Chapter 7

On the Major Class Features Elisabeth Selkirk

and Syllable Theory

1. Major Class Features in a Theory of the Syllable

Developments in the theory of phonological representation have pro¬

gressively chipped away at the set of distinctive features presented in The
Sound Pattern of English (SPE). The new understanding of the nature of
stress patterns and their representation that has been gained in metrical
phonology has meant the elimination of the feature [stress] from that
repertoire. The autosegmental theory of tone has made it possible to do
without contour tone features. Given the autosegmental theory of the
syllable, it is also possible to do without features relating to the im¬
plementation of segments in time: [ + long] segments may now be viewed as
single segments associated with two terminal positions in syllable structure,
[ +delayed release] segments (affricates) as a sequence of two segments
associated with a single position m syllable structure, and so on. It has
also been suggested that the major class feature [ +syllabic] might be
eliminated, given that syllable structure forms part of a phonological
representation,^ though the consequences of eliminating that feature have
not been fully explored, as we shall see. In each instance, an enrichment of
the theory of representation has meant a reduction in the need for certain
features in the representation of distinctions between particular forms. In
this paper I will explore yet further the consequences for distinctive feature
theory of the theory of hierarchical representation in phonology, and the
theory of the syllable in particular.
I will present evidence pointing to the conclusion that all the major
class features—[ +syllabic], [± consonantal], and [±sonorant]—should
be eliminated from phonological theory. Specifically, I will show that
characterizing segments in terms of these features is an obstacle to a
descriptively adequate account of syllable structure in language, and thus
Elisabeth Selkirk 108

that in a truly explanatory theory of syllable phonotactics they must be

given no role.
The major class features that have been standardly assumed since SPE
characterize as follows the natural classes of segments listed on the left in
(1):
(1)
[syllabic] [sonorant] [consonantal]

Glides —
+ —

Vowels + + —

Sonorants (consonants) — + +
Syllabic sonorants + + +
Obstruents — — +
Syllabic obstruents + — +

(The feature complexes of the last two lines characterize no classes of

segments, presumably because the assignments [ — son] and [ — cons] are
somehow “contradictory” and therefore universally impossible in com¬
bination.) Examples of segments belonging to these classes are listed in (2):

(2)
a. Glides: j, w, ^
b. Vowels: i, u, y, ui, e, o, 8, ae, a, etc.
c. Sonorants: m, n, g, etc. (nasals); 1, r, etc. (liquids)
d. Syllabic sonorants: m, n, g, etc. (nasals); 1, r, etc. (liquids)

e. Obstruents: s, z, f, etc. (fricatives); b, p, t, etc. (stops)

f. Syllabic Obstruents: s,??2

(Throughout this paper, the reader should note that by using the terms
glide and vowel, I am not committing myself to their theoretical validity.)
In the taxonomy provided in (1) and exemplified in (2), the feature
[± syllabic] suggests itself as an especially obvious candidate for elimina¬
tion from phonological theory. Given that segments are organized into
syllable structure, but are independent of it, if “syllabicity” is to be repre¬
sented with a feature, that feature has the peculiar property of being
syntagmatic: whether a segment is “syllabic” depends on its position in a
syllable, not on any inherent phonological property of its own. Every
Major Class Features 109

sonorant consonant has its syllabic counterpart, every glide has its com¬
panion vowel, and 5 (and perhaps others) may stand alone as syllabic.^ It is
not clear that anything is lost by eliminating [± syllabic] from the feature
repertoire. On the contrary, it would seem that a great deal is gained. If
this feature is eliminated, then the property of being “syllabic” can be seen
simply as the property of having a particular place in syllable structure, or,
more exactly, a particular relation to other elements in the same syllable. I
will argue below that this is the correct interpretation of what it means for
an element to be “syllabic.”
If [ ± syllabic] were indeed eliminated from the repertoire of major class
features, then the natural classes defined by the remaining ones would
simply be as follows;

(3)
Vowels (2a,b): [-1-son,—cons]
Sonorants (2c,d): [+son, + cons]
Obstruents (2e,f); [—son, + cons]

There would be no class of glides to be opposed to vowels, and there

would be no distinction between syllabic and nonsyllabic sonorants and
obstruents. Yet the natural classes so defined are unable to provide the
basis for certain generalizations concerning possible syllable structures
that must be made in language. Consider statement (4), the likes of which is
needed in the phonotactic description of many languages:

(4)
The onset of a syllable in L may be occupied by any consonant or glide of L.

