongolic /i/ is transparent to labial harmony. Ko (Reference Ko, Cathcart, Chen, Finley, Kang, Sandy and Stickles2013, Reference Ko2018) proposes that this difference is also due to feature ordering. For Tungusic, here exemplified by Oroqen, he adopts the ordering of Zhang (Reference Zhang1996), [low] $ \gg $ [front] $ \gg $ [round] $ \gg $ [RTR], as in (25).Footnote ²² He proposes that Mongolic, represented by Khalkha, uses the same features, but ordered [front] $ \gg $[low], as in (26).

On this analysis, the key difference between Oroqen and Khalka /i/ is that Oroqen /i/ is specified [$ - $low, $ + $front] and Khalkha /i/ is specified as only [$ + $front]. Ko (Reference Ko, Cathcart, Chen, Finley, Kang, Sandy and Stickles2013) observes that labial harmony in Tungusic and Mongolic is stratified by height, operating only in the domain of vowels that are [$ + $low]. It is thus reasonable to suppose that adjacency in labial harmony is relativized to vowels that bear a feature for [$ \pm $low]. Thus, labial harmony in both Oroqen and Khalkha is blocked by the [$ - $low] specification of /ʊ/ and /u/. In Oroqen, it is also blocked by the [$ - $low] specification of /i/; an example is shown in (27). Khalkha /i/ lacks a [$ \pm $low] feature, and is thus transparent to labial harmony (28).

5.2 Diachronic change

The same kind of variation in feature ordering can also be seen in diachrony, as different stages of the same language may have different contrastive hierarchies. A growing literature has made use of contrastive hierarchies in diachronic studies (see Dresher, Harvey & Oxford Reference Dresher, Harvey, Oxford, Hyman and Plank2018 and the references therein).

A recurring idea that has been applied in a number of studies is that diachronic mergers tend to involve segments that have similar contrastive specifications. This idea has been formulated more specifically by Ko (Reference Ko, Maezawa and Yokogoshi2010, Reference Ko2018) as the Minimal Contrast Principle, given in (29a); Oxford (Reference Oxford2015) reformulates it as the Sisterhood Merger Hypothesis (29b).

Here we will briefly review two such studies: Oxford’s (Reference Oxford2012, Reference Oxford2015) account of a contrast shift in the vowel systems of some Algonquian languages, and Krekoski’s (Reference Krekoski2017) study of the evolution of Mandarin tone systems from Middle Chinese.

5.2.1 Contrast shift in Algonquian vowel systems

Based on what is known about the vowel system of Proto-Algonquian, Oxford (Reference Oxford2012, Reference Oxford2015) posits that it had the contrastive feature hierarchy shown in (30).

The hierarchy in (30) is motivated by feature activity that can be recovered as having been present in Proto-Algonquian. Thus, */o/ triggers rounding, an indication that it has an active, hence contrastive, [$ + $round] feature. Similarly, */i/ triggers palatalization, indicating a contrastive feature we call [$ + $front]. Patterns of partial neutralization relate */ɛ/ and */i/, suggesting that they are contrastive sisters by (29). Finally, */a/ does not trigger any processes, consistent with its being assigned no positive (marked) contrastive features.Footnote ²³

The Proto-Algonquian vowel feature hierarchy continues unchanged in the Central Algonquian languages and in Blackfoot (see Oxford Reference Oxford2015: 319 for details). It accounts for two recurring patterns: (a) palatalization always includes */i/ as a trigger; and (b) */ɛ/ regularly merges with */i/.

On the eastern and western edges of the Algonquian area, the vowels derived from Proto-Algonquian */o/ and */i/ began to pattern together as high vowels, contrary to the predictions of the Proto-Algonquian hierarchy (30), in which */o/ and */i/ share no vowel-specific features in common. Oxford posits that in these areas the feature hierarchy shifted: the feature [low] was reinterpreted as [high] and moved to the top of the vowel feature hierarchy, as shown in (31).

Subsequent developments in the eastern and western daughter languages follow the predictions of the new hierarchy (see Oxford Reference Oxford2015: 319 for details). The patterns consistently differ from those of Central Algonquian: (a) palatalization in these languages is triggered by */ɛ/ but excludes */i/; and (b) */ɛ/ merges with or shifts to */a/ (not */i/).

