1 
AUTOSEGMENTAL PHONOLOGY 
BY 
JOHN A. GOLDSMITH 
A. B. , Swarthmore College 
1972 
SUBMlTTED IN PARTIAL FULFILLMENT 
OF THE REQUIREMENTS FOR THE 
DEGREE OF DOCTOR OF 
PHILOSOPHY 
at the 
MASSACHUSETTS INSTITUTE OF 
TECHNOLOGY 
JUNE 1976 
- -Signature redacted 
Signature of Authon •• 0 ~~;~;~~~~. ~jJ/~;~j_~~- ~!~~;~~~;~;. ~~d 
· Linguistics, April 30, 1976 
. . Signature redacted _ 
Cert if 1ed by ......................... -· ., ....... -.- ..,,. .• .,. ,1 ... • • ... • • • • • • • • • • / ,,.--··· 
1 Tbesis Superviso.,..r.,,,.,, 
Signature redacted 
Accepted by ••••.••••••••••••••.•• • 1 ••••••••••• • , .• ,. •.• -••••• -•••••••• 
Chairman, Departmental Committee 
on,Graduate Students 
2Autosegmental Phonology
John A. Goldsmith
Submitted to the Department of Foreign Literatures and Linguistics
on Apri 130, 1976, in partial fulfillment of the requirements for
the degree of Doctor of Philosophy.
A modification of the theory of generatrve phonology is suggested.
in this thesis in the introduction of parallel tiers of segments
(or "autosegents"). This is shown, in the first chapter, to
resolve certain formal and substantive problems in the current theory.
A detailed analysis of Igbo, a tone language of Nigeria, is
presented in autosegmental terms in Chapter Two, as well as a
new analysis of the phenomenonvof "downstep" found in most
African languages. In Chapter Three, these notions are develope.
to account for stress and intonation patterns in English, and
various accentual and non-accentual. systems are dealt with as
ways of co-ordinating the tonal and syllabic tiers of autosegments.
Work by G.N.Clements on vowel harmony is cited to support a more
general account of autosegmental phonology. Chapter Four presents
a hypothesis for The origin of autosegmental phonology, suggesting
that the inherent geometry at the phonetic level is"autosegmentat",
but that language acquisition will include the task of
"de-autosegmentalization", which tends to collapse the multi-
linear autosegmental geometry to a linear one at the lexical level.
Thesis Supervisor: Morris Halle
Title: Professor of Linguistics
P11.111 Ill 3lhMW!IhI.jAiIl1 I!V#,-
ih'EIMMII.iI...' urns
3
TABLE OF CONTENTS
Abstract . 2
Epigraph. ..... . ......... . . . .4
Preface................ ... . .......... 5
Prelude ,- The Question of Suprasegmentals . . . 11
Chapter One - A Overview.... .. ......... 28
Notes . 93
Chapter Two - Igbo Tonology .. ......... 97
Notes . . 185
Chapter Three - Melodic Association Rules . 193
Notes . . 260
Chapter Four The Autosegmental Index 264
Bibliography .. . . . . . . . . . . 276
Biographical Note . . . . . . 280
I lii IL  H I
From the continuing saga of Guru Rug, the 34-year-old perfect
Rug Salesman:
Student: 0 Master, I have had a revelation!
Guru Rug: And what is this revelation, my son?
Student: 0 Master, I have seen that you can
spell "Guru Rug" -- backwards!
Guru Rug: Yes, my son, that is true.
You can spell anything backwards.
-- WCAS Radio 1974
"For in the science of harmony, as you probably know, the same thing
happens. The teachers of harmony compare the sounds and consonances
which are heard only, and their labor, like that of the astronomers,
is in vain.
"Yes, by heaven!" he said; "and 'tis good as a play to hear them
talking about their condensed notes, as they call them; they put
their ears close alongside of the strings like persons catching a
sound from their neighbor's wall -- one set of them declaring that they
distinguish an intermediate note and have found the least interval
which should be the unit of measurement; the others insisting that the
two sounds have passed into the same -- either party setting their
ears before their understanding."
-- Plato: The Republic Book VII
IL I I1~[Ij,
Preface
"It all started about two Thanks3iyings ago....
It was two years ago on Thanksgiving."
The idea behind this thesis began with a reading of Will Leben's
thesis on suprasegmental phonology, in which he argued that in some
languages, even short vowels could bear two successive tones.
Impossible!
.,.Because inexpressible, in segmental terms. And yet the
facts persisted; Leben's case was clear. How could we express two
successive tones on a single segment? In a little paper at the end of
1973 called "Tonemic Structure", I suggestErd a motation for falling
toned vowels: V . The observation was made then -- by Paul Kiparsky,
H L I think -- that this would account for
what, in this thesis, is called "stability" of tone melodies.
As a notation, this was simple enough; as a theory of anything,
it was quite inadequate. Igbo, a tone language with an excellent
descriptive grammar, Green and Igwe's, seemed a good testing ground
for the theory. Edwin Williams had looked at a floating tone in the
Igbo system, and his analysis was insightful, though it was formulated
in the same theoretical framework as Will Leben's thesis.
It proved fruitful. The work in Chapters 1 - 3 of this thesis
is essentially the product of 1974. The last chapter, Chapter Four,
is of more recent vintage, and remains quite a bit more speculative,
but submits a hypothesis which accounts for the origin of auto-
segmental tiers, and also explains why the notion of distinctive feature
and the notion of an underlying inventory of atomic units --
call them phonemes -- why these two notions are not incompatible,
but are rather twc aspects of a more general strategy in language
acquisition, what I call "de-autosegmentalization." It is, in
any case, a step towards such an explanation.
There are several topics more or less directly connected
with autosegmental phonology that I shall, more or less, eschew.
The first is the question of the syllable. In a paper on
tone in Sanskrit by Robert May and myself, I suggested that the
syllable might well be considered an autosegmental level. In this
sense, the string of familiar C and V segments could be broken up
into an autosegmental representation where the second tier was
composed of syllables:
(i) C VCc v cvc V C
The Well-formedness Condition would permit structures like (i),
where a single consonant occurred in two syllables, but none like
(ii):
(ii) C V C C V C
Furthermore, all segments would occur in at least one syllable.
['P VIIIlkN 11.1 MIlh l ,it..,
This proposal has certain immediate merits, beyond merely
formalizing the traditional notion, and beyond the advantage
gained in talking about syllables without placing syllable
boundaries in various places between the Cs and the Vs, a by now
familiar approach that is always tripping over its own feet.
Most importantly, this approach would remove the glaring restriction
in the theory of awtosegmental tonology that says that only syllabic
segments can associate with tonemes. Given syllabic structures
like (i), CV CV would become (iii).
(iii) CV CV (a)
(b)
H L (c)
Here these is no a>symmetry among levels; all segments in levels
(a) and (b) enter into a Well-formed autosegmental relation, as do
levels (b) and (c).
This notation suggests further that the familiar process
sending, e.g., i -+y is a syllable merger process, as illustrated
in (iv) .
(iv) i a i a
If this is true, then an immediate consequence of this is that
i's tone would have to reappear on a, i.e., as in (v), but not
as in (vi).
a! I
S
(v) b 1ia = ba li a
A /
(vi) *ba lia = ba lya
Such a process in (v) could
the circling of two (necessarily
indicate an "identification", as
of two syllables into one.
(vii)
ba 1ha
T'L
= V y
-ba lya
be expressed as in (vii), where
neighboring) syllables would
mathematicians say, or joining
i a
Such a theory would then remove the feature "syllabic" from
the highest tier in (iii) or (v), instead develop a more articulated
notion of "being in a syllable", indicating how a sonorant could be
the nuclas of a syllable with which it is associated. Various
possibilities and notations come to mind here - like (viii) -- but
this is not a simple matter by any means, and as I say, I will not
be pursuing it here.
(viii) b:a i a k = balyak
"N" for"nucleus"
We may ask, further, what the elements on the syllabic tier
are. I have indicated them with subscripted sigma's; but on all
II I
other tiers, the units were segments with features specifications.
What are these syllables?
Such a question is interesting, even important, within the
context of the discussion of Chapter Four below, but I shall leave
the question untouched. In a forthcoming thesis, Dan Kahn
presents a theory of syllable phonology as sketched in (i) and (ii)
and obeying the Well-formedness Condition.
Two other recent theses bear directly on the material here, and
I shall say little about them. One is Mark Liberman's account of
English intonation, stress, and timing; the other is Shosuke
Haraguchi's analysis of Japanese dialects. As I indicated above,
the first three chapters below predate these works, and at my risk,
perhaps, I have left them basically unchanged. Some of Mark's ideas
and criticisms have made their way -- no doubt, unrecognizably --
into Chapters Three and Four. I don't doubt that a firmer and more
adequate synthesis is yet to be achieved. Of Sho's work, its bredth
and depth speaks for itself. For that reason, I have restricted
my attention in Chapter Three to different types of systems from
those in Japanese. The interested reader is urged to consult
Haraguchi's analysis of autosegmental melodic association rules
(Haraguchi (1975, which will be published soon); see also, regarding
Japanese, Goldsmithf(1975a) ). Again, a deeper understanding of
universally possible systems will require a synthesis of Haraguchi's
work on Japanese dialects with an analysis of African accentual
tonal systems.
This thesis is addressed generative phonologists; my back-
ground and training has been in this tradition. Autosegmental
, I~rk
10
theory is, I believe, an improvement to this system. A fair amount
of what distinguishes autosegmental representation from, say,
that found in Sound Pattern of English bears a resemblance to
certain non-generative systems. This relationship is discussed in
the introductory section, following immediately below.
I have one last chance now to thank all of the people who
helped me during the writing of this thesis -- and there are
surely more than I know who did. I must thank Will Leben, Mark
Liberman, Ivan Sag, Dan Kahn, Frank Heny, Mary Clark, Erich
Woisetschlaeger, Nick Clements, Sho Haraguchi, Michael Cohen,
Robert May, and all the rest of the linguistics community here;
and to Noam Chomsky and Paul Kiparsky, who were on my thesis
committee, and who improved several earlier versions of this
thesis, I am very grateful. But by far most of all, my deep
gratitude goes to Morris Halle, who for several years now has borne
with me, helping with the work that went into this thesis, and
providing ideas, and support, and criticism. He claims he is
responsible for neither the good ideas nor the bad in this
thesis, but I know for a fact that much of what is good here would
not have been written without his prodding, and his own suggestions.
And I know, too, that he still doesn't believe that there is a
star that works as I suggest in Chapter Three. So caveat lector.
:11 I I ' N llm I.11il.6" j -, !.1
ll
Prelude: The Question of Suprasegmentals
Ever since there have been segments in phonology, there have
been phenomena that evaded segmental classification, and so there have
been suprasegmentals. As Rulon Wells noted in 1945,
In effect, the term 'suprasegmental phoneme' As
used in recent discussions simply denotes phonemes
which are neither vowels nor consonants.
There have been attempts, of course, to give a definition of
"suprasegmental" that carries more information than this -- attempts
to characterize "suprasegmental" in a positive rather than a
negative way. This thesis deals with a certain subpart of what have
traditionly been called "suprasegmentals"; it attempts to give
a specific characterization:dfthese types of prosodic phenomena.
The general framework of this "autosegmental" approach is
generative phonology, and we are, therefore, concerned with phono-
logical rules that relate phonological and phonetic levels as
much as with providing a formal characterization of these two levels.
Much of the literature concerning suprasegmentals, some of which
we shall look at below, has been formulated within a structuralist
framework; the concern there is primarily with methods of
analyzing initial phonetic data into elementary units which may then be
btiilt up into higher level units, such as phonemes and morphemes.
Despite this basic difference, there are a number of similarities
between'the questions that arise within a generative and a struc-
turalist treatment.
If we look at the American linguistic traditin in the decade
of the 1940's, we find a fair amount of discussion regarding supra-
I --
12
segmentals, with contributions by, among others, Bloch, Harris,
Haugen, Smith, Trager, and Wells, All begin with the initial
assumption -- spelled out clearly, for example, in Harris (19144)
and Bloch (1948) -- that logically the first procedure in linguistic
analysis is to "slice" the linguistic signal vertically into a
number of pieces -- setments, we may call them. Care must be
taken, of course, to make sure that the signal is sliced finely
enough to find the truly minimal units of the utterance. One
this process of "segmentation" has been completed, the classification
of these slices. may proceed. The formal representation of a
signal sliced into segments is, of course, a linear sequence of
elements, but this is due to the nature of the initial process
of segmentation, rather than the inherent nature of the speech
signal.
We shall look at some of these structuralist analyses, but an
initial distinction must be borne in mind. We must samrate the
issue, on the one hand, of how English stress, accentuation, or
intonation is to be dealt with, from, on the other, the defining
characteristics of suprasegmentals. There are two tendencies here
in defining "suprasegmental", either one defines a process or
feature as suprasegmental from the outset -- in effect, making it
something independent of the linguistic analysis of any particular
language; or else "suprasegmental" is a name we give to any process
we find occurring in a particular language when it displays certain
general properties, either of distribution of, in generative terms,
of behavio rwith respect to rules and rule applications.
10111iHilil! ACAE11hl-li ill'! -
I l 4 1 i ID I ik il
13
More specifically, someone taking the first approach might
say, with Haugen (1949):
The term psodeme has been used in recent writings
on phonemics as a stylistic variant of prosodic pho-
neme to describe such modifications of the basic
speech sounds as tone, stress, and duration. (278).
Juncture, as Haugen also notes, was frequently classified as
suprasegmental, but again -the grounds for-, the classification
are not language-particular, but rather theoretical.
On the other hand, some treatments -- of components, for
example, by Harris (1944) and by Bloch (1918) -- develor definitions
or criteria which may be applied either language-specifically or
even utterance-specifically to determine whether a certain feature
is being used suprasegmentally. As I have implied, this is the
approachoof the present work: to develop a concept (here, autoseg-
mental tier) whose particular application to a language must be
determined by the facts of that language. Thus, for example, the
diagnostics, or indications, of autosegmental behavior that are
developed in Chapter One below will be indications to the linguist
that a certain feature is, in fact, autosegmentalized in a particularr
language.
If we return to the structuralists' assumptions, we must
rephrase this question, of course; we ask not what indications there
are concerning what the language itself is doing, but we ask, rather,
what the appropriate methods are for analyzing data from a particular
language, and under what circumstances these methods will lead,
in effect, to suprasegmentals.
One factor that led to consideration of stress as being non-seg-
I IIR. l! d IA I llI
mental -- and thus, under one interpreattion, "suprasegmental" --
was the belief (cf. Trager and Smith pp. 35-6) that the very
meaning of stress was relational, and therefore could only be
defined in relation to neighboring segments in the utterance --
a property, it was held, that distinguished stress from "normal"
features like voicing and vowel height.
A more accurate way to say this would be that if we abstract
away from stress levels, we can (it was assumed) identify a segment
for wha.t it is independent of the context it is found in; if
however, we do not make this abstraction, and the stress level of a
vowel determines in part just what sort of segment that vowel is,
then identifying that segment as a (say) S (an /a/ with 2stress)
requires looking at the context which the segment is found in.
This motivation for calling something "suprasegmental" is not
the only possible one by any means. As Haugen puts it,
Some sounds are thought of as occurring one after
another, like bricks in a wall, while others occur
simultaneously with these and usually span a number
of the individual bricks at a time. (278).
These longer-than-usual bricks, or suprasegments, may be either
(i) homogeneous in time, or (ii) non-homogeneous ( or rathernot
necessarily homogeneous) in time. Both theoretical options can be
found in the literature; the clearest statement of the second position
is found in Harris (1944), which we shall return to; the closest
to the first position is Bloch (1948).
A third possibility arises if one works within a framework which
provides a fixed inventory universIdly of phonetic features. This
third possibility is that the longer-than-usual bricks, the supraseg-
ments, may be changing in time, or dynamic, just in case there is a
Ill i W IAJ idl? 10
15
single phonological feature that, by itself, specifies this dynamic
phonetic character. The strutturalists we are looking at did not provide
themselves with such a universal phonetic system, and so this option
was not available to them. It is the position that we shall take, how-
ever, in this thesis, beginning in Chapter One. If there are no dynamic
features -- whihh is possible , but not yet known for sure -- then
this position is equivalent to (i) in the previous paragraph.
Yet the availability of phonololical features is of more
consequence than this matter of dynamic suprasegmentals. For in what
is perhaps the most successful treatment of suprasegmentals within the
structuralist framework, Zellig Harris', the two affairs (on our
account, distinct) of "features" and "suprasegmentals" are reduced
to the same thing. In"Simultaneous Componetts in Phonology", Harris
maintains that, in effect, phonemes -- or "components", rather
may occur at the same moment in time. These "simultaneous components",
if all of the same length in time, play the role of Trubetzkoy and
Jakobson's "features"; if one lasts longer than another, the longer
one may be viewed as a suprasegmental.
The distinction between features and suprasegmentals will be
oentral in this thesis, and so the attempt to reduce thet two to one
is of some real importance to us.
In his review of Trubetzkoy's Grundzuge three years earlier (1941),
also in Langtage, Harris had rejected phonetic (read: distinctive)
features as irrelevant to phonemic analysis. The only relevant
criteria in determining the bredth and measure of a linguistic system,
he maintained, were distributional.
The types and degree of phonetic contrast (e.g.
whether all the consonants come in voiced and
unvoiced pairs) have nothing to do with the
classification of the phonemes....It may indeed
be interesting to work out patterns of the phonetic
relations between phonemes and see how they compare
with the distributional pattern. But that would
be a new correlation, interesting for diachronic
lingulstics and for linguistic psychology,e.g., for
the question: How do the physical (phonetic) differences
within thesrans of phonemes (events to which people
conventionally react uniformt.) compare with the
difference between different phonemes (events to
which they react differently)? In synchronic linguistics,
it is only the distributional pattern that would show
what work each phoneme can do, what operations can
be performed upon each, i.e., what its place is in the
structure. (348).
This position changes somewhat by 1944, though by no means completely
or always explicitly. In "Simultaneous Components", a method of
analysis is introduced which "breaks all or most of the phonemes into
new sub-elements (components). Each of the old phonemes will be
a particular simultaneous combination of one or more of these new
elements." (181). "It will be possible," Harris continues, "to
select and symbolize the components in such a way as to show immediately
the limitations of distribution,and in mayy cases -the phonetic
composition, of the phonemes in which they co-occur." (emphasis added).
In this tack towards the phonetic are included "components which are
precisely the length of a phoneme", i.e., "simultaneous breakdown of each
phoneme by itself." This may "indicate the phonetic composition of
each phoneme," as well as -"eliminate phonemes of defective distribution."
This technique, it is suggested, is of particular importance in
separating off tone in tone languages. But if the techniques is useful
in distributional statements about tone-languages (182), it may well
be "profitable to continue this extraction [of components, i.e.,
simultaneous breakdownsof segments -- JG ] until all phonemes have
16
44 I 
I i'
17
been reduced to combination sof components." (200). Such components
may include "front middle, back", the components which distinguish
n, n, andq (200); they may include that which /p,t,k/ have in
common (Harris calls it " ", adding that it "may well be the combin-
ation of the unvoicing and the stop components") (200).
The crucial step comes next. "The physical movements of
articulation may offer certain absolute factors common to various
phonemes: /p,tk/ are generally voiceless,fortis, stopped. Since
the components will in the last analysis have to identify
articulatory as well as acoustic events [one wonders why? JG], it
is desirable to reflect these as closely as possible." And it
is on grounds of "phonetic simplicity" that tonal features are
separated from vowels in tone languages (201). This remark might
be contrasted with Harris' footnote, some pages earlier (p. 183) that
"the Prague Circle more closely approached the technique of dividing
elements into simultaneous components, but purely on arbitrary
phonetic grounds....
On the whole, it would not be unfair, I think, to interpret
Harris' attitude toward Trubetzkoy's "features" (and Jakobson's - he
is thanked in Harris' introductory footnote) as amivalent and
ambiguous. Harris' program is aimed at showing that the distributional
techniques of traditional American phonology, if carried to what he
suggests is their logical extentions, componential analysis, would lead
the linguist to the useful, "non-metaphysical" part of the Prague
conception of distinctive features. By resting with a familiar
technique of classification by distributional restrictions, Harris
A di kw I bI j. 1 I N
could, he was suggesting, arrive at the useful part of the Prague
school's feature-theory.
If finding simultaneous components is a discovery procedure
for simpltfying grammatical description (in Harris' basic sense
of reducing the number of constraints that need be stated in a grammar
on the distribution of basic elements), then the character of these
components is essentially arbitrary, not phonetically conttrained.
These "long components" could then express relations between successive
segments as well as internal composition of particular segments; these
relations between successive segments, furthermore, can logically be
most anything at all - what might today be called morpheme-structure
constraints, or the effects of phonological rule.
If this interpretation is right - and I shall cite some of
Harris' passages in support of this below -- then two quite different
notions of simultaneous components are being presented at the same
time. One conception, presented earlier in Harris' paper, sees them
as phonetically meaningful; the second, as phonetically arbitrary.
Thus at the beginning of the article.& Harris suggests that
components have a "statable physical characteristic", and that they
are "time stretches of sound (sound-waves), or sets of simultaneous
motions of the vocal organs [footnote: e.g., vibrations of the
vocal cords, giving 'voice'....)]" (186). Again, he says, "since
components are to be physical elements, we must consider the phonetic
value that they can have" (187).
This step toward phonetic, or feature, orientationis strikingly
illustrated in the very first paragraph, in fact, where Harris gives
an example of a componential analysis of /b/, consisting of "voicing
! I I l 0 14 : 1 I dI [ i , 1-: .o 61 W
19
plus lip position plus stop closure, all occurring simultaneously."
And then, in a footnote to this apparently phonetic bent, Harris corrects
this impression: "This example of phonetic components is given
here only for introductory simplidity. The analysis presented
below is primarily distributional rather than phonetic."
Thus, by section 4.1 (187), "It is not required that components
have a constat phonetic value throughout their duration. A compone 'ter
a rhneic value which changes in a fixed way in respect to its end
points: e.g., Falling tone, increase in nasality, voiceless beginning
and voiced ending." Harris, parenthetically, makes it clear that what
have been called "intonational melodies" (see Chapter Three below,
for example) are, for him, single components, although he does make
reference to their status as lexical elements (in present terminology).
This is irrelevant, nonetheless, for the componential analysis, according
to Harris.
In the most detailed part of the article (sections 5.2 to 5.4,
pp. 191 - 201 ), Harris spells out the uses of long components as
restrictions on distribution, He considers several examples, including
Grassman's law in classical Greek (permitting only one aspirate, under
normal conditions, per word), devoicing of consonants in German before
juncture, and the presence of clusters like /rtr/ (partridge) in
English, but not */trt/.
In the last example,.he posits a long component
extending over the length of a cluster and having
the following phonetic values:
in first position, general consonant
value (serving indidentally to distinguish consonants
from vowels; this because any consonant may occur here);
20
in second position, continuant
or /t,d/ if it follows a stop, otherwise general
consonant value;
in third position, continuant if it
follows a stop (but if the stop is /td/, this
value only if a continuant precedes it), otherwise
vocalic value.
The value [Harris continues] of the
consonant-componant permitt any cluster of two
except stop plus /p,b,k,g/, and then permits
the third place to have continuant value (and to
remain a member of the cluster) only if the pre-
eding two are continuatt and stop; otherwise the
component has non-consonaht value and thus changes
the third position into a vowel. This statement
does not allow for clusters with middle Is!, as in
sexton, and omits several details which would be
taken care of in the other components for the
individual phonemes. However, it is included here
to show that even fairly complicated clusterings can
be described by single components."(19T198)
I have quoted at length from Harris' work to show the phonetic
or feature-like constraints on long components are, in reality, nil.
Harris was aware of the ambiguity of his position, apparently; in
a footnote (numbered 22a, perhaps added after the first version), Harris
points out that there are indeed differences between the kinds of long
components he has been dealing with. But nothing comes of this
observation.
A few years later, Bernard Bloch developed "A Set of Postulates
for Phonemic Analysis" (1948), in which he axiomatizes a method for
determining 1K phonological system of a language from a set of
utterances. His system is of some interest to us because he
proposes that "a feature that is common to all the segments of a span
(or to all the segments that are open to it) is a feature of the span
itself, not of any segment in the span." (52.4, p.36). His particular
statement is, curiously, inconsistent; if we define a feature of a
II lI[I lW!# WI I 8 I ,10 '
21
span in terms of the features of the segments of the span, we cannot
then say that the segments in that span do not have those features.
Nonetheless, Bloch's intent is clear; it would be more clearly
stated if Bloch had had recourse to another level in the grammar,
for exanple, at which spans of segments "had" features; his definition
would then be a defintion of the relationship between this new level
in the gra mar and the phonetic segmental level.
But while this line of analysis is temptingly similar in some
ways to the autosegmental we will develop below, Bloch's treatment
is ultimately quite different from ours, for the fact that there may
be spans with features is a fact, in his system, determined completely
by surface phonetics, and does not correspond to his notion of
suprasegmentals (it should be noted, too, that his use of the term
"feature" does not correspond to present use; hit "aspects" are
rather more like our "features", but the difference is not crucial
here.) With respeet to suprasegmentals, BloCh says,
48.2 Definition. Every distinttive aspect of stress
or of pitch is a prosodic feature of the segment in
which it occurs....
(The term "prosbdic feature" is intended
merely to name one kind of feature, as defined in
section 47.2).
This kind of definition (admittedly
ad hoc in its phrasing) allows us to abstract the
qualities of stress and pitch from 'the totality of
distinctive aspects in a segment, without obliging
us at the same time to abstract certain other qualities
that we do not usually wish to treat in the same way --
for instance voicing and vowel color. For a language
like Hindustani, where nasalization acts very much like
pitch or stress in the more familiar languages, Postulate
4)can be rephrased or supplemented so as to allow this
quality also to be abstracted as a prosodic feature.
For still other langugges, where still other qualities
act in a similar way, still other modifications or
extentions of our postulates will be required.
11,1 '0 I
22
If we abstract away from Bloch's disclaimer (p.46) that
his set of postulates is not intended to reflect any truth,but
simply to axiomatize an arbitrary but useful procedure ("To ask
whether this is a 'correct' or a 'true' transcription of the given
utterance is meaningless..."), we come quickly to the conclusion
that Bloch has not provided a universal characterization of
suprasegmental features. Indeed, it is the linguist's intuition
when dealing with a new langutage that justifies adjustment of the funda-
mental postulates of phonemic analysis,as Bloch says in the quotation
above. But the real task for the linguist must surely be rather
to propose a set of criteria that determines when and whether certain
features are or are not suprasegmental.
Our object, ultimately, will be to produce a notion of "supra-
segmentals" which makes specific predictions about the behavior
of these elements. We shall try to show real differences between features
per se and. supra- or (auto- ) segmentals, such as the automatic
spreading of autosegments overtke longest domain possible (in a sense
defined below in Chapter One). Features by themselves do not
spread; they merely identify a segment for what it is. The domain of
association of an "autosegment", on the other hand, does spread,
quite automatically.
As I suggested earlier, it is 'difficult to compare this type of
approach with the struturalists', since there were no phonological rules
in their system. Certain things comparable can be found. In his
discussion of the suprasegmental pitch phonemes of English (1945), Wells
k I.
23
says (p. 31):
Considered phonetically in relation to the segmental
phonemes, the length of a pitch phoneme is variable.
It can coincide in time with any number of segmental
phonemes. Its effect will continue until replaced
by another pitch phoneme.
This last statement comes close to our Well-formedness Condition
(see Chapter One below); Wells is suggesting that the domain of the
suprasegmental specification spreads itself until another supra-
segmental specification is found. Wells places a left/right
asymmetry on the spreading, since it only occurs to the right in this
system, but this asymmetry is not argued for. He continues directly,
A corollary of this statement is that a pitch phoneme
cannot immediately follow itself. For instance, 22
would be merely a notational lapse for 2.
We shall meet this idea again as Leben's (1973) Obligatory Contour
Principle. It is important to recognize, however,that this is not,
in fact, a corollary of any notion of spreading such as Wells suggests.
Its truth, as an empirical hypothesis, is independent of how the
domain of a suprasegmental is determined.
More broadly, the question of how suprasegmental elements interact
with each other, and interact with the segemental phonemes, was not,
to my knowledge, rigorously addressed. Most importantly, from our point
of view, what factors determined the placement of coordination of
suprasegmentals with respect to segmental phonemes -- and especially
the changes that these coordinations may undergo -- was not
seriously dealt with in the strutturalist tradition. There is one
exception to that, and this concerns the connection of English
intonation patterns with stress.
In this domain, it was frequently explicitly recognized
that there were, in Pike's terms, "contour points" where the intonational
patterns had special points of connection with syllables, or more
generally, with elements of the semental phonology. Bolinger -- who
consistently evades classification as a structuralist -- most
clearly (1955, 1958) separated out the characteristic a syllable
may have as being the potentially prominent syllable from the point
of view of the shape of the intonation curve (called "stress") from
actually being prominent ("accent").
But having recognized that pitch oontours in English are associated
with syllables that are stresse- not, for example, with a syllable
whose vowel is a schwa -- the structuralist analyses went no
further; nor could it, as far as I can see. There is no room for.
a distinction, for example, between a rule of feature-assimilation,
on the one hand, and true spreading of a suprasegmental element -
what we will call a "Flop" rule below-on the other. Again as far
as I know, such a distinction is not necessary for the analysis of
English, at -. least to 4e -level of detail that has been devoted to
it so far. But the distinction becomes vital in the analysis of
African tone languages, such as Igbo, in Chapter Two below.
As this type of distinction suggests, we are more interested
here in phonological, as opposed to phonetic,kinds of analysis.
Again, it will be a feature's behavior with respect to phonological
rules that gives it away as suprasegmental, not any phonetic facts.
Ultimately, it would be extremely interesting to pursue the extentions
of autosegmental representation to phonetics, but this will not be
our goal here. Once again, the nature of the criteria used to prove
!I n
Wli[d k,.1 I :4 i
25
a feature "autosegmental" is something which translates very hard
into structuralist terms.
For our part, as I have indicated, we shall take our problem
to ind de the determination, for any particular language, of what features
are suprasegmental. By the criteria and within the system developed
in this thesis, "stress" --- as a multi-level, or as a binary system --
does not satisfy the requirements for autosegmentalization. If stress
works as described in, for example, Sound Pattern of English, then
stress is not autose gmental and not, in our sense, suprasegmental.
And yet much of the discussion of suprasegmentals in te American
structuralist tradition of the Forties and Fifties included long
disquisitions on stress as a suprasegmental. If stress is not
"degree of loudness" (as has been one understanding) but is rather
a means of modulating pitch, as Bolinger proposed, then this new
sense of stress qua accent is a strong caddidate for suprasegmental
representation (see the section in Chapter Three below on English as
a tone language).
The important literature on this subject is reviewed by
Bolinger in section 3 of his (1958) "A Theory of Pitch Accent in
English". Bolinger remarks,
According to Bloomfield, "stress that is, intensity
or loudness -- consists in greater amplitude of sound
waves." Jones gives more attention to sestural accompani-
ment, but he too insists on "the objtvJ impression of
loudness."
In a footnote at this point, Bolinger continues,
Pike (Intonation of American Einlish) gives first place
to intensity, but allows that duration may be a factor.
d billIIH pdi
26
Trager-Smith correlate stress with loudness, and
Edith C. Trager says unequivocally, "there is only
one component that matters - loudness." (p.21)
Our own position regarding the analysis of English is largely in
agreement with Bolinger, and therefore in disagreement with the
analysis of the American structuralists regarding the existence
of several levels of stress as distinct from pitch.
If we accept, however, the view prevalent in the Thirties and
Forties that stress is something closely akin to loudness, then of
the classical three suprasegmental traits -- stress, length, and junc-
ture -- none are suprasegmental wiin the autosegmental system. One might
choose to conclude, therefore, that autosegmental and suprasegmental
phenomena, as their names suggest, describe entirely different things.
As the citations above from Harris, Wells, and Haugen suggest, this
would be simply wrong from a historical point of view, for tone,
vowel harmony, and nasal spans are thorough-bred suprasegmentals. But
as the discussion in Chapter One will show, there are clear restrictions
on what features may be autosegmental in a language., and we intend
through these restriction to produce a stronger theory, a theory which
is not a theory, for example, of juncture.
Finally, the question we must ask before we begin are:
Q1) Why another theory of suprasegmentals?
For two reasons: first, because no totally satisfactory theory of
suprasegmentals has been proposed within any framework yet, and second,
because even the rudiments of a successful theory of suprasegmentals
isn't toThefourd in generative phonology. Generative systems have not
to Uate been able to deal with the kind of prosodic systems Harris'
system (or its most similar relative across the Atlantic, Firthian
110 I PI lW II. j |d i lIR 0 ., V
27
prosodic analysis) could handl&.
(2) What will count as a successful generative theory of
suprasegmentals?
The theory will have to do two different types of things:
(a) make new (and presumably true) claims about the structure
of representations at the phonological and/or phonetic levels; and
(b) explain how these different structures undergo phonolgical
rules.
The bulk of Chapters One and Two is devoted to answering these
questions.
(3) Abstracting away from the difference between structuralist
and generative phonological analysis, how does the proposed
theory of suprasegmentals differ from the structuralists' theories
of suprasegmentals?
There is no satisfactory answer to this question, because
ultimately it is not possible to abstract away from the different
frameworks in whickthe theories have been put forward. In any case,
autosegmental phonology, as it is developed here, attempts to deal with
a proper subpart of the domain that was covered by the term
"suprasegmental" in structuralist treatments. Oar intention is
thereby to give an empiitically more contentful theory of these
suprasegmental phenomena. Tn the end, the test of our success will
be found in detailed language-specific studies, like that in Chapter
Two, and in comparative studies, in line with some of the proposals
in Chapter Three. For the linguist already committed to generative
phonology, the arguments for our system are those in Chapter One.
28
CHAPTER ONE
AN OVERVIEW OF AUTOSEGMENTAL PHONOLOGY
1. Introduction
The subject of linguistics being the relation between
sound and meaning, its first question must be what the nature
and form of the phonetic and semantic levels are. For a
generative linguistic system, this task begins with hypotheses
about the type of formal representation that counts as a
faithful rendering of the phonetic- or the semantic aspect of
a sentence, =word, discourse, and so forth.