If [± syllabic] is included among the distinctive features, such a common¬

place restriction can be stated quite simply.

(40
The onset of a syllable in L may be occupied by any [ — syll] segment of L.

But without [± syllabic], the restriction must be stated as a disjunction, as

in (4"):

(4")
The onset of a syllable in L may be occupied by any segment of L that is
either [+cons] or [ — cons, -l-high].

(Without [± syllabic], glides and high vowels have the same feature com¬
plex [-cons, Thigh].) Or consider the equally commonplace phonotactic
statement (5):
Elisabeth Selkirk no

(5)
The rime of a syllable in L may end in either a glide or a nasal.

With [± syllabic] in the feature repertoire, such a restriction can be ex¬

pressed quite simply:

(5')
The rime of a syllable in L may end in a [ —syll, +son] segment of

Without [± syllabic], it must be stated as follows:

(5")
The rime of a syllable in L may end in a [ + son, — cons, + high] or a [ + son,
+ cons] segment of L.

The dilemma, then, is this: the feature [± syllabic] appears to be necessary

to a straightforward characterization of (natural) classes of segments that
play a role in phonotactic descriptions of the syllable, but at the same time
appears to be rendered unnecessary by the mere existence of syllable
structure as part of phonological representation.
There are a number of possible responses to this dilemma. One is based
on the view that it is entirely appropriate to cast generalizations about the
natural classes involved in phonotactic description in terms of complexes
of binary distinctive features. It would involve eliminating [± syllabic],
for the reasons given, and hypothesizing some other binary major class
feature(s) that would permit a simple characterization of these natural
classes.^ Another response is based on the view that the problems en¬
countered by a theory using only the major class features [ ± sonorant] and
[± consonantal] to characterize the natural classes of phonotactic descrip¬
tion are symptomatic of a more general problem with the theory, and that
the natural classes involved must be characterized in some entirely different
way. In this paper, I will offer a response of the second sort. My proposal is
that the major class features be eliminated entirely from a theory of the
phonotactics of the syllable (and, perhaps, from phonological theory as a
whole), and that they be replaced in effect by the sonority hierarchy and the
assignment of a sonority index to individual segments that reflects the niche
they occupy in that hierarchy. In other words, I propose that there is a
single n-ary feature, call it [n sonority], that is at play in language, where the
feature specification n is the sonority index. In what follows I will show that
the notion of natural class that is required for an insightful expression of
phonotactic generalizations must be cast in terms of sonority indices, and
not in terms of complexes of binary distinctive features.
Major Class Features 111

The point to be made about the major class features is a bit different,
then, from the one made earlier about the features [stress], [long], etc. The
“work” done by the latter features in earlier phonological descriptions is
now done by the hierarchical representation itself:/“stress” is the alignment
of syllables with the metrical grid,® “length” is the association of a single
segment with two positions in syllable structure, etc. My proposal is not
that the “work” of the major class features be done by any aspect of the
hierarchical representation. Rather, I am suggesting that an understanding
of that hierarchical representation, and of the the theory required for
describing it, simply shows that their “work” must be done in a different
way, by something else. That something else, I submit, is what may be
thought of as a feature representing the phonetic dimension of sonority,
the sonority hierarchy, and the assignment of a sonority index to every
segment of the language.^
In the general case, any segment of a language may be more or less
sonorous than any other, so that a continuum may be estab¬
lished, wherein is the least sonorous segment type and a„ the most
sonorous. The subscript integer i is the sonority index of the segment.
Moreover, it seems that members of certain natural classes of segments,
defined in terms of nonmajor class features such as [± continuant],
[±voice], [± nasal], [±high], etc., are so alike in sonority as to make
distinctions among them irrelevant for most descriptive purposes. For
example, the nasal consonants appear to pattern alike, as do the high
vowels or the class of voiceless stops, when it is degree of sonority that is at
issue in phonological description The members of these classes, and some
others, will therefore be assigned the same sonority index.
A new definition of natural class is available in terms of this sonority
continuum, or hierarchy. Any set of segments with the same sonority index
or with consecutive sonority indices within designated limits forms a natural
class from this point of view. The discussion here will show that it is natural
classes defined in just these terms that appear to be at play in phonotactic
description.
I will not offer a definition of sonority here. There is clearly a phonetic
basis for it, probably corresponding in part to simple “loudness.” But
just what the relevant acoustic parameter is cannot be determined in¬
dependently of linguistic analysis. Just what the natural classes of segments
are is an empirical question, whether they are defined in terms of the n-ary
feature for the sonority dimension or in terms of features for place and
manner of articulation, for example. And only once phonology has pro-
Elisabeth Selkirk 112