Both of these patterns provide striking support for Oxford’s analysis. In the hierarchy of (31), */ɛ/’s contrastive sister is */a/, which accounts for the new merger patterns. Perhaps more striking is the restriction of palatalization to */ɛ/, which is unexpected in light of the general cross-linguistic tendency for /i/ to be the preferred trigger of palatalization (Kochetov Reference Kochetov, van Oostendorp, Ewen, Hume and Rice2011), as is the case in Proto- and Central Algonquian. In those languages, both */i/ and */ɛ/ have the feature [+front] (or [+coronal]), which tends to be associated with vowels that trigger palatalization; */ɛ/ also has the feature [$ + $low], which tends to inhibit its ability to trigger palatalization. However, in the shifted hierarchy of (31), only */ɛ/ is [$ + $front], leaving it as the only potential trigger of palatalization. In sum, a single contrast shift accounts for the patterning of a large number of phonological changes across the Algonquian family.

5.2.2 Contrastive diachronic stability and change in Mandarin tones

Krekoski (Reference Krekoski2017) investigates the tone systems of a number of languages that descend from Middle Chinese, focusing on the patterns of activity that they display in the form of tone sandhi. He assumes that, where possible, tones related by a sandhi rule differ minimally, that is by only one feature (this is the synchronic counterpart of the diachronic principle in (29)). Paying attention to the contrastive relations among tones, rather than to their phonetics, reveals a surprising degree of stability in the contrastive structures of tone systems that descend from Middle Chinese, despite significant changes over time in the phonetic realizations of individual tones.

For example, Beijing Mandarin has the four contrastive tones in (32), which participate in two robust sandhi rules (33). Rule (33a), known as ‘third tone sandhi’, is obligatory, and states that a 214 tone becomes 35 before another 214 tone. This rule suggests that tone /214/ differs by one feature from tone /35/. The rule in (33b) is optional, and states that a lexical 35 tone becomes 55 when it immediately follows either 35 or 55, and precedes any other tone. This rule suggests that /35/ differs by one feature from /55/.

A tree satisfying these relations is given in (34). Krekoski (Reference Krekoski2017) then names the features ex post facto according to their phonetic correlates. The Roman numerals under each tone indicate the Middle Chinese tones from which the Beijing tones derive. (Middle Chinese tone I split into two tones designated Ia and Ib.)

Following the same procedure for the Jin language Pingyao, Krekoski arrives at a hierarchy that is isomorphic to the one in (34), but whose features have different phonetic content. The tones in these languages have kept their places in the phonological system despite extensive changes to their phonetic realizations.

There is a limit, however, to the phonetic changes that the system of oppositions can withstand. Following the same methodology as for Beijing and Pingyao, Krekoski (Reference Krekoski2017) posits the tree in (35a) for the Mandarin language Tianjin. Surprisingly, these tones do not correspond as expected with their cognates in Beijing and Pingyao. Tones /21/ and /53/ are in the ‘wrong place’ relative to the other dialects that descend from Middle Chinese.

Why did a contrastive shift occur in the history of Tianjin? An answer can be found by considering what the contrastive feature tree of Tianjin would look like if the tones had maintained their relative positions from Middle Chinese (35b). Krekoski (Reference Krekoski2017: 66) observes that it is difficult to find plausible phonetic correlates for the features [$ \pm \alpha $] and [$ \pm \beta $] in this system, concluding that

Tonal drift likely accreted changes in height values until the system of contrasts reached some critical inflection point which precipitated the reanalysis of specifications. If one conceptualizes systemic change as occurring with subsequent generations of learners formulating new hypotheses about a phonological system from available data, a contrast shift such as this represents a new hypothesis of a system that is more phonetically natural and possesses a lower complexity than the preceding one.