Autosegmental phonology is an attempt to supply a more
adequate understanding of the phonetic side of the linguistic
representation. Viewed in this light, it is a proposal at the
same logical level as the idea that a phonetic representation
is a linear sequence of atomic units--call them segments; it
is at the same level as the suggestion that these atomic units
are cross-classified by distinctive features. Autosegmental
phonology is a particular claim, then, about the geometry of
phonetic representations; it suggests that the phonetic repre-
sentation is composed of a set of several simultaneous
sequences of these segments, with certain elementary con-
straints on how the various levels of sequences can be inter-
related--or, as we shall say, "associated."
To say that autosegmental phonology is a hypothesis about
the geometry of phonetic--and ultimately phonological--repre-
sentations is rather abstract at best. From another, more
29
down-to-earth, vantage point, autosegmental phonology is a
theory of how the various components of the articulatory
apparatus--the tongue, the lips, the larynx., the velum--are
coordinated. That is, at the most superficial, observable
level, the linguistic signal is split up into a large number
of separate information channels. Viewed from the production
side, this consists of the specific commands to the larynx,
the velum, the tongue, and so on. At an "abstract" level, this
information no doubt comes about from splitting up a more uni-
fied representation. All this becomes clearer with an example.
Consider the word "pin." As the orthography suggests, the
linguistic representation of the word consists of three seg-
ments linearly ordered, as in (1).
(1) iconson +syllabic +conson
-nasal -nasal +nasal
+labial -labial -labial
-coronal -coronal +coronal
' i' 'n
The production of this word, however, involves separate,
though coordinated, activity by the velum, the lips, and so
forth--expressed roughly in (2).
(2) A Score for the Orchestration of 'Pin'
Lips: .. .Close up.. .. open...........*.............
Tongue: .. .High and front ........... touch the palate
Velum: . .Raise... ...........Lower ...................
30
The standard linguistic assumption regarding the nature
of phonological representations--that they look much as in
(l)--implies that the process of language acquisition and of
perception includes the development of the ability to take a
representation much as in (2) and slice it vertically into
columns, assigning the appropriate feature specifications to
each column, ultimately deriving a representation like (1):
P-I-N. In short, the normal assumption about phonological
representations implies that in processing a signal, we learn
to shift around slightly the horizontal alignment of the
commands in (2)--we "justify" it--and patch it up so that it
may be sliced up vertically into the phonologically, and hence
psychologically, real segments. Let us call this assumption
the "Absolute Slicing Hypothesis."
In the example we considered--saying "pin" as in (2)--'the
process described above is surely essentially correct. The
Absolute Slicing Hypothesis is adequate, then, to the level of
detail we have considered in (2). Suppose, however, that we
add to the orchestral score in (2) the activity of the larynx
that gives rise to pitch. If I am uttering this word "pin" in
isolation, and disregarding the actual sluggishness of the
vocal folds for the nonce, the syllable will be uttered at a
rapidly falling pitch. For reasons that will be developed in
detail in Chapter 3, I will represent the Falling pitch as the
sequence of a High pitch and a Low pitch. A more faithful
version of (2), then, would be (3).
1 1:111I4 'i ' * ikii  o A I b1 V,
31
(3) A Revised Orchestral Score for 'Pin'
Lips: .. .Close up... .Open.... ...................
Tongue: . . .High and front .... ...... touch the palate
Velum: . .Raise... ... *.........Lower.......... .......
Larynx: .. .High Pitch.......Low Pitch...............
The assumption of the Absolute Slicing Hypothesis fails
now, and this failure is in no sense a trivial one. The
laryngeal pitch specifications are, we shall argue below, and
in detail in Chapter 3, not the result of specification of any
of the segments in (1), the phonological representation. The
Falling pitch of this utterance is not part of the phonologi-
cal segments in the same sense that the other commands in (3)
(those for the lips, the tongue, or the velum) are.
The speaker of English thus factors out the pitch, and
does not attempt to include the pitch features in the huge
musical score which is sliced up into the more abstract seg-
ments as in (1). In this sense, then, the Absolute Slicing
Hypothesis fails: the slicing is not absolute or complete,
but rather may exclude some parts of the linguistic signal.
This failure of the Absolute Slicing Hypothesis is non-
trivial in two separate, important senses. First, it is only
an accident about English that the laryngeal pitch features
are excluded from the Great Slicing. Other languages may well
include pitch as a part of the signal which is sliced up into
successive segments. Conversely, a language may exclude some
i lA1 140j14 w Il.) : I ii. .
32
features from the Great Slicing that English happens to in-
clude. Guarani, as we shall see below, happens to exclude the
nasalization feature, while English includes it. Whether a
particular channel of articulation is included cannot be
specified universally once and for all.
The second sense in which the Universal Slicing Hypothesis
fails non-trivially will be the key to the Euccessful formula-
tion of its successor. The articulatory levels that are
excluded--in the present example, pitch--are themselves sus-
ceptible to a "slicing" or segmentation. TIus, while the
laryngeal pitch commands in (3) do not correspond to the stan-
dard phonological representation in (1), they do in fact
correspond to a more abstract segmented level, as in (4).
(4) H L where H =High and L=Low
In the present example, of course, the empirical significance
of the representation in (4), including the meaning of the
star *, remains open for the present. The important point is
this, however: while a language may exclude an articulatory
level from the Great Slicing that leads to the principal seg-
mentation (corresponding to P-I-N in (1)), the excluded arti-
culatory levels form themselves segmented domains, domains in
which the segments are linearly ordered, and where the segments
are cross-classified by feature-specifications. In general,
the formal properties of the "phonological" representation, as
in (1), will be mirrored on each level. Just as (1)-
33
corresponded to (2), then, (5) below--the revised and now
autosegmental representation corresponds to the more adequate
orchestral score, (3). And (5), we see, is a synthesis of (1)
and (4). The precise significance of the "association lines"
linking the tonological and the phonological levels will be-
come clearer shortly.
(5i) +si-06 on-c"S
To develop this idea,,however, let us turn our attention
from these rather general comments on the nature of phonologi-
cal segmentation to some more pressing problems in generative
phonology. A proposed theoretical revision must in generalbe
shown to resolve a dilemma in the standard theory; we shall
therefore proceed to several general types of phenomena that
are all difficult or impossible to account for in the standard
framework, which assumes only one string of segments in a
phonological representation. We shall show how each phenomenon
of this type individually leads toward the same solution, auto-
segmental phonology. In the process of doing so, we shall see
that these phenomena in general, though not invariably, cluster
34
together, as predicted by the autosegmental solution.
We shall consider:
(i) The existance of contour-valued features, such as
rising or falling tones on vowels, or pre-nasalized stops,
n M
such as d , b, etc. Here the argument must be made that in
some languages, short vowels bear these dynamic or contour
tones, and that furthermore such a contour tone is linguis-
tically the concatenzation of two level tones, rather than
being an atomic unit itself.
The existence of such short vowels with contour-tones
has been argued for recently by a number of linguists, most
forcefully by Will Leben (1973). No latisfactory theoretical
account has been given, however.
(ii) From contour-featured segmefts we shall turn to the
phenomenon of "stability," or the tendency of a feature value
to persist despite the erasure of the major segment (generally,
vowel) which appeared to have born that feature. Roughly,
that is, we find in tone languages that when a tone-bearing
vowel is deleted by a phonological rule, the tone that was
being borne does not delete also, but rather shows up elsewhere
on a neighboring segment.
(iii) We shall consider next melody levels in the grammar:
that is, linguistically significant levels in the grammar which
refer to just one or two features in the utterance. We have
already seen a foreshadowing of this in the discussion of the
difference between the pitch features in English and the other
I 4II ls l lb l~
35
features. If we view the musical scoring for the utterance
"pin" as in (3), and many more like it, we will notice a dif-
ference in the different lines (or "voices," in the musical
sense). Some, like the line of instructions to the velum
legislating nasality, will appear rather uninteresting when
viewed alone. There are no generalizations to be made by
looking at nasalization alone in English. On other lines--
such as the level of instructions for pitch--there are linguis-
tic generalizations to be gained by restricting our attention
to just those features. There are, that is, strictly pitch-
feature regularities;--which is not to deny that further in-
sight is to be gained by observing how the pitch-level is
coordinated with the other levels, of course. The point here,
however, is that certain subsequences of features do form
linguistically significant "melody levels," while others do
not.
(iv) Then we will see how the heuristic notion of "float-
ing tone" can be rigorously defended in autosegmental terms.
The floating tone has served well in practical terms for lin-
guists dealing with tone languages; it has, however, remained
with a tainted reputation because of its apparent anomalous
nature from a theoretical point of view.
(v) Lastly, we will consider some processes of automatic
spreading of features both to the left and right, over seg-
ments unspecified for those features. The accepted notation
for phonological rules implicitly predicts that spreading
36
should be simpler if it occurs only to the left or only to the
right; we find, however, many cases of bidirectional spreading.
When viewed from an autosegmental perspective, the reasons
become clear.
This is, perhaps, an appropriate moment at which to point
out some differences and similarities between autosegmental
phonology and other treatments of what have been called
"suprasegmentals." Much of what is covered in this present
analysis would be called "suprasegmental" by the criteria im-
plicit in the literature. What is suggested by the term
"suprasegmental"? Two things, I would submit, which are en-
tirely different but which have been continually confused in
2
the recent linguistic tradition. Calling tone "suprasegmental,"
firstly, immediately distinguishes it from the "segmental"--
that is, it correctly views pitch as different from and not
part of the phonological segmentation as in (l)--the segmenta-
tion into phonemes, roughly speaking. But this first observa-
tion has wrongly led to the assumption that "suprasegmentals"
could not be segmental in their own right--the second, more
fundamental, sense of "segmental," which is overlooked by the
term "suprasegmental." If the "suprasegmental" of pitch, for
example, does itself form a sequence of tonal segments, then
"suprasegmental" is a misleading label. A more accurate pic-
ture, we are suggesting, is parallel sequences of segments,
none of which "depend" or "ride on" the others. Each is
11,14ki T! Itill I I -lUHLA
37
independent in its own right: hence the name, autosegmental
level.
2. Contour Tones
Let us proceed, then, to particular problems within the
theory of phonology. Our first task is to make quite clear
the reason why two tonal specifications on a single vowel is
incontrovertibly in contradition with the standard theory.
I shall first sketch the argument heuristically, and then
redo it in more technical language.
Suppose we analyze the. tonal workings of a language and
decide that where contour tones occur, they are really the
3
concatenation of level tones. That is, Rising tones are
actually composed of the sequence Low tone and- High tone;
Falling tones are composed of High tone and Low tone. Suppose
further that such contour tones occur on short vowels. We
shall consider several actual examples shortly. A long vowel,
we know, may be analyzed as a sequence of two short vowels,
but a short vowel has no such analysis. How can we represent
the contour tone of the short vowel?
Our conclusion will be simply that we cannot represent it
if we stay within the assumptions of the standard theory. Let
us assume, for purpose of exposition, that tone can be repre-
4sented in terms of binary features (although the assumption
of the binary nature is not strictly relevant at this point).
Suppose, then, we try to represent /a/ with a Falling tone--
38
that is, /a/--a High tone and a Low tone.
(6) +syllabic
+constricted
pharynx
-high
-round
+Highpitch
-Lowpitch
b{ M-Highpitch
+Lowpitch
Now this clearly won't do. This curious segment (6) is
both +Highpitch, and -Highpitch; it is both +Lowpitch and
-Lowpitch. The features in the part labeled (a) seem to indi-
cate a High tone; those in (b). seem to indicate a Low tone.
But putting one on top of another does not order them, as we
require. In particular, we must distinguish a Falling tone
from a Rising tone; but to do so, a new and theoretically
momentous set of conventions must be introduced in order to
make vertical ordering of features within a segment play the
role of left-to-right (temporal) ordering of segments else-
where.
Or one could attempt an equally radical revision of the
notion of segment with the introduction of a notation as in
(8).
(8) +syllabic
+CP
-High
-round
[+HighpitchY[-Highpitch
[-Lowpitch +Lowpitch
A A BB
One has here, however, what is to all appearance a cate-
gory error. The non-tonal features have one relationship to
the entire segment: they are, let us say, "features-of" it.
The tonal features are features of a sub-segment (there are
two subsegments here, A and B); and the sub-segments bear some
entirely other relationship to the entire segment: they
"subcompose" it, let us say. Consequently, tonal features are
"features-of a segment which subcomposes" the entire segment;
in particular, the tonal features are no longer features-of the
segment. Here, too, then, is a radical suggestion intended to
make possible the representation of a contour tone.
Now these two attempts--the second of which is rather a
commonplace in the literature--are not misguided; quite the
contrary, even though both are wrong. They recognize that the
existence of contour tones is a problem for the standard repre-
sentation, and problems are the stuff of which advances are
made. Our task is to outline a third proposal, one which is
conceptually simpler and which has a number of direct empirical
consequences: autosegmental representation. First, however,
I would like to review the argument, in more technical terms,
4o
that demonstrates the dilemma the standard theory finds itself
in.
Segments are atomic elements ordered linearly left to
right; this simple and elegant hypothesis lies at the founda-
tion of standard phonology. A feature-specification of a
segment is, then, a property or an attribute of the segment.
The property of being voiceless "belongs," we may then say, to
the first segment in "pin, " just as the property of being
"+front" belongs to the second segment. More technically, we
would say that "pin" has a representation at the phonological
level as in (9).
(9) -S S S
7 5  31 33
The subscripts indicate nothing but that these are primi-
tive elements in the phonological vocabulary. We would fur-
ther specify that there are various "feature-projection" maps
-- characteristic functions, in mathematicians' language--that
tell us what the feature specifications for each segment are.
For example, there is an F function which maps S 5 in the
voice 75
present example to - (minus). F . (S 5 ) = - . All this
voice 75
amounts to saying that the first segment in the representation
of pin in (9) is voiceless.
Having made precise our notions of segment and feature-
specification (following essentially Chomsky's Logical
Structure of Linguistic Theory), it is clear that there is no
way that a segment can be specified both + and - for some
141
particular feature. In particular, no single segment can be
contour-toned unless there is a single feature which is "Fall-
ing" or "Rising," contrary to hypothesis. If a single segment
were both +Highpitch and -Highpitch, then F would not
highpitch
be a function, again contrary to assumption, since it would
be two-valued.
If we wished to resolve this formal conflict by saying
that there were two separate characteristic functions associa-
ted with each phonological feature, one mapping to the
specification + (plus) and the other to the specification -
(minus), we would avoid the immediate contradiction, but lose
the formal representation of the fact that the two values are
values of the same feature; such an approach would clearly be
the wrong direction to proceed in.
No, the correct position to take here is that, indeed,
the formalism demands that no segment may be doubly specified
for a feature. If tones are features of the vowel segment,
then, there may be no Low-High sequence on a single vowel, and
that's all there is to it. Some other basic axiom must be
abandoned.
The formal approach we will take toward a resolution of
this problem is to deny that tonal features are features of
the vowel in the case of the problematic contour tone. Rather,
the tonal features are factored out to another level; feature
specifications on the other level constitute segments, but
their relation to the vowels with which they are associated is
614 AIfli- NAk df,,:iqI r.A"""
merely one of simultaneity in time. We represent this dynamic
element by association lines. The Falling-toned /a/, then, is
represented as in (10).
4- oy Lr
0 1k
\/VA more complex example, like the word akala would be represen-
ted as in (11).
(11) a k a 1 a
L H L H
Before proceeding with the development of this notation,
let us turn to an example from a tone language.
2.1
Our first example is from Oiuhu Igbo, as described in
Green and Igwe (1963). These facts will be discussed in much
greater detail in Chapter Two. For the present we will con-
tent ourselves with finding a contour tone on a single segment,
and showing that it is indeed the concatenation of two tones;
we shall then see why the notation as in (10) or (11) does more
than represent these facts, but predicts and explains the
43
intuitively clear sense of left-to-right order of tones that
is revealed inside of contour tones by the tonological rules
involved.
There is a simple tone rule that occurs in the "I Root"
form. Throughout the Igbo language, the particular tone of
the verb stem, suffixes, and prefixes is determined by the
"form" of the clause. We shall devote Chapter 2 to an inves-
tigation of the details of this system; for the moment, we may
focus our attention on the I Root form, in which the verb-stem
is Low-toned, in non-exceptional cases. Consider, for example,
a simple sentence in .the I Main form where the subject is a
pronoun:
(12) I Root form
(a) -o ci akhwa
he carry eggs
(b) za ula
he sweep house
(c) M ci anu
I carry meat
"He must carry some eggs"
or "He was carrying some eggs
"He must sweep the house"
"I was carrying meat"
We see that in the I Main form, the prbnominal subjects
are H in tone, and the verb stem is uniformly Low in tone.
When the subject noun phrase (NP) --not a pronoun--ends in a
low tone, the tone pattern of the sentence is just what we
would expect on the basis of the sentences in (12). The verb
stem is Low; the subject bears it inherent (isolation) tones.
(13) Eze ci akhwa
Chief carry eggs
Uwa ci akhwa
Uwa carry eggs,
"The chief was carrying eggs"
Eze chief (HL)
"Uwi was carrying eggs"
Uwa (a name) (LL)
However, when the subject NP would normally (i.e., in
isolation) end in a High tone, then it undergoes a slight
tonal change here As we see in (14)", the final H becomes a
Falling tone.
(14) Ekwe ci akhwa
Adha ci akhwa
"Ekwe was carrying eggs"
Ekwe (a name) (HH)
"Adha was carrying eggs"
Adha (a name) (LH)
What is happening is clear if we watch the tonal melody.
In (14), the drop from the High on the last syllable of the
subject NP to the Low of the verb stem is shifted leftward, or
"anticipated" on the last syllable of the subject. Using the
autosegmental notation, this change is represented as in (15),
where the dotted line indicates that the association line was"
added. (16) separates the Before and After states; it says
the same thing as (15) in more familiar but less perspicuous
notation.
NP
EkW
S
V
cL kkt1UGt~
'Al r' i-i I-. p
structural
change
(16) Ekwe ci akhwa.
I H L L H1
H H L L H
Ekwe ci akhwa
H H L L H
The point to observe is that the final syllable in Ekwe is
associated with two tones--H and L--in the derived structures;
this is what is identified as a contour tone.
Before proceeding to the behavior of the irregular verbs
in the I Main form, let us see what this autosegmental notation
has provided us with. By setting the tonal segments or tonemes
off on a separate autosegmental level, we were forced to in
troduce "association lines" to coordinate the two levels in
time. We hoped, thereby, to represent contour tones, which is
prima facie a phenomenon purely internal to the vowel in
question. The notation captures precisely, however, the fact
that the tone associated with a vowel on the right--here, the
verb stem--may associate with a vowel neighboring on the left,
115
i. I iR Iii M i [41 I 41i; .
causing a change in the right-hand side of the latter vowel.
Thus we find processes like (18a), but not as in (18b), as the
notation predicts.
(1?L) V V 'V
The difference between two guch rules is intuitively plausible,
and in fact borne out in empirical work in tone languages
throughout the world. (18a) is in some sense a proper "assi-
milation"; (18b) is not. The representation in (16) makes
clear what that sense is; alternative proposals, such as per-
haps (6) or (8),, do not.
Having called the tonal phenomenon in (14) an assimila-
tion, and having characterized it as a "flop" rule as in (16),
we have committed ourselves to a view with certain predictions.
If the verb stem should be on a Mid tone, for whatever reason,
then the falling tone of the subject's final syllable must
fall to Mid, rather than to Low. This is in fact true, and is
borne out in three separate cases, discussed in Chapter 2.
Here I shall just mention one representative example.
A Low-toned suffix immediately following the stem causes
the stem to raise to Mid, as in (19a). The pitch of the
utterance is as drawn; the autosegmental representation is as
in (19b).
'I Ih.mIIfl iii I.
47
(19a) Ha dir n'yru (19b) Ha diri n'oru
They were at farm H M L H H
where V means "High falling to a Mid which is level with the
following vowel, which must be Mid." The reason for this
complicated statement lies in the nature of Downstepping
tonemes (the Mid here is one); these are discussed in Chapter
2.
2.2 English
Our next example is from English, a language that may
feel more familiar. The neutral intonation pattern in English
* *
is HL or MHL, with the High tone associated with the accented
syllable; this, in fact, is the, significance of the star *.
Consider a polysyllabic word like "archipelago." It has
a tone pattern roughly as in (20).
* *(20) archipelago archie
H L
This is derived by linking the starred elements in either
level; the Well-formedness Condition of autosegmental theory
then comes into play and does the rest. Thus we start with a
representation as in (21), with no linkage between the levels;
then rule (22) comes into ,effect, creating (23a). The Well-
formedness Condition creates (23b) from (23a).
ii 11 . * :! -Ml
48
* *
(21) archipelago (22) V
* *
H L T
In (22), as elsewhere, "T" stands for any arbitrary toneme,
and "V" for any arbitrary vowel; a dotted association line
indicates that the rule adds that line.
* *
(23a) archipelago (23b) archipelago
H L H L
(24) Well-formedness Condition (initial statement)
(1) All vowels are associated with at least one tone;
All tones are associated with at least one vowel.
(2) Association lines do not cross.
Note that the Well-formedness Condition is in the indicative,
not the imperative. A derivation containing a representation
that violates the WFCondition is not thereby marked as ill-
formed; rather, the Condition is interpreted so as to change
the representation minimally by addition or deletion of
association lines so as to meet the Condition maximally. In
the case at hand, (23a) is changed to (23b) minimally by addi-
tion of four association lines.
Holding aside one or two small points until the extended
discussion in Chapter 3 (Section 3.2), the description just
given accounts accurately for the assignment of neutral into-
nation to English words. So far no contour tones have been
1[9
encountered, but that is because the accented syllable was not
final in the example considered. If we consider a word like
balloon, we get a structure as in (25), corrected to (26) by
the Well-formedness Condition.
(25) balibon (26) balloon
11 L H L
What the theory is predicting about English now is something
curious, something more than a bit reminiscent of tone
languages, but something nonetheless true: non-final accent
is realized as a High tone (as in (24).); final accent is
realized as a Falling tone.
This conclusion is true independent of the length of the
final vowel, too; so our search for contour tones on short
vowels has led us home to English with a clear example. See
(27) (the reader will recall that were the final vowels in
(27) long, they would undergo diphthongization and Vowel
Shift).
(27) Japan hat pin.pun
vA
A L H L H1,L
50
2.3 Excursus On Formalism
With the initial statement of the Well-formedness Condi-
tion (24) for tones given, the central aspects of the auto-
segmental theory are clear. In this excursus, I would like to
restate some of these provisions formally. A formalization of
these ideas is simple, almost too simple, in that it requires
a certain amount of familiarity with the concepts involved
before the axiomatization makes as much sense as the simple
diagram notation.
Be that as it may, let us proceed. Each autosegmental
i
level is a totally ordered sequence of elements, a.: this is
th th
the j element on the i level. Call the set of segments
th i
on the i level L We shall restrict our attention to the
case with two levels for now. So we have:
a1 a1 1 a
1 2 a a
2\12 2 2
a a2 a3 a
In addition to these two sequence of segments, there is
a totally ordered sequence of pairs--essentially the associa-
tion lines, from the geometric point of view.
(a1 ,a2) (a2 ,a2 ) (a1,a2 )
Call the set of these pairs "A."
It is clear that A in a sense organizes the other levels.
A L
Define projection y from 2 (the set of subsets of A) to 2
III j+ I Ili iqlI,1, AM I 114 '. ArIp lw I !: - I! w I i'j'fj ]I
51
(the set of subsets of L ) in the natural way:
12 12 12 111
Tr1 ({(a.,ak)(azam), (anap)......) =...
That is, the 1st projection rr picks out the set of first
elements of the pairs. Likewise, the second projection Tr2
picks out the second element of each pair, and so forth.
Define the inverse of these projections: r 1 may be defined
-1
as follows (or any other way that makes clear that T1  is the
inverse of fr in the natural sense):
-l
T (a) is the largest subset a of A such that
Then the Well-formedness condition is:
-1(WFC) Tr. and m. preserve connectedness.
Recall that we have a notion of connected already defined on
each level L and A since we assumed they were linearly
ordered. This ordering naturally induces a connectedness in
which the set circled in (26) is, connected, but the one in
(27) is not, and so forth.
(26) a b& d-ef (27) a® c f
I will briefly run through why the statement WFC above has the
effect of the Well-formedness Condition in (24).
IH l jl 3 1 - j 4iI i M M q? _ J
1-] I I 4l i ll -,I ['4 l I , 1 5 g -,- .~l o ,1 6[
52
Suppose we have a situation as in (28).
(z-) A b&
This violates (24-2); we shall see that its corresponding
formal representation does not have a proper projection opera-
tor preserving connectedness.
L for (28) is, evidently, A B C. There are, rather
trivially, 7 connected subsets: 0, A, B, C, AB, BC, and ABC.
Similarly, L2 is a b, with four connected subsets: 0
a, b, and ab. We will see that there is no formulation of
the set A with the correct properties. There are only two
candidates: (29) and (30). Consider (29) first.
(29) (A,a) (Bb) (C,a)
-l
Since {a} is a connected set, T'2  (fa}) should be
connected. r2 (fa}) = {(Aa),(C,a)}, which we see is not
connected in (29). Contradition.
If we try (30), we get a similar result.
(30) (A,a) (Ca) (Bb)
-1Now T (a) = {(Aa),(C,a)I is connected.
2
However, in (30) {(Aa) ,(C,a)T} is connected, so
Tr ({(A a),(Ca)}) = {A,C} should be connected--i.e., A and C
should be neighboring. But they are not; contradiction.
11 I I I [ I M kAi..H ,ll IYO t
53
Thus, by extention, any situation like (28) where lines
cross must be ill-formed; and so (24-2) is derived.5
Similarly in (31), which violates the first part (24-1) of the
Well-formedness Condition, we find WFC violated also.
(31) A B C L
2
a b L
-l -
If f2 and fr 1 preserve connectedness, then 1 Ta 2 must
-l
preserve connectedness too. But Tr a Tr (ab) AC, which is
not connected (although ab is); contradiction. The same
6
argument holds if a = b. Thus WFC implies 24-li
3. Stability
The second type of phenomenon we shall consider in
motivating autosegmental representation is what I have called
"stability." In tone languages, we find that when a vowel
desyllabifies or is deleted by some phonological rule, the
tone it was bearing does not disappear--rather, it shifts its
location and shows up on some other vowel. The toneme, or
more generally, the tone melody, has a stability which main-
tains it independently of the other aspects of the signal, and
thus the tone melody preserves itself despite modifications to
the syllabic structure.
Reference to this type of phenomenon in the literature
has generally been associated with the notion of "conspiracy"
or derivational constraint: in this case, a derivational
41 1 11.1141 I JiAH -, 1
54
constraint or conspiracy to move around the tonal specifica-
tions from vowel to vowel in order to find, on the surface,
the same tone melody that was there underlyingly. This is not
to say that there are no tonal rules that delete or modify the
tone melody: there surely are such rules. But the normal
case is where the tone melody survives the effects of
phonological rules.
If the tone of a vowel is specified by its features, then
the pitch of a vowel is just like any other of its character-
istics, like its tenseness or its roundedness. If a phonologi-
cal rule should delete that vowel, then its tonal specifica-
tions are deleted along with all of its other properties.
Suppose we have a phonological rule deleting a vowel as in
(33).
(33) V + / - V V-Deletion
(A common Bantu rule--see, e-g., Spa, 1973:78 rule (32),)
However, we need to save the -tonal information of the deleted
vowel, because it shows up on the surface. Looking at tone as
a feature of the vowel, we could do this in one of two ways,
similar in intent.
(1) We could posit a special "Tone Copy" rule which
copies the tone of the to-be-deleted vowel onto its neighbor--
we could do this, I should add, if we permit two tonal feature-
specifications inside a single vowel segment, the position I
have just argued against in section 2. But in order to
55
consider the logic of "stability" independently of that of
"contour specifications," let us permit the latter for the
moment to be expressed as in (34) below.
(34) Tone Copy
[V] V] V] V ]
Ihi yhiI +ahi ahi] yhi]
[ [610j [o Slof 6101]
A typical derivation applying Tone Copy and V-Deletion would
be as in (35).
(35) ... a i...
a i ... Tone Copy
VV
... i... V-Deletion
(2) Second, we could posit a general "derivational con-
straint" to apply to all tonal rules either particular to the
language, or as a general condition on the application of
rules--this is the approach Spa takes in his grammer of Enya,
a Bantu language. He suggests, "when a segment carrying a
High tone is delected or becomes incapable of carrying a tone,
the High tone is transferred to the nearest syllabic segment...
((This constraint)) applies each time any rule whatever meets
its structural description" (Spa, 1973:139 et passim). In
fact, the correct statement of his derivational constraint
should apply to preserve equally both High and Low tones.
This modification both simplifies his phonological system and
generalizes the derivational constraint.
F IF
56
Solution Number 2 is explicitly global if taken as a
language-particular rule, and therefore suspect within received
generative theory: a theory counternancing global rules
approaches vacuity. This solution introduces a general global
condition on vowel-affecting rules, and while this seems like
an improvement, in that it is a generalization, it is nonethe-
less worse theoretically because we now permit not only
global rules, but a whole new kind of object which is global
and applies anywhere during a derivation, outside the set of
ordered rules. As a general condition on rules, it meets no
such criticism, but it stands as an admission that the original
framework does not adequately model the behavior of the seg-
ments under the application of rules. Solution (1) holds out
promise, and yet what we find in actual work is that for every
case of vowel-deletionor desyllabification, we must set up
another case of tone-copying, and what we have then is a miss-
ing generalization. The generalization is precisely solution
(2), the general derivation constraint. A paradoxial situa-
tion: to effect a satisfactory linguistic solution, we need
to state a generalization; but inclusion of this generalization
within the standard theory amounts to either a serious weaking
of the theory of phonology, or an inadequacy in the standard
representation.
We might note that even if we did include the derivation-
al constraint, in the belief perhaps that constraining a theory
must always play a more minor role than stating a generaliza-
1 ' 1 hd i l [1 .i I'qI.' N-!
57.
tion, three important questions would be left unanswered:
first, why are tonal features copied, but not the other fea-
tures? What makes them special? Second, what is meant by a
representation with two feature specifications inside a
single vowel (the issue pursued in section 2)? Is there in-
deed a connection between the fact that there can be contour
tones and the phenomenon of stability--apparently so different?
Third, and most telling, the "conspiracy" to preserve tonal
melodies extends past a derivational constraint that whisks
the tone off of a sinking vowel: in fact, in a tone language
where the Derivational Constraint seems to generally hold,
what we find is that vowel assimilation rules like (36) copy
all vowel features up to, but not including, tone features.
(36) V- +aligh aHigh
fBack Back
yRound yRound
6ATR 6ATR
A rule like (36) certainly exists in Igbo and Yoruba, and
in Enya according to Spa (pp. 47,57). So if two vowels come
together, each with its own tone, then either one vowel is
deleted and its tone gets retained, as in (33), or one assimi-
lates in quality in every regard save tone. The only empiri-
cal difference lies in the length and syllable quality of the
remaining vowel(s). From the point of view of tone--and its
conspiracies--the same fate has come to pass. Yet the deriva-
tional constraint speaks only to the case with deletion, not
C d1 1, W N -iF A I -
the case of nearly-complete assimilation: thus missing the
generalization.
This is the logic of the situation. Let us look at some
actual cases in more detail.
(A) Consider two articles by Julie Lovins (1971a,b) on
Lomongo, whose tonological rules, she suggests, "conspire,
individually or in concert, to derive surface tone patterns on
words and phrases without changing the underlying melody."
Central to the analysis is what Lovins calls "tone composi-
tion," in which "the tones stay where they are when segmentals
are deleted." She continues with an example, "if two vowels
are juxtaposed, within a word or across word boundary, it is
usual for the first vowel to be elided. Its tone remains and
combines with that of the following vowel." For example,
(37) balongS> bakae + balongakae 'his book'
bana bamo + banamo other children'
VV \ / / VA/ /
bomo botamba + bomy tamba another tree'
V / N \ V A/
batswa la. emi + batswemi you who lead me away'
With a number of similar examples, Lovins concludes,
"The only derived forms that occur are the ones that preserve
the underlying melody.., and the only way to get these derived
forms is to posit a species of rule application that many
linguists find objectionable." She is certainly correct, given
the standard framework, andshe isexceptional among writers on this
subject in recognizing the implications for phonological theory
of the type of rule she posits.
59
The existence of the tone melody's "stability" is our
concern: how can it be that a tone refuses to be deleted when
its vowel is deleted? In autosegmental formalism, this is
precisely what is predicted. In any theory of generative
phonology, and -this one is no different, a deletion rule
deletes a segment. Now, if a rule--(33), for example, V-Dele-
tion--should delete a vowel, it does not delete any of the
tone segments that the vowel is associated with, since those
tone segments are quite separate segments. The worst that can
happen is that the ,tone segments will be left "orphaned" or
free, without a vowel associated with it. That will be the
interesting case to look at in detail.
The point we have just seen should be emphasized: the
stability phenomenon, formerly paradoxical, is a natural con-
sequence of the autosegmental system-not by proposing a
constraint on rules, but rather by proposing in effect a new
geometrical shape (in a somewhat abstract sense) for formal
representations.
Let us consider in more detail the reduplication treated
in Lovins papers. Verbs are lexically marked for tone, H or
L; the sten is reduplicated and an /a/ infix is added between
the two copies of the stem. An L or H desinence then follows.
i 
kI:4. II tlI W 1 [ I '. mr :'d
6o
(38) L-toned sten /sik/ 'stop
H-toned desinence
sik + a + sik + V
underlyingly
L H
sik + a + sik + V bytheWell-,
1 formedness
L H Condition
L-toned desinence
sik + a + sik + V
L L
sik + a + sik + V
L L
sa sik + V
L H
Phonological
rules
sa sik + V
[sasikV][sasfkv]
The last stage is reached by pure phonological rules:
k-+0 and ia-+a.