vided sufficient information about the hierarchy can the precise phonetic
character of sonority be determined.
A number of proposals have been made concerning the sonority hier¬
archy,® based on various sorts of evidence, including the place segments
may occupy (with respect to each other) in syllable structure. In (6) I
suggest a provisional version of the hierarchy, to be used as a working
hypothesis. What I will say below bears only on the relations between the
sound types represented in (6); just where sounds that are not represented
in (6) are to be introduced into the hierarchy will be left an open question.

(6)
Sound Sonority index
(provisional assignment)
a 10
e, o 9
i, u 8
r 7
1 6
m, n 5
s 4
V, z, 6 3
f, 0 2
b, d, g 1
p, t, k .5

The right-hand column lists the hypothesized sonority indices of the seg¬
ments on the left. It is not clear whether the absolute integer value of the
sonority indices assigned to each of these segment types is important. I
assign absolute values for expository convenience, though for the moment
I will assume that only the sonority relations expressed by the indices are
important. Later we will see that in fact a purely relational characteriza¬
tion of the sonority hierarchy is inadequate and that some indication of
absolute sonority values is needed after all.^
It is now clear how to express certain natural classes involved in phono-
tactic descriptions. The class “glides plus sonorants” is simply the set of
segments whose sonority indices range from 8 (/, u) to 5 (nasals). The class
“glides plus consonants” includes segments whose indices are less than or
equal to 8. The class “vowels” includes those whose indices are greater than
or equal to 8. And so on. The claim here is that natural classes defined in
this way, and only these, are relevant for characterizing syllable structure
in natural language.
Major Class Features 113

It is a systematic fact that many of the natural classes defined in terms of

sonority indices have no simple expression in terms of the major class
features and can be designated only by a disjunction of feature complexes.
Thus, compare the sets of segments that can be treated as a natural class in
terms of conditions on the sonority index n and the designation of that class
with binary distinctive features (but without the feature [± syllabic]), as
illustrated in (7).

(7)
Natural class Conditions on Binary feature
sonority index complexes
1, m, n, T +son 1 '
obstruents j+latj ^
L (+nasJJ
, [ — son] ,
r, 1, m, n, s 7>« >4“ ' r + son
+ cons]
f— son 1
+, cor r
+ cont
— voice

Like those mentioned in (4) and (5), these natural classes appear in phono-
tactic descriptions, and they show the difficulty of using the major class
features as vehicles for their expression.
The general claim, then, is that the theory of sonority indices provides
the basis for just the right characterization of natural classes that is
required in a theory of syllable structure.
It should be noted that this proposal offers a more restrictive theory of
natural classes than the proposal that they be based on the major class
features. The major class features alone are inadequate to the task and
must be supplemented by features appealing to, among other things, place
and manner of articulation, as illustrated in (4"), (5")^ ^nd (7). And in the
absence of a theory of sonority, no explanation is provided for why some
complexes of place and manner features, and not others, enter into dis¬
junctive statements like those above. For this reason, the new theory
appears to give a better account of natural classes (for phonotactics) than a
theory cast in more traditional terms.
Elisabeth Selkirk 114