5.3 Reduction and neutralization

Contrastive hierarchies also offer insight into patterns of reduction (Spahr Reference Spahr2014; Nevins Reference Nevins2015). For example, Spahr (Reference Spahr2014: $ \S $2.1) shows how the hierarchical organization of Bulgarian vowels in (36) identifies pairs neutralized in reduction.Footnote ²⁴

In unstressed syllables, vowel reduction neutralizes contrasts at the lowest levels in the hierarchy. All dialects and registers reduce unstressed /ʌ/ and /a/ to [ə]; some combinations of dialect and register additionally reduce /o/ and /u/ to [ʊ]; and some also reduce /e/ and /i/ to [ɪ]. As Nevins (Reference Nevins2015: 51) puts it, ‘vowel reduction is a process of pruning the lowest node in the hierarchy’. Eliminating [$ \pm $low] from /ʌ/ and /a/ produces a vowel [ə] that is [$ - $front], [$ - $round], and indeterminate in height. Similarly, pruning [$ \pm $high] from /o/ and /u/ or /e/ and /i/ produces vowels that are no longer distinctly high or mid phonologically, and are intermediate in height between [o, e] and [u, i] phonetically, but which retain their other contrastive properties. Reducing all three pairs yields the truncated surface system in (37).Footnote ²⁵

Nevins (Reference Nevins2015) takes a similar approach to the reduction of vowel contrasts in whistled speech, and Rischel ([Reference Rischel, Grønnum, Gregersen and Basbøll1966] 2009) uses a feature hierarchy to account for an otherwise puzzling reduction in distinctions connected with the Scandinavian runic reform.

This approach is consistent with the notion that reduction involves the elimination of formal structural complexity (not just articulatory effort), and offers an additional kind of evidence that can be brought to bear in identifying the feature hierarchy of any particular language. In doing so, it is important to keep in mind that phonetic neutralization does not always need to eliminate all surface evidence of the underlying featural contrasts to which it applies. For example, Compton & Dresher (Reference Compton and Elan Dresher2011) discuss Inuit varieties in which ‘strong i’ triggers palatalization and ‘weak i’, historically descended from an earlier schwa, does not. Although both vowels surface as [i] phonetically, the underlying distinction between them is seen in their different effects on adjacent consonants. This neutralization applies too late in the computation to bleed palatalization, whereas a reduction pruning the contrastive hierarchy earlier on would cause segments to behave alike rather than merely sound alike.

The pruned hierarchies in these cases of neutralization are epiphenomenal: the underlying segments are specified with all the features assigned to them by the full contrastive hierarchy for the language. Neutralizing the lowest features in the hierarchy produces a reduced surface inventory in certain contexts, but the underlying system of hierarchical contrasts remains constant. Another type of reduced inventory can be found in morphologically defined contexts, as in the case of many Romance languages, where inflectional suffixes (desinences) use a reduced inventory of vowels as compared with roots. Dyck (Reference Dyck1995) and Frigeni (Reference Frigeni2002) have proposed analyses of such cases that involve separate contrastive hierarchies for the desinential domain. Here, different hierarchies are posited underlyingly, not just at the surface.

5.4 Phonetic enhancement and the shapes of inventories

Another benefit of the contrastive hierarchy approach, as Hall (Reference Hall2011) argues, is what it can contribute to our understanding of cross-linguistic patterns in inventories. It has long been known that phonemes tend to be spread out through the available phonetic space – for example, a vowel system is unlikely to have /y/ contrasting with either /i/ or /u/ unless it also has /i/ and /u/ in contrast with each other. From Liljencrants & Lindblom (Reference Liljencrants and Lindblom1972) through more recent work such as Kirchner (Reference Kirchner1997), Flemming (Reference Flemming2002), and Padgett (Reference Padgett2003), Dispersion Theory has accounted for this pattern by positing mechanisms that explicitly enforce auditory distancing between sounds. What Hall (Reference Compton and Elan Dresher2011) shows is that the contrastive hierarchy, combined with the independently motivated process of phonetic enhancement (Stevens, Keyser & Kawasaki Reference Stevens, Keyser, Kawasaki, Perkell and Klatt1986 and subsequent work), generates dispersedness automatically, with no need for constraints that evaluate auditory distance.