(39) H-toned stem /lomb/
H-toned desinence
lomb + a + lomb + V
H H
H H
lymb + a + lomb + V
H H
*H H H
'be shy'
L-toned desinence
lomb + a + lomb + V
H L
*
lomb + a + lomb + V
H L
l + lomb +
*
H L
[lalombV] [alombV]
61
In short, from the notation we get the "conspiratorial"
results automatically by keeping the syllabic and the tonal
levels formally separate. It may be noted that we get the
desinence-tone spreading automatically, too, as well as a
formal understanding of the notion "contour tone." Further-
more, the process of total vowel assinilation--construed as,
e.g., (40)--has the desired property of copying all features up
to, but not including, tone features, since tone features
aren't features of vowels.
(40) X - +syllabic - +syllabic - Y
SD: 1 2 3 4+
SC: 1 3 3 4
The resolution of the paradox in tonal stability was
derived from viewing tones as segments on an equal rank with
"phonological" segments. This parallelism can be pursued;
in fact, we find in general perfect formal symmetry between the
two levels. The "dual," then, of vowel deletion would be
tone-deletion, followed by reassociation to another tone by
the vowel that had been associated with the deleted tone.
This in fact occurs; see the discussion of the II Main form in
Igbo (Chapter 2), and Sanskrit (see May and Goldsmith (1975)).
4.1 Melody Levels
The third type of argument for autosegmental theory comes
from the existence of "melody levels." As explained above on
62
p.M, these are linguistically significant levels in the gram-
mar which refer to just one or two features in the utterance.
Faute de mieux this has been taken sometimes to indicate tone
features as "features of an entire morpheme," in some unex-
plained sense.
Let us begin with an example from Mende, a Mande tone
language. The analysis is due to Leben (1973).
On short vowels in Mende, we can find Low, High, Rising,
Falling, or Rising-Falling tones. Morphemes are one to
three syllables long, and if the distribution of tones over
these syllables were random, we would expect to find five
tonal classes of 1-syllable words, 52 or 25 classes of 2-syl-
3lable words, and 5 or 125 classes of 3-syllable words: 155
types in all. In fact, there -are 5 classes for each, not
n5 , and they are of a very particular sort. Leben explains
this by proposing that there are only 5 available uhderlying
melodies in Mende, and that the melody is mapped from left to
right onto the word. The five possibilities are:
(41) H pele, ko
L bele, kpa
t A
HL kenya, mbu
LH nika, navo, mba
LHL nikili, nyaha, mba
Such an analysis, we might note, supports the contention that
contour tones are the concatenation of level tones, and that
Il 4l Ii ! II I- l[iI .Il I- I! 11J I-. 1; II j
63
short vowels (and these in (41) are) may bear several ordered
level tones.
4.2 Tiv Verbal System
The next example is somewhat more complex, and more
interesting. The Tiv verbal system has been the subject of a
sequence of more and more refined analyses, including
D.W. Arnott's exposition (1964), McCawley 's reanalysis (1970),
and Leben's proposal (1973).. The present reanalysis departs
from Leben' s in one particular way: I shall indicate that the
inclusion of Leben's principle (which I shall call the
Obligatory Contour principle) leads to unnecessary complica-
tions, and that it should be abandoned. This leads to certain
particulars thataare different than their counterparts in
Leben's analysis; it should be kept in mind that although I am
using autosegmental notation here, the principal empirical
differences between Leben's analysis and mine result from the
inclusion or rejection of the "Obligatory Contour principle,"
which may be stated as:
Obligatory Contour principle (Leben): At the
melodic level of the grammar, any two adjacent
tonemes must be distinct. Thus HHL is not a
possible melodic pattern; it automatically
simplified to HL.
As I say, I shall argue against the inclusion of such a prin-
ciple within phonological theory at the phonological or
A j 1 v N ii IJ ~ l D
64
8
tonological level.
In describing the Tiv verbal system, Arnott begins by
explaining that the tonal pattern for each tense need only
be illustrated by a small number of examples: six, to be
exact. There are twelve "tenses": General Past
Future
Imperative
Habitual 1, 2, 3 & 4
Recent Past A & B
Subjunctive
Continuous
Past Habitual.
The verb stem can in general be one, two, or three syllables
long; and--it is important to note--for each length there are
at most two (generally two) possible tone patterns. Thus,
for example, the General Past is illustrated as in (42).
(42) GENERAL PAST ye 'they'
High Low Drop; see text
ve va ve dza
ve ungwa ve vende
ve yevese ve ngohor
(As of yet, the reason for calling the first column "High"
and the second "Low" is unexplained.) Normally two successive
High tones are at the same pitch; sometimes, however, there is
a "Drop" of one or two semitones between the successive Highs.
This drop is indicated by the raised exclamation point.
The point has been emphasized (in the three analyses
cited above and elsewhere) that the six tonal melodies in (42)
65
are not unrelated; that the language learner need not memorize
six separate patterns for each tense. While this is true and
important, we should emphasize equally the following point:
whether the speaker knows six formulas for the General Past
or just one, these are melodies of the General Past, not of the
word. For each tense, there is a corresponding chart as in
(42), and from tense to tense the particular tone pattern on a
particular verb varies wildly. All verbs of the same syllabic
structure have the same tonal pattern in a particular tense,
however. The conclusion is that there is a level in the gram-
mar which contains as entries in it various melodic patterns
independent of the words or vowels; these entries can be
assigned to particular tenses.
This then is the argument for autosegmental representation
here: the tonal pattern found on each verb is linked in the
lexicon not with the verb but with the verb tense; thus the
melody must be a separate entity in the lexicon. The only
system capable of merging into one simultaneous utterance two
9lexical (phonological) entries is autosegmental phonology.
We might then conclude that for each tense, the speaker
of Tiv has learned six tonal melodies, two for verbs of one
syllable, two for those of two, and two for those of three
syllables, as in (43). Which of the two possibilities a
particular verb selects must be arbitrarily marked in the
lexical entry of the verb.
III xj '. cl l:I!A,-'o1, - ll Oli IV 0-1 dl
66
(43) 2. General Past 1 syll: H L
2 syll: HL LL
3 syll: HLL LLL
Such a tentative conclusion, though, is implausible, for the
six forms in (43) have too much in common for it to be an
accident. Arnott points out that in general a "drop " is
imperceptible before L; one could ,as well say that the drop
occurs before all forms, and lessen the amount of arbitrari-
ness in the tone formulas.
Similarly, Arnott points out, it is only the first
syllable whose tone must be arbitrarily memorized, and the
distinction here is between H and L. Modifying Arnott's
notation slightly, we could summarize both 2-syllable tone
patterns as BL. B here is a variable that ranges over H and
L. Thus the abbreviation stands for HL, or LL, which, as we
noted, is indistinguishable from simply LL.
If we adopt this formula for the 2-syllable forms, the
Well-formedness Condition accounts for the tones of the
3-syllable verbs starting with the same formula. That is, if
the first two syllables of yjvjse were treated as a bi-sylla-
bic verb, the third syllable would automically become L (see
44).
(44) y e v e sre
H L
Through the use of the B variable, we can generalize
across the L and H tonal forms (as Arnott points out):
through the use of the Well-formedness Condition, we can
generalize over the 2- and 3-syllable cases. One would hope
that the 1-syllable forms would be expressible by means of the
same formula, for then we could conclude that the Tiv speaker
need memorize only one tonal formula for the General Past.
The formula "BL works, we observe, for the L-verb, since
LL = L in effect; however, for the H-verbs this formula
predicts HL (i.e., Falling) rather than H, the correct form.
In a sense, the difficulty has been created by the Well-formed-
ness Condition, which causes the L to associate with the
single vowel.
(45) va
H L
The theory now requires us to set up a language-particular
rule to delete the L in this context.
There are two reasons to believe that this is, in fact,
a language-particular rule and not a failing in the theory.
First, recall that in the preceding example from Mende, the
same melody HL does in fact map onto a monosyllable as HL;
thus the simplification to H is certainly not unversal. In
fact, we have seen the same thing in English ((27) above)
where HL does not automatically simplify.
!I I d 4I -hI' IK;o
Secondly, we need a process to delete the L toneme in
(45) anyway, assuming that BL is the correct formula. As
Arnott suggests, and McCawley argues, the Drop is really a
Low toneme. Independently of whether such. a claim is true in
all cases, it is superficially true that a pitch drop does
occur on the second of two Hs when they are separated by an L.
Thus an unattached L (diagram (45) if the dotted line were
not present) should cause a drop in the pitch of the next
tone if that tone is an H. In fact, no such drop occurs;
therefore the L has been delted. In sum, the fact that the
Well-formedness Condition in this case forces us to draw the
association line in (45) is not unfortunate in any sense, for
we need a rule to delete the L here independently of any con-
ventions like the Well-formedness Condition. In fact, it is
the inclusion of the Well-formedness Condition that makes the
Deletion rule simple to state, as we shall see.
Our rule, then, will be (46), Fall-Simplification.
(46)V V
which abbreviates+A
H +1H L H
With the inclusion of this rule, all the forms in (42) or (43)
fall under the single formula 'BL.
Our formula needs one more modification before it can be
adopted finally. With a theory of downstep as described
below in Chapter 2 (Section 1), we expect a downstep or Drop
toneme that has the properties of the sequence "H" (we have
69
already seen this briefly in Figure (19) above). If we call
this toneme "Drop" or D, then the formula written 'BL actually
stands for DL (when B = High) and LL (when B = Low); again
using a suggestion of Arnott's, we summarize these two formu-
las as B*L, which now replaces the formula *BL.
To repeat: B is a variable that ranges over H and L;
B* is a variable that ranges over D and L;
in particular, B* = D if B = H.
Let us now look at the rest of the Tiv verbal tone
patterns, a bit more rapidly. It is not our intention to
prove that in all tenses, the formulas reduce to one, as they
did in the General Past. To succeed in doing so would, it
appears now, require a fairly powerful set of rules, and to
claim that the learner of Tiv prefers to learn, say, two basic
formulas for a tense and stay with formally simple tonological
rules permits us to posit an evaluation metric with which we
can come very close to our ultimate goal: that is, explaining
why this particular solution was arrived at on the basis of
the data.
For purposes of exposition, I shall give for each of the
remaining tenses only the tonal patterns (as in (43)) rather
than the actual spellings of the verbs with tones marked dia-
critically.
(47) 2. Continuous High Low
1 Syll: HL HL
2 Syll: HLL HLL
3 Syll: HLL HLL
TO
In the 1 and 2 syllable forms, a suffix is added to the stem,
so two tones are indicated for the "1 syll" forms, and three
for the 2. But the conclusion is rather evident here: there
is only one tone formula, HL.
(48) 3. Future High Low
1lSyll D L
2 Syll HL LL
3 Syll HLL LLL
On the basis of the 2 and 3 syllable forms, the tone formula
should be BL; given this formula, in fact, the 2 and 3 sylla-
ble forms work exactly as expected, and so does the Low 1
syllable form. As in the general past, the problematic form
is the 1-syllable H form; here, the sequence HL reduces to D.
But given the revision that the General Past formula for
H-toned verbs is DL, the revised form of (46) is D L + D.
Using the features suggested on p. 9tL eloc (and identifying
the Drop toneme D with Mid tone), we write the revised Fall
Simplification rule (49).
(49) Revised Fall Simplification Rule
{D3L -+ D, or [+High] [+Lof [+LO]
Thus, having simplified the already-needed rule of Fall Simpli-
fication (simplified in that one feature specification has been
eliminated), we are left With BL as the Future tonal melody.
I.j11k :,Id lfll ' 411 k~ll! P1 1 1k e s" I :! , J!4* T
Ti
(50) 4. Habitual 1 High Low
1SylliD D
2 Syll DH LH
3 Syll HHL LHL
The 3-syllable forms reduce to one formula, utilizing the B-
10
notation, BHL. Maintaining left to right mapping as in
Mende, the Low-toned 2-syllable verb works as expected, then,
for LHL is mapped as VA. The HL contour is simplified to
L HL
D, and, as we know, the sequence L D and L H are indistinguish-
able.
Nonetheless, this formula (BHL) does not work for the
remaining cases, nor is there a simply-formulatable rule that
can derive the correct surface form from underlying BHL.
Rather, we posit the melody B*H for the 1 and 2 syllable
forms. This in turn requires positing a complementary contour
simplification rule (51).
(51) LH Simplification
L H+D
(51) 5. Habitual 2 High Low
1 Syllable D ID
2 Syllable DH' LH
3 Syllable DHH ?
Maintaining the same principles we have seen so far, the
tone formulas for this form are straightforward. Arnott notes
that it is not clear to him whether there is, in fact, a
III l!vI 111 "E'll :111 LJ Ptllh dll i I - 1 ,1411m, -o - ., -JI ml! :j - I N'q. I
72
3-syllable Low form in the Habitual. The other tri-syllabic
form inexplicably does not have a following ; aside from
that curious feature, the five forms cited above reduce to the
formula B*H'. The reader will recall that for the Low-toned
monosyllabic form, this formula gives LH and that this re-
duces to D by (51) LH Simplification. If in fact Arnott is
correct about the absence of downstep following the tri-sylla-
bic form, then either a minor rule must be posited to delete
the Drop there, or two formulas--B*H and.B*H--must be posited.
At present, these two cannot be compared on empirical grounds,
and our evaluation metric is not subtle enough to decide for
us.
(52) 6. Habitual 3 High Low
1 Syllable DH DH
2 Syllable DHH LHH
3 Syllable DHH LHH
Again, as in the continuous form above, the 1 and 2 syllable
forms have a suffix, and so an extra tone-bearing element.
With no further ado, this form simplifies to the melody B*H#H.
(53) 7. Habitual 4 High Low
1 Syllable HL HL
2 Syllable HHL HHL
3 Syllable HHHL HH(H)L
In this form, the suffix /n/ appears to bring and bear its own
tone; thus the + in the tonal formula. We find, then, forms
like vaan, dzaan, ungwan, vendan, yevesen, ngohor. I shall
tentatively presume that the.suffix, preceded by a phonological
A H 110 1 'r Ik. 1 1 io 4J
73
boundary (represented here by +) has a corresponding toneme
(L) and the boundary on the phonological level is matched by
one on the tonological level. Thus we set a pattern like:
(54) yevese+n
H +L
But, with this assumption, the tonal formula for this form is
just H#L.
(55) 8. Past Habitual High Low
1Syll D +L D +L
2 Syll DH+L LH+L
3 Syll DHH+L LHH+L
In this form, we have the same appearance of a Low-toned
suffix as in the form preceding. Again, with the set of
simplification rules posited above, these six forms reduce to
the formula B*H#L.
(56) 9. Recent Past A High Low
1Syll D D
2 Syll DH LH
3 Syll DHL LHL
The 3-syllable forms, considered alone for the moment, reduce
immediately to the familiar formula B*HL. This formula,
however, does not apply to the 1 and 2 syllable forms, which
derive from the formula B*H. We shall not posit any ad hoc
rule to derive one from the other; such a move does not, in
fact, seem possible in this case. Some sort of parenthesis
notation may well be available for these lexical melodies:
the two would then be expressed B*H(L) 3 , where the subscript
"3" indicates that the L is chosen when the word is tri-sylla-
bic.
The next two tenses have the same tonal melodies, a
puzzling set of tones:
(57) Recent Past B/ Subjunctive High Low
1Syll H H
2 Syll HH HD
3 Syll HHL HHL
What strikes us as immediately odd is that there is no tonal
distinction between the two columns for the 3-syllable forms,
but there is for the 2-syllable forms. On the basis of the
3-syllable forms, we would rather naturally expect a melody
HHL; this generalized correctly to the Low-2-syllable form,
but not to the High-2-syllable form or the 1-syllable forms.
The 2-syllable forms, considered along, would derive from only
one of two possible melodies, given our assumptions: HHB or
HBH. Neither of these, however, generalize naturally to the
1- or 2 syllable forms. Adopting the parenthesis notation
that seemed plausible in the previous form (Recent Past A),, we
see that the melody HHB that is appropriate for the 2-syllable
forms can extend to the 1-syllable forms thusly:
(58) H H (B)2
When the subscript option is not taken--that is, in the case
of a 1-syllable stem--the formula is HH, which is the same as
75
H on one vowel, as required.
Pursuing this, we would attempt to see if the formula
(58) has any natural relation to the formula for the 3-sylla-
ble forms, HHL. The answer seems to be clearly yes; the
relationship may be stated as in (59).
(59) Recent Past B/Subjunctive
H H B
L3
This shall be the formula we posit, then, for the Recent Past
B and the Subjunctive,
(60) 12. Imperative High Low
1Syll H H
2 Syll HL LH
3 Syll HHL LHL
This is the last of the 12 tenses to be considered, and the
one for which we can offer the least satisfactory explanation.
Given our rules so far, the formulas underlying the 1 and 3
syllable forms is immediately dictated to us, fortunately:
the 1 syllable forms are simply H (or conceivably HH, ebut
there seems to be nothing supporting this); the 3 syllable
forms are derived from the formula BHL.
The problem is the 2 syllable forms. As Arnott points
out, the generalization here, if there is one, is that the
tone of the form is the "Base tone" (b) followed by its
opposite--thus L follows H, and H follows B. If we accept this
I I j I i I I I [;!!. ,: 1 1;011.1 411 [1 I 1 li 111I I IN ,i
hi dli
76
proposal, then the 2-syllable formula is, in Arnott's terms,
B-opp (B).
This is less than satisfactory for two reasons. First,
it is the first and only time such a device--"Opposite"--need
be introduced in this statement of the Tiv tonal system. From
this we could not conclude, of course, that a notion like
"tonal opposite" does not occur in tone languages in general;
in fact, we shall see a clear example of such a thing in Igbo
below (Chapter 2). Thus, while it is not impossible as a
solution, we would hope to see more reason why the "opposite"-
function could play a role in Tiv.
Second (and leading out of the first point), all the
other forms showed some coherence among the 1, 2, and 3 sylla-
ble forms; the adoption of the "opposite" notation breaks this
generalization.
The question, then, is to determine when a language-
learner, presented with data as in (60), will memorize inde-
pendent melodies for each syllablic structure (61), and when
she or he will adopt a simpler melody supplemented by tonal
adjustment rules. To the extent that the tonal adjustment rules
necessary for the second approach seem very complex, the
independent memorization position is supported; to the extent
that they are fairly simple, the second, general melody
approach is supported. However, the notion of "simple" is
internal to a theory of tonological rules, not external to it,
and so we shall not be able at this early state to reach a
S ;. Ii 4 1 1-j I I .I I I,-A 1
77
conclusion about the actual form of the Imperative in Tiv.
(61) Tiv Imperative a la memorization:
1 Syll: H (essentially
2 Syll: B + opp(B) Arnott's
3 Syll: BHL solution)
If we attempt to fit all these patterns into one melody, we
observe that the 3-syllable formula, BHL, predicts the correct
form for the 2-syllable Low-toned verb, for the melody LHL
spreads over a bisyllabic word as L it, which becomes L D,
which is (we have observed) indistinguishable from the reported
LH. However, such an approach does not work straightforwardly
for the H-toned bisyllabic forms, for HHL would give HD,
rather than the required HL. Sdppose we could resolve this
problem; the formula for the 1-syllable forms could (rather
trivially) be derived from the .- syllable formula with paren-
thesis notation, leaving us witi the Imperative Formula as in
(62):
(62) Tiv Imperative with Readjustment Rule
(B) 2 3 H(L) 2 3
The readjustment rule that is necessary, as we have observed,
is for the H-toned bisyllabic form, as in (63).
(63) Minor rule for H-bisyllabic imperative:
V VI N
H &L
11  Ii 4 ' i 0 iI II N H,I: .,
To repeat, the adoption of solution (61) or (62) or some
intermediate solution .must await further theoretical work.
To sum up this section, we have considered in some detail
the verb tone system of Tiv, an African tone language, indica-
ting how the tonal melodies must be described at a particular
tonological level in the grammar.
5. Floating Tones
Next we will consider the nature of "floating tones," a
device that has proven useful in working with tone languages
but whose theoretical status has always been suspect. A
floating tone is, in essence, a segment specified only for
tone which, at some point during the derivation, merges with
some vowel, thus passing on its tonal specifications to that
vowel. This ris, in any event, the traditional view; and this
traditional view, framed within the standard theory, fixes the
floating tone as one of the segments, and therefore linearly
ordered amongst all the other, more completely specified,
segments of the phonological representation.
Thus it has been suggested that certain affixes are
purely tonel--Bird (1966), for example, posits a floating L
tone for the Bambara definite marker on nouns.
We shall say--and we shall see--that floating tones are
melodic levels, much as in the previous section, that map onto
the syllabic structure in a slightly more complex way than the
Tiv forms. Often floating tones associate after some
79
phonological/tonological rules have applied, though this is
not always true.
Igbo presents several interesting cases of floating tones,
which we shall look at in detail in Chapter 2. For the pre-
sent, let us consider the general outlines of their behavior.
We will look at the pre-verbal floating H tone, a mark
of subordinate clauses. In Chapter 2 we will consider in de-
tail when the H tone occurs; for now, we will content our-
selves with observing its behavior when it is there. If a
tone "floats" when it has no vowel associated with it, let us
say that the process of associating a floating tone is "dock-
ing. "
We shall see four separate cases of "docking" of this
preverbal H tone--four kinds of docking that are predictable
on autosegmental grounds, and on no other.
To describe these effects, we must review some of the
basics of the structure of Igbo. It is a Subject-Verb-Object
(SVO) language; singular subject pronouns (though not plural)
have two forms, non-cliticizing or Strong, and cliticizing or
Weak. The cliticization of the subject pronoun, however,
includes a syntactic movement. In certain tenses, that is,
the verb stem is normally preceeded by the prefix /a+/ when
the subject is a full noun phrase (NP). Thus, if the subject
is anyi ('we,' not a clitic) and the stem is /za/, we find
(64).
(64) anyi azaa ala
we sweep floor
'we swept the floor'
8o
/za/ = sweep
(stem followed by
suffix /+a/)
If the subject cliticizes, the prefix /a+/ disappears, because,
we shall see, the cliticization puts the clitic subject into
the /a/-prefix's position. Thus we get (65).
(65) 6 zaa ala 'he swept the floor' /o/ = he, she
We may formalize this operation as in (66),, illustrated in
(67)
Pronoun Verb
Weak [Prefix
1 2 +
'0 1
Subject Cliticization
S
INP V
Pro VP
Weak Prefix Stem. Suffixes
(66)
(67)
81
Such a rule must be a local rule, in Emonds' sense, given the
general structure-preserving framework (see Emonds (1970) and
(forthcoming)). Therefore, if any element could appear in
between the Weak subject and the prefix position, it would
block this cliticization process, and thus keep the /a/ prefix
from disappearing.
The /na/ relative-clause marker does precisely that. In
a structure such as in (68), the clitic subject stays where it
starts, and it cooccurs with the /a/ prefix.
68) tP
'the market he should
$ have left..
NP
VP
The pictre in (6) is imprtant; roud t il rvov
ezhya,0 n'AL
f IW
mahket lit /ea-vt
The picture in.(68) is important; around it will revolve
our argument. The tone attached there with the Relative Clause
marker /na/ is the floating tone H. We shall see that under
other circumstances., that same H tone docks on other vowels,
depending, we might say, on what is "closest" to it in derived
structure.
A AMAA A
82
We have seen first, then, that the /na/, when present, is
H-toned. The/na/ may be not present for two reasons: first,
in a relative clause, the /na/ is optional; thus we may
simply delete it and see what the derived structure is. As we
observed in Section 3 above, when discussingtone stability, we
expect the H tone to appear elsewhere. Second, and quite
differently, while the /na/ always marks a relative clause when
it appears, the H-tone appears in other subordinate clauses
indicating adverbial dependency (that is, in sentences trans-
latable as "Lest X happen,..."). In this second case,, the H
tone is not a remnant of a deleted /na/, for the /na/ could
not. have been there in the first place.
In both such cases, when the subject of the clause is a
normal noun (non-cliticizing, that is), the final tone of the
subject is raised. For example, the noun oni (a type of yam)
shows up as cntk in the relative clause "onu rrerer" (r r tr6 is a
comples predicate meaning 'is rotten'). Here, as elsewhere in lgbo,
the contour tone LH -generally simplifies to what is called a Mid tone
in Igbo, but is precisely the same as the Drop tone (D) in Tiv. Its
tone is slightly lower than the first preceding High tone. But since
any High tone is slightly lower than a preceding High tone if the two
are separated by a Low tone, High andflid (or Drop) are indistin-
guishable after a Low tone. Thus in this position, LL nouns
become L+Rising or L Mid; HL nouns become H+Rising, or more
commonly H+Mid, as in (69).
83
/,I I A~(69) azu rere ere "the fish that was rotten"
azu 'fish'
Nouns that end in a High tone do not change; see (70).
(70) amu rere ere anu 'meat'
akwha rere re akwh'a 'eggs'
In all these cases, the /na/ could have been present, captured
the High tone, and then these nouns would have displayed their
isolation tones as cited. In the following sentences, the na
may not appear, since these are not relative clauses, and the
same subject tone raising occurs as noted above.
(71) Khwachie uzo
Shut the door lest. ..
aghu egbuo eghu /aghu/
leopard/kill/goat 'leopard' (a) HH
at" akhu/ke/
rat eat / palm kernels 'rat' (b) LH
'U""'Z"4 ataa ya /uze/
squirrel eat them 'squirrel' (c) HL
epwo ataa ya. /egwo/
monkey eat them 'monkey' (d) LL
If in one of these "lest..." clauses, as in (71), the subject
were a weak cliticizing pronoun, it would undergo the clitici-
zation expressed in (66) and (67) above. Looking at the tree
in (68), we see it would then be to-the right of the H tone,
I kl lll 4111:14,W441 I i I il all A dh .:.i
iflUIIU t
in a certain sense; in as much as the cliticized pronoun par-
takes of the verb stem's L-tone just as the /a/ prefix in (68),
the H-tone docking rightward now will give rise to a Falling
tone! All this is clear geometrically: see (72).
(7-2) a
NP VP
V~NP
Pro Pre stem suf
Weak
gbu eghu
H LI
he kill leopard
'lest he kill the leop'rd'
G gbuo eghu
Lastly, when the relative clause or subordinate clause is
constructed appropriately, the H-tone can dock right out of
its clause. Thus, if the subject is a clitic, and moves by
rule (66) over the floating H (much as in (72)), the floating
H may then dock leftwards onto what is to its left. If we
have a relative cluase where the head deletes into object
position in the relative clause, we get a situation as in (73),
and one can say either (74a) maintaining the /na/, or (74b),
where the floating tone docks leftward onto the head mgbe.
(73) NP
NP S
NP VP
V
Pre Stem Suf
mgbe zu ia anu
LL L HH
85
mghe ozula an,...
'whenever he buys meat,
(74a) mgbe o naaz\ula anu.
(74b) mgbe ozula anu
In summary, then, the floating H tone appearing before the
verb in certain tenses in Igbo docks onto various vowels
depending on the derived syntactic structure: the end of the
subject of the clause it is in, or the beginning of the sub-
ject clitic pronoun of that clause, or the grammatical parti-
cle /na/, or the head of the relative clause. The further
behavior of this floating tone, and its interaction with other
processes in Igbo, will be covered in detail in Chapter Two.
kl I I I 'It III lbu ' I I h i- 61 ; I lv I -- ! -j J F- ". :- vIf
86
6. Automatic Spreading
The fifth and last argument for autosegmental phonology
here comes from the phenomenon of bidirectional spreading and,
we would suggest, its ungoverned nature in these cases--that
is, the spreading is not due to a specific phonological rule,
but rather to the geometry of autosegmental representations,
and its Well-formedness Condition (or WFC as expressed above
in Section 2.3).
It should be clear by now from the examples in each sec-
tion how the Well-formedness Condition creates the spreading
of tonemes over various syllables automatically; we have seen
examples from English, Mende, and Igbo. In this section, I
will look at a particularly interesting example, one in which
the autosegmentalized level is not tone, but rather nasaliza-
tion. The language is Guarani; the data and much of the ini-
tial arrangement of it comes from Lunt (1973) and especially
Rivas (1974,75).
Let us begin by considering the forms that morphemes may
take. We observe in (75) that there are basically two forms
for the affixes /no/, /ro/, and /i/: each have a basically
nasal form and a basically oral form.
(75a) ndo + ro + hahu + i (oral stem)
b) no + ro + he du + i (oral stem with nasal
in it)
) a r 4 M+ nupA+
c) no + ro + nupa + I (nasal stem)
87
The prefix nI thus alternates with do, as oes ro with ro,
and i with i. Which alternate appears is determined by the
stem, the morpheme which contains the accent in the examples
above. In general, a vowel with an accent in Guarani is also
one endowed with a nasality melody or specification. That is,
if we say nasality is autosegmentalized in Guarani, we specify
that the nasality melodies (essentially just Oral or Nasal)
are initially associated with accented vowels, just as we saw
in Section 2.2 that the tone melody in English was initially
associated with the accented vowel. In addition, true nasal
consonants are endowed with an inherent Nasal specification on
the autosegmental level (these true nasal consonants are m, n,
W
n, 5 , 9). We shall represent a stem as in (75a) by (76a);
a stem as in (75b) by (76b).
(76a) h a i h u (76b) h e D u
o N O
The symbol "D" represents all the normal feature specifications
of an /n/ except for nasality--raised tongue, and so forth.
When associated with a [+nasal] autosegment, then, the two of
them are an /n/. "0" symbolizes [-nasal]; "N" symbolizes
[+nasal] (read "Oral" and "nasal," respectively). As above,
the star * indicates an accent; corresponding autosegments are
accented (starred) also.
I 61kl I illlkI4 I ' :l :t., R
~l. Iii
88
The Well-formedness Condition must apply to the forms in
(76). The form in (76a) becomes all oral, as in (75a); so
much is correct. The form in (76b) becomes the form in (77),
which would be transcribed as henu non-autosegmentally.
(77) h e D u
N t
The empirical difference between the form created by the
Well-formedness Condition and the actual form is that the
Condition does not create forms like /nd/, the prenasalized
stops. This is not, however, a bad result; the prenasalized
stop is a complex configuration, and the universal theory
should not produce them automatically. Rather, the language
particular rule (78) creates the form as in (75b) from the
form in (77). When we apply (78) Postoralization to (77), we
get the correct form.
(78) C\V Guarani Postoralization Rule
NO0
(The reader will recall that the dotted line in an autosegmen-
tal rule indicates the .addition of an autosegmental line by
that rule.)
With this autosegmental interpretation of the data, let
us derive the forms in (75) precisely. We note that when the
Well-formedness Condition applies here, if there is an ambig-
uity as to whether a starred or an unstarred autosegment
I4f-h llk! . Ei1 4 . ,D 10- . ,1,,,-n-
89
spreads, it is the starred element that does.
* * *
(y9) Do + ro + haihu + i Do + ro + heDu + i Do + ro + Dupa + i
N 0 N NO N N N
(79) represents the underlying forms. The Well-formedness
Condition changes these to (80); rule (78) Postoralization
creates (81), the correct output (compare (75))
(80) Do + ro + haihu + i Do + ro + heDu'i + 'Do + r + Dupa
N 0 N NO N N Nif / *
(81) + + thu + i Do + + heu + i Do + ro + Dupa + i
N 0 N NO N i
As expected, we have seen that prefixes and suffixes adopt
their nasality from the stem they are attached to. These pre-
fixes are found with the usual + morpheme. Note that this
bidirectionality of spreading is expressed only with addition-
al specification if handled by a phonological rule; within
autosegmental theory, the bidirectionality of the spreading is
the result, in a sense, of the fact that there are prenasal-
ized stops! That is, the existence of the prenasalized stops
required the autosegmental analysis, which in turn brought with
it bidirectional spreading. In summary, borrowing from Rivas
I I~I! ill
90
(1974) #8.3: suffixal nasality spreading is as in (82).ll
(82) Stem + suf-+ stem + suf (nasal)
Stem + suf -> stem + suf (oral)
Stem + suf --> stem + suf (nasal-oral)
There are also suffixes separated by word-boundary (#)
endowed with their own nasality-specification, according to
Rivas (1974,75), as in (83).
(83) ## Stem # Suffix ##
## aNas # SNas ##
We see in (83) what has been only implicit up to this
point: .the universal convention that corresponding word-
boundaries, but not +boundaries, are connected by association
lines.
Another logical possibility is predicted by the notation
so far: why couldn't there be an accented suffix--that is, one
with its own nasality melody--but one which has only a +bound-
ary, not a #boundary. With a +boundary, there will be no
association line, and the nasality autosegments of the stem
and the suffix will interact. We predict, then, the follow-
ing type of suffix--one which meets the description given by
Rivas for /+r 4/ suffix.
/+re/, endowed with an oral melody on its stressed vowel,
gives [irure], as in (84). Autosegmental theory does not
predict whether the second r is nasal or oral; when surrounded
91
by vowels with conflicting specification as in (84), such a
determination is not possible.
(84) iru + r
* * -
1
AC
N + 0 from [irul
When /+re/ is placed on a nasal stem with a nasal con-
sonant to the right of the accent (such as /mena/) we get the
correct result as illustrated in (85), a remarkable form.
(85) B e D a + r e BeD a + r e
N NN 0 N NN 0
n/[me dare]
7. Conclusion
We have traveled a long way since our early remarks
about the abstract geometrical shape of phonological repre-
sentations, but in another sense we have surely never left
that topic. In these five general cases reviewed, phenomena
that are puzzling on the standard view of phonological repre-
sentations become quite tractable and reasonable when viewed
from an autosegmental perspective: and we see, furthermore,
connections berween these that we never saw before. If we may
draw some morals from this, perhaps they are that advances in
phonological theory may start from an interested attitude
towards low-level articulatory facts; that if they are interes-
ting, they go beyond these superficial facts to unexpected
92
phenomena; and most importantly, that we do not begin our
research with an understanding of the nature of the most
elementary linguistic observables. The most astounding revela-
tions may be those that change our conception of what we
thought were the observables, either in phonology or seman-
tics. We would not restrict our attention to constraints on
rules--phonological, syntactic, or semantic--at the risk of
missing the very nature of the items involved.
l i 4 1.6C Ai jI I k!b-. N-, Iw.I EuI, J jrn. I
93
FOOTNOTES FOR CHAPTER ONE
1. The requirement of "short" here is meant to avoid the
possibility that a long vowel might actually be two
successive short vowels, each with a level tone. This,
of course, does happen frequently; it is, however, irre-
levant to the present context. The situation is much
the same for pre-nasalized stops: in some languages,
these stops are clearly one segment, even though they
have two successive feature-specification for nasality.