2. Sonority Indices in a Theory of Syllable Phonotactics

According to the autosegmental theory of the syllable that has been

proposed recently,the terminal positions of hierarchical syllable struc¬
ture are “empty positions” of sorts. Phonological segments (a distinctive
feature matrix) are represented on a separate segmental melody tier, and
are associated with the terminal positions of syllable structure by univer¬
sal conventions and/or language-particular rules. One advantage of this
theory is its ability to properly represent “long” segments. A long vowel or
geminate consonant is a single segment on the segmental melody tier and
may be treated as such by rules applying to segments on that tier, but it is
also double, in the sense that it is associated with two terminal positions in
the core syllable structure. Hypothetical kappa would have the represen¬
tation (8a), and hypothetical tuuli the representation (8b):

(8)
Syllable structure a. a o b.

Association lines
\ / \ /
I I \ / I \ / I I
,1 I V I V
Melody tier k a p a u

Given an autosegmental theory of the syllable, the phonotactic descrip¬

tion of the syllable has at least three parts: (i) the characterization of
possible syllable structures, (ii) the characterization of possible (or im¬
possible) sequences on the melody tier, and (iii) the characterization of
possible associations between the two.^"'' Each of these is to be viewed as a
set of well-formedness conditions. For a syllable to be ruled well formed, it
must be well formed with respect to (i-iii).^^
For each of the three sets of well-formedness conditions on syllables in
an autosegmental framework, there will doubtless be some that are uni¬
versal and some that are language-particular. Included in the conditions of
type (i), which define the possible syllable trees for a language, are (a) a
characterization of the internal structure of the syllable (perhaps only a
(universal?) division into onset and rime), (b) a specification of the mini¬
mum and maximum number of terminal positions in the syllable, and (c) a
set of conditions on the terminal nodes. I propose to view this as a syllable
template, as in Selkirk (1982), and will require that every syllable tree of an
utterance be nondistinct from it.^^ Included in (ii) are specific filters
(collocational restrictions, in the sense of Fudge (1969) and Selkirk (1982))
that rule out particular sequences of segments. Included in (iii) is the
Major Class Features 115

universal condition that association lines not “cross,” as well as the uni¬
versal condition that a segment a may be associated with a terminal
position fi in syllable structure only if a is nondistinct (in a manner to be
made precise) from jS. Other language-particular ponditions on these as¬
sociations may exist as well.^^ Clearly, it is not only well-formedness
conditions of type (ii) that contribute to defining the possible segment
sequences of a language. In fact, as we will see, types (i) and (iii) are even
more important. For a segmental melody to be a possible melody of a given
language, it must be capable of being mapped onto a (sequence of) possible
syllable structure(s) of the language. The template, which defines this class
of structures, itself specifies what sort of segment will be permitted in what
position in the syllable, and in this way puts severe constraints on possible
segment sequences.
The nature of the terminal positions in syllable structure has come under
debate in recent years. McCarthy (1979), Halle and Vergnaud (1980), and
Clements and Keyser (1981) have argued that those terminal positions are
either C or V (where what C and V stand for is not always explicit). ^ ® In the
spirit of Selkirk (1982) and Harris (1982), which are couched in a non-
autosegmental framework, it might be argued that those terminal positions
are characterized by a complex of major class features. According to
Kiparsky (1979), also a nonautosegmental account, the terminals are
marked s or w’ and given integer values of strength according to the
so-called Liberman and Prince algorithm. What I wish to propose here
is that those terminal positions are characterized in terms of sonority
indie es.^^
The syllable template of a language indicates the maximum and mini¬
mum number of terminal positions in the syllable and identifies the ter¬
minal positions with names. The template structure is also (universally)
divided into onset and rime, though this may not be crucial. (9) is an
example of such a template:

(9)
a

O/ and Ry are terminal position names, convenient mnemonics for onset

position and rime position, respectively. The subscript integers stand for
first position, second position, etc. It is in fact possible to view template (9)
as a template schema, standing for the set of templates in (10):
Elisabeth Selkirk 116

(10)

Accompanying the template schema is a set of conditions on its terminal

positions, which are expressed in terms of sonority indices (SI). The fol¬
lowing are some examples of language-particular conditions on the ter¬
minal positions:

(11)
a. If jc is associated with Oj, then SI(a:) < 8.
b. If X is associated with O2, then SI(x) < 3.
c. If X is associated with Rj, then SI(x) > 8.
etc.