For example, consider the vowel inventories in Figure 5. The widely attested three-vowel inventory in Figure 5a is spread out through the phonetic space. The unattested inventory in Figure 5b, on the other hand, is tightly packed into a small region of the available space, with very little acoustic difference between any two vowels.

Figure 5 Three-vowel inventories.

If the vowels of Figure 5b are assigned contrastive feature specifications in a hierarchical ordering, the features can only mark how they differ; they cannot specify all the things these vowels have in common. The only two features that can be assigned here are [$ \pm $high] and [$ \pm $round] (or features with different names that mark the same distinctions, such as Jakobson et al.’s (Reference Jakobson, Gunnar, Fant and Halle1952) [compact/diffuse] and [flat/plain]). These may take scope in either order, producing either of the hierarchies in (38).

In either case, the features assigned to these vowels are ones that could equally represent the vowels of Figure 5a: /a/, /i/, and /u/. While it is possible to assign specifications to /a, i, u/ that could not correspond to /ɘ, ɨ, ʉ/, the reverse is not true.

Enhancement, in turn, means that a vowel inventory specified as in (38) is more likely to surface phonetically as [a, i, u] than as [ɘ, ɨ, ʉ]. As noted in Stevens et al. (Reference Stevens, Keyser, Kawasaki, Perkell and Klatt1986) and subsequent work, phonetic properties that are not phonologically contrastive tend to be deployed in such a way as to reinforce the acoustic effects of contrastive features. For example, in (38) /ɘ/ is specified as [$ - $high]. The elevated F1 frequency associated with [$ - $high] can be enhanced by making the vowel [$ + $low], so enhancement favours realizing /ɘ/ as [a]. Similarly, /ɨ/ is [$ - $round], which causes a higher F2 that can be enhanced by realizing the vowel as front [i], while the lower F2 associated with the [$ + $round] feature of /ʉ/ can be reinforced by making the vowel back [u].

The features of any individual segment can be enhanced without reference to the rest of the inventory. Because contrastive feature specifications encode differences, enhancing them makes the inventory as a whole more phonetically dispersed, with no need for constraints that explicitly evaluate acoustic distances between segments. As Hall (Reference Hall2011) shows, the combination of contrastive specification in the phonology and enhancement in the phonetics accounts for the typological patterns that Dispersion Theory was designed to explain.

6.1 Universality

The approach to contrastive feature hierarchies that we presented in Section 5 assumes that feature hierarchies are language particular. This was true of Trubetzkoy’s (Reference Trubetzkoy1939) theory, but through the 1950s and 1960s there was a movement, beginning with Jakobson (see Toman Reference Toman1995: 119; Battistella Reference Battistella, Dresher and van der Hulstto appear), to put as much as possible of feature theory on a universal basis, and SPE gave universalism a central place in generative phonology. By the time the notion of organizing features hierarchically returned with the introduction of autosegmental feature geometry, the hypothesis of universal features was well established, as reflected in works such as Clements (Reference Clements1985), Sagey (Reference Sagey1990), Clements & Hume (Reference Clements, Hume and Goldsmith1995), and Halle, Vaux & Wolfe (Reference Halle, Vaux and Wolfe2000). In the context of this broader trend toward universalism, an attempt to entrench language-particular feature hierarchies might have had difficulty gaining traction in the first decades of generative phonology.

6.2 Syntagmatic sources of underspecification in SPR

The branching tree shown in Figure 1 is not the only source of unspecified feature values in SPR. All languages have constraints on sequences of phonemes, and these constraints give rise to additional feature redundancies that could serve as a further source of underspecification, particularly if Minimality is the guiding principle. Thus, the features left unspecified by the tree in Figure 1 (what we could call paradigmatically underspecified features) are supplemented in SPR by many more features that are lexically unspecified because of sequence structure constraints (syntagmatically underspecified features) that Halle (Reference Halle1959: 56–62) calls morpheme structure (MS) rules.