2. Leben (1973) wrote, "Suprasegmental tones are by defini-
tion independent of any segments--rather than being
expressed as features on segments, they are features
on larger linguistic units" (p. 26).. Compare this with
Troubetzkoy (Principes de Phonologie, p. 95-6): "Les
particularit. es phoniques qui forment des oppositions
distinctives dans les diverses langues peuvent 'tre
reparties entrois classes: particularites vocaliques,
consonantiques, et prosodiques. Les phonemes vocaliques
consistent en particularites vocaliques distinctives et
les phonemes consonantiques en particularites consonan-
tiques distinctives. Par contre il n'y a aucun phoneme
qui consiste exclusivement en particularites prosodiqueS.
Ces particularites sont plutt liees, selon les langues,
a un phoneme vocalique determine , ou a un phoneme con-
sonantique determine, ou enfin, a toute une suite de
phonemes.
* : It I SUMC1 illdjI mr1, 4 I m i m ..1, ,1qt :!., ,olv Is 'I;u
94
3. One could, in this regard, accept the suggestion of Woo
(1969) that all contour tones are necessary such conca-
tenations. The logic of the situation would not be much
changed; we would simply not need to investigate the
particular language in question to come to the conclusion
mentioned in the text.
4. We shall--more for the sake of being definite now than
for anything elese--assume a theory of tonal features
that is defended at more length in the following chapters.
We adopt two binary pitch features: + Highpitch and
+ Lowpitch.
+ Highpitch.
Lowpitch] will represent High tone (H)
-Highpitch.K Lwpitch jwill represent Low tone (L)
Highpitch[+ Lw pitch] will represent Mid tone (M)
the fourth possible combination, H Highpitch], will[- Lowpitch J
tentatively be excluded as a universal principle. The
distinction between tone and pitch will be drawn in
Chapter Two below.
5. Strictly speaking this is not true: the one case we
.r AB
cannot rule out is: aXb. There are, however, good
reasons not to want to rule this out, for if we could
rule it out we would have given an inherent sense to each
level, not just total ordering. In fact, in every
95
possible linguistic case, there will be at least three
segments on some line, and there WFC above works correct-
ly.
6. Again a proviso must be made. (32) is not in violation
of WFC; in general, we find, when we inspect the
definitions provided here, that if segments on the end of
isome L are not associated, WFC is not violated. In
fact, this odd formal characteristic appears to be
exactly reflected in the behavior of certain floating
tones. That is, to the extent that WFC differs from
(24) above--WFC making curious predictions--it is WFC
which is more correct.
(32) Q A B
b a
7. The phenomenon of stability is the subject of my "Tone
Melodies and the Autosegment, " from which this section
draws heavily. Further examples not mentioned here are
treated there (it appears in the Proceedings of the 1975
African Languages Conference, in the Ohio State Working
Papers in Linguistics No. 20).
8. In particular, I claim that it is not true at levels in
the grammar where phonological and tonological rules
apply. Leben's system calls for there to be sequences of
/,/ /identically-toned vowels (CVCVCV) after his "tone map-
ping"; thus, for him, the Obligatory Contour principle
R AtL I LJ L
96
holds in the abstract levels but not at the superficial
levels. I shall suggest that rather the opposite holds,
in Chapter 4.
9. One could imagine logically the position that each verb
has its tone marked on it in the lexicon, arbitrarily
selecting one of the tenses' tonal patterns as the under-
lying one; the other tonal patterns might be then derived
by tonological rules sensitive to the semantic mood and
tense of the clause. While not impossible within current
phonological theory, such an approach does not in fact
work.
10. Here is the first case where the Obligatory Contour
principle is violated, for HHL does not satisfy the
principle.
11. See Section 6 of Chapter 3 below for a further discussion.
97
CHAPTER TWO
IGBO TONOLOGY
1. Introduction
Our course so far has been to consider a series of argu-
ments, each of which focused on a recurrent phenomenon in
phonology that points toward autosegmental representation.
Adoption of autosegmental tiers, we argued, resolved these
fundamental problems in phonological theory.
In extending autosegments to linguistic theory, we have,
it seems, made it an offer it could hardly refuse, once aware
of the problems discussed in Chapter One. Still, even if we
adopt autosegmental representation, we will want to know what
it is that we have just taken on. What will a grammar look
like now? Our most fervent hope would be that the inclusion
of autosegmental levels will simplify our conception of phono-
logical rules, rather than render it more complicated; but
this cannot be decided in advance.
There is only one way to settle the question, and that is
to investigate in depth a system that has a non-trivial auto-
segmental component. This chapter is devoted to such an inves-
tigation, wherein we shall look at the tonological system of
Igbo, a Kwa language spoken in southeastern Nigeria.
This chapter is ,intended, at the same time, to begin the
development of a more articulated autosegmental theory. Auto-
segmental theory as presented in Chapter One is a family of
theories, rather than just one. Because it represents such a
98
fundamental change in phonology, the ramifications of autoseg-
mentalism lead to a whole series of new questions when we try
to implement it in a specific linguistic theory. Just what
features may be autosegmentalized, for example? In a given
language, may a certain feature be autosegmentalized at one
point in a derivation, but not at another point? Once we be-
gin to consider the questions that arise, each one virtually
leads to ten more.
This is, I believe, a highly desirable situation for
phonology to find itself in. For it is only against the back-
ground of a choice among hypothesis that the truth may emerge;
and only when alternative hypothesis may be expressed, enter-
tained, and argued against can we make a scientific advance.
The point, then, is that although the decision to revise
phonology to include autosegmental levels opens, in turn, a
series of further question, the autosegmental proposal is no
sense vacuous; it is, rather, theoretically fruitful. It
leads us to see, pose, and answer questions that we were not
aware of before.
The remainder of this thesis is devoted to my own view of
the answers to some of the questions that autosegmental theory
raises. It would be fair to say that of the number of lin-
guists who have begun to work within the autosegmental frame-
work, the major area of agreement rests within the proposals
made in Chapter One. From here on, we consider one autoseg-
mental theory among several.
III v 41, I h4t[ 1,444 !11 Ail " .. A[ 1:-.Ji , Si, j, A -!-I I -i 11 w,j F, !"
99
This chapter develops a number of specific claims. For
example, I shall argue that a representation as in (1) is
empirically distinct from that in (3), and that when a repre-
sentation like (1) occurs (produced, perhaps, from (2) by the
Well-formedness Condition), it cannot be reinterpreted as (2).
(1) C V C V (2) C V C V C V
M M
(3) C V C V C V
L 1k
Beyond that, in this chapter, I shall illustrate how
virtually the entire class of rules that are distinctly
autosegmental do show up in the autosegmental grammar of Igbo,
and we shall see that these rules interact in the interesting
ways we have come to expect of phonological systems. In sum,
we shall be primarily concerned with the nature of autoseg-
mental derivations and the rules found therein.
In Chapter Three, we shall turn our attention from the
nature of the derivation to the bootstrap operation that be-
gins it: the linking of autosegmental tiers, a process I have
elsewhere called "melodic association rules." Our attention
there will be focused on problems of accentuation and the na-
ture of the "star," already outlined in Chapter One.
Chapter Four will deal with the question of whether the
number of autosegmental tiers is constant throughout the
14J :1 -; 1 W II I 1 1, '1 .14 11 " R - I ': , ! -1 [, F', . ,
100
derivation. The rather speculative remarks made at the begin-
ning of Chapter One would suggest that the answer to this
question is "no"; we shall elaborate an autosegmental theory
consonant with this negative answer.
2. Guideposts
Our attention here is, admittedly, directed by our
interest in finding in the tonological facts of Igbo charac-
teristics that will tell us about the nature of autosegments,
rather than about just anything bearing on phonology in
general.
Considered in this light, the phenomena that will in-
terest us resolve into three classes:
(i) rules that insert or delete segments on one auto-
segmental level, but not the others;
(ii) rules that add or delete association lines; and
(iii) rules operating on a single autosegmental level
which would be inexpressible (and thus, in theory, impossible),
were the remaining features not autosegmentalized. In this
last group, only one example occurs; it is a very interesting
one, however, pertaining to the interaction of the cliticiza-
tion of /a/ with the docking of the floating H tone, a process
already briefly discussed above in Chapter One, Section 5.
We have already considered the deletion of tone-bearing
vowels in an autosegmental framework in Chapter One, Section
3. We shall see below a case of what we would expect equally
: i11 1 ilI I i m : .4 1A 1It
101
to find: deletion of a (vowel-bearing) toneme and consequent
reassociation of the now-stranded vowel, as in (4).
(4) C V C V tone C V C V by C V C V
t/deleUion C
T1T2 2 2
We shall, furthermore, find a rather large collection of fair-
ly idosyncratic tonal melodies in the different "tenses" or
"verbal forms" of Igbo. Which of these underlying tonal
melodies is in fact appropriate is determined by the interac-
tion of semantic and syntactic factors in each sentence.
Negative sentences thus choose different tonal patterns from
affirmative; and the tone pattern of relative clauses depends
on whether the element that is relativized in the lower clause
is the subject Noun Phrase or not. Subordinate clauses quite
generally have different tonal patterns from non-embedded, or,
in Emonds' terminology, root sentences. We shall propose to
account for this much as we did for the case of Tiv in Chapter
One: the melodies are constructed by the morphological compo-
nent--in the lexicon--and inserted as any lexical material is.
We shall require, furthermore, that this process of lexical
insertion be sensitive to derived syntactic structure rather
than deep structure. This leads naturally to the hypothesis
that lexical insertion more generally occurs at surface struc-
ture, or at the end of the cycle. We shall return to this
hypothesis in Section 4. As we shall see, although the
A11; I T u 1 tII h IJ II
102
"surface lexical insertion" hypothesis is not a logical conse-
quent of our other conclusions, it does permit a much tighter,
and quite plausible, theory of possible tonal rules.
3. The Tonal Inventory
Igbo tones display a three-way contrast superficially,
and we shall adopt what is therefore the most "concrete"
analysis: there are three tones in Igbo, which we shall call
High, Lowy and Mid (or, as we called it in Tiv, "Drop"). This
three way contrast is illustrated in (5).
(5) in 'nose (b) f mI 'to crawl' (c)IiM i 'to bear
fruit'
The Mid, or Drop, tone illustrated in (5b) is a curious one,
however. Two successive Mids, unlike two successive Highs
or two successive Lows, are not at the same pitch; we see this
illustrated in (6).
(6) o d h u o k e 'tail rat' = rat's tail
odhu oke
H M ML
An understanding of this "downstep" phenomenon requires that
we draw rigorously a distinction between tone and, pitch.
Although virtually all of the ideas in the immediately follow-
I IlL II i ~I
103
ing discussion can be found in various places in the litera-
ture going back two decades, a strange misunderstanding, it
seems, in recent theoretical Africanist analyses has led to
many a pointless discussion of the supposed problems that the
existence of downstep causes for the theory of binary features,
phonological rules, and so on.
Tone is a characteristic of segments determined by a
small set of features (minimally two, but probably larger).
These tonal features may (or may not) be autosegmentalized.
We shall maintain the hypothesis that all features are binary.
Pitch, on the other hand, refers to a much more super-
ficial analysis of an utterance. It is only a bit more
linguistically abstract than the instrumentally-observable
fundamental frequency. Unlike pitch, however, the fundamental
frequency itself is affected by such factors as consonantal
characteristics (voicing, for example), vocal chord sluggish-
ness, and so on.
We shall return to the traditional notion of pitch regis-
ters; the word "pitch" is used here advisedly. We will con-
sider languages with two pitch registers, for simplicity's
sake, but languages may have more than two registers; Yoruba
comes to mind as an example with three.
The formal output of the pitch register 'component' is
not a segmental representation, but rather a graph like those
we have already seen--in (5), for example. The pitch registral
component will predict where the graph will be level, where it
1I RI !0 1 i J I I. f +
104
will dip and drop, and so forth, on the basis of the segmental
representation input to it.
Considering, as we have said, a language with two pitch
registers, like Igbo, we construct the pitch diagrams as in
(5) out of slightly more abstract pictures which include both
the High and Low pitch registers (see (7)),.
(Ta)
(Tb)
(To)
--- High
Low
High
Low
Low
The task of the pitch registeral component is, then, first to
determine what causes movements of the pitch registers--pul-
ling of the High Register down, for example, as in (8); and
second, to assign each syllable (or in general, tone bearing
unit) to a place on the chart. In general this will be done
by placing certain of the syllables on pitch registers, as we
have with the x's -in (7a,b,c) , and then drawing lines connect-
ing these fixed pitches. In ,tone languages, all syllables are
fixed on pitch registers; in non-tone languages, it may be that
i I:I I I! IId [l k 4 !1 ,1 , i
105
only certain ("accented") syllables are fixed on pitch regis-
ters, while unaccented syllables' pitch is determined simply
by the curve that connects the determined pitch of accented
syllables on either side of them (see (9) ) (This last
proposal concerning non-tone languages is due to Mark Liberman;
see Footnote 1 of Chapter 3.,) For now, we may rest with the
simpler case wherein each syllable must be placed on one of
the pitch registers or the other, as is the case in Igbo.
------- high High Register pulled
-lowdown
high
lol
CV CV CV CV
Some of the examples so far have suggested what is, in
fact, true: that throughout African languages, a very common
pitch registral effect is the "pulling down" of the High
Register illustrated in (8). This. Drop is rule-governed, and
we shall need a notation for it. Let us denote it as in (10):
(10) High Register +>-1 / .. . . ..
That is, we shall use "-l" to mean Dropping of the indicated
register. This drop is about one to two semitones.
106
We will specify for Chuhu Igbo that High and Mid tones
are realized on the High Register, and Low tones are realized
on the Low Register (see (9a)). Using the feature notation,
we may say that a toneme [aHi] is placed on the alHigh Register,
if we call the High Register the "+High Register" and the Low
Register the "-High Register." Whether or not such a conven-
tion is adopted, the point is that the feature High determines
which register a tone is realized on.
(9a)
High Register
~ ~ Low Register
H M L[+ L+ -ijhi1
The High-mid sequence is realized as in (7b). We need
minimally, then, a rule approximately as in (12) which performs
the High Register lowering. This rule is not a phonological
rule in the strict sense. It applies in the pitch registeral
component, and reapplies iteratively, left-to-right, as do all
rules in this component.
(12) High Register + -1 / [+hi] . +hi
I-lo 1+1
'H' 'M'
107
We must ask whether rule (12) could not be expressed more
simply, and by the same token whether it is not part of a
larger phenomenon. The answer to both questions is yes.
There is too much specification in rule (12). A Mid-Mid
sequence, as we have seen in (5b) and (7b), should trigger the
rule as well. The feature "-lo" in the rule made it therefore
too specific. Amending (12) to (13) by dropping this feature
specification, we observe that (13) produces the correct pitch
register modifications for all the examples we have considered
so far.
(13) High Register + -1/
[+hi] +hi I
H {M}
It is generally agreed that the existence of Drop tonemes,
like the Mid we have been considering, is intimately associa-
ted with the existence of the phenomenon of "downdrift," which
causes the second of two Highs, when separated by a sequence
of one or more Lows , to be slightly lower in pitch than the
first High. This is illustrated in (14).
(14)
CV CV CV CV CV
108
The lowering that occurs between the two Highs is precisely
the same amount as occurs between the High and the Mid in
la(5b). This intuition is justified when we observe that
this downdrift phenomenon is included in the operation of
downstep rule (13) by simplifying that rule--by dropping the
feature specification "+hi" on the second segment. Thus (13)
becomes now (15), our final form of the downstep rule for
Uhuhu Igbo.
(15) High Register -* -l / [-lo]
+10j
If we take an input like that in (14), rule (15) applies as
shown below in (16), lowering the High Register at the se-
quence H - L. This produces, as we see, the correct final
result.
CV CV CV CV CV
rule (is)
applies
Conclusion: From this discussion of the tonal inventory and
the operation of pitch registers, two general points important
to the tonal analysis of Igbo must be remembered.
First, the sequence Low + Mid is phonetically indistin-
guishable from the sequence Low + High. This follows
A i I I [1111! 1, 1 .-I I I I; I 4 a di 1. Ir 'k , I
. .1 .!uO I " i N It 4
109
immediately from what we have seen so far. The Mid tone and
the High tone differ Only in their specification of the fea-
ture "low"; and since it is the other feature, high, that
determines which pitch register the tone is realized on, the
specification of the feature lo can influence only the Register
Rule (15). Put differently, tonal feature specifications can
do only one of two things in the pitch system developed so
far: either determine which pitch register that tone is
realized on, or affect pitch registral rules. But the feature
low, the only feature which distinguishes Hicjh tone and Mid
tone (High tone is -low and. Mid tone is +low), has no effect
on which pitch register the tone is realized on; that is
determined by the feature high. Furthermore, if the preceding
tone is Low (with the features -hi,+lo) then Rule (15), the
Downstep/downdrift rule cannot apply since its structural
description is not met. Under these conditions, then, High
and Mid are distinct when they follow either High or Mid tone.
The second point that must be clear is what the phonetic
shape of a representation like (15) is.
(15) CV CV CV
H M
If a Mid tone is associated with two syllables, as in
(15), then these syllables will be level in pitch., The first
of the two will be downstepped with respect to the High-toned
ii |I M II II4 i &I
110
syllable in (15). In short (15) will produce a pitch regis-
tral graph as in (16).
(16)' b
(a) i-e.
H M CV CV C7
CV dV CV H \M
Looking at (16b)-, we see that this is the pitch contour that
would have been created by the sequence in (17) (compare this
with (15), repeated here).
(17) CV CV CV (15) CV CV CV (17 ) CV CV CV
I I IIV I I I
H M H H M H M M
The point, then, is that a Mid tone spread over two or more
syllables is phonetically indistinguishable from a Mid follow-
ed by a sequence of Highs, though it is distinct from (171
(see (6) above).
We shall find, throughout this chapter, that these two
surface neutralizations, though they may cause trouble to
linguists, do not present problems to the language learner if
the learner is in fact endowed with the autosegmental linguis-
tic faculties we are positing. Information about the general
systematic structure of the language makes it clear to the
language learner which form is occurring, as we shall see.
I I I l IIe| l III | M I:
4. The Derivation
Accepting the essentials of Chomsky (1965, 1970) and
Chomsky and Halle (1968), we view the structure of a deriva-
tion as in (18).
(18) P Deep structure. P characterized by phrase-
structure rules.3. application of
cyclic rules
n
n+1
application of
post-cyclic rules
P
,n
. application of
P r phonological rules
Each pair (PP i+l) corresponds to the application of a
linguistic rule, where, in particular, P. meets the structural
description and Pi+l satisfies the structural change of that
rule when P. is the structural description.
P . is a phrase-marker, or tree, with no lexical items
as terminal elements. Lexical items are inserted by the pro-
cess of "lexical insertion," whose place in the derivation we
return to in a moment.
Some phonological rules may be best viewed as applying
inside the lexicon, and the domain of the application of such
rules would be at most one lexical item (it might, of course,
I ll I In
112
be smaller--just one syllable, for example). There are other
phonological rules which operate across lexical items--there
may be an interaction, that is, between the phonological
material of successive words. Let us call this second sort
of phonological rule, which cannot be construed as a lexical
process, wide-domain phonological rules. The discussion below
concerning the ordering of lexical insertion in the derivation
will be concerned with these wide-domain (or sandhi) phonolo-
gical rules, rather than with lexical phonological rules.
The application of rules can be grouped into syntactic
and wide-domain phonological (which includes tonological),
with the application of all syntactic rules preceding the
2application of all phonological rules. Syntactic rules can
similarly be divided into cyclic and post-cyclic rules; some
would argue that the latter class is empty.
Our concern, as we have noted, is the section of the
derivation from P to Pr, the phonological part. We may notem r
that the insertion of the phonological information for each
lexical entry in the derivation must have been inserted by the
time of the fitst application of a phonological rule. We must
ask if we can determine the point of lexical insertion in the
derivation any more precisely.
In Chomsky (1965, 1970), Chomsky proposes that lexical
insertion consists of a set of transformations applying before
all other transformations--applying in (18) before the cyclic
syntactic rules noted there. For our purposes, there are two
113
empirical criteria for deciding where in the derivation lexical
insertion occurs. First, the phonological information which
lexical insertion provides must occur before P , as we have
m
seen, the first phrase marker in the phonological derivation.
The earlier phonological information is inserted before P
however, the more ad hoc conditions must be placed on the
functioning of the grammar to prevent interaction of this
phonological information with the functioning of syntactic
rules. No transformational rule, that is, is conditioned by
the phonological character of the items involved. We want to
find it no accident that phonological information becomes
necessary to the derivation, then, at the point where it in
fact becomes available to it. As I say, this is what we would
hope, in the best of all worlds, to find; it may or may not be
true.
The second empirical criterion for determining where
lexical insertion occurs depends on the type of contextual
information needed by the rules of lexical insertion. Again,
in the best of all worlds, lexical insertion will crucially
have to be stated either at essentially deep structure level
before the application of transformations--or at "shallow
structure," defined either as P or P in (18), or some -other
n m
well-defined linguistic level. That is, if the correct point
for lexical insertion is at deep structure, we will find evi-
dence that the context for lexical insertion must be stated
before the application of transformations; if the correct
i- i || II ,p, I I 1 ld '
locus is at P in (18), we would find cyclic rules which must
n
create the context for lexical insertion, but post-cyclic
rules would in turn destroy the necessary context. In each
case we would find a level in the linguistic derivation where
the contexts for lexical insertion were optimally stateable.
Again, if ours is the best of all possible worlds, we
will find that the two empirical criteria--one based on the
necessity for phonological information, the other based on
proper context--will point to the same level in the grammar
for lexical insertion. This, of course, need not be so; if
it were not true, one might, as a second best, hypothesize
that lexical insertion occurs necessarily at one level, but is
contextually defined by some other level, in a sense that
would then have to be made more precise.
Our conclusion based on the evidence from Igbo tone shall
be that the second best need not be settled for, that both
criteria are compatible with lexical insertion occurring at
the point labelled P in (18)--in short, that lexical inser-
m
tion occurs at surface structure.
Our argument goes like this: considering a very wide
range of Igbo tonal data, we are led to posit a set of about
ten simple autosegmental rules. These rules relate the super-
ficial forms with a more abstract level. The tones for nouns
at this more abstract level are simply the lexical tones. The
situation is more complex for verbs and their affixes. We
find that the underlying tonal patterns for the verbal system
,I I l iI IL!, 11, -1 IPHY '. 11 ii . -R,
115
are determined by a complex--and essentially idiosyncratic--
interaction of lexical, syntactic, and semantic factors.
We are led--much as was Chomsky in Remarks on Nominaliza-
tions--to propose a reason for the difference in character
between the well-motivated phonological/tonological rules, and
the essentially arbitrary rules necessary to derive the tonal
formulas for the verb and its affixes. The verbal tones, we
shall say, are inserted at the point of lexical insertion, and
thus are the product of the morphological component.
This lexical insertion, however, must be sensitive to the
derived syntactic structure--the structure after relativiza-
tion, in particular. We would be led by our second criterion
to posit lexical insertion at the frontier between the syntac-
tic and the phonological components of the grammar, or at the
end of each cycle.
One consequence of this is that there will never be found
phonological rules which have left traces in the segmental
phonology of the word-order that existed prior to the applica-
tion of some syntactic transformation. 3 For example, there
could not exist (the claim goes) languages like the following,
which leave the trace of the pre-passive-rule form on the seg-
mental phonology.
a) Suppose there was a language much like English as
we know it, with a rule palatalizing t -+ c before y. This
phonological rule, furthermore, applies before the passive
transformation. Thus the underlying form "Did she meet you"
'I 'll 1111 II . ir [ 1 i + --,
becomes, by this palatalization, "Did she meec you?"
If the passive transformation were to apply, this would
produce a derivation like (A).
(A) She past meet you underlying
She past meec you palatalization
You were mee6 (by her) passive
b) Consider another example, again an odd dialect of
English with a rule devoicing word-initial voiced stops after
a voiceless stop:
Devoicing [+stop] + [-vcd]/[-vcd ]##
+stop
Such a language would produce derivations like (B), if Devoic-
ing could apply before the Passive rule.
(B) You met Betsy Underlying
(a) [You met Petsy] Devoicing
(b) [Betsy left last night.]
Someone met Betsy at the station Underlying
Someone met Petsy at the station Devoicing
(c) [Petsy was met at the station (by someone)] Passive.
As I say, it is language systems like these that are ruled out
as impossible by the ordering of wide-domain phonological
rules by the ordering of phonology after syntax.
The question of the placement of lexical insertion with
respect to the application of transformations is a subject
deserving of more than these few pages here. Indeed, the
117
extent to which cyclic rules should be viewed as transforma-
tions rather than "lexical redundancy rules"--the term is per-
haps less than apt--is called into question if lexical inser-
tion is carried out after cyclic transformations. The remark
calls to mind certain recent suggestions made by Joan Bresnan
(1976), but they are equally reminiscent of the older proposals
of such diverse linguistis as Bowers and McCawley.
Some speculations on this issue are worth dispelling.
It has been suggested, for example, by Bresnan that there is
a clustering of rules that are (i) structure-preserving,
(ii) bounded, and (iii) lexically-governed (Bresnan suggested
this, for example in a public address (1976)). If this is
true, it could be explained, as she suggests, by saying that
these (cyclic) rules are not transformations at all, in
Chomsky's sense, but are rather lexical rules relating the uses
of the same or related lexical items. Thus there would be no
transformational relation, on this account, between the active
and the passive forms of a sentence.
If this was true, furthermore, my suggestion that lexical
insertion occurs after all cyclic transformations becomes tri-
vially true, since there would no longer be any cyclic trans-
formations.
The notion of "structure-preserving" transformation would,
the reasoning continues, be eliminable, since the "reason"
that the direct object is moved into the subject position by
passive--a paradigm example of a structure-preserving rule--is
-Il [ 1111 i woiE~ i ll H! ii V ".
118
simply that the phrase-structure rules determine the positions
into which lexical insertion may place NPs, whether these NPs
be logical subjects (as in active sentences) or logical ob-
jects, but now placed in subject position with a verb marked
in the passive voice.
Such a proposal, while attractive on the face of it, is
ultimately not acceptable. Even if we ignore the question of
how to deal with idiom chunks--while this has traditionally
been an argument for a passive transformation, Higgins (1974)
presents evidence that this is a non-argument--the claim that
rules cluster in the way noted above is highly dubious. In
fact, given the eight categories that the three binary dis-
tinctions listed above give, either seven or all eight have
attested members.
Structure-Preserving Bounded Lexically Governed
Str Pre Bounded Lex Governed Rules
No No No ?Topicalization (see note 1)
No No Yes "So-preposinig," "the more the
merrier -mbvement, (but see
No Yes No Subj-Auxiliary invetsion,
Verb-fronting in' German
Yes No No ?Topicalization (see note 1)
No Yes Yes "Evidemment"-movement in
French 'notsee e 2)
Yes No Yes (see note1)
Yes Yes No Stylistic Inversion (French)
Yes Yes Yes Passive, Raising tosubJect
Note 1: Whether Topicalization is a movement rule is un-
clear, so its placement in this chart is doubly unclear.
Furthermore, whether Topicalization (if a movement rule),
119
So-preposing, and so forth, are structure-preserving or not
depends on one's analysis of the putative Complementizer posi-
tion (see Bresnan (1972) for one position on this).
Note 2: This is the rule that changes (Que) Il est evidemment
venu to Evidemment qu'il est venu in French. See Dubuisson
and Goldsmith (1976) for an analysis of this rule.
The conclusion of this is simply that if lexical inser-
tion occurs after cyclic rules apply, it is not in the trivial
sense that Bresnan' s hypothesis would lead to. We maintain
that there are transformations that are structure-preserving,
that are lexically-governed, and that actually apply before
lexical insertion. However, we cannot pursue this topic here;
we return to Igbo and its tones.
5. Verbal Forms
The discussion so far in this chapter has been prepara-
tory and anticipatory; we turn now to a detailed analysis of
Igbo tone.
Most nouns in Igbo are bisyllablic, with the form VCV.
There are also a fair number of monosyllablic nouns of the
form CV, and some with more than two syllables. We shall
4restrict our attention, as Green and Igwe do, to one and
two-syllable nouns.
Monosyllabic nouns are all High in tone; bisyllabic nouns
may be High-High, High-Low, Low-High, or Low-Low. The tonal
A j l ll l 'li f I 1. 41 A Ii" A I
i Ihll~ M 41 '61 +I e
120
behavior of a noun is fully determined by its lexical tone,
in the sense that what tonal rules a noun undergoes is deter-
mined by its tone only and not by its non-tonal characteris-
tics.
Verbs divide into two categories, the Low and the non-Low
(or High). This distinction is a lexical one, and it fre-
quently occurs that the distinction is neutralized on the
surface. When the distinction is maintained on the surface,
the Low class is usually Low, while the non-Low class may be
either High or Mid, depending on the specific context. All
verbs are CV, except compounds, which are CVCV.
There is a highly versatile set of verbal suffixes which
mark various aspectual and other core semantic functions.
These verbal suffixes divide into two categories, which,
following Williams, we shall analyze as toneless and tonally-
specified (Class I and II, respectively). The toneless Class
I is much more common that the Low-toned Class II.
There are two verbal prefixes which occur in a fixed or-
der when they co-occur. The first is the optional relative
marker na, which marks the clause as being a relative clause.
The second is a (or its counterpart e when the verb stem
triggers tense vowel harmony, changing the a to e), which is
again a sort of aspectual marker. The presence or absence of
the prefix a, however, also marks which or two series, the I
Complete or the II Incomplete, the tonal pattern is selected
from. That is, the tonal pattern of a verb and its affixes is
121
determined essentially by whether the sentence is embedded or
not, and, if so, what sort of embedded clause it is. For each
of these types, though, there is a I Complete form, and a II
Incomplete form. The II Incomplete form, furthermore, is
marked underlyingly by the a prefix.
The tonal melodies of the verb and its affixes consists
of four components, each of which has a specific syntactic
locus (see Figure (19) below). The first is the floating H
tone that was discussed in Chapter One, Section 5. The second
is the tone of the a prefix; the next is the verb stem tone,
and the last is a suffixal tone.
The Chart (19) below summarizes our analysis of the tonal
melodies of the eight simple affirmative forms. As we have
observed, the tonal melody depends in part on whether the
clause is embedded or not. Non-embedded sentences are called
here "Root," followed Emonds' terminology (Green and Igwe call
these "Main" forms). Among embedded clauses, we must distin-
guish relative clauses and non-relative clauses. This is
simple; relative clauses are those embedded clauses which are
preceded by a head, or antecedent, noun; and in just these
cases the head noun will also semantically correspond to an
empty position in the embedded clause. We must furthermore
syntactically distinguish between relative clauses where the
function of the antecedent in the lower clause is as a subject
(as in the English the man who was here) and those where the
antecedent plays any other role in the relative clause (e.g.,
'I  II' ll il .'I M 10 ra 1'1 :
122
as in the man I saw -- ). We shall call the first type
"Relative/subject," and the second type "Relative/non-subject."
Note that there are no words in Igbo corresponding to wh-words
or a complementizer like that in relative clauses.
The Chart below is divided up into two major halves, one
for the I Complete and one for the II Incomplete. The numbers
I and II are Green and Igwe's; I have added the terms "com-
plete" and "incomplete" to suggest part of the semantic
difference between the two types of sentences. These English
words are perhaps more misleading than helpful, but there are
none better that I could find, and it is helpful to have some
mnemonic label to distinguish the two series of melodies.
(19) Simple Affirmative Melodies
I Complete
Floating/a-prefix/Stem/Suf fix
Root
Subord.
Relative/
Subj ect
Relative/
non-subj
Root
Subord.
Relative/
Subj ect
Relative/
non-subj
- - L -
- - B H
H - M H
H 
-L
II Incomplete
Floating/a-prefix/Stem/Suf fix
- B B* H
H - L
H - M H
H 
- L-
# i i I i III l i M di 6
123
"B," here as in Chapter One, stands for "Base"; see Chapter
One, Section 4.2.
6. I Complete/Root
The normal underlying tone pattern for this form consists,
as (19) indicates, of just a Low tone on the verb stem.
A simple sentence in this form, constructed from the
V"'\ /lexically High verb ci 'carry' and the object akhwa 'eggs' is
(20).
(20) o ci akhwa
he carry eggs 'he must carry some eggs'
A lexically Low verb also appears with a Low tone in this
form--for example, za sweep."
A \ / \(21) o za ulo
he sweep compound 'he must sweep the compound'
In general, then, verbs are Low in tone in this form, despite
their lexical tone specification. There are a small number of
exceptions to this which we shall return to below.
When Class I suffixes, the toneless suffixes, areadded to
this form, we expect a situation as in (22), which, by the
Well-formedness Condition, should become (23), which is what
we find. Class I suffixes attached to the verb stem in the I
Root form are Low in tone, all other things being equal.
I ill llh I 11% BlV H-H
124
(22) o zu + u + ru (23) o zu + u + ru
I I I I
H L H L
o zu'uru i 'he bought himself yams'
he yams zu buy'
u (here) 'for self'
ru 'past action' suffixes
No tonal information particular to the I Complete Root form
need be specified regarding the verbal suffixes; the Well-
formedness Condition is all that is necessary for the type of
forms considered in (22) and (23).
Something further does need to be stated, however, to
derive correct surface tone pattern for the noun phrase sub-
jects in the I Complete Root form. Consider the pattern in
(24). When the subject ends in a Low tone, it displays its
lexical tone; when, however, we would expect it to end in a
High tone, the subject in fact ends in a Falling tone, as dis-
cussed in Section 2.1 of Chapter One. (24) repeats (13) ,and
(14) of Chapter One.
(24) Eze ci akhwa Eze chief' H L
Uwa ci akhwa UwaL (a name) LL
(NP) 'carry' 'eggs'
Ekwe -ci akhwa Ekwe (oL nome) H H
Adha ci akhwa Adha (a name) L H
As analyzed in Chapter One, these tonal modifications are the
results of the I Root Flop rule, expressed in (25).
I I ||| I |H|I I
125
(25) V C V I Root Flop Rule
\T
This causes the tone of the verb stem to be shared by the
final syllable of the subject noun phrase when that subject is
outside the VP--that is, not cliticized onto the verb.
This simple rule can act as a useful example to illus-
trate the linguistic claims of this notation. Although one
may say, as a first approximation, that the degree of linguis-
tic unpredictableness df .a rule is proportional to the number
of symbols in the rule, or, more accuratley, to its measure of
complexity, this is only a first approximation.
The rule claims, first, that the language would be sim-
pler if there were no such rule, and if the subject Noun Phrase
in (24) always displayed its underlying tones. The existence
of the specification "VP" (verb phrase) in the rule claims
that, formally speaking at least, the rule would be simpler if
it applied to any successive pair of words, rather than only to
subject noun phrases. Functionally, it seems clear that such
a generalized rule would, however, be useless, since its
application would mark nothing at all.
Given, furthermore, that this rule changes the derived
tonal pattern, it must do this either by deletion of tonemes,
addition (epenthesis) of tonemes, or addition or deletion of
association lines. We shall assume that changes of association
lines are always linguistically simpler than addition or
I 1111111 dii di -~
126
deletion of segments (a rigorous statement this assumption
will be made below). In this case, this means that the tonal
rule creating the contour tones in (24) must add an associa-
tion line; and since the rule must specify which segments are
involved in the addition of the association line, the four
segments mentioned in rule (25), I Root Flop (two Vs, one C
and one T) are the minimal number possible. In sum, the
conclusion is that although no rule is simpler than some rule,
the simplest rule is a Flop rule (a rule adding an association
line); furthermore, such a rule requires minimally the speci-
fication of V's, C's, and T's as in (25).
We remarked above that the Flop rule must be made formal-
ly slightly more complex by specifying some syntactic informa-
tion about the environment (the expression "VP" in the rule).
This addition, it was pointed out, -was necessitated by the
simple fact that the Flop rule did not apply to any arbitrary
sequence of words, but only at certain subject-verb boundaries.
We should note now, however, that this additionally explains
why the Flop does not apply in (20), (21), or (22), where the
subject is a cliticizing pronoun, and thus, as we saw in
Chapter One, Section 5, is moved by a cliticization rule into
a position inside the VP. Thus there is no VP boundary in the
appropriate place, and the Flop rule cannot apply.
In Chapter 1, Section 5, a Floating H tone was discussed
which, under certain circumstances would "dock" onto the right-
end of a subject NP. Looking at the data in (24), a linguist
i1 1 '4ll A I 'H
127
or a language-learner might be tempted to impute the Falling
tone that is induced in the second two examples to a floating
L tone. We should like to propose that, although the linguist
may be led to that, the language-learner will not, for the
reason that a Flop rule (adding an association line) will
always be preferred over (be evaluated as less costly by the
evaluation measure) when compared to either an epenthesis rule
or a floating tone as discussed in Chapter One, Section 5, all
other things being equal. In short, the Falling tone in (24)
can be analyzed as a Flop or assimilation, since the following
tone is Low, on the verb stem. Therefore, it will be inter-
preted as an assimilation. Such an analysis is not available
for the type of facts discussed in Section 5 of Chapter One
(discussed in more detail in this chapter), and therefore a
floating tone must be posited.
The way to test such a claim is to consider the two al-
ternative sets of rules that are consistent with some set of
data that we may assume is learned earlier by language-lear-
ners. The claim that the Flop rule is preferred is tested
by determining whether the Flop rule makes more correct pre-
dictions outside the original data set than does the alterna-
tive solution. We have already seen that the Flop rule
correct predicts that cliticizing pronouns will not undergo
the Flop rule. The alternative account which proposes a
Floating Low tone suggests that (24c) arises from a structure
like (26). If the subject were a pronoun, as in (20), the
rl !I Pll  Al , d . 0 1 N
128
Floating Low analysis would predict cliticization as in (27)
(parallel to (72) in Chapter One), and a derived tone pattern
with a Rising (LH) pattern on the pronoun subject, parallel to
the deriving Falling in (72) of Chapter One.
(26) E k w e c i a k h w a
S HL L LH
(27)S
NP VP
pre stem
C i akhva
L~ L
In fact, this is wrong, as we know; the subject clitic tone is
High, not Rising.
A second type of empirical test was mentioned in Section
2 of Chapter One. If the tone of the Verb stem should be any-
thing other than Low (which is all that we have considered so
far), then the Flop analysis predicts that the subject NPs
tone should assimilate partially to whatever that tone was
(note that the tone specified for the verb stem in the state-
ment of the Flop rule in (25) is simply "T"--i.e., any tone at
all). If the verb stem should for some reason be Mid, then,
the Flop analysis predicts that the subject should take on a
I'II I H.!fi" I F I I I -
129
contour tone whose second half is level with the following Mid
tone, as illustrated in (28). This repeats the point made in
examples (19a and b) in Chapter One.
(28) Ha d i r n ' or'u Ha d i r i n o r u
H M L H
they were at farm
This type of raising of the stem tone from the normal L of the
I Root form to a Mid tone occurs in three cases: when the
stem is immediately followed by a Class II suffix, as illus-
trated in (28); when the stem has no suffixes (which is
unusual in itself) and is followed by the pronoun anyi ('us')
as an object; and for a small class of exceptional non-Low
verbs which appear in the I Root form as Mid rather than Low.
We will review these three exceptional cases in order.
The Class II suffixes, unlike the Class I suffixes, have
an inherent tone: Low. They furthermore affect the tone of
the syllable that precedes them, raising it to Mid or High.
We shall ascribe this to the influence of a floating High tone
on the left side of the suffix. The general underlying form
of the Class II suffix, then, shall be as in (29).
(29) r ri
H # L
Hull :1 114 RI ,II ,qHddIl:I ,
130
(Whether the ri and the L should be associated underlyingly or
not we leave open for now.)
When such a suffix occurs on a normal verb stem in the I
Root, then, we get a situation like (30). The first L is the
L constituting the I Root melody; the other H and L correspond
to the lexical entry for ri as illustrated in (29).
(30) # no # ri # /no/ = verb stem (be (+location))
# L H# L #
(I have assumed in (30) that an initial process has linked the
verb stem with tonal melody L; the illustration works as well
without this assumption. Note the use of association lines
to associate corresponding word boundaries, as mentioned in
Chapter One, Section 6.)
The Well-formedness Condition adjusts (30) to (31), and
a late rule of LH-simplification (32) in Igbo converts (31) to
(33), the correct output.
(31) #n # ri
# L H # L #
(32) V V LH-Simplification
L H M
(33) # no # ri #
I / I I
# M # L #
131
If a Class II suffix should be appended not immediately
following the stem, as in (34), we get the expected form.
(34) kwu = speak; ru = CLass I (toneless) suffix
#kwu + ru # ri #
# L H # L Output: kw ruri
(Note that we can be sure that the L tone of the stem does
not spread to the first suffix, giving. us (*35), because (32)
LH-Simplification would change this to (*36), rather than the
correct form (34).
(35) # kwu + ru # ri #
I L---- 7  I / I
# L H # L #
(36) #kwu + ru # ri #
# M #L # (*kwururi)
The second case raising the verb stem to M is found when
the object is anyi (1'us') and where there are no suffixes.
Consider the paradigm below (37).
rA \(37) a) Oyi ii m.
b) Oyi j 1gi
c) Oyi ji ya
d) Oyi ji anyi
/0 A A /\
e) Oy ji unu
f) Oyi ji ha
"I am feverish"
"You (sg) are feverish"
"He is feverish"
"We are feverish"
"You are feverish"
"They are feverish"
1l 11ITli f .Al 1 HA T l MI+
;kll i'141411 .A A H ,,i ,
132
(Each of these sentences literally means "Fever grips me (you,
him, etc).)
(a) 0 y i jiM
H H LH
Rule (25) Flop
(d) O y iJjii a n y i
H H L H
by LH--,M(rule (32))
01Y 1l i a n Yi
H H M M H
As we see, the tone of the verb ji is low everywhere
except where the object is anyi. And just there, the tone of
the final syllable of the subject falls to Mid rather than
Low, as predicted by the Flop rule.
I I , ll-KIII IIII V, J 1
133
Third and finally, two verbs--wu, the copula of identity
and ka 'surpass'--take the Mid tone in this form normally.
We thus get 0 ka mma rather than 0 k mma ("it is better"--
"it surpasses good,," literally).
I have no examples of non-clitic subjects preceding ka or
wu in this form, but the prediction is clear; we expect
Subjects ending on a High tone to fall to Mid rather than Low.
So much, then, for the I Root form, whose melody consists
just of an L tone on the verb stem. We turn our attention
next to the I Subordinate form.
7. I Subordinate
This form appears only in subordinate clauses, intro-
duced by an element (a complementizer) ka or ma. The tonal
pattern here differs in several ways from that of the I Root
form. First, the tone of the subject NP is unaffected here in
the I Subordinate; second, the tone of the verb stem depends
directly on the underlyingly verb. An underlyingly Low verb
will be Low; an underlyingly non-Low verb will be High. I
shall denote this general occurrence by "B," for "Base," as in
Chapter One, and as in Figure (19) above. We note, informally
for the moment, that suffixes in this form are High-toned. In
the following two examples, the preposed clause with the
meaning "if" contains the verb in the I Subordinate form.
I i l R, I l p 1,1 I oH
(38) Low-verb: Anyi fee unu unu adu anyi
we pass you, you will follow us
"if we pass you, you will follow us."
(G&E, p. 81)
fe is here the verb stem: a Low-toned verb,
e is a suffix.
(39) High-verb: AnyL cir akhu / ndi rwe ya ga ijidhe
we pick kernels
up
"if we pick up kernels, the owners will
arrest us" (G&I, p. 79)
verb is ci (High toned); ri is a suffix.
We observed on page t20 above that there are two classes
of suffixes (I and II) and that, following Edwin Williams, the
more common Class I suffixes are toneless. This corresponds
to a formal representation entirely on the non-tonal tier,
with no tonal segments in the lexical entry for the Class I
suffixes. In the I Subordinate form this assumption leads
directly to the postulation of suffixal H tone which is lexi-
cally specified only on the tonal autosegmental tier, but not
on the nontonal tier. Thus the H-tone and the Class I suffixes
are formally parallel: each is specified only on one tier.
We are led to this suffixal H-tone most obviously by the
fact that Class I suffixes are High-toned across the board in
the I Subordinate form. The tone of the Class I suffixes is
I II1~ III I
135
arrived at differently here than in the I Root forms, where
the tone of the verb stem was passed onto the Class I suffixes.
Thus for Low-toned verbs, we get forms like (40); for High-
toned verbs, we get forms like (41).
(40) (i) Verb (41) Verb
stem suffix
ke IStem suffix suffixkce eI
L Hgba u 1
H H
(ii) verb
stem suffix
stem suffix suffix
ka a la
Igbat
L H
Given that the Class I suffixes are inherently toneless,
the language-learner will analyze the High-tone on the suffixes
as an autosegment, which here is a High-tone particular to the
I Subordinate form.
As Williams furthermore correctly observed, the postula-
tion of an H-tone suffix--a "floating tone "--provides an ex-
planation for what seems to be quite a different tonal process.
This is what I shall call "Object tone mutation," and refers
to a change of tone undergone by the noun phrase immediately
following the verb in several of the verb forms (this process
of object tone mutation applies in the genitive construction
ii414 A1ill 1 m1 ! I - k l
136
also, as we shall see).
In a word, it is the presence of a floating H tone
immediately preceding an NP that triggers Object Tone Mutation.
Object Tone!Mutation causes the following changes in mono- and
di-syllablic nouns:
(42) Tone Before Object Mutation After Object Mutation
H Ml
HH H M
LH M H
H L H L (no change)
L L L L (no change)
If we attempt to formalize this tonal process, we note first
that three forms undergo some change; two undergo none, the
two that end in L tone. We might make a first stab at a rule
like (43), leaving off the triggering environment, the
floating tone.
(43) H + M/ -- ; or, more precisely, (a)
T + [+10] / -- # (b)
Such a rule is literally as simple as a rule can be,
containing just one feature mentioned. 6 We observe that it is
empirically satisfactory for four of the five forms: it ex-
pressed the fact that forms ending in L = [-hi,+lo] undergo
no change; it also correctly states the change undergone by
the forms H and HH. It wrongly predicts, however, that LH
137
will become LM. Recall, too, that LM1is superficially indis-
tinguishable from LH: thus (43) as written predicts a
"systematic" change for the LH forms, but with no distinguish-
able surface realization of that change. As (42) shows,
however, the actual result is "MH."
Looking at the surface description "MH," we remark that
this pattern may reflect either of two structures--(44) or
(45)--in the autosegmental framework.
(44) CV CV (45) CV CV
M M
As noted above, rule (43) applied to the LH forms gives (46),
which, obviously, is neither (44) nor (45).
(46) CV CV
L M
If the effect of rule (43) is augmented by a rule delet-
ing the L in (46)--thus, in effect, permitting (43)'s change
of the H to M to be heard on the surface, no longer masked by
the L--then (45) results, for the M immediately reassociates
with the first syllable once that syllable loses the tone it
was associated with.
The rule required to state such a deletion is again
extremely simple; it is of cost 1, as we see in (47).
(47) L+$/ ## --
:1 -1 1|| if iji . J. a ] d I, Ib ,
138
'Note that the other place where (47) could operate
besides on the LH form--that is to say, on the LL form--its
effects are not noticeable. If the initial L is deleted by
rule (47), the remaining L reassociates, spreading over both
syllables, as in (48).
(48) CV CV (48) CV CV WFC CV CV
I I I -F
L L L L
Summarizing, the following analysis for the Object Tone Muta-
tion is selected by the theory:
(49) Object Tone Mutation
TM / 1-low] .affixT --
L 0 [/-Flow] affix h ff
What is important to us now, however, is the existence of the
floating H tone suffix, rather than the precise formulation of
the Object Tone Mutation rule. We must furthermore keep two
goals separate: arguing that the correct analysis for Igbo
contains a floating H tone here, on the one hand; and, on the
other, arguing for the present autosegmental theory on the
grounds that it directly forces the grammar of Igbo to contain
this floating H tone. The former task must await more data;
the latter is contained in each discussion where this H tone
appears. We have, for example, already seen that the Class I
suffixes High tone in the I Subordinate form have required
I : 1 d# Iu[ i 
139
positing the floating H tone there. That the Class I suffixes
were toneless, the reader will recall, was justified on the
grounds that the Class I suffixes appeared Low-toned in the I
Root form, but High-toned in the I Subordinate; and more
generally, either take on the tone of the verb stem or the tone
of the floating H-suffix.
Let us consider a prediction made by the system developed
so far. The verb stem is underlyingly assigned its lexical
tone, either Low or High; a floating H tone is placed as a
suffix. We have not considered any cases so far where there
are no suffixes on the non-tonal tier, which I shall hence-
forth call the syllabic tier. A strong consequence of the
autosegmental hypothesis, of course, is that the floating H
tone may continue to exist whether or not there are syllabic
segments associated with it.
We expect to find, then, structures as in (50) for High-
toned verbs, and as in (51) for Low-toned Verbs.
(50) High-toned verbs
Verb
stem
CCA = CV
H I
stem suffix
vrb
I- 1 i llbl I Mld# 4 Ik
14o
(51) Low-toned verbs
verb
stem
cvVCCV
L H
stern suffix
verb
However, we have already postulated a rule (32) of Low-High
simplification:
(32) CV CV Low-High Simplification
L H M
This rule would apply in (51), changing the rising tone to a
Mid tone. Thus we expect the forms in (52) in the I Subordi-
nate without suffixes.
(52) Hightoned verb: CV
Lowtoned verb: CV
In fact, however, the tone of the verb in the I Subordinate
form without suffixes is more complex than hypothesized by
(52). High toned verbs are, as predicted, High; however, the
more interesting Low-toned verbs are Low before a High tone,
and Mid otherwise, as illustrated in (53).
I I I ~~udI I
(53) (a) anyi ke anu (HH) ... meat
(b) anyi ke akwha (LH) . .. eggs
(c) anyi ke a zu (HL) ... fish
(d)) anyi ke ala (LL) ... land
we share out (direct
(verb) object)
The tones of the direct objects in (53) have been changed from
the tonal form indicated in parentheses to the indicated tones
by the process of Object Tone Mutation discussed immediately
above. From the data in (53) it is clear that on any account
there must be a rule affecting the tone of the verb stem, a
rule that is conditioned by the tone of the direct object.
Our system so far makes it clear that, all other things being
equal, the Mid tone seen in (b) and (d) is the' expected tone;
all is not equal, however, in (a) and (c), where the presence
of an initial High tone in the direct object has caused dele-
tion of the floating H tone before rule (32), Low-High Simpli-
fication could cause the Low-High sequence to become a Mid.
In short, we have a rule of "Floating H Deletion" (54).
(54) V Floating H Deletion
T n'H (f) H
1 2 3->
1 0 3
Floating H Deletion says that a contour tone followed by
an H is simplified as shown above. Note, crucially, that such
an H-Deletion does not occur in examples such as (40), where
the H-tone is associated to its "own'' syllable(s).
In summary, we have seen that the tonal formula for this
tone is B-tone for the verb stem, and a floating H-tone suffix.
A rule of Floating.H Deletion has been postulated to deal with
more complex tonal alternations, and the Object Tone Mutation
phenomena have been introduced.
8. I Relative/Subject
This form is used for relative clauses where the head of
the clause is the subject of the embedded sentence, as in "the
woman who won the prize," but not as in "the man I saw." The
latter would be treated as a Relative/non-subject form.
The tone of the verb stem in this form is invariable M7
while the suffixes are H, caused by a floating H-tone suffix,
just as in the I Subordinate form considered immediately above.
As we would now expect, given the formulation of Object Tone
Mutation in (49), direct objects undergone Object Tone Muta-
tion--the consequence, as we have seen, following Williams, of
the floating H-tone suffix. This can be seen in (56); (55)
is intransitive. Some examples:
(55) Low-toned verb: Ndi no n'ahya
people be at market
"the people who are at the market"
11 '] I Ili [1 ]IN1114 IT I Ima A.  h-H,,
143
N d i n o n a h y a
H H H MI ' H H
explained
below
(56) High-toned verb: Igbo rri ji
Igbo eat yam
"the Igbo people who eat yams..."
Igbo riri ji
I fII by (49)
L LtI M H H\f
M,
I shall not pursue the matter of the Floating H-tone
suffix; its effects on the Class I suffixes and its triggering
of Object Tone Mutation are sufficient evidence forvits
occurrence here. An example of Object Tone Mutation occurs in
(56) above; mutation of ahya is regularly blocked in (55) by
the preposition na (or n).
A new tonal phenomena arises in this form. The tone of
the final syllable of the subject of the clause is raised, as
we see above in (56), where the underlyingly tones of the
subject is Igbo, which becomes Igbo in this form. In brief,
the changes undergone are as follows:
| I | H I 1 | T .|.1| aP
IiiH Hlll llla 1]] 41- 5 .-.a ,E
(57) Underlyingly Subject Mutated Form
-j (H) ii (H) (no change)
anu (HH) anu (HH) (no change)
akwha (LH) kwl'a (LH) (no change)
onu '(LL) onu (L LH) or onu (LH)
azu (HL) azu (HM)
We see clearly what is going on if we look at the change
onu > onu; an H-tone is being added on. As for the L LH/L H
alternation, we have already posited (32) LH-Simplification,
which will change L LH to LM, which is itself indistinguish-
able from LH. This same rule explains the H L + H LU + H M
change (the last line of (57))].
In fact, we are rather pleased to see surface alterna-
tions justifying rule (39). That the forms ending in H don't
change tonally follows, of course, if we accept this analysis
of tone-raising as t resulting from a leftward docking of a
floating H-tone, since V is indistinguishable from V.
H H H
If we step back from the analysis a moment, we may note
that tonal alternations like those displayed in (57) are
precisely the sort predicted by a theory with floating tones:
unidirectional (raising or lowering) tonal changes on the left
or right side of a word, sometimes realized by contour tones
themselves. In addition, we would expect a systematically
motivated syntactic or morphonological orginal- for the floating
1 ! 1 1 N l H: -
tone. On this last point, see Section 13 below.
In summary, then, the I Relative/Subject form is:
Floating Stem [Suffixal
H M H
The precise formulation of the rule that docks the floating
H tone here onto the subject will await the discussion below
of the underlying position of this H-tone. It is, of course,
the floating tone discussed in Section 5 of Chapter One above.
9. I Relative/Non-Subject
Roughly speaking, this form is used for relative clauses
where the head of the relative clause is not also the subject
of the relative clause--"roughly,'" because we find cases where
the I Relative/Non-Subject is used where there is no particu-
lar head at all; rather it is used as an adverbial clause.
We shall consider these below. There are no cases, however,
where the I Relative/Non-Subject is used as a relative clause
where the head of the relative is the subject of the relative
clause; in all such cases, the Relative A form is used. Con-
versely, I know of no cases where the Relative A is used as
an adverbial.
The tone of the verb stem here in the I Relative/Non-
Subject form is always L. The suffixes follow on L tone, too,
indicating that there is no suffixal tone in this tense. As
we expect, then, there is no object tone mutation in this
II I I4I*III~iIJII 14
tense, either.
Just as we found Subject Tone Mutation in the I Relative/
Subject form, we find such a tone mutation here in the I
Relative B form. Some examples:
(58) Underlying Mutated Form
Hji Hji
L H ubhe + L H ubhe
H H eghu H H eghu
HEL ukhwa + HM ukhwa
/ \S HL Ladha H LI Ladh (personal name)
L L ovu L H (=LM) ovu
In the examples in (41), the form with surface alternations
was LL. Here, in the texts, we find occasional alternations
in the HL form, as indicated above. Considering the sparsity
of data, I have no way to hypothesize concerning the option-
ality or obligatoriness of LH Simplification in this context.
Again, we attribute this to a floating H tone docking onto the
subject as in the I Relative/Subject; its precise history will
become clear in Section 13.
There is a new tonal rule that occurs in the I Relative/
Non-Subject form that we have not considered yet. Recall
that the I Relative/Non-Subject form is used when a. relative
clause follows the head NP which is not the subject of the
relative clause. Since Igbo is an SVO language, this juxta-
poses the head of the clause with the subject of the clause
, I1 1!1 ill III1 v ! .l
like this:
(59) . . . Head NP/Subject NP Verb Etc.
RelatIve Clause
In the following discussion, I shall refer to the first and
second NPs in (59) as the Head NP and the subject NP respec-
tively. From the point of view of eitherluseful prepositions
or case markings, Igbo is quite impoverished; in place of such
things, Igbo uses word order and tonal modifications. In
order to help make clear the grammatical relation between the
Head NP and the Subject NP--to wit, none--ther e is a rule,
henceforth "Disjoint Tonal Sandhi"--which operates on the head
NP's tone. The point of the rule, again, is to point out that
the two nouns are not closely related, as distinguished, in
particular, from the two nouns in the genitive relation; on
this construction, see Section 15 below.
The change that occurs is the following:
(60) Disjoint Tonal Sandhi: If the subject NP begins
with L, then the Head NP
mutates:
From To
H L
HH LL
LH LL
HL HL
LL LL
i H I 4k,
The syntactic structure to which this rule is applying is
essentially (61).
(61) NP
NP S
NP VP
V NP
We shall formulate the Disjoint Tonal Sandhi Rule, but this
is tentative, and based on some applications we have not
considered yet.
(62) Disjoint Tonal Sandhi
H + L / -- X # [ L Condition: X contains
no # and no L
This produces the correct results according to (60), and
applies to a structure such as (61), but in no other struc-
tures encountered so far. This process of Disjoint Tonal
Sandhi, however, plays a rather marginal role in the entire
grammar of Igbo; in this way it is quite different from the
other tonal rules, which form a very tight and inter-connected
system.
In summary, then, the tone form of the I Relative/Non-
.9Subject is:
Floating Stem Suffixal
H 10:
li, i, I-
149
10. Summary
We have considered four forms to this point: Green and
Igwe's Affirmative I Root, I Subordinate, I Relatives/Non-
Subject and Subject. We shall next consider the corresponding
II Incomplete forms, which will give us a clearer idea of the
structure of the phrases we are investigating.
(63) Form Floating Stem Affix Special Rules
I Root 
-- L -- I Root Flop
I Subord 
-- B --
I Rel/subject H M H
I Rel/non-subj H L -- *
*This form is subject to the Disjoint Tonal Sandhi rule, but
this is not due to a special marking of any sort, but because
the I Relative/Non-Subject satisfies the structural descrip-
tion of the rule.
Note that the only specially marked rule is one which
applies only in "root" sentences, in Emonds' sense.10
11. The II Incomplete Forms
The form of all the affirmative II forms is similar to the
I forms except that there is a verbal prefix /a/ in the II
form. (/a/ becomes /e/ if the verb stem is in the "tense"
harmony class.) In addition, there may appear before the /a/
a relative marker /na/ in the II relative forms.
I, hun hh~I 1
ALb H.
150
Thus, schematically:
(64) II Incomplete -- Schema
NP VP
(na) V NP
n
Prefix stem suffixn0
a cv' (t)V
When one of the short forms of the singular pronouns--the
cliticizing pronouns--is subject, the a/e prefix is suppres-
sing--unless the /na/ relative marker intervenes, in which
case the a/e remains. On this important process, more below.
12. The II Incomplete Root
Initiating use: This form, the II Root, is divided by
Green and Igwe into (i) the initiating, and (ii) the non-
initiating, forms; by this is meant (i) discourse-initial use,
or (ii) elsewhere.
In the initiating use, the clitic pronouns are used. The
following chart indicates the morphological shape of the
pronouns throughout the forms. The cliticizing forms are given
first, the strong forms second.
(65) Pronouns
Person
2
3
indef
Singular
M/mu
i/gi
o/ya
a/-
Let us consider the tonal characteristics of this form
when there is a non-cliticizing subject, thus not deleting the
a/e prefix.
(66) High-toned verb c 'carry'
Prefix Sem Suffix
a ci a
Egbe acia khwa
Xdh Acia akhwa
Ez acia anu
Ugho acfa akpA
Egbe carried eggs
Adha carried eggs
Ege carried meat
Ugho carried bags
(67) Low-toned verb za sweep'
Prefix Stem Suf fix Egbe azaa ama
a za a Adha azaa ama
Uce azaa amA
Ugho aza ama
Egbe swept the street
Adha swept the street
Uce swept the street
Ugho swept the street
Apparently the final tone of the subject affects the tone of
the prefix-frhere, a--for it changes, depending on whether the
subject ends in an H or L tone. We can summarize these tones
in (55). I shall write "H/M" where we cannot determine which
it is because of an L tone immediately preceding.
151
Plural
-/anyi
-/unu
-/-a
ellll'!lml 91|?!'l El  ThillNI!Tlit il a
152
(68) Subject ends in H-verb L-verb
Prefix Stem Prefix Stem
(a) H M H L L
(b) L L H/M H/M L
Note that the H following the M in (68a) (e.g., as in
Egbe acia. .. ") is actually on the same pitch as the prefix
which is indicated by M; in general, as we remarked above (44)
and (45) are phonetically indistinguishable.
(44) CV CV (45) CV 7C
H M
Indeed, we see that the prefix and the stem are on the
same pitch if and only if the subject NP ends in H (the cases
in (68a)). Furthermore, this tone is found on the stem in the
other case (68b), when the subject ends in L.
The autosegmental analysis of this is straightforward.
A final H on the subject NP deletes the tone of the prefix,
which automatically reassociated the stem tone to the prefix.
For example:
(69) # egbe # # a + ci + a
# H H # # M H #
# egbe # # a+ci + a #I Ih I I %1 / I I
# H # # M H #
added by Well-formedness Condition
w u I I 
T liI ! :I !i 1: A l H! , * K
,T xn fl 1! ;I AI I.i i [1 21 ] ;Il
153
The underlying tonal form, then, is what is given in
(68b). The tone of the stem is L with an L-verb; it is M with
a non-L verb. We shall continue with the convention given in
Chapter One, and abbreviate this dependency by "B*."
The verb stem has tone B*, then, in the II Main form.
The prefix for non-Low verbs, as we can see in (68b) is L; the
prefix tone for Low-verbs is either M or H--we cannot tell
which. I shall arbitrarily pick H, but nothing rides on this
choice. Thus the prefix has a dissimilatory tone on it with
respect to the stem; I shall denote this as "B ."
(70) II Main Tone
Prefix Stem
B B*
Let us formulate, then, the prefix-deletion rule illustrated
in (69).
(71) W H - T VeX
affix verb
2 3
1 0 3
Tonal Behavior of Pronominal Subjects in II Root Form
Cliticizing pronouns behave in the II Root as illustrated
in (72). They are uniformly H in tone, be the verb H or L.
Note that the a/e prefix is suppressed.
w yltll -:I lll IIi ll] !H ,IN4 dA ::l
154+
(72a) ci 'carry' H-toned verb
1 cia anu "you have carried meat"
o cia anu "he has carried meat"
a cia anu "meat has been carried..." or
"one has carried meat.'"
(72b) za 'sweep' L-toned verb
i a ama you have swept the street"
o 6zaaILma "he has swept the street"
/ k/ /
a zaa ama "one has swept the street"
The H-tone on the suffixes, and its corollary object tone
mutation, of. course, remains constant from (66-67) above.
The disappearance of the a-prefix in the presence of the
cliticizing pronoun is due, as we have already said, to the
fact that the cliticization process substitutes the pronoun
for the a-prefix. The rule was introduced in Chapter One (66),
reproduced here.
(73) Subject Cliticization
Pronoun] [ Verbx]
weak ] prefix]
1 2+
0 1
The reader is urged to reread Section Five of Chapter One here,
because the floating tone process sketched there pulls together
a larger number of the processes that are being introduced one
155
by one in the present chapter.1 2
By way of review, consider the tone of the underlined
verb in the following paradigm (74). The superficially com-
plex tonal alternations reveal themselves as quite natural,
given what we have seen so far. We shall show how these
follow from the rules we have already posited.
(74) (i) Ya jikhere, anyi afu ama (from ama 'street')
he ready we go out street
'if he is ready, we'll go out to the street'
(ii) ahu khowe, unu afu ezi (from ezi 'outside')
body swells you go outside
'when it gets hot, you will go outside'
(iii) A bha yA, ya abha ede
one grips him he grips cocoyam
'when one grips him, he grips the cocoyam'
(iv) A bha ya, ya abha ala
one grips him he grips ground
'when one grips him, he grips the ground'
In all four cases, the underlined verb--prefix plus stem--is
in the II Main form, and in all four cases is a Low-toned
verb. But due to different environments, each has a
different tonal pattern. These are derived as follows:
I Ii
156
anyi af u a m
H L H L H H
a
by Object Mutation (49)
H-- M
by Floating H Deletion (54)
an y
I
t afu ama
L H L H
u n u a f u
H L H L H
M
by LH
yields: u n u a f
H L H
(74111) y a a b h a e de
Hi L H L
by Floating H Deletion (54)
by II Root
Rule (105)
yields:
11
14
e
by Object Mutation (49)
L H -M
Simplification (32)
e z i
M
y a a b h a
H L
e d e
1 I
H L
(741)
yields:
157
(7hiv) y a a b h a a 1
HIT L L
0 Mt by LH Simplif ication (32)
by II Root
Rule (71)
yields ya a b h a a 1 a
H M L L
added by Well-formedness Condition
These derivations illustrate our rules:
(49) Object Tone Mutation
(54) Floating H Deletion
(32) LH Simplification
(71) I Main Rule
We also have:
(25) I Root Flop
(62) Disjoint Tonal Sandhi
(73) Subject Cliticization
ll 1 H 1 : 1 :i ll I fH
158
13. The II Subordinate and the II Relative/Non-Subject
The reader will recall that we observed above that there
was some overlap between the uses of the I Relative/Non-Subject
and the I Subordinate: sometimes the I Relative/Non-Subject
was used adverbially rather than as a relative clause. None-
theless, a glance at chart (63) indicates a substantial dif-
ference between the two forms.
The parallel II forms--that is, the II Subordinate and
the II Relative/Non-Subject--bear an even closer affinity to
each other. In fact, .we shall conclude that they are under-
lyingly identical from a tonological point of view.
The verb stem and its Class I (toneless) suffixes are L
in the II Relative B, just as was seen in the I Subordinate
form; the same is true of the II Subordinate. From this .we
conclude directly, as we did, for the I Subordinate, that the
stem tone in these two forms is L, and that there is no
suffixal floating tone; the Low tone that appears on the Class
I suffixes is the stem's L tone spreading rightward, as in
the I Main form. The Direct object NP does not undergo Object
Tone Mutation, as we would expect, since there is no suffixal
H tone to trigger it.
Furthermore,, in both the II Subordinate and the II
Relative/Non-Subject forms the non-clitic subjects undergo the
tone-raising on their right-end side as we have observed, for
example, in (57), which we attributed to a floating H-tone to
the left of the verb stem that docked leftward. Thus on the
. I + i' i I -H ,,
159
three major tonological dimensions we have defined for Igbo
verb forms so far, the II Relative B is indistinguishable from
the II Subordinate: both have the floating H tone to the
left of the verb stem, and both have an L-tone associated with
the stem (and no suffixal toneme). Schematically:
(75) Pre-Floating Stem Suffix
H L
Thus along the three tonological dimensions we have defined
for Igbo verb forms so far, the two forms are indistinguish-
able. Furthermore, when the II Relative/Non-Subject is used
as an adverbial, it can appear with the suffix mo--which
appears elsewhere in Igbo only in the II Subordinate form.
Why distinguish the II Subordinate and the II Relative/
Non-Subject at all, then? Why give it two names, for that
matter, if it is really only one form?
There are three reasons, really: first and most simply,
because we are following Green and Igwe's classificatory sys-
tem, which deals with these two forms quite separately.
Second, because these two forms, as I have indicated, play
quite separate roles; a relative form has one function, an
adverbial Subordinate clause another (although, admittedly,
there is some overlap). The distinction between these two
functions is kept separate in the I Forms, as we saw in chart
(63).. So it i just an accident, really, that the two forms
have the same underlying melody. In short, the forms
llI ii bfill I I J IA I -
i6o
themselves are different; they simply have the same underlying
melodies.
But the real reason for pursuing the differences between
the II Subordinate and the II Relative/Non-Subject is this:
our theory has so far led us quite directly to a specific way
of analyzing the tonal system of these forms; it forces the
conclusion that the melodies of these forms are as in (75).
When we look now for differences between the two forms--and
there are some--we are hamstrung by the lack of distinctions
permitted by the theory; and so the differences that appear
must be "superficial" differences arising out of something
entirely different. Indeed, this is the case, and it has al-
ready been foreshadowed in the discussion of the floating tone
in Chapter One.
Consider the paradigms below, which illustrate the tonal
forms in the II Subordinate and the II Relative/Non-Subject.
First are given the forms with non-clitic subjects, and these
forms are essentially identical. Note the presence of the low
toned /a/ or /e/ prefix. Below are the forms with pronominal
subjects, and there we find a sharp constrast: the II Subor-
dinate has a Falling tone on the cliticizing subjects, while
the II Relative/Non-Subject has Low-tone.
1:11I10I .1A . Hlih- Ilm 
161
(76) II Subordinate
Subject:
(i) /aghu/ agl
leop
hu egbuo
ard kill
eghu
goat
'lest the leopard
kill the goat. ..
(ii) /oke/ oke ataa akhu 'lest the rat eat the
palm kernels'
rat eat palm
kernels
(iii) /uze/ uze ataa ya 'lest the squirrel eat
them'
squirrel eat them
(iv) /ew/o w atXa 4 'lest the monkey eat
them'
monkey eat them
(77) I Relative/Non-Subject
Subject:
(i) /eghti/ Ubhe eghu ataala.., 'The pears eaten by
the goats'
pears goat eat
(ii) /Adha/ Uwe Adha etiele. . 'The clothes Adha
wore...'
clothes Adha wear
(iii) /6z Uwe eze azuola... 'The clothes the
chief has
clothes chief buy bought...'
(iv) /Ala/ Ok al eriele... 'The shares Ala has
taken...'
shares Ala take
I r ll Il iii lll Ll A!i
162
(78) II Subordinate
Pronoun Subjects:
A N\/
... m ghuo eghu
A
.. i "i "
anyi egbuo eghu
unu egbuo eghu
hb \ \gN / /
h& egbuo eghu.
(79)
(a)
(b)
(c)
II Relative/Non Sub
Mgbe m zula anu,
Time I buy meat
Mgbe i zula anu
Mgbe o zula anu
lest I kill the goat
lest you kill the goat
lest she kills the goat
lest the goat be killed
lest we kill the goat
lest you (pl) kill the goat
lest they kill the goat
ject
Whenever I buy meat
Whenever you (sg) buy meat
Whenever he buys meat
(d) Mgbe ztla nU Whenever one 'buys.meat
(e) Mgbe anyi azula anu Whenever we buy meat
(f) Mgbe unu azul& anu Whenever you (pl) buy meat
(g) Mgre ha azula anu Whenever they buy meat
We are thus faced now with the need to explain the falling
tone on the cliticizing pronouns in (78)--the II Subordinate
form. In doing this, we will come to understand the origin
of the floating H tone to the left of the verb stem that we
163
have postulated to account for the raising of the final tone of
the subject Noun phrase in several forms (see (68) and (72),
(73) of Chapter One, for example).
The Falling tone on the cliticizing pronouns in (78) is
distinctly different from the only other Falling tone we have
encountered so far--that is, the Falling tone created by the
Flop rule in the I Main form, as illustrated in (80).
(80) Ekwe ci akhwa
H H L L H
The dotted line represents the association line added by the I
Main Flop rule.
In the case of the Falling tone created by the I Main
Flop rule, the cliticizing pronouns were precisely the forms
that did not take on a Falling tone. In the II Subordinate
form, however, it is rather the reverse: here, only cliticizing
pronouns take on a Falling tone. And the origins of these two
falling tones are quite different: similarities in appearance
may be deceiving.
We shall show that the Falling-toned clitic pronouns in
the II Subordinate are just the Low-toned clitic pronouns of
the II Relative/Non-Subject (see (79a-d)) plus a Floating H-
toned docked onto their left-side. This H-tone is precisely
the same H tone what causes a raising of tone on the subject
NP. This Floating H-tone begins, in a certain sense, in
between the subject and the verb, and may dock either leftward
-mff ' 1 4-1 l 1 I1 i yIo: A rjli; 0 1tIR
or rightward.
Thus we will be arguing that the raising of the subject
NP's tone operates as in (81).
8 ) Subject Verb
CV CV CV
L H L
But if a pronoun were just like any other noun phrase, we
would then expect a Rising tone on it; as in (82).
(82) V CV
V
L L = V CV
I (82), we have not taken into account the cliticization of
the pronoun; we shall argue that cliticization brings the
clitic pronoun into the verb phrase, as in (83a); the docking
of the floating H tone, now, if it be rightward, causes the
tone of the clitic pronoun to be a Falling tone, as in (83c).
(83) (a) Subject Verb (b)
V CV VCV
tIkI added by
H L L Well-formedness
Condition
(c) VV
= VCV
I I: p i||4 % dilr~]aH
165
We now precede to argue these two points.
In the discussion of the raising of the tone of the
subject NP in the I Relative/Subject (following illustration
(57)), we argued that, given our conception of autosegments,
and thus floating tones, such a tone mutation must come from
a floating tone, docking from the right. In the II Relative/
Non-Subject form, we find confirmation of this as direct as
one could hope to find.
In the II Relative forms, it is possible--indeed, pre-
ferable--to have an /na/ affix in front of the a-prefix. Thus,
for the II Relative/Non-Subject both (84a) and (84b) are
possible.
166
(84a) Ahya eghu naala eci "The market the goat
should have left
market goat leave yes- yesterday"
terday
(84b) Ahya eghu ala eci (same meaning)
Na, when present, is always H-toned; and, note well, when na
is present, the Subject Raising process does not occur. This
was not apparent in (84); see however (85), where the subject
/oso/ becomes /oso/.
(85a) Ahya oso naala eci The market the
pepper should have
market pepper leave yesterday left yesterday
b) Ahya- oso ala eci (same meaning)
The Subject tne raising we have observed all along appears
now to be just a particular example of the more general
"stability" phenomenon discussed in Section 3 of Chapter One:
when the na deletes, its High tone shifts leftward onto the
subject. When the na does not delete, as in (85a), of course,
the subject displays its isolation tones.
More generally, this H tone may appear in forms where
there is no possibility of the na occurring--as in the I
Relative forms and the II Subordinate. The correct way to
phrase the generalization we saw above linking the na and the
H tone is to say that lexical insertion for strictly tonolo-
gical material and strictly syllabic (non-tonological)
i iii Iili H F i 11; d I 1 0 1, IA - , : -
167
material occurs on parallel syntactic phrase-markers, and if
corresponding elements occur, they will be associated. Thus,
for a sentence like (85a), we get a tree like (86). Note that
the same phrase structure rules govern both the tonological
and the syllabic structures.
(86) S
NP VP
1stem
,CV CV t- na a CV
TT2 H L
v1rb
NP
VP
S
However, when the na does not appear--for what ever reason--we
have a "floating tone" on the tonological tier: the H tone.
It will then ultimately dock onto some vowel.
We will now see why cliticization of the subject pro-
nouns should be viewed as a movement rule, as in (83) (or
better--(67)--of Chapter One). This analysis has already, we
note, made certain correct predictions regarding the I Main
Flop rule. This rule we have formalized as (73), repeated
here.
168
(73) Subject Cliticization
Pronou Verb
Weak Prefix
12
0 1
Adopting Emonds' framework, this rule must be a local (or
minor movement) rule, since it is neither a root rule nor a
structure-preserving rule. This implies that there can be no
material intervening between the two factors in the rule. If,
therefore, the na should be present, as in (86), rule (73)
cannot apply, and furthermore, the rule cannot (according to
Emonds' principle) be generalized to (87), for example:
(87) Pronoun] (na) Verb Impossible Variant
[Weak [Prefixj of Subject
Cliticization
1 2 3
0 2 1
In fact, this is correct; when the na is present, subject
pronouns do not cliticize, and so the Verb prefix /a/ remains.
Compare (88a) and (88b).
(88a) Mgbe o a zula anu
IRule (73)
[Mgbe ozula anu]
(88b) Mgbe o na a zula anu
[Mgbe o na a zula anu]
Prefix does not delete
,,, -; , J HI 111: 1[ 1 '1 ,Jjh[ IIA"] al 11: : -1] r 111 '., I :
169
Parenthetically, this provides us with a whole new
argument for autosegmental theory, for if the presence of na
blocks the cliticization rule (70), the presence of the H-tone
between the subject and the verb does not. Thus (88a) pro-
ceeds as in (89), regardless of the presence of the floating
H-tone; therefore, the H tone is not in the string to which
rule (70) applies, and must be on another (autosegmental)
tier--but this is exactly what was predicted.
(81)
NP
t4VP
L La L
S
III -II i 1 1:1 1 '  : . i:w 4::I 1 1 s, I :-
170
Purs/ing this a bit further, we see that rule (73) is a
rule applying to the syllabic (upper) autosegmental tier, and
the moved pronoun subject takes up the tone of the /a/ prefix
that it replaced, as indicated above in (89). The unmoved
clitic form, as in (88b), is H in tone, but whether this is
from the H-tone or some other source, we cannot presently say.
The rule that docks that floating H tone in (89) and
similar structures is an autosegmental rule; the rule that
clicitizes the subject pronoun (70), is a syntactic rule,
since it falls within the domain of Emonds' hypothesis.1 3
Given our general assumptions about rule-application ordering
and the derivation, the cliticization rule must always apply
first. We assume, as discussed in Section 4 above, that
syntactic rules--minimally, cyclic syntactic rules--precede
insertion of phonological material, and again, preferably
before insertion of all lexical material. Therefore, an
autosegmental rule must follow a syntactic rule in application.
We will close this section with a formulation of the rule
that "docks" the floating H tone. The point of this section--
as of the previous section--was a positive and optimistic
answer to the question posed in the second paragraph of this
chapter: we see that inclusion of autosegmental analysis has
led to a simple conception of the underlying tonal form in
Igbo, with a small number of elementary rules. Furthermore,
the complementarity between such processes as Subject Tone
Raising, high-toned na and falling-toned subject clitics is
- 't III I 'FI II  I ..K~ iell
straight-forwardly explained.
In the I Relatives, the II Subordinate and the II Rela-
tive/Non-Subject, we have seen the H-tone dock leftward in the
following circumstances.
(90) II Subordinate, I, II Relative/Non-subject
NP
NP S
NP VP
CVCV V
T T () T
(91) r Relative/Subject
NP
S
NP VP
cC'MVV
T TH M
(got) II Subordinate, II Relative/Non-subject
NP
NP S
PN V
CVCV V CV
TT T
(if and only if the subject is a cliticizing pronoun)
me:ee:lltN! fl||1101'1RIll!H&ltidill'N P "I
172
Our main problem in formulating the rule that docks the
floating H tone (circled in (90), following Mary Clark's
notation) comes not from the autosegmental approach, but
something quite different. Present generative theory is quite
silent about the way in which syntactic boundaries may be
placed in phonological rules, and more importantly, given a
certain small data set which indicates that a certain phono-
logical process is syntactically conditioned in part, which
aspects of the syntactic structure will be most highly-prefer-
red by the language learner to condition the rule. In short,
we the linguists have too many choices in viewing a structure
as in (90)-(91.) when we want to write the syntactic conditions
for the docking rule. The question is what conditions strike
the language learner as simplest--what will he or she guess
the conditions to be on the basis of (90)-(91). And this
question is quite independent of any autosegmental proposal,
for the same issue could have arisen in a purely segmental
phenomenon.
This much said, let us look at the relevant parts of the
structures in (90)-(91) in more detail: (see (92)). Here we
make explicit references to the convention that corresponding
word-boundaries are associated to each other, above and beyond
all tone/vowel associations. There is a word-boundary to the
left of the floating H-tone that has no corresponding word-
boundary on the syllabic tier, and thus remains unassociated.
P 7 T 1 1dl 'Till i  111 4 1H]!111' 14
S 1 Hill~ l. J~ I H H 1
(92) AVP 17
T4fT H:T it
VP
We may note that there is no inconsistency in having an
operation insert an association line that crosses a previously
present association line. The Well-formedness Condition
prohibits such a structure, and the interpretation we have
given to the Condition will 'change the representation minimal-
ly by addition or deletion of association lines so as to meet
the Condition maximally" (p. 4  above). Thus the old associa-
tion line will delete in the process of putting the new
(dotted) line in (92). The same type of operation occurred in
the I Root rule (above, Section 6).
The absence of a corresponding syllable (na) for the H
tone in (92) is crucial for the operation of the docking of
the H-tone. Our present notation appears to have an arbitrary
asymmetry, in that the presence of association lines is easy
Ii4l 141 i 4I,  aI14 1'1V 1 i
174
to write in as part of the structural description or change of
a rule; the corresponding absence is, so far, impossible.
This asymmetry, however, is unprincipled, as far as I can
see, and we may rectify this by proposing that a notation (i)
for deleting association lines as a structural change, corres-
ponding to the dotted line which indicates insertion of
association line; we shall use an 'x' in an association line
for this, and (ii) a notation indicating that a tone or
syllable is associated to nothing at all--circling the seg-
ment.1 4
With these conventions, two solutions are unfortunately
available. We could write the rule as (93); we could also
write it as (94). Note that the circle around the H means it
is a floating tone--more precisely, associated with no sylla-
ble at that point in the derivation.
(93) M
$P DOCKING RULE
(94) where this means
4fn
The effect of (94) is perhaps not immediately obvious; but
by deleting the association line in (92), we get a structure
as in (95), which, by the Well-formedness Condition, leads to
the association of the floating tone on the final syllable of
175
the subject NP.
(95) 4 CVcv
T T T '
by Well-formedness Condition
CVCV I $VCV 4
4 T* H 4- T4,
Rather arbitrarily, I shall choose (94) rather than (93)
as the appropriate formulation of the docking rule. It must
be emphasized that detailed investigation of other tonal sys-
tems is the only way to determine whether this choice was
correct on the basis of the data.
This formulation of the rule, then, works appropriately
for the cases laid out in (90)-(91). We have looked at (90)
15in detail in (92); the same rule operates in (91).
As the docking rule is written in (94), it docks the
floating tone onto whatever is to its left with no further
conditions as to what is found there to the left. This is
largely, but not entirely, true. We have seen examples as in
(90), where the H docks onto the subject NP in the subordinate
clause; we have seen examples, as in (91), where it docks onto
..llh Ii ll P 11- vi OA ,14H 1 11 -.
the head NP outside the relative clause. In (96), we see
examples where the H tone docks onto a complementizer (ka,
ma) to the left of the clause. There are no examples, however,
of the H tone moving outside the clause and associating with
an NP external to the clause that was not grammatically
related to the clause.
(96) $ma o si ya gara ya ozi$ = ma osi ya gara ya ozi
L@L L L H M HH
comp she send him go him errand
'lest she should send him on a mission'
Thus we find no examples like (97); instead, we find the
the floating tone docking rightward, as in (78); this is
illustrated in (89). Although this process is clearly
related to the leftward docking rule (94), it is not identical
to it. Another rule must be formulated that applies after
(94), or when, due to the structural configuration (94) cannot
apply. This second rule we shall call Rightward Docking (99).
It is the mirror image of (94).
(97) *Khachie uzo mg o eghu
L H HL L H H
shut door I kill goat
'Shut the door lest I kill the goat'
ANT I N Il 'm
I I 
ll li I$+ih IA . H il1 EI-k
(98) K h a c h i e
L H
z 0
HLjL
177
e g h u
H H
S
NP VP S
NP
m gbuo eghu
(99) Rightward Docking
And yet we have no formal principle why the 'floating H
tone docks rightward in a structure like that in (98)., and
not leftward. There is apparently no intonatio~ial pause here,
unfortunately, which would account for this. I shall leave
this question open.
I I H I hi II
178
14. The II Relative/Subject
This form is simple to understand, given the analysis so
far. The tone of the verb stem is Mid; there is no underlying
tone for the verbal prefix a, so it associates with the stem M
throughout the Well-formedness Condition.
This form is used only for relative clauses where the
head noun plays the role of subject of the relative clause.
This head noun undergoes Subject Tone Mutation, if the na pre-
fix does not occur, just as we have seen in the previous sec-
tion. Consider forms like (100) in the II Relative/Subject:
(100) Eghu na ala ahya n'Ekhe vGoats which were
expected to leave
the market on EkheH MHH M L H day'
goat leave market on Ekhe
(a day)
Oso na ala ahya eci 'Pepper which was ex-
II I \'I LI pected to leave the
H L H M H H M H H market yesterday...'
Pepper leave market yesterday
Oso ala ahya eci (same meaning)
H L H MH H M H H
M
As I have noted in (100) above, there is a suffixal H tone
which raises the tone of any suffixes and triggers Object Tone
Mutation. In short, the tonal formula for this form is:
179
(101) Floating Prefix Stem Suffix
H 0 M H
Summary
There is still more to be said about the verbal forms of
Igbo. We have looked at only half of them so far; we have not
yet approached the negative forms. These present slightly
longer tonal melodies, making considerably more use of the
"B*" notation. They do not, however, present much more
information about the functioning of an autosegmental system,
and while the system has a great deal of inherent interest, it
is not germane to our present questions regarding autosegmental
rules.
We should not go without observing that there is some
order in the chart (19). For example, there is a connection--
somehow--between the functioning of rule (72), the rule which
deletes the prefix tone in the II Root form, and the fact that
in the subordinate and relative forms, there is a floating H-
tone and no prefix tone. It is as if the floating H-tone is
playing the same role as the final H-tone of the subject NP
that deletes the prefix tone in the II Root form. Yet in the
subordinate and relative forms, the floating H is always there,
and so the prefix tone is always deleted. How much a regular-
ity should be stated in a grammar--indeed, whether it should
be stated--I leave as open questions.1 6 ,1 7
I I I IIIi~~ !I~I I'* H'
180
(19) repeated
Root
Subord
Relative/
subj ect
Relative/
non- subj
Root
Subord
Relative/
subj ect
Relative/
non-subj
I Complete
Floating/Prefix/Stem/Suffix
- - L -
- - B H
H - M H
H - L -
IT Incomplete
Floating/Prefix/Stem/Suf fix
- B B* H
H - L -
H - M H
- L -
15. Nominal Compounding
We may pursue the empirical validity of the notational
convention suggested above for tones not associated with any
syllables at a certain point (we circled them at that point).
The significance of this move was that we would then expect
to find "being circled" to be a simple condition for a
language to place on a factor in a rule, and that certain
rules should become quite difficult to express if we do not
have access to this notional convention.
The process of compo nding nouns in Igbo is quite produc-
C-,
tive, .and corresponds to a genitive or possessive constru ion
in other languages. Thus "John's book" is expressed in Igbo
as "book John," with the two words undergoing certain regular
tonal changes. Edwin Williams (1971) originally pointed out
that the tonal processes in this construction can quite
naturally be explained in terms of the Object Tone Mutation
and Subject Tone Raising processes we have already seen.
These together point toward an H tone intervening between the
two nouns of the compound construction, an H tone which plays
much the same role as the floating H tones we have seen in
the verbal system.
Chart (102) presents the relevant data, and (103) sum-
marizes the tonal forms. However, in the terms we have
developed so far, these alternations can be summarized easily:
the second noun undergoes Object Tone Mutation uniformly; the
first noun undergoes Subject Tone Raising just in case the
second noun does not begin with an H-tone.
P 11 +41 1 PA PAI I I I
-l I I I I 'l l l 4 IA 1 1., i
(102) First noun
Second Noun H HH LH HL LL
H A \i ji isi ji ahu ji odhu ji mkpa ji
HH eghu 9wa eghu isi eghu ahu eghu 6oeghu abha egu
LH oke wa oke isi oke ahu bke odhu oke abha oke
HL /me wa mbe Ist mbe ahu mbe ulo mbe ala mbe
LL e woewa is1 e WO ahu e o /dhu /ewo abha ewo
(103)
Secnnd First noun
noun H HH LH HE LL
H H M HH M LH M HLH M LL M
HH H HM HH HM' LH HM HL HM LL HM
LH H MH HH MH LH MH HM MH LH MH
HL H HL HH HL LH HL HL EL LL IMl
LL H II HHLL LH LL HM LL LHMLL
In (103) 1 have underlined a form where some
tonal modification occurs.
A ;1Ii iJ II II .HI
183
Indeed, let us take (104) as the underlying structure of
the possessives.
(104)
NP ?
Gen- NP
itive
odhu H oke
tail rat' 'rat's tail'
or more accurately (105)
( 05)
KJ
The rule of Object Tone Mutation (see (40) above) is trigger-
ed by the H affix in (105). The floating H tone will cliti-
cize leftward onto the first noun in (105) by the Docking Rule
(94) , repeated here.
(94) #
Docking Rule
111 ,1- j -
This H tone must be deleted if the second noun begins with an
H, however. We could write this rule as (106):
(106) - # H
1 2
0 2
Indeed, this might be correct. However, Williams argues that
the deletion of the H tone here is to be identified with the
process summarized in the rule of Floating-H Deletion (54) on
page/// above. The formulation given there makes (106) super-
fluous; for once the Docking rule (94) applies, (54) Floating-
H Deletion applies, undoing the effects of the Docking rule by
deleting the H-tone. In short, no new rules are needed at all
to deal with the tonal process summarized in (102) and (103).
vA II III 1.'I 44JjII AI
I I II~ II~II V
185
FOOTNOTES FOR CHAPTER TWO
1. That is to say, rules that are formally and technically
"simple" in the autosegmental framework, but which would
be relatively complex (and thus unexpected) if expressi-
ble within a standard theory.
la. Some take this as evidence that the High Mid sequence
is "really" High Low High. Of course this does not
follow, although in some particular cases it may be
correct. We shall see in a moment that the parallel
dropping of the High pitch register is due to the fact
that the same rule applies in both cases (essentially
a generalization of (13)).
2. We may say that an application A1 of a transformation
occurs before an application A2 of a transformation if
application A1 corresponds to the pair of PMs
(Pi,Pi+l) and A2 corresponds to pair (PjPj+,), where i
is less than j.
3. Two types of possible counterexamples have been suggested
to this ordering of phonological rules (this ordering is
the Sound Pattern conception, of course, not my own).
The first is Bresnan's suggestion that the Nuclear Stress
Rule applies cyclically; the second is the Bresnan-Selkirk
analysis of the blocking of liaison phenomena. The
Selkirk analysis--which explains this blocking in terms
of word-boundaries left behind by transformations in
their operations--is, in one form or another, not strictly
speaking a counterexample to the Sound Pattern conception
of ordering syntactic and phonological rules. Th
Bresnan Hypothesis is clearly contrary to this, h wever,
and would require abandonment of the picture in (18) if
her analysis of the Nuclear Stress Rule were correct.
A I I li Ildi]8 ! I
3-86
However, I do not find her examples convincing, and the
replies in the literature--especially Bolinger's--are
telling responses. If the Nuclear Stress Rule were
a formal rule responsible for the placement of intona-
tion (the star, in the sense developed in Chapter Three
below), then Bresnan would have adduced strong evidence
for its applying cyclically; the question, however, is
whether there is such a rule at all that places senten-
tial accent. If there is no such rule that looks just
at structure--and I do not believe that there is--the
question of when it operates does not arise.
4. All the data from Igbo, unless otherwise specified, come
from Green and Igwe's grammar of Chuhu Igbo.
5. In fact the case here is stronger than I have suggested,
for the connection between the cliticization of the pro-
noun, as in (27), and the non-application of the Flop
rule is supported by some further facts. It is only the
singular, weak forms of the pronouns that undergo cli-
ticization; the diagnostic for cliticization is deletion
of the a/e prefix in the II Incomplete forms, as discussed
above in Section 5 of Chapter One, and below in Sections
11-14 of Chapter Two. Thus neither the singular strong
pronouns nor the plural pronouns cliticize. The singular
strong pronouns do not occur in the I Main form, but the
plural ones do, and they do undergo Flop, as illustrated
in (28). Thus the non-application of the Flop rule is
I I il 4Il I II IT
187
directly connected, not with pronounhood, but with
cliticization.
6. Throughout this thesis, I am using the following evalua-
tion metric implicitly:
1. Boundary elements, major category names (syntactic,
N,A,V,N',A',V',...or phonological, C,V,T) and
association lines are counted as equally costly;
these elements, however, contribute a great deal
less complexity than other feature specifications,
which are counted on an equal par: one feature, one
cost unit. Boundary elements thus have a privileged
position in the writing of phonological rules.
2. Epenthesis of a segments costs as 1 unit above and
beyond the features required in the statement of the
rule.
3. Elements left unchanged by a rule are not counted
twice, of course, in counting the complexity of a
rule, but only once--hence the notational convention
of writing rules that insert association lines in
one step by using dotted association lines to indi-
cate that line is part of the structural change, not
the structural description (see rule (25) for
example).
7. The statement that Low-toned verbs have a Mid-tone is
made by Green and Igwe. Their few examples, however, of
Low-toned verbs here are all without suffix. The present
II I H I 1 TI A Hidi -
188
analysis, therefore, is indeterminate as to whether the
stem tone of the Low-toned verb is really M or L, given
this limited data base. I shall accept Green and Igwe's
statement that the stem tone is in fact Mid; comparison
with the II Relative/Subject form strongly supports
Green and Igwe's point, as we shall see below.
8. Where the LH-Simplification rule is obligatory and where
optional cannot be decided on the basis of the limited
data available to me at present.
9. As mentioned above, there are also uses of the I Relative
B without a head, but rather preceded by the complementi-
zer ka which appears, as usual, with an L tone.
ex: Ahuru m unu ka unu vusi nghughwu n' isi
I saw you as you were parcels on head
carrying (your)
10. The lack of system in (63) should neither surprise nor
disconcert us. It is upon the arbitrariness of such
patterns that the argument in Section 4 of this chapter
was based, arguing for lexical insertion of tonal melo-
dies at surface structure, or after cyclic rules.
11. "B" stands for base, or underlying tone, as in Section 4
of Chapter One. This is meant in the following sense:
there is a two-way underlying distinction for verb tone.
Those we call L-verbs are generally realized on an L-
tone: . The other class, the non-low or High verbs,
iii Il .ll  A li i 10i 1 1 NI k
189
+hi]are generally realized either as M = +10]or
H -[t . In either case, we see the underlying dis-
tinction determines the feature "high"; the specification
for "low" is either simply plus, whence an M/L opposition;
or the specification for "hi" is the opposite sigh from
that of "lo"--that is --whence an H/L opposition.
12. There is a non-discourse-initial form of the II Root.
It differs from the discourse-initial form only with
respect to the behavior of singular subject pronouns. In
the non-discourse-initial form, the non-clitic or Strong
forms are used, and they behave like any other lexical
item.
(i) I bha ya, ya abha M
you grip him 'he grips me'
'When you grip him,' gi Abha h
'you grip me'
mu abha ya
'I grip him' etc.
So far we have no reason to distinguish systematically
(as Green and Igwe do) between the initiating and the non-
initiating forms; all we have seen is that in the II
Root form, cliticizing subjects are used if and only if
the sentence is "discourse initial." All the rest would
follow, and there would be no need for this footnote.
||IGHI II I M
190
But there is one point remaining. There is no non-
cliticizing form of a, the impersonal subject pronoun; it
is permitted in the II Main non-initiating. And rather
than being H, as in 73a and 73b, it is L for both L and
H verbs, as in (ii).
(ii) A 'sweep' (L-verb)
ya bya, a zaa ulo when he comes home, the
house will be swept'
he comes one sweep house
NB: a za a is II Main
form.
__ 'see' (H-verb)
Ebe a gala, a hu ya 'wherever one goes, one
sees him'
place one goes one sees him
NB: a hu is II Main form.
This L-tone on the a is connected, no doubt, with the
fact that cliticization of the subject pronoun moves the
pronoun into the place of the prefix, thus taking on its
tone (see (73) above, the cliticization rule); similar
tone shifts occur below in the II Relative/Non-Subject.
Yet this leaves open the question as to why the clitic
pronouns are in fact high-toned in (72). I have no
answer at this point.
13. Preferably one would have the "because" running in the
other direction.
14. Such decisions are ultimately empirically testable, al-
though with difficulty. The system of rules in Igbo are
I I NI
191
the sort of evidence we need to decide upon which set of
notation conventions most adequately reflect a language-
learner's preferred grammar.
15. If the relativization rule left behind word-boundaries
in the subject NP position, some modification would need
to bemade. While we are not using Selkirk's system, if
one did, it would require that a succession of contin-
guous association lines attached to word-boundaries be
deleted. Nothing particular rides on this, as far as I
can see.
16. This particular example seems to be related to the more
general problem of how to account for the fact that
underlying forms in a generative phonology often have as
morpheme-structure constraints (i.e., redundancy measure)
conditions that mimic the application of the phonological
rules. As I say, I have nothing to add on this question.
17. For those interested in Igbo, the negative forms go
essentially like this:
Floating Prefix Stem Suffix
Root (i) H B* -
(ii) -L HB* -
Subord - H B*
Relative/ (or else stem tone
H- M M .
subject is MM and suffix
Relative/(i) is H)
.u -H B*
non-subj (i.
IIT I ' I L III1 1
192
The two Relative/Non-Subject forms suggest very strongly
that there is indeed a synchronic connection between the
presence of a floating H-tone and the possibility of a
tone in the prefix position. There is another significant
fact here, but just how it should be interpreted is not
clear to me at all. We have equated the position of the
relative marker /na/ with the tonal position of the
floating H tone. In relative clauses, the /na/ is
optional, though preferred. Green and Igwe remark,
however, that they know of no cases where the /na/ is
used in the Relative/Non-Subject (i) , which, under our
analysis, does not have a floating H-tone (it is the
only relative clause form which does not, in fact).
41 HlPA O 4 ~ I k .
193
CHAPTER THREE
MELODIC ASSOCIATION RULES
1. Introduction
In the last chapter, we considered the nature of the
autosegmental derivation, and so we looked at a set of rules
that modified the associations between two autosegmental
tiers--the tonological and the syllabic, or non-tonological.
The particular underlying forms chosen were aruged for, but
the discussion of the nature of the rules that link the two
tiers underlyingly--the "melodic association rules"--took a
back seat to the discussion of the subsequent rules in the
derivation.
The purpose of this chapter is to fill in part of this
gap, and to consider possible ways for autosegmental tiers to
be linked. In this sense, the present chapter considers the
nature of the grammar at a point just slightly "deeper" than
that part considered in Chapter Two. Chapter Four below, on
the other hand, will deal with processes more "superficial"
than those of Chapter Two. Thus while in Igbo we considered
the number of autosegmental tiers--let us call this the "auto-
segmental index"--to be fixed at two, in Chapter Four we shall
consider the possibility of the autosegmental index increasing
during the derivation. As we shall see, furthermore, a stage
in a derivation with more autosegmental tiers is always more
"superficial" than a stage with fewer tiers.
194
In the present chapter we shall consider the way in which
two strings of segments are selected from separate entries in
the lexicon and then associated with each other, thus permit-
ting the autosegmental derivation to proceed.
We shall assume that such a linking process, or "melodic
association rule," is ordered in the derivation at the earli-
est point possible. Such a hypothesis is consistant with all
languages I am aware of, and of course, serious constraints
are necessary for the theory regarding that phonological
processes may occur before any melodic association rule, and
thus before the Well-formedness Condition comes into effect.
This last point--the effects of the Well-formedness Condition--
is our major empirical diagnostic for determining the ordering
of a phonological rule before or after the major association
rule.
In general, two types of melodic association procedures
are available: accentual and non-accentual. Accentual sys-
tems distinguish themselves by assigning the tonal melody to
the syllabic tier by means of an abstract marker, which we
shall denote by an asterisk or "star." In the simplest case,
there is exactly one star in the domain corresponding to each
tonal melody; and on the tonal level, one tonal segment bears
a star also. If the language then contains the simplest
accentual melodic association rule--(l)--then the two tiers
will be matched as in (2).
M ih 14 I ] !:fl .114144; Al I i I! k k
I ii 4
195
(1) S (2) S S2 23 S4 5
T T T T
We may, however, find either too many or too few stars
on the syllabic level. Thus it may be that no star appears,
and some provision must be made for aligning the two tiers.
On the other hand, there may be more than one star available,
and some decision must be made to indicate which star has
precedence in the application of rule (1) or whichever melodic
association rule occurs in the language.
The effect of accentual association rules is to make a
certain vowel or syllable prominent or "distinguished," and
this apparent purpose has certain consequences for the Well-
formedness Condition. While the Well-formedness Condition
requires certain restructurings of the autosegmental repre-
sentation at times, it may leave open more than one way to do
this. The nature of accentual systems leads to a lessening of
this ambiguity in ways that preserve the prominence of the
starred segments.
The alternative to accentual systems is the type of
system we .have already seen in Igbo and, more briefly, in Tiv
and Mende, in which only categorial information was involved
in aligning the tonal and the syllabic tiers. In the simplest
case, the tone melody is mapped left-to-right, one-to-one,
onto the syllables. This pattern of association has been
I 1114 HIM
196
considered with respect to Etung by Edmundson and Bendor-
Samuel (though rejected by them), Margi by Williams, and Mende
and other languages by Leben. We shall consider several cases
of this and other principles in this chapter.
2. English as a Tone Language
We shall begin with a look at the tone patterns of the
simplest part of the English pitch system, the utterance of
words in isolation, as at the beginning of a dictionary entry
-- reserving for the moment judgment as to whether this reflects
tonal or pitch registral rules, or something else still dif-I9
ferent. These citation forms, furthermore, are the forms
that standard phonological treatments of English concern
themselves with.
A number of characteristically autosegmental traits show
up in this system. First, we find contour tones on short
vowels which must be analyzed as sequences of level tones.
Second, we find independent tone melodies in the lexicon.
Third--and here there is a caveat1--the tone melody spreads
itself over the syllabic structure in a characteristically
autos egmental way.
All of this will lead us to the conclusion that tone is
autosegmental in English. We have already seen some of the
tonal analysis for English in Chapter One, Section 2. There
we found that the neutral melody for English was H L (or MHL).
The starred toneme is associated with, the syllabic which, in
11 ii 4
197
the Sound Pattern of English, Trager and Smith system would
receive a '"stress." In Liberman (1975),, it is called the
"designated terminal element.."
In Chapter One we looked at the example of archipelago,
sketched in (3),; the same point--that "lstress" is the sylla-
ble associated with the starred H toneme--is clearly brought
1 3 3 1 1 3 3 1
out by such minimal pairs as magazine/magazine, cigaret/cigaret,
and so forth. What could be represented as the difference
between a "1" and a "3" is simply that the pitch contour falls
on the lstressed syllable, as sketched in (4).
(3) rehipelag
HHL
(4) (a) m a gaz n e
ilL
(b) in g a z i n e
ilL
1 I
198
For my speech, these frequencies are physically approximately
the following: in initially accented magazine (4a), the first
syllable is at 100 hz.; the second two at about 84 hz. When
the accent is final (46), the pitch of the first two syllables
rises from an initial 92 hz to a peak of about 102 hz, falling
sharply on the final syllable to about 87 hz.
The theory being proposed here is making heavy use of
what is called the "lstress"; the other levels of stress
inside the word, 3stress and no stress, are playing no role.
In (4a) and (4b), the H toneme and the L toneme spread them-
selves through the graces of the Well-formedness Condition,
but irrespective of the nature of stress on the syllable level.
What is noted as "3stress" is a segmental characteristic which
has an important consequence for vowel quality, but little
effect on such suprasegmental matters as pitch or intonation.
A stressed vowel will tend, perhaps, to be a slight bit higher
in pitch due to the air pressure, but not significantly.
Both the lstressed vowel and the 3 stress vowel differ
from the unstressed vowel by two major processes that Chomsky
and Halle demonstrated in Sound Pattern of English. Put simp-
ly, vowels divide into those that are tense and those that are
lax; and among lax vowels, those and only those that are un-
stressed reduce to schwa. Among tense vowels, only those that
are stressed undergo vowel shift.
The assignment of the "lstress," or star, is given by the
provisions of Schane's "detail" rule, cited in Halle (1973):
I ITI11 I 1ilV l ,
199
(5) V + * / - Q (VC 0 (+y)) ## Condition: Q $ [+stress]
Chomsky and Halle, following Trager and Smith, produce a
description of English with several stress levels, where these
stress assignments are essentially segmental features. Liber-
man (1975) proposes a revision that maintains a binary seg-
mental stress feature, but provides an organization of neigh-
boring stresses, an organization of relative weaker and
stronger, so that given a word of arbitrary length, there can
be stressed vowels subordinated to any arbitrary degree with
respect to the mainstress.
Our primary interest here rests in showing that pitch
phenomena in English display certain autosegmental character-
istics. Thus Bolinger's insight--that main-stressed syllables
are manifested as pitch extrusions, either up or down--is
clarified and extended in an autosegmental analysis, which
furthermore provides that the extrusion be essentially level
in tone, except in the case where the mainstress is phrase-
final, in which case an "extra" toneme may be associated by
the Well-formedness Condition, creating a gliding tone. We
have already seen examples of this in Chapter One, and we will
return to it below, when we consider other melodies.
The existence of non-primary stresses is, therefore, not
our central concern, nor their organization in time; but we
may ask , if main stress manifests itself as a pitch extrusion,
do secondary and tertiary stresses do likewise? If not,
200
should this surprise us?
Chomsky and Halle, who defend a multiple stress-level
analysis in Sound Pattern of English, suggest the following,
concerning the reality of stress-levels:
We do not doubt that the stress contours and other
phonetic facts that are recorded by careful phoneti-
cians and that we will study here constitute some sort
of perceptual reality for those who know the language
in question. In fact we are suggesting a principled
explanation for this conclusion. A person who knows
the language should "hear" the predicted phonetic shapes.
In particular, the careful and sophisticated impression-
istic phonetician who knows the language should be able
to bring this perceptual reality .to the level of aware-
ness, and there is ample evidence that phoneticians are
capable of doing this. We take for granted, then, that
phonetic representations describe a perceptual reality.
Our problem is to provide an explanation for these facts.
Notice, however, that there is nothing to suggest that
these phonetic representations also describe a physical
or acoustic reality in any detail. For example, there
is little reason to suppose that the perceived stress
contour must represent some physical property of the
utterance in a point-by-point fashion; a speaker who
utilizes the principle of the transformational cycle and
the Compound and Nuclear Stress Rules should "hear" the
stress contour of the utterance that he perceives and
understands, whether or not it is physically present in
any detail. In fact, there is no evidence from
experimental phonetics to suggest that these contours
are actually present as physical properties of utterances
in anything like the detail with which they are
perceived (26).
It is to be expected that determined [fi.e., not freely
chosen--JG] phonetic features should be quite difficult
for the user of the language to learn to identify, whether
they involve stress or degree of aspiration (where
undoubtedly there are many levels, predictable, at least
roughly, by general rules). The apparent ease with which
phoneticians trained in the same convention can, to a
large extent, agree on the assignment of four or five
stresses in utterances may very well be traceable to
their ability, as speakers of the language, to grasp the
syntactic structure of utterances and to assign to them
an "ideal" stress contour by the rules of the transfor-
201
mational cycle...
To summarize this discussion of phonetic representation,
we do not doubt that representations of stress contours
and similar predictable phenomena correspond, up to a
point, to some perceptual reality that can be brought to
consciousness with training and care... the representation
of the perceptual facts is likely to be governed in part.
by arbitrary convention or irrelevant cognitive limita-
tions after a certain degree of complexity is reached.
Thus it is impossible to expect (and, for purposes of
investigating linguistic structure, unnecessary to attain)
a complete correspondence between the records of the
impressionistic phonetician and what is predicted by a
systematic theory that seeks to account for the percep-
tual facts that underlie these records (27).
It is difficult to agree with this passage for the follow-
ing reasons: (1) Chomsky and Halle assume a concensus among a
group of linguists ("careful and sophisticated impressionistic
phoneticians," in particular) regarding the representation of--
if we can find a theoretically neutral term--syllabic promin-
ence. In fact, there is no such consensus or uniformity.
There are traditions with their notations; but there certainly
is no general unanimity.
Vanvek (1960) , for example, says, "One is inclined to
think that it is possible to perceive secondary stress since
phoneticians have written about it. But often, I think, it is
really inherent prominence that one hears" (i.e. non-reduced
vowel quality--J.G.) (p. 73). Later, "It seems doubtful
whether it is at all necessary to mark secondary stress after
a full stress in the same word, e.g., foot-passenger, make-
believe. The inherent prominence of vowels following the
fully stressed syllable seems to settle the question automati-
I IVI PT I ,!- I I OL I
202
cally in the above cases. Finally, there are cases like
subcommittee, telerecord (noun), unliquidated (attributively).
I wonder whether English people really hear a secondary stress
here, or whether D. Jones is influenced by the words committee,
record, and liquidate when he puts the secondary stress mark'
(75)
On another line, when dealing with forms like light-house
keeper, and light house-keeper, Bolinger and Gerstman (1957)
conclude "Since the disjunctures transparently supply a
physical separation whose width corresponds inversely to the
semantic bond, it follows that the disjunctures function
directly to carry the information, and not indirectly as
components of a hypothetical stress. Rather than attempt to
salvage the stresses it would seem better to reconsider the
juncture complex, for it is apparent that the uniform "plus"
juncture with which Trager and Smith have marked the examples
has overlooked the function of disjuncture."
(ii) The issue is--and has always been--not whether
certain differences exist, but rather that the linguistically
or psychologically correct way is to represent these differ-
ences, at the physical, the phonetic, the phonological, and
whatever other levels may be relevant. If the nstress system
has a "perceptual" reality--and a reality dependent on the
observer's familiarity with the language at hand--then the
"careful phonetician," whose training is intended to overcome
the particularities of a language and attain a language-inde-
9-11h Ili I I A:, 111 jrIIj:;,1 o" ]:.'[, r 'I
203
pendent description, would be in a poor, not privileged,
position to report on these "perceptually real" stress levels
-- qua phonetician.
(iii) Whether the predicted stress contour matches the
physical description of the signal "in a point-by-point
fashion" is not the issue; it is certainly possible that the
most adequate treatment of English stress contains of a level
of nstress, as Chomsky and Halle suggest, and then translates
this in a complex way into a combination of physical cues.
lstress would be physically realized in a certain way; 2stress
by some other way, 3stress by yet a third, and so on. While
this would be surprising, it would not affect the validity of
Chomsky and Halle's position. Their position would rest
supported if, at a certain level in the derivation, the n
stress levels were the appropriate representation, even if this
level is not closely related to either the phonological level
or the acoustic level (if we admit an acoustic "level" in our
conception of grammar).
If, however, the relevant properties of the acoustic
signal--syllabic length, intensity, and pitch--can be predic-
ted from other considerations, like autosegmental spreading of
tone, and placement of word-boundaries, then the nstress level
system would not be empirically supported. In fact, the
question remains an open one; there is as yet no successful
system that explains English syllable length, intensity and
pitch. However, we shall look at some cases below, and the
1111 I Irl 1 q i .: IJ L
2o4
facts tend to the conclusion that whatever difficulties remain
in present theories, the analyses would not be improved by
having access to a level with nstress features in the grammar.
It would no doubt be worth stepping back from this
particular issue for a moment to see the broader consequences
of this issue, for whether there are two levels of stress in
English or four is not of great significance to anyone in and
of itself., I hope. One related issue is the existence of the
transformational sycle; I refer the reader for a discussion of
this to Chomsky and Halle (1968), Chapters Two and Three, and
Liberman (1975) Section 4.1. Another issue is whether there
is a linguistically real distinction between pitch-accent and
stress-accent languages. To the extent that paradigm examples
of pitch-accent languages like Japanese (see Haraguchi (1975))
can be dealt with in the same way as so-called stress-accent
languages like English, the distinction is untenable.
Returning one last time to Chomsky and Halle's remarks:
"the representation of the perceptual facts is like to be
governed in part by arbitrary convention.., after a certain
degree of complexity is reached." On this we are agreed; the
question is, now, whether inclusion of "3stress" levels, and
so on, is arbitrary, or part of the linguistically real system.
We shall argue that (i) the perceived difference between
the 3stress in (6) and the 2stress in (7) below is due to the
difference in length; that (ii) this difference is not a
general fact about syllables designated as 2stress and 3stress,
205
and not even a general fact distinguishing compounds from
adjective-noun pairs (as in (6)/(7)), for it occurs in only
a subpart of these forms; and that (iii) therefore the
difference in length is not attributable to a difference in an
abstract nstress feature, but to certain straightforward
lengthening processes sensitive to word-boundaries (#); we
shall see that such contrasts as those between black-board
(that one writes on) and Black board (a committee presided
over by Justice Hugo Black) show that certain actual differ-
ences would not be describable if length were predictable from
an nstress feature.
1 3Consider first the compound black-board, an object that
may as well be green or white as black; it contrasts with
2 1black board, which is black but likely has nothing to do with
chalk (we shall consistently use the hyphen to distinguish
1 .3 1 3these). Black-board's pitch pattern is like that of magazine
2 1(see (6)); black board' s is as in (7).
(6) blick-b ard (7) black rd
H L H L
The "2stress" on black in (7) is supposed to indicate a promi-
nence that board--a 3stress--does not have in (6)-. Although
the representations in (6) and (7) do reveal a difference
2 3between the two syllables--black is H-toned,, board is L-toned--
206
a glance at (4b) shows that this could not be the relevant
difference, since 3-stressed syllables can be H-toned. An
analysis of the duration of the syllables shows, however, a
clear difference, a difference which is not hard to hear. We
can make the pa ir in (6)-(7) minimally different by giving
them the same tonal pattern as in the context, "But that's my
-- ," which is roughly as in (8).
(8) But that's my black board.
For some reason--which we shall return to--the black in (7)
black board is considerably longer in duration than the black
of (6) black-board.
Now we shall see that this difference of length in (6)
and (7) is not due to a difference in stress patterns, if
stress patterns is taken as in Sound Pattern. Other 2stresses,
for example, do not lengthen as the black in (7). Thus (9),
when the cyclic application of stress in the SPE system applies
to give a 1 2-4 stress pattern, displays no lengthening of
black in black-board.
(9)
first cycle
1 3 second cycle
1 3
1 2 4 third cycle
-Id I d.. ;I JI I 11! i '11, rl uT 11! iq 1;1 ml I I -N 1, 711.
Les I l| I | Im!ln I 1
207
Conversely, an adjective-noun sequence as in (10), where the
adjective is not monosyllabic, displays no lengthening of the
first syllable corresponding to the black in (7).
2 1(10) hacking cough
In general, then, being marked as 2stress is not a sufficient
condition to undergo the lengthening displayed in (7).
Nor is it a necessary condition. Monosyllablic lstress
1
words undergo this lengthening, as in The board was black.
Yet this occurs, again, only if the word is monosyllabic. In
short, the difference in length observed in (6)/(7) arose from
something quite other than a 2stress/3stress distinction. All
the facts observed to this point concerning length follow from
the hypothesis that lengthening occurs to monosyllables
surrounding by double word-boundaries (11) , if we assume that
lexical compounds like black--board have the structure##black#
board##.
(11) Monosyllabic Lengthening Rule
V + lengthened /.##C - C ##0 0
The assumption that lexical compounds have just the sin-
gle word-boundary, rather than the double word-boundary pro-
posed in SPE, is a natural consequence of the lexicalist
hypothesis, not yet adopted in Sound Pattern. Given that the
lexicon can construct words and place morphological boundaries
between morphemes, it would be surprising if there were no
208
constructions that did not place single word-boundaries inside
a word; in effect, we are suggesting that the difference
between a compound like black#board and words like govern#ment
is that both halves of the former could be words, something
not true of the latter.
The Monosyllabic Lengthening Rule (11) mentions nothing
about stress, and so we would expect it to apply equally to
lstressed ("starred") vowels and to the 2stresses we have
seen. This is, as we observed above, true (The board was
1
black); the distinction is equally manifested in the pair in
(12)
(12) The lion turned on his trainer (no lengthening)
JF4En* hs t&nEr*
The teacher turned on his students (lengthening)
Furthermore, there are corpounds that would be described in
the n-stress system as having 1-3 patterns which display a
different lengthening. If there were a committee chaired by
Justice Black, such a Black board, though starred on the first
209
syllable, thus displaying a tone pattern as in (13), is clearly
1 3
pronounced differently from black-board.
(13) B la c k- b o a r d
H L
The difference here has nothing to do with tone or stress
in the binary sense: Black is simply lengthened. Whether it
is by the Monosyllabic Lengthening rule, suggesting two
word-boundaries in (13), which I suspect, or whether this is
due to some other process, is irrelevant: the point is that
describing both as 1-3 is inadequate, and attention to syllable
lengthening that the autosegmental analysis forces is necessary
here.
It would follow as a consequence of the strong assump-
tion that stress is inherently binary that the difference
between 4stress and 3stress that has been cited word-inter-
nally must be due to some other factor. In an attempt to
measure this difference, I made recordings of the two pronun-
ciations of the word "Ticonderoga," one with a "34010" stress
pattern, and one with a "43010" pattern. The difference refers
to whether the first syllable is more prominent in some sense
210
then the second, or conversely. I was rather liberal in
trying to emphasize either the first or the second syllable
at the expense of the other, but the recordings sounded
natural.
Below in Table 1 are the results of the analysis of the
recordings for intensity, pitch, and syllable length. I
could not measure the length of the final syllable, and its
pitch measurement is probably inaccurate, but it is not
relevant to our present question, fortunately.
Table 1 T34eog1 A %%o
Ticonderoga T41c n deroga
Pitch (hz) 104/100/88/88/62? 92/108/92/96/72?
Amplitude 6.8/6.5/2.1/4.8/1.8 6.8/6.3/1.8/6.7/0.8
approx. 160/280/100/180/? 130/260/110/150/?
syll.
length (msec)
In short, what we find as a difference between the two
utterances such as it is, is that the second syllable is
perceived as more prominent when the pitch goes up rather
than gradually down. There is no significant difference in
amplitude of the syllables in the two pronunciations, and our
hypothesis is that the difference between these two pronuncia-
I e n Il I B H. 0.15 l 111 '
211
tions is one that is not linguistically significant. It could
not be used to differentiate two morphemes, for example; the
two are in free variation. Why? Simply because pitch may go
up and down before the main accent of' a word; it is this free-
ly varying pitch that has been interpreted in the Ticongeroga
case as the reflex of stress. But we must emphasize that the
discovery of pitch differences between the two ways of utter-
ing the word does not support the nstress analysis; there was
no reason in that system to expect the 3/4 stress difference
was purely a matter of a slight rise in pitch for the 3stress,
and no other prosodic effect. On the rather more simple-
minded "only binary stress" view, th6se minor pitch variations
are to be expected, but they are also expected to play no sig-
nificant linguistic role. Which, in fact, seems to be the
case.
All this does not disprove the validity of the nstress
system; but it does call into question what the empirical base
is that it is supposed to rest upon. Pitch, however, not
stress, is what is our concern here.
Let us turn to some more complex intonational examples
from English, where the tonal melody is not the simple H L.
Consider the disjunctive question, the sort where the asker
assumes that the person asked wants one of the choices: "Do
you want coffee, tea, or milk?" asked in such a way that the
person asked feels it incumbent upon him or her to take one
of the three. This could be sketched as in (14).
212
(14) Do you want coffee, tea, or milk?
This constrasts sharply with the question form where the
person is simply asked if he or she would like any of the
choices offered, and where the answer "no" is perfectly all
right, as in (15).
(15) Do you want coffee, tea, or milk?
(15) is the intonation of a normal, polite yes/no question,
which we shall not consider here.
The choices in a disjunctive question may number any-
where from two up, since to make a choice, at least two op-
tions must be present. If the question has N options--in
(14), N = 3--the first N-l options are spoken with a rising
intonation on each option, arid the last option has a falling
intonation.
In particular, the rise and fall are star-linked. The
starred element in each disjunct or offer is the final low of
that phrase, in the first N-1 options. In short, the pitch
213
pattern is exactly what would derive from the autosegmental
formula L H for each of the first N-1 disjuncts. When the
disjunct has the star on the final syllable, this is realized,
as we would expect, as a rising tone ("tea", e.g., in (14)).
Precisely the same type of observation leads to the
conclusion that the tone.melody for the final disjunct is H L.
To derive (14), the autosegmental representation is (16),
where corresponding boundary elements ($) have been linked.
(16) $ Do you want coffee $ tea $ or milk $
LH $ L H $ HL $
In each domain defined by the boundaries, the major
association rule (1) applies, giving (17), which becomes (18)
through the Well-formedness Condition.
(rr) $ Do you want coffee $ tea or milk $
L H L H1 H L
!IlF " ,,I lllh 1 11 i1l ' I t . 11  jIm
214
(18) $ Do you want coffee $ tea $ or milk $
$ L H $ L H $ H L
This (18) describes formally what we sketched in (14).
The formula for the disjunctive question, then, is as in (19).
* n *
(19) $ (L H $) H L $
There is another possible intonation contour for the
disjunctive question in (14), one which is slightly less
polite, it seems to me, which looks like (20).
(20) Do you want coffee tea or milk?
The formula for such an intonation is (21); we leave off the
derivation, for it procedes just like the others we have seen.
* n *
(21) $ L (H $) H L $
Compare this with (19). It is more than slightly remarkable
that these two contours, used in about the same sense, have
tone contours which,, in this formalism, differ precisely in
the parenthesization. Note that the hopping of the star from
1!,1 I h 'A r 1. M
215
the L, in (19) to the H in (21) is in effectly predictable;
every phrase must have a star so that it may be associated
with the phonolgy in English, and if the L had taken the
star with it when it left the parenthesis in (19), this would
not have occurred.
The Well-Formedness Condition in Accentual Systems
In the discussion thus far of the application of the
Well-formedness Condition to accentual systems, it has been
made to do more work than we might rightfully expect of it.
In the first example that we considered--archipelago--the
Well-formedness Condition revised (22) to (23).
(22) atchipelago (23) ripelago
H L H L
While it is true that (22) does not satisfy the Condition and
that (23) does, the revision to (23) was not, strictly speak-
ing, forced. (24) and (25) also satisfy the Condition; why
was (23) select d rather than either of them?
(24) Archipelago (25) h 1 go
H L H L
216
Why (24) was not created form (22) is clear; it contains (23)
as a subpart, and (23) does satisfy the Condition. Therefore
(24) would not be a minimal way to modify the structure (22)
in order to satisfy the Condition. But any difference to be
found between (22). and (25) must make reference to the star,
for (25) would have been the correct structure, rather than
(23), had the star been on the fourth syllable rather than
the third.
The principle that we need is one that insures that the
presence of the star on a syllable on the syllabic level finds
its way into the association structure. One simple way to
state this is:
(26) Given ambiguity in ways to fulfill the
Well-formedness Condition, do not reassociate
a starred segment.
That is, if there is a choice about associating a segment with
a starred or an unstarred element, associate it with the
unstarred one. The point is to preserve the prominence of the
star, or accent. A structure like (27) will reassociate by
this interpretation of the Well-formedness Condition to (28),
rather than (29) or (30). If we mark each association line
associated with a starred element, we see more graphically
that the output (28) is preferred because the starred associa-
tion is most unambiguously identifiable.
21T
(27) A B O.D E (28) ABC DE (29) ABD E (30) A B3 CD
1 23 4 5 12 3 45P23 12 34 5
We will modify this point somewhat below; see also the discus-
sion of Sanskrit in May and Goldsmith (1975).
It is worth noting that this is the first use of the
star in autosegmental rules so far other than for the initial
major association rule.
3. Non-Accentual Melodic Association Rules
Determining the conventions on the interaction of the
star and the Well-formedness Condition is intimately connected
with determining the types of possible melodic association
rule, since in general the more powerful the Well-formedness
Condition is, the less will the language-particular rule need
to do. We have thus far considered only a simple accentual
system--that of English--and proposed a strengthening of the
interpretation of the Well-formedness Condition in line with
this.
In general we would not expect all melodic association
rules to be as simple as (1), the "star-to-star" association
rule, if only because not all languages have a star (accent).
In Section 4 of Chapter One we reported on Leben's analysis of
218
Mende, where he showed that the tonal melodies are mapped
one-tone-to-one-vowel in a left-to-right procedure. Repeating
the tonal classes here, we find five groups:
(31 H pele, ko
L bele, kpa
A
HL kenya, mbu
LH nika, navo, mba
LHL nikili, nyaha, mba
Whether we say that the one-to-one mapping is a productive
process or not--essentially depending on whether we take these
words to arise from two separate lexical entries, one syllabic
and one tonal--nonetheless it cannot be an accident that the
pattern associates with the syllables in a particular, non-
arbitrary way. And we must propose a formalism that expresses
this intuition. In Luganda, below, we shall suggest that such
a formalism is necessary. for a productive part of the grammar,
for the melodies and the syllabic morphemes are underlying
separate. The same point has already been observed in the
case of Tiv in Chapter One, Section 3.2.
What factors must be stated in the association rule that
are not part of the universal theory so far? The fact that the
mapping is left-to-right, for one thing, follows from nothing
we have seen so far. Yet all the examples we have seen, or
will see, of non-accentual systems bear this property. We may
observe, then, without any principled explanation, that the
I " , i I , V ll ' l1 1 dli 4  !-4i.; I:cli -
219
unmarked pattern is left-to-right mapping.
We can approach this, however, with certain principled
hypothesis. Consider, for example, the Local Environment Con-
dition, proposed in line with Sanskrit in May and Goldsmith (1975).
(32) Local Environment Condition (tentative): No
autosegmental rule may mention in its structural
description a segment in a syllable, or a toneme,
unless that segment is (in a syllable which is)
affected by the rule, or associated with a syllable
or toneme which is affected by the rule.
The local environment condition proposes that elements that
condition an autosegmental rule must be affected by the rule,
in a certain sense. As it stands, the LE Condition is probably
too strong (see, e.g., rule (54) of Chapter 2, a counter-
example),, but it is, I think, in the right direction.
One situation the LE Condition rules out is where
judicious use of symbols permits a phonological rule to count.
It rules out, for example, the possibility of a language like
Mende but where the melodic association began on the second
syllable. This would look like (33), and such a rule would
contain within it something like (33a) (leaving aside details
we have not considered yet).
(33) 4 CV cv cv cv cv
dh T T
220
(33a) 4 CV0
One cannot speculate too much at this point, given our limited
knowledge of devices employed by existing languages. The
strategy illustrated in (33), for example, is surprisingly
similar to one that might propose for Kikuyu.
If these suggestions are correct, however, they imply
that non-accentual mapping must start from the extreme left
end or the extreme right. Once again, while this is intuitive-
ly plausible, it is in no sense the null hypothesis.
A case that we wish to rule out--a language like (33)--
would reveal itself, as we see in (33), by having tone
melodies like Mende's but which began on the second syllable,
the first syllable matching the second because of the Well-
formedness Condition. The question of the possibility of
melodies which begin with two identical tones then arises.
The task of falsifying the hypothesis suggested here would
then be simpler if we could be sure that tone melodies never
begin with two identical tones. This would be a consequence
of a proposal made by Leben, cited above in Chapter One as the
Obligatory Contour Principle: .Adjacent underlying tonemes
-,q I, nr ; , r . '10 1 Wi I i'll"41' ilimilflol 1-1 1
-4 . I , - 41 I I I 1I I
221
must be distinct. The system of Etung, however, appears to
show conclusively that this hypothesis, while extremely attrac-
tive, is too strong.
The Etung system is sketched by Edmundson and Bendor-
Samuel. They observe,
In an initial analysis of Etung, four contrastive
pitches are found: high, low, rising, and falling...
In examining the occurrence of the rising and
falling tones it is found that these usually occur on
the final syllable of the phonological word. In
addition, these tones very rarely occur on three
syllable words, but rather on words of one or two
syllables. Thus, the rising and falling tones are
generally found on the final syllables of one or two
syllable words.
.. .An analysis was therefore set up stating tone as a
feature of the phonological word, and this made it
possible to eliminate the two glides [Rising and falling
--JAG] as separate tone units and so to describe
Etung in terms of high and low tones plus downstep.
With this analysis, a limited number of tone patterns
are set up for the phonological word.
They then consider two ways to state the generalizations
more precisely. The first would be to describe "each tone
pattern as consisting of a sequence of three tones," and then
to express how this pattern is spread out over one, two, and
three syllable words.
"A more satisfactory type of statement," they continue,
"would set up these patterns as contours marking the phonolo-
gical word rather than as a succession of three tones which
are compressed in certain circumstances [i.e., two level tones
compressed to a glide in mono- or bi-syllable word, JAG]. In
this way, pattern 1 [L, LL, or LLL--JAG] would consist of a
222
level, low contour.. Pattern 2 [LH, L H, L H H--JAG] would
consist of a contour starting low and rising to high...
Statements could easily be made for the patterns and in addi-
tion such statements could show how each contour was realized
in words of three, two, or one syllables."
Let us look at their tone pattern classes (I exclude
three because they do not seem to be underlying tone classes,
2
but this is irrelevant to the argument at hand) .
1 Syll
kpS 'even'
bem 'softly'
k~a 'to,into'
v
no 'how'
A
tik (future
particle
na 'it is'
kpa 'first'
tfen 'all'
2 Syll
hm 'juju'
egu evening'
nsi fish'
ekat 'leg'
egom 'jaundice'
oda platform'
gban servant'
nse 'father'
cbo'arm'
etl 'head'
nsi 'mud'
ebin 'farm'
/ AI
efo cloth'
eA
erop 'spear'
3 Syll
eyuri 'dress'
nkdi 'corn'
bisope 'spoon'
'wife'
akpuga 'money'
bir9ma 'me'
$kim 'prosecu-
/ / ter'
ekue '-forest'
mbut 'rain'
mb4u 'goat'
orobe 'beam'
njae 'hunger'
pgare 'pepper'
e see ''sand'
(34)
1) L
2) LH
3) HL
4) H
5) LHL
6) LLH
7) HHL
I -t l A!. I'll I 1 If Ti I I : m I ul U A ! I
223
8) HLH - abo 'they edimba 'pot'
The conclusion is clear: we have in Etung a small class
of tone formulas that may be spread over words of one, two,
or three syllables, proceding from left to right. These
melodies are L, LH, HL, H, LHL, LLH, HHL, and HLH. The
occurrence of contour tones that motivating Leben's left-to-
right mapping in Mende occur here, but the melodies HHL and
LLH make it clear that the Obligatory Contour Principle is too
strong in fact.
In an attempt to preserve the Obligatory Contour Principle,
one might reply that there is no melody L L H: if you find
forms like orobe, then that must come from L H. The form we
find just proves that, to our surprise, Etung is an accentual
language: the melody L H has an accent--on the L--and orbe
is accented on the second syllable. Thus:
(35) or/ b e r o b e
L H L H
Why is this wrong? While this move is less interesting apriori
since a highly counter-intuitive ply had to be called on to
preserve the Obligatory Contour Principle, this does not
prove it wrong; it is conceivable, of course, that (35) is
right. Nor does it prove (35) wrong that this line of
11 I k H, 114.- ElP 21 41-1
224
explanation leaves unexplained what was explained before:
'~V
that there are no words with the tone pattern CV CV CV, which
would have, in the non-accentual account, come from the
putative pattern LLLH. On the non-accentual scheme, this does
not occur because there are no tonal-patterns with four
tonemes;on theaccentual scheme, this is an accidental gap,
because "final-accented" words with the tonal-pattern L H
should be CV CV CV (see (36)). So the accentual scheme can't
explain this gap; this too does not prove it wrong. There
are such things as accidental gaps. The real question is:
how does a language-learner decide if the language is accentual
or not?
(36) cv cv cv CV CV CV CV CV CY
L H L H
One plausible initial hypothesis might arise from the
observation that while English, an accentual language, has
only one neutral intonation contour, Igbo, Etung, and Mende
appear to have several tone contours which are essentially
neutral--they may appear in isolation, for example. We might
speculate, then,, that an accentual system may not, in princi-
ple, have more than one neutral tone melody. (A non-neutral
melody is one used to express a question, or in some other way
convey a speaker's particular attitude.)
225
This speculation, while it appears to contain a germ of
truth, is wrong. Goldsmith (1975a) and, in much more detail,
Haraguchi (1975) shows that in many Japanese dialects,
accentual systems coexist with more than one neutral tone
melody.
I believe we should look in a different direction for a
deeper understanding of the interaction of the Obligatory
Contour Principle and accnetuation. The Obligatory Contour
Principle has going for it the fact that it is almost always
true: this hardly seems accidental. The examples from Etung,
however, disconfirm it in its strongest form.
I would suggest that the Obligatory Contour Principle is
not a condition on underlying forms, but a reflex of the way
in which underlying forms are learned. In Chapter Four below,
we shall consider certain related questions in more depth; for
now, let us observe that a word orobe has as its "tonal
projection"--that part of it that is purely tonal--simply the
melody L H; see (37) .
(37) (a) .be syllable (b) orobe
Low High tone LH;
(38) orobe
L L H
We shall suggest that it is a more complex operation for the
language--learner to change (37b) to (38) than to leave the
226
tonal melody as it stands in (37).
Nonetheless, in order to map the melodies (LH, LHL, etc.)
in Etung onto the syllabic structure, the language-learner
constructs a one-to-one mapping procedure. On the basis of
this, the more marked--more complex--operation of expanding
the one phonetic L into two underlying Ls is called for,
violating, in the final grammar, the Obligatory Contour
Principle.
In short, the Obligatory Contour Principle is a condition
not on possibl underlying forms, but on simply-learnable
grammars; not 11 grammars, however, need be simple. Further-
more, the Principle arises as an artifact of the way in which
tone is processed (though this more important point has not
been argued here; we consider this in Chapter Four).
The one-t&-one tone-to-syllable mapping, as we observed
above, starts from the left. Suppose we had a rule, corres-
ponding to the accentual rule (1) above, that simply attached
the first tone to the first syllable:
(39) # CV (1) repeated V
# T
T
The would operate on (40) to give (41).
(40) # e dimba # (41) # e dimba #
# H L H # # H L H #
I -I P-, I.idl III I I I Tv! 41141119vC N 1 41 1 i
pl II I I IiI.
227
Now there is only one way for the Well-formedness Condition to
be satisfied, as defined sin Chapter One; (41) must become (42).
(42) # e di mba #
# H L H
What is missing from this account is a notion of left-to-right.
We have posited the simplest possible non-accentual major
association rule (39) (compare it with (1)), and it says
'left,' but it doesn't add "-to-right"; the rest seems to be
taken care of by the Well-formedness Condition. We can see
by inspection that (42) adds two association lines to (41),
that there are no ways of satisfying the Well-formedness
Condition by adding only one, and that there is no other way
of satisfying it by adding only two. Therefore (42) is
unambiguously derived.
The only cases where the Well-formedness Condition as it
stands does not suffice to give the correct output on the
basis of rule (39) is for cases like a tri-syllabic word
associated with a two-tone melody, as in (43), or the parallel
case of a bisyllabic word with a three-tone melody.
(43) # bi so ne #
H L H #
(43) could be corrected minimally either to (44) or to (45);
in fact, only (44) is correct.
228
(44) # bi so ne # (45) # bi so ne #I I V[ ."' I 1I 1 /I / I
# L H # # L H #
We could take this as evidence either that the language-parti-
cular rule is more complicated that the simple (39), or that
the Well-formedness Condition is more powerful. Under parallel
circumstances in Section 2 above, we took the second option
There we said that to derive (23) from (22) rather than (25),
the Well-formedness Condition would, in effect, tend not to
reassociate a starred segment.
(22) archiplago (23) archipelago (25) archipql go
H L H L H L
The situation is parallel in (43-45), but there is no
accent. We may simply say (and thereby render the accentual
modification (26) of the Well-formedness Condition unnecessary
in its specificity) that
(46) Given an ambiguity in ways to fulfill the
Well-formedness Condition, do not reassociate
an already associated segment.
Methodology, the attempt to make the Well-formedness
Condition stronger in exchange for simpler language-particular
rules must be preferable to the alternative course of making a
more complex language-particular rule. The first option may
ultimately prove wrong, but it is the only one which offers to
ededujls|JE WMMMIMIR&Il l! i
229
tell us something about the nature of tone languages in
general, and thus about language in general.
This stronger option--which says that "left-to-right"
spreading is not a rule-governed operation--predicts further,
given our present understanding of the Well-formedness Condi-
tion, that there could not be two free unassociated tones on
the same side of an associated -tone, as in (A)--or if this
arose, there would be any number -of ways to associate this, as
in (B) or (C). Only in the case where there were only three
syllables would the arbitrariness disappear, as in (43). Now,
in fact it seems to be correct that tone languages implying
the so called left-to-right spreading for melodies of three
3tones are restricted to three syllable words. It is also the
case that in the extended analysis of Hargguchi for numerous
4
Japanese dialects (1975) and Liberman for English (1975),
there are tone melodies of the form (T) T2 (T3 ) , but not of
the form (T T T T --that is, a situation where there would
be two free tones on the right side of T2 after the application
of rule (1).
230
*(A) T1 T2 T3
(B) V Ivvv
T T12 3(c) V V V y
C) V V V V V V
T T T
1 2 3
It is not difficult to imagine how these observations
might be made to follow in a principled way from languages
using major association rules of the type in (39) or (1)--
that is, major association rules without variables.
I shall not, however, pursue this here. Such a direction
would be extremely attractive if one could claim that (1) and
(39) were essentially the only choices a language had in the
way of major association rules. The next example, though,
-~ U I I
231
appear to revoke the possibility.
4. Ganda
The following paradigm comes from Ganda, or Luganda, a
Bantu language (the particular example was called to my atten-
tion by Frank Heny). It is one of a series of several
"tenses"; they have received a number of analyses in the
literature, always essentially in accentual terms, though
never, to my knowledge, within a reasonably constrained accen-
tual theory. We shall propose a non-accentual solution. (See (47).
This is clearly a rather complex series of tonal forms,
and one might expect that the rules mediating the underlying
and the surface forms would need to be rather complex. We
shall suggest that a simple solution exists, involving one
essentially morphological rule, and a melody association rule.
First, the relevant morphemes. On the syllabic tier,
they appear, in the following order:
(48) a te ta a gu VERB
e ba
rel Neg subj relat. far objectMain
marker cause mark clause past marker
clause negat'n marker
We assign the following tones to each morpheme:
(49) a te j / ba ta a gu VERB
e L ) (see
L L H H text) L
HH(H)
~* I d-ITT411114., ! d
232
(47) CV
(A)'Grind' /sa/
\ A(i) yasa
he ground
(ii) b'asa
they ground
(i)eyasa
he who ground
(iv) absa
they who ground
(B) lya 'eat'
(i) y!tlya
(ii) baaly~a
A
(iii) eyalya
CVCv
(A) gula 'buy'
(i) yaguIa
(ii) baagula
(ii)) eyag&la
(iv) abaagu a
Yagusa
he ground it
baagusa
they ground it
eyagusa
he who ground
it
N // t
abaagusa
they who
ground it
yagulya
/ / \ *N
baagulya
eyagulya
N /1 \ y
abaagulya
\ //A
yagugula
// / /' A
baagugula
eyagugula
N // / / A
abaagugula
teyasa
he didn't
grind
tebaasa
they didn't
grind
ataasa
he who didn't
grind
abataasa
they who
didn't grind
\\A
teyalya
tebaaly~a
ataalya
abataalya
%.N - /\
teyaguLa
Cebaagi la
ataagLa
\ / // \ N
abataagula
\ \ /A
teyagusa
he didn't
grind it
tebaag sA
they didn't
grind it
1% .1/ / A
ataagusa
he who didn't
grind it
\ / /"/ I/ A
abatasgusa
they who didn't
grind it
teyagulya
tebaagulya
ataagulya
\ // t \
abataagulya
teyagugula
tebaagugula
\ // / /A
ataagugula
t /// /// A
abataagugula
Al. ! I I  i dlI! p. .I. 'IIlb il
(n) laba 'see'
(i \/N \ / NN
(i) yalaba yagulaba
(ii) baalaba baagulaba
(iii) eyalaba eyagulaba
N /r \b\ \b//\ N
(iv) abaeLaba absagulaba
ovvcv
(A) kweeka 'hide'
(i) yakweeka yagukweeka
(ii) bakAweeka baagukweeka
(iii) eyakweeka eyagukweeka
(jv) abaakweeka abaagukweeka
(B) leeta 'bring'
(i) yaleeta yaguleeta
4X/ \ \ If\ /\ \ .(ii) baaleeta baaguleeta
(iii) eyaleeta eyaguleeta
(iv) abaaleeta hbaaguleeta
N\1\ / N
teyalaba
\ //\N
tebaalaba
\ /1/ \ NS
ataalaba
abataalaba
V /\\
teyakweeka
tebaakeeka
\Kr, \XN\
ataakweeka
abataakweekA
teyaleeta
\ / / \\
ataaleeta
abataaleeta,
ty\l/Kb\
teyagulaba
\ // \\ KN
tebaaguaba
ataagulauba
abataagulaba
\K / //\
teyagukweeka
tebaagukweeka
\// / /1re\
ataagukweeka
\ /' /I., . /1N
abataagulcweeka
teyaguleeta
tebaaguleeta
\ // \ 'A \\
ataaguleeta
\ / //N \ \\
abataaguleeta
233
234
First note that (49) is not written with association
lines; these are meant as lexical entries, before any associa-
tion rules have applied. Thus "te/L" describes two, as yet
unassociated, sets of segments.
The tone assigned to gu requires a further statement.
The difference between the "A" verbs and the "B" forms of (47)
are attributed, under this analysis, to gu's tones being
deleted when a B-verb follows (which will, note, be immediate-
ly following)-. Furthermore, in the other case, of an A-type
verb, gu has the tone HHH in relative clauses or with plural
subject; elsewhere it is HH in tone. This relative complex
specification appears to arise from pressure towards paradign
parallelism (observe the point at which the tone falls in
CVCV and CVVCV verbs A-type with gu). Rather than go through
all the examples, I shall rest content with some illustrative
cases.
(50) y a gu la
LU HL L
235
(50) is an underlying form; no associations are present
yet. If we associate from left to right, one-to-one--(50)
becomes (51); note that y is not syllabic, and so does not
associate with a tone, even though there is a tone in the
lexical entry for this morpheme.
(si) y a gu la
L HL
yaguLa
(52) works the same way.
(52) a ba ta a kweeka
IL H H H IL
(53) a ba ta a kweeka
L H H L
If there are syllables or tones left over after associating
one-to-one from the left, the Well-formedness Condition effects
their association, as in (53)-(54)e
(53) a ba ta a gu gula
L H H H HHH L
a ba ta a g9gula
L H HH
a ba ta a gu gula
I H H HAL
abataagugulaBY WFC
j a ll| t |Ein I 'K
236
This particular analysis is of some .interest because it
violates the intuition one might have had, upon observing the
initial data in (47). There we would observe that all the
forms bore the tonal ''envelope" L H L; all that varied among
the examples was where the tone rose to H, and where it fell.
Fixed on this, several analyses have been offered, but all are
quite complex, and do not bear out the value of the initial
intuition that a single melody LHL is involved.
Within the theory developed so far, such an accentual
approach would not be stateable, since both the H and the L
would be star-linked, presumably, and we have not needed to
have recourse so far to such apparatus. In sum, what may
appear to the linguist to be a single accentual melody, as in
(47), may turn out to be quite different.
This analysis further requires a way to state formally
the procedure: associate tonemes and vowels one-to-one start-
ing from the left. In Goldsmith (1975a), I proposed the
following algorithm; although the arguments given there I no
longer find very persuasive, it is the best formalization I am
aware of for this procedure. To formalize the "one-to-one"
mapping, we shall appropriate the notion of rule schema as set
forth in Vergnaud and Halle's work on variables in phonology.
We set up the phonological segments at the top of a
sequence of columns, and the tonological segments at the left
of a sequence of rows, like (D).
CV CV CvrCV, Cy 237
The assignment rule then takes the form of (E)
V
(wh ee'W'J means "is associated
( E ) V f _ - 1 w h e rw i t h " ) .
This is to be interpreted in what I take to be the natural
way, given the way Vergnaud and Halle's "real" variables work:
we produce an algorithm that checks off a certain number of the
boxes in the matrix (D), These boxes then represent pairs of
vowels and tonemes (column and row), and we apply the struc-
tural change of the rule to these pairs of toneme and phoneme,
which means in this case to hook them together.
In particular, the algorithm is to check off the box
(i,j) (think of the induces as coordinates as indicated in
(D)) where (i,j) is the minimal pair such that i > i' and
> J' for all other check-off boxes of the form (i',j').
This algorithm is reactivated successively until it cannot do
anything anymore (Goldsmith, 1975a).
Whether this is correct or not, however, does not force
the question considered earlier as to whether Etung or Tiv
utilize a rule like (39) or like the one quoted in the passage
immediately above with a variable.
h iI P, ' 'i , I iT : - : i.I -I JO-k .1 -
238
5. Tonga
In his article "Some Tonga Tone Rules," James McCawley
deals with some problems in the tonal system of Tonga, a
Bantu language. In most analyses of Tonga and related
languages--McCawley cites Meeussen and Stevick, for example--
the tonal characteristics of particular morphemes or syllables
has seemed elusive. One couldn't simply report that such and
such a morpheme was High-toned, or Low-toned--whatever inherent
tonal characteristics it had seemed to show up elsewhere,
frequently on the following syllable. Indeed, McCawley says,
"In these languages it is often far from obvious which member
of an underlying tone contrast should be designated 'underly-
ing high' and 'underlying low.' Many scholars... have avoided
such terms, preferring DETERMINANT and NEUTRAL, where deter-
minant is that underlying tone which (under at least some
conditions) causes deviation from the "least marked" tonal
contour (which is generally-level low pitch) " (p. 140).
Proceeding from the assumption that the Determinant/
Neutral contrast must correspond to either to a High/Low or
Low/High contrast, McCawley develops a system of tone rules
for Tonga, although he recognizes from the start that it is
nigh onto arbitrary whether we choose Determinants to be under-
lyingly High or underlyingly Low. Although historically the
determinant morphemes derive from High tone, McCawley opts in
favor of a synchronic analysis which analyzes Meeusen' s
"determinants" as underlyingly Low; this choice is made so
11H Ij I - i i '1
, I j I II P ia hi
239
that global rules can be avoided. At the end of the article,
McCawley notes another analysis might be possible which was
based essentially on accent or pitch-fall.
In an unpublished paper (1974), Michael Cohen has
developed an analysis of Tonga tone which follows this second,
accentual, approach,, using an accentual autosegmental frame-
work. Because his work is not generally accessible, and is
highly instructive regarding the functioning of a complex
accentual system, I shall review a part of his findings in
some detail. In some of the conclusions drawn I differ slightly
from Cohen, but I shall note explicitly where any parts of
the analysis below are mine and not Cohen's.
Cohen's first tack is to attribute the Determinant/Natural
distinction not to a tonal High/Low distinction but rather to
equate it with an Accented/Non-Accented distinction. Thus
Determinant morphemes will have a star on them. The neutral
*
tone melody for Tonga, furthermore is H L. Thus we get (55).
* *
(5) i + ma +ka ni + ma + ka ni
H L H L
i + mu + su n + mu + su ne
*H*
H L L
. I I- l : 4 14 [T INI!MIF M P '
240
i = noun prefix
ma, mu = class marker
kani = 'news, 'sune = 'ox'
Cohen',s own analysis of the major associa
similar to that described in Williams (1972) a
but I shall proceed using the principles devel
present chapter. Thus the first stage in the1
in (55) results from the application of rule (
formedness Condition unambiguously results in -
output in the right.
(1) v
:ion rule is
id Leben (1973),
)ped in the
:wo derivations
L); the Well-
:he correct
T
Noun stems that Cohen cites are mono- or bi-sy
are commonly as illustrated above in (55),. Mc(
gives trisyllable forms that work according to
pattern. A small, class of monosyllable nouns c
ced either as in (55), or with an alternate tor
which all the syllables are High. For this cli
say there is the optional melody variant, H. S
(56) + ma *+ 1i+ M
variant of
H
li 'money'; ma (concord)
lablic, and
awley (p. 150)
the same
an be pronoun-
al melody, in
ss we shall
ee (36) .
. + 1
H L
241
There is a further set of nouns whose tone follows neither
the pattern in (55)--accented, with the melody H L--nor that
in (56)--accented with the melody H. This class is pronounced
in isolation with all syllables Low-toned, and is analyzed by
Cohen as unaccented. Some process--and about this there is
here little detail to specify, unfortunately--will insert a
low-tone for a word totally devoid of tonemes. See (57).
(57) tonge 'ruins' i + ma + ton e
L
The accentless nature of these forms is illustrated by the
paradigm in (58) (a modification of Cohen's (4)). See
especially (58e,f). Cohen says:
Neglecting the second and only considering the
behavior of the first word, we see its behavior
is quite regular. High tone spreads over the
first word until the first accented syllable is
encountered. If none is encountered, high tone
spreads over the whole word.
I , iIA w k iW. -l I H4 Am! F T. 0 1-
(58) isolation form
*
(a) iku bojk
H L
boko 'arm'
(c) mu su ne
H L
ox
(e) mu ii o
L
'work'
2412
renttive
(b)i ku bo ko wa mu ka in
H L H L H L
'arm of woman'
(d) musune wa mu ne ne
H L H L HL
ox of old man'
7 *
rnu m wa mu asim 14
HI \ HA IH L H L
twork of girl't
(Note that the derived genitive forms are intermediate
forms and not final forms. I have noted the surface forms
with diacritic accent marks over the syllables; they will not
correspond exactly to the autosegmental notation because they
result from the application of further rules. We shall
p irl I I, r i, I 'I PH I;uIP, !]I r i: ! I 'I
I ni I i* ii I
return below to the tonal characteristic of wa, the posses-
sive marker.) Although the forms in (58) involve some
complications, Cohen's primary point--that the nouns that
display Low tone in isolation are accentless--comes through
clearly upon comparison of the tone in (58b, d, and f), or
comparison of (58e) and (58 f).
The same argument applies to compounds, Cohen argues.
The facts are much as in (58) except that there is no connec-
ting particle parallel to the wa in (58). Thus an accented
word like imukaintu has an H L contour in isolation; in a
compound, it is affected as in (59), and uttered ultimately
imukaintu mobutu.
()* *(59) i riu ka in tu + Ma ha tu
HL H L
'woman good'
Note that this spreading of the H-tone follows automatically
from the Well-formedness Condition if unassociated tones
have precedence in spreading over already-associated tones.
Thus (59) derives from (F) directly.
(F) i muk in tu +mob tu
H L L
I I l lith I
I have no independent justification, I might add, for the
nature of the boundary in (F). Accentless forms, like
iciindi (an isolation form) are all H in compound; see (60).
(60) ici in + ci lam fu 'time" + 'long in-duration'
In another connection, Cohen notes another class of
exceptional nouns. in Tonga. Rather than being exceptional by
having no accent, as in the preceding forms, this next class
is exceptional in that words in this class have two accents.
Corresponding to each accent is an instance of the tone melody
HL. Thus we get forms like (61).
(61) ba +acis ya ba +a cis ya ba +a cxsyT
* * Vt*
H L H L H L H L H L H L
The final form (61c) is later subject to a rule that
realized a single L-toned syllable surrounded by H-toned
N 5
syllables as a downstep; thus (61c) is uttered Va a cis ya.
Even if the prefix ba is not present, as it is in (61),
we get a Low tone on the first syllable of a cis ya; an
initial HL (falling) tone is simplified by rule to L tone.
(62) a cis yA
H L Initial Simplification
SL H L
In Cohen's analysis this deletion of H results from a
more general provision of the major association rule involved
here. The difference is not crucial here.
Pursuing this, Cohen cites the following "polysyllabic
nouns," not indicating any morphological division. Whether or
not there are any morphological boundaries internal to the
form is essentially irrelevant, however, to the point that
the conditions we have lately suggested for the Well-formed-
ness Condition (statement (46) above, or equivalently here
statement (26)) provide for the reassociation between (63a)
and (63b) automatically. (64) is not generated by the Well-
formedness Condition, correctly enough, and all this is
accomplished without recourse to talk about "left-to-right" or
"right-to-left" spreading.
(63a) i nk? ong ol e kwa 'butterfly' i nko ong k
HLH L H L
246
(64) *inko ong ol e kwa
Cohen's analysis, as I alluded to above, for these forms
depends on a specific right-to-left conception of tone-mapping,
rather than the interpretation I am giving.
Cohen proceeds in his analysis from possible noun tone
patterns to the tone patterns ,displayed by affixes. Perhaps
the most interesting class has the following three character-
istics:
"1. They are always high even if not followed by an
accented syllable.
2. If the next syllable would have been realized as
a high, then the following syllable becomes a
slipped high.
3. Otherwise, the following syllable becomes a low."
(p. 8)
Cohen continues, "segments with this behavior seem to
occur more frequently in Tonga than other exceptional seg-
ments." Cohen's own analysis for these forms is that these
are forms accented and individually specified for the melody HL.
I I i L'l1IiHJ,,I1 1:1 . I, d%4 Ili !I Him I M D i %I I
We shall propose, to the contrary, that the melody here
is H L, but that these forms are post-accenting forms: that,
under specifiable conditions, they place a 'star on the
following syllable prior to the melodic association rule.
Examples of this behavior are as in (65) = Cohen's (15),
slightly modified.
(65)
(i) nggmu kain tu 'she's the woman' i mu ka in tu 'woman'
HLH L H L
(ii) nge mi sa mu 'they are the trees' i mi sa mu 'trees'
H L H L L
(iii) nje zo k 'it is the snake' in zka snake'
H L 1 L
(iv) mbi ba ntu 'they are the pe ple' .i ha n 'people'
H L L
Post-accenting morphemes: (i) ngo; (ii) nge; (iii) nje; (iv) mbi.
r ,l l ill W 4 1 !N-I V q: I d I
248
Note that by calling these elements 'post-accenting,"
the theory makes predictions for structures of the form (66);
these should become by the Well-formedness Condition (67).
If instead of (66) we had the structure (68), the Well-formed-
ness Condition would make no prediction as to whether (67),
or, say (69).
(66) cv C Cv Cv Cv (67) Cv Cv Cv Cv Cv
H L H L H L L
(68) cv cvcv cv cv (69) CV CV CV Cv
H L H L H L H L
I do not have examples that meet these specifications, but
they can certainly be found, either in Tonga or in related
languages with similar systems, such as Bangu-Bangu, another
Bantu language, where the underlying tone melody is L H.
The notion "post-accenting" should be a basic one in
any adequate theory of accentuation. We would expect that
the post-accenting property of morphemes, for example, should
be maintained despite a shift in the underlying tone melody.
Thus if HL is the basic tone melody, a post-accenting morpheme
I : 11 *+-A * t I - III i P I j -I "I '.. A I 1.! 9 1;: J:. j ,Id l I 1 1
I - . I - Or Ilk llld 11 I AIlMI1 I
appears to bear an H tone and put a Low-tone on the following
syllable. The same facts would be stateable within a theory
which attributed an exceptional melody HL to that morpheme.
The post-accenting theory predicts, however, that should the
normal tone-melody shift to, say, LH, post-accenting morphemes
will suddenly no longer look like they did before, but will
rather put a Low-tone on the following syllable and a High
tone on the syllable two syllables to the right, as in .(69).
(69) CV + CV + CV
L H
Although the data there is not sufficient to provide a com-
plete analysis, McCawley's discussion of Bangu-Bangu ("Global
Rules and Bangu-Bangu Tone") suggest very strongly that this
is precisely what has occurred there. The questions requires,
of course, further investigation.
6. Further Non-Accentual Questions
In the past few sections, we have been trying to separate
out the functioning of the Well-formedness Condition--the
universal contribution of the autosegmental theory--from the
language particular major association rules. The attempt has
been to keep the language-particular rule as simple as possible
-- or rather to show that this, the best result a priori, is
consistent with a wide range of facts. We have worked with an
250
interpretation of the Well-formedness Condition (46) (repeated
here):
(46) Given an ambiguity in ways to fulfill the
Well-formedness Condition, do not reassociate
an already associated segment.
This formulation appears to supercede (26) above, which
gives precedence to non-accented segments in spreading due to
the Well-formedness Condition.
Condition (46) explains what was left unexplained in the
discussion of Guarani in Chapter One, Section 6. There we
derived, for example, (71) from (70).
(70):Do + ro + hamhu +# (71)0+ o+hathu + i~
The association of the nasal segment, we asumed, was listed
in the lexican, while the association of the starred Oral
melody 0 was rule-governed. The association effected by the
Well-formedness Condition in (71) appears to violate form (26)
of the Condition, which gives precedence in spreading to non-
starred segments. (26) would predict (72) rather than the
correct (71).
(72) Do +*+ u +
N
I r I -I A ,Ml-I HIIF I, A 1 L, III] 4v fld T11 IiI: 'I
I I '111 i I k .1 l,- . 2t,' tJ
However, if we adopt the more general condition (46) rather
than (26), we see that (71) derives not from (70), which has
putatively undergone the effects of a major-association rule
like (1), but rather has undergone no major association rule.
Thus the stage preceding (71) directly is not (70) but (73).
(73)Do+ rc +haihu + i
N 0
(modifie d)
The correct output results, then, from an even simpler
system than that suggested in Chapter One: there is, in fact,
no major association rule in Guarani.
One might ask how forms like (76b) of Chapter One, are
derived. That is, the question might arise how it is insured
that the 0 melody starts off to the right of the N autosegment
if there is no major association rule to govern where the
autosegment is.
This is a fair question, but it is not televent to the
question as to whether there is a melodic association rule.
Even in the analysis given in Chapter One some means was
necessary to place the Oral or Nasal Melody in the proper place
in the nasality tier, and it was not the melodic association
rule that determined the structure of that tier. Some steps
toward a deeper understanding of this question arise from
G. N. Clement's analysis of vowel harmony, where the simple
answer is given that the melody is entered in the lexicon as
such. We turn to this now.
Vowel Harmony
In its simplest form, vowel harmony is a phenomenon
wherein the vowels of a language can be divided into two
classes, and where within a certain domain--generally the
word--all of the vowels must be chosen from one of these har-
mony classes or the other (it may happen that a vowel, like
/i/, occurs in both classes--it is then called "neutral").
In addition to merely constraining the form of possible under-
lying forms to those containing vowels entirely from one
class or the other, vowel harmony also determines the surface
form that an affix will take. When an affix is joined to a
stem, its harmony class is generally determined by the harmony
class of the stem; thus, depending on the particular stem that
the affix is joint to, the affix will have one of two possible
phonetic forms.
For example, there is a "tenseness" harmony in Igbo.
Adopting the feature "Advanced Tongue Root" (ATR), this har-
mony system, we may say, insures that all vowels within a
certain domain of the Igbo word agree in the value they sustain
for the feature "ATR." There are four vowels marked +ATR: i,
A! q: I I
253
e, u, and o; and four marked -ATR: i, a, u, ando (these
correspond to the more traditional "tense" and "lax" groups,
respectively). Thus /madhu/, /h'chara/ (lax), and /ochie/,
/ogbu/ (tense) are possible (indeed, existent) words in Igbo,
whereas */madhu/, for example, is impossible.
We noted above, in Chapter Two, that the prefix /a/
that marks the II Incomplete form will appear as /e/ if the
verb stem to which it is attached begins with a tense (=+ATR)
vowel. In general, we can find affixes--prefixes and suf-
fixes--which illustrate this kind of vowel alternation. Thus,
/kwa/, a suffix, has the form/kwa/ after a lax stem, and the
form /kwe/ after a tense stem. Similarly, /ghi/ may also
appear as /ghi/, depending on whether the verb stem is marked
+ or -ATR (depending, as is said, on the "harmony class" of
the stem).
For example, with the tense stem /le/ we get the form
/hwe/, as in LEkw ya; (see (i) ), with the lax stem /di/, we
get dtkwa (see (ii)).
(W) Le + kwe ya
look suf it *Look at it2'
at fix
254
(ii) di + kwa inma 1Are you well?' ('How are you?')
you are suffix well
Viewed from the standpoint of vowel harmony, this alternation
of the suffix /kwa/ and /kwe/ is caused, we may say, to
preserve the harmony of the vowels in the domain governed by
the stem vowel. Alternatively, we might view the vowel in
this suffix as being an "archivowel," in that is has no
specification for the ATR feature underlyingly, and receives
this specification from the verb stem.
G. N. Clements has approached the phenomenon of vowel
harmony from an autosegmental point of view. He suggests,
following Ultan, five properties distinguishing vowel harmony
systems:
(i) a phonological feature, or set of features, is
determined by the root or stem (with certain qualifications we
will consider in a moment);
(ii) the affixes on this root then agree with the stem,
255
and the domain over which the agreement holds is unbounded,
that is, not limited to one or two affixes;
(iii) the harmony occurs both to the left and to the
right of the stem; it is, that is, bidirectional;
(iv) vowel harmony is not optional;
(v) the phenomenon of vowel harmony is "phonetically
natural," in that the feature which appears to spread over,
the harmony domain is phonetically and phonologically motiva-
ted, from among the class: back, high, round, advanced tongue
root, constricted pharynx.
Clements procedes with the assumption that the mettle
of a theory of vowel harmony is ultimately tested not by how
it deals with the non-exceptional case--for the non-exception-
al case can be treated, to some extent or other, by several
formal approaches, including the "features spreading" conven-
tions (Lightner, Chomsky and Halle), iterative rule applica-
tion, or general variables (Vergnaud and Halle); rather, the
test comes from the way exceptional cases are treated, and how
the exceptional cases are related to the regular forms.
Clements argues that autosegmental theory provides an
extremely natural way to subsume both the normal case of- vowel
harmony and the exception "opaque" vowel, the vowel which does
not "go along with" the harmony specification of the stem, and
which imposes its own harmony specification on those vowels
further away from the root than it itself, as in (74).
I ! I -j -4 4" i : 11 --1, 111 :1 ! , I 11: V, I I j q i I . I , A -I -. I
0( 0 P( . k256(74) (a) CV CV CV + CV + CV normal case
harmony class
K A opaque vowel(b) CV CV CV + CV + CV opaque case
harmony har
class oc cl(,
This opaqueness, Clements suggests, arises from a vowel
being underlying associated with a harmony autosegment.
Non-exceptional vowel harmony behavior is the surface mani-
festation of vowels underlyingly unassociated with any harmony
autosegment. Thus the two cases of (74) become (75) in auto-
segmental representation.
(75)C CV CV + CV + CV
+ ATR3
CV CV CV + CV + CV
[+ATRJ &ATh
The notion of "deviance" or "amount of exceptionality" of a
form then becomes easily quantifiable, for it is, as a first
approximation, the number of vowels underlyingly associated
with harmony autosegments.
We may remove the provision of "first approximation" if
we note that there are some "exceptions" which are, indeed,
predictable. Thus, Clements says, the Advanced Tongue Root
257
(ATR) harmony in Akan is consistently blocked by the vowel
/a/. "We assume, then," he writes, "that in the phonology
of Akan there is a statement to the effect that every lexical
occurrence of the low vowel is bound to an occurrence of the
feature -ATR." Thus the form /bisa/, with a +ATR first
syllable, causes prefixes to become +ATR, but the second
vowel, /a/, blocks this from spreading to the suffixes;
rather, the suffixes become -ATR, as illustrated in (76).
(76) 0 + b I s a + I
underlyinglY
+TR -ATM
t0 + b I s a + I
[obisaii
There may be, furthermore, suffixes in a system like that of
Igbo described above, whose vowel harmony specification is not
determined by the stem, but is, rather, specified underlyingly
-- once and for all--by the suffix. Such a suffix would not
alternate in form depending on context; such a suffix should
additionally block vowel harmony from extending past it, by
the Well-formedness Condition. Clements pointed out that this
is the kind of behavior we do in fact find in a suffix like
/si/ in Igbo:
258
(77) prefix stem suffix suffix suffix
E +vU + tE + sI + ghI
[+ATRAA
The capital letters stand for vowels unspecified for the
feature Advanced Tongue Root; the second level is an autoseg-
mental level consisting of the feature Advanced Tongue Root.
The Well-formedness Condition changes (77) to (78).
(78) fE +VU + tE+ sI*ghIK
4[iATIi -AThRk 41
This is the correct output, in fact: /evutesighi/.
We shall not consider in any further detail the autoseg-
mental treatment of vowel harmony; the reader is referred to
Clements' work, cited in the bibliography. Two points are
relevant to us, though: first, a clear account of certain
vowel harmony systems can be obtained through an autosegmental
treatments in which the autosegmental level corresponds to the
feature being harmonized for; and second, such an account is
able to account for exceptions of vowel harmony without
recourse to any notion of "rule feature" or "rule exception";
what makes "exceptional" segments that, within this framework,
is additional underlying specification.
259
7. Conclusions
In Chapter Two, we used the tonal system of Igbo to get
a clearer idea of how an autosegmental system operates. In
the present chapter, we have used autosegmental techniques
to gain new insights into traditional questions, beginning
with the nature of stress subordination, and proceeding to
other questions of accentuation and the nature of underlying
forms. By attempting to make the mapping of tone melodies
to syllable structures as simple as possible in any particular
language, we were led, step by step, to a view where the very
simplest system had precisely the properties of the well-known
phenomenon of well-known phenomenon of vowel harmony, in the
case where the autosegmental level was the level corresponding
to the features being harmonized for (advanced tongue root,
backness, and so on).
What we have is insidious autosegmentalism.
When we began this study, we seemed to be showing that
certain paradoxes for the standard segmental theory that tonal
systems manifested were resolvable by making a slight change
in our assumptions of the "geometrical shape" of underlying
forms. What we are finding now is that this modification
extends, at least potentially, to virtually all the familiar
features, not just the tonal ones: nasality, backness, ATR,
rounding, and height all can participate in vowel harmony sys-
tems. Why should this be? Why doesn't autosegmentalism stick
to the tonal domain? Why does it arise in the first place?
This is the subject of the final chapter, Chapter Four.
260
FOOTNOTES FOR CHAPTER THREE
1. The caveat comes from Mark Liberman's (1975) discussion
of the spreading of tonemes after the melodic association
rules of English. He argues that in spoken English, a
form like (i) (his3.1/1QO) does not become (ii); rather,
the shift from the L toneme on spec rises gradually to
the High toneme on cau.
(i) especially elaborate prec utions
L H
esp ally te ecutions
H L
Such a structure as (ii), however, would be correct,
he continues (p. 102) for changed English.
"Thus we may attribute the following property
to chants, but not to unchanted intonation
in English: Every syllable must be associated
with some tone."
The other conditions of the Well-formedness Condition,
as presented here, are explicitly adopted on p. 120:
3.2.3/la All tones must be associated with some
syllable.
lb Association lines may not cross.
Liberman is'clearly right in his analysis of (i) and
(ii)--the full Well-formedness Condition would be too
261
strong, and thus empirically wrong, in requiring (i)
over (ii).. This raises serious questions, but of detail,
I believe, rather than principle. As Liberman himself
says, in considering the relation of speech and chant,
"It is hardly necessary to point out tha the introduc-
tion of unchanged speech.. .into a primarily changed song-
structure means that the child is prepared to accept
speech as the (so to speak) aesthetic equivalent of the
chant. We will argue shortly that this aquivalence is in
some ways a very deep one" (p. 44). I take it that his
reference here (to "shortly") is to the citation above
from his p. 102. The relation between chant and speech
remains, then, a baffling one, if chant should be a mode
that imposes a three-part Well-formedness Condition, and
speech a mode that imposes only a two-part Condition,
leaving open whether syllables associate with a toneme.
Furthermore, languages like Igbo, which we considered in
detail in Chapter Two, obey the "chanting" principles--
why? I believe that an answer may emerge out of the
considerations in Chapter Four. For the moment, observe
that in all of the discussion so far--except in the
Introduction--not even passing reference has been made
to the fact that consonants intervening between the
vowels, or more generally, syllabics, do not associate
with tonemes. There is thus an important sense of
"opaqueness to the Well-formedness Condition" that has
I I T .. .1111) !1 , J , I,,. - T .17 ': I ..
262
been implicitly brought in from the beginning. The
provision of the Well-formedness Condition might read,
more exactly:
All segments which are tone-bearing (by some
criterion independent of the structure at
hand) must be associated with some toneme.
This notion of "tone-bearing" is, clearly, the operative
one, although we won't have much to say about it here.
In general, however, the string of segments on the
upper, syllabic tier, appears to partition into disjoint
domains, each containing exactly one tone-bearing segment;
the other segments are then unspecified for the feature
on the other autosegmental level. We shall have more to
say about unspecifiedness in Chapter Four.
2. Morris Halle points out that there may well be a
generalization here to be observed, if instead of listing
the melodies as in (i), we list them as in (ii).
i. H ii. HHH
HL HLL
LH LHH
L LLL
LHL LHL
LLH LLH
HHL HHL
HLH LHL
The generalization is then clear: all melodies expres-
sible as T1 T2 T3 appear in (ii). Whether this is a
linguistically real generalization, we can only conjecture
1 .1 1 I k, illil . J
I I-- -4 I:I!,II IIII I I,, - ! t v , t I:
263
at this point; the observation, however, is striking.
3. The exceptions I know of are borrowings.
4. We ignore here what Liberman calls "boundary tones,"
which are irrelevant since they are not free, unassocia-
ted tones.
5. The fact that a single noun may be associated with two
different copies of the H L melody distinguishes
Tonga from Indo-European and Japanese systems, where the
domain of a single tone melody is always at least as
large as a single word. Thus if a single word contains
more than one accent, some language-particular principle
must select which accent associates with the starred
toneme in the tone melody. Tonga, as we see, is not
restricted in this way; in fact, there are as many copies
of the tone melody H L inserted as there are accents.
For discussion of the Indo-European system, see Halle
(1975), Kiparsky (1973), for the Japanese, see Haraguchi
(1975)-.
264
CHAPTER FOUR
THE AUTOSEGMENTAL INDEX
We have managed, with only a momentary lapse or two,. to
overlook a certain basic question for most of the development
so far: just why are there autosegmental