Cast in terms of sonority indices, these conditions state in effect what

classes of segments may be associated with particular positions in the
syllable structure of the language in question. The condition on SI(Ri), for
example, states that the first position of the rime must contain a “vowel.”
The condition on Oj states in effect that this position must be filled by /, u,
or any other segment with a sonority index less than 8. In an autosegmental
framework, it is assumed that a segment on the melody tier may not be
associated with a particular terminal position of the syllable unless its
sonority index falls into the range specified by these conditions.
It is well known that syllables conform in general to what may be called
the Sonority Sequencing Generalization (SSG):^*

(12)
Sonority Sequencing Generalization
In any syllable, there is a segment constituting a sonority peak that is
preceded and/or followed by a sequence of segments with progressively
decreasing sonority values.

The existence of (12) as a universal of syllable structure gives some plausi¬

bility to the sonority-hierarchy-based approach to phonotactics being
advocated here (though it is consistent with other theories as well). Such a
condition could be easily formalized in terms of sonority indices, but I will
Major Class Features 117

not do so here. The SSG can be viewed as imposing universal constraints

on the possible form of language-particular sets of conditions on syllable
structure. It in no way constitutes on its own a theory of syllable phono-
tactics, however, for languages will differ precisely/ in their choice among
the various conditions on terminal positions that are consistent with (12).
The hypothetical list (11) is one such set, and I will give examples of others
below.
The advantage of the sonority index theory of conditions on terminal
positions is not only that it properly characterizes the natural classes of
segments that can be associated with particular positions in particular
languages, but also that it allows a straightforward expression of the
relations between particular positions that need to be stated in grammars of
particular languages. As Harris (1982) points out, a theory of syllable
phonotactics must have a way to specify a minimum sonority difference (or
dissimilarity, to use Harris’s term) between two adjacent positions in a
syllable. This sort of relation is easily stated in terms of sonority indices
(and impossible to state directly without them, as we will see). For example,
the generalization that in an onset nasals may precede glides but not liquids
might be expressed as the requirement that for a sequence O2 Oi,
SI(02) < SI(Oi) - 3, where “3” is the minimum sonority difference re¬
quired. Such a condition would also rule out the sequences *//', */w, *//, *rw,
*rj, and so on. As a final point, note that all of these (hypothetically) ill-
formed onset sequences would be consistent with the SSG. Clearly, in¬
dividual languages impose even greater restrictions on sequences of seg¬
ments, marking off particular spans of the sonority hierarchy that may be
realized in one position or another, and in one position with respect to
another.
The theory I have outlined may be referred to as a relational theory of the
syllable. There are three distinct senses in which it is relational, (i) The
specification of individual syllable terminals is cast in terms of theoretical
constructs that are themselves inherently relational: sonority indices and
conditions on sonority indices, (ii) It permits the formulation of explicit
statements of conditions involving relations between adjacent terminal
positions in the syllable, (iii) The set of possible conditions on sonority
indices and terminal positions is presumably restricted overall by the
essentially relational SSG.
In recent articles on syllable structure, Kiparsky (1979, 1981) has also
offered a relational theory of the syllable. There are two major differences
between his theory and mine. My theory stipulates the SSG (as part of
Elisabeth Selkirk 118

universal grammar), while Kiparsky’s attempts to derive it from yet deeper

principles of universal grammar. Also, mine is a theory of language-
/

particular phonotactics (within a universal framework), while Kiparsky’s

is not.
Kiparsky proposes that, universally, the syllable has a relational struc¬
ture that is represented in terms quite analogous to the relational (metrical)
representation of stress, that is, with binary branching trees having nodes
labeled 5 or w. As Kiparsky points out, given certain stipulations con¬
cerning the nature of (a) the branching structure assigned to syllables, (b)
the s/w labeling of that structure, and (c) the interpretation of that labeled
structure in terms of integer values, along with one additional assumption
concerning the relation between segments and this tree structure, it is
possible to make something like (12) follow as an automatic consequence.
Kiparsky’s enterprise is an interesting and important one, to be sure. But I
do not think it is entirely successful.
Specifically, Kiparsky proposes that it be stipulated that syllables uni¬
versally have the branching structure in (13) and that their nodes be labeled
as shown there. The strength relations among the terminal nodes tree (13)
can be straightforwardly translated into (relative) integer values, as written
below them.

(13)

...4 3 2 1 2 3 4 ...

Then, assuming (as I do) that segments have integer values corresponding
to the relative sonority associated with them (Kiparsky suggests that
complexes of binary features, including the major class features, determine
the integer value of a segment in the sonority hierarchy) and that the
relations between integers in the tree (which are either “greater than” or
less than ) are matched by the relations between the sonority-determined
integers of adjacent segments, the SSG follows automatically.
My theory of syllable phonotactics based on sonority indices is perfectly
consistent with Kiparsky’s tree proposal and the theory of (12). If the tree
Major Class Features 119

proposal were right, the only consequence for my proposal would be that
the SSG would not have to be stipulated. However, there is an important
reason for questioning KiparSky’s assumptions about syllable structure
and its interpretation from which the SSG is considered to follow—namely,
the relational tree theory of stress that provides the analogy on which
Kiparsky’s tree theory of the syllable is based is quite possibly wrong. Both
Prince (chapter 11 of this volume, 1983) and Selkirk (forthcoming) argue
that stress patterns are not to be represented by trees, but only as the
alignment of syllables with a metrical grid. If this theory is right, there is no
motivation independent of the syllable for branching structures of this sort
in phonological representation (though onset and rime may remain), and
no independent motivation for the labeling conventions required. More¬
over, another explanation must be sought for the SSG. I have no such
explanation to offer, and so for the time being will have to leave (12) as a
mere stipulation.
I should also add that, as the later examination of Spanish syllable
phonotactics will show, deriving the SSG from a uniformly labeled
branching structure of the syllable runs into certain serious problems and
gives reason to question an approach like Kiparsky’s that bases the SSG on
syllable geometry.
In addition to his proposal concerning the SSG, Kiparsky (1979, 1981)
points to the need for viewing syllable phonotactics in terms that are
relational in the second sense that I mentioned. Specifically, he points out
the impossibility of specifying absolute conditions on the terminal posi¬
tions of syllable templates, but does not elaborate on just what such a
relational theory of phonotactics might be. This is in fact what I am doing
namely, offering a theory of the phonotactics of particular languages, one
that is couched in terms of sonority indices and conditions upon them.

3. Case Studies in Phonotactic Description

3.1 English Rimes

The English syllable template schema is shown in (14);

(14)
Syl

It specifies that the maximum number of positions in the onset is two, and
that there may be none. It also specifies that the maximum number of
Elisabeth Selkirk 120

positions in the rime is three, and the minimum one. Elsewhere it has been
shown that, given two assumptions, this schema correctly characterizes
both English onsets and English rimes.
I will examine the English rime, looking first at the smallest syllable
template schematized in (14) and then at the larger ones. It turns out that
conditions stated on positions in smaller templates remain valid for the
larger ones. This is an interesting result, which supports the view of the
maximal template as simply a schema that “collapses” all the templates
together. Consider first the template (fragment) in (15):

(15)
Syl

(16)
Ifx is associated with Ri, the SI(x:) > 5 (equivalently, SI(x:) > SI(m,n)).

There is only one condition to be stated, (16). This condition says that
an English rime may consist of a vowel or a sonorant (in such a case
“syllabic”) on its own. This generalization distinguishes English from
French or Spanish, for example. As for the rime template (17), the next
larger in size, the same condition (16) on Rj obtains, along with the
additional condition (18) on R2.

(17)

(18)
If X is associated with R2, then SI(x) < 8 (equivalently, SI(x) < SI(i,u)).

(18) says that R2 must be a glide or something less sonorous. Conditions

(16) and (18) together allow for the sequences in (19), which are permitted
in English.

(19)
VV: cow, bye, toy, etc.^^ rN: pattern^^
VL: pal, far, etc. rO: mallard
VN: run, sing, slam, etc.
VO: cut, tap, pick, etc.