Whereas paradigmatic underspecification is entirely determined by the branching tree (that is, by the contrastive feature hierarchy), syntagmatic underspecification is much less well defined. As Halle (Reference Halle1959: 57) admits, ‘It is not possible to give a complete procedure for discovering the most economical representation in every case. The best that can be done is to formulate the sequential constraints as rules specifying certain contexts’ (emphasis in original). SPR lists a series of some 30 MS rules numbered 1a–11e. A sample of a few of these, shown in (39), illustrates how context-specific they are (in part because Halle made no recourse to syllable structure). In these rules, ‘&’ represents a boundary, ‘V’ stands for any vowel, ‘R’ for a liquid, and ‘C’ for any consonant. These rules apply in a block prior to the phonological (P-) rules.

Morpheme structure conditions have been the subject of much discussion since the writing of SPR (see Dresher & Nevins Reference Dresher and Nevins2017 for a recent treatment); the issue here is whether underspecification is the best way to account for them. The arbitrariness and incompleteness of syntagmatic underspecification – unlike paradigmatic underspecification, which is governed by an explicit and complete procedure (the branching tree) – suggests to us that it is not. However, the existence of syntagmatic underspecification made it problematic to propose a model like the one in (21), where non-contrastive features are filled in after the contrastive phonology. Moreover, syntagmatic underspecification posed difficulties that made contrastive underspecification more vulnerable to the criticisms of Lightner (Reference Lightner1963) and particularly Stanley (Reference Stanley1967), to which we now turn.

6.3 Stanley and the critique of underspecification

While Stanley (Reference Stanley1967) devotes one section to a critique of ‘branching diagrams’ (contrastive feature hierarchies) and another to the Distinctness Condition, much of his discussion is aimed at Halle’s (Reference Halle1959) morpheme structure rules, which he finds (Stanley Reference Stanley1967: 435) have ‘a certain kind of arbitrariness’ that cannot be remedied. With respect to the branching tree and paradigmatic underspecification, Stanley makes a number of technical arguments that we cannot review here, but which are discussed in detail by Dresher (Reference Dresher2009: 96–102). Here we would like to highlight one fundamental assumption that colours and vitiates Stanley’s entire discussion of underspecification, and this is his understanding of what an underspecified feature (what he calls a ‘blank’ or ‘0’ entry in a feature matrix) represents.

As should be clear from our discussion of the connection between contrast and activity and from every case reviewed in Section 5, we understand an underspecified feature to be phonologically different from both the $ + $ and $ - $ values of that feature, irrespective of what its phonetic interpretation turns out to be. For example, the implosive /ɗ/ in Ngizim is unspecified for [$ \pm $voice], and is treated by the phonology as not having a voiced feature; hence, it is exempt from the co-occurrence restriction that affects /t, d/. The phonetically similar /ɗ/ in Hausa is specified differently and consequently cannot co-occur with /d/. That is, contrastive underspecification has phonological significance because the logic of contrastive underspecification is a ternary logic.

This is not how Stanley understands underspecification, however. For him, a ‘0’ in a feature matrix is merely an unexpressed $ + $ or $ - $, as the case may be: ‘we must take care that the feature values remain binary, and that “0” in a dictionary matrix is never allowed to function as a third value, distinct from “$ + $” and “$ - $”’ (Stanley Reference Stanley1967: 410; emphasis in original). Therefore, the phonology must not be allowed to take advantage of a 0 value to achieve a simplification or generalization of a rule. For example, he would consider it improper to describe Ngizim as disallowing segments with differing values of [$ \pm $voice] to co-occur in a word without further stipulating that laryngeal harmony is restricted to segments specified [$ - $constricted glottis]; this would be ‘a specious simplification, an improper use of blanks’ (Stanley Reference Stanley1967: 416).

As Dresher (Reference Dresher2009: 98) has remarked, ‘What Stanley considered specious was later considered a result!’ That Hausa implosives are unspecified for [$ \pm $voice] does not just allow us to simplify the statement of voicing harmony, it explains why they do not participate in it. More generally, the Contrastivist Hypothesis does not make sense unless we assume that unspecified values have phonological significance and cannot be manipulated by the phonology. Nevertheless, Stanley’s interpretation of blanks appears to have been widely shared, and was not contested until Kiparsky (Reference Kiparsky1982) reintroduced underspecification into generative phonological theory (Dresher Reference Dresher2009: 117–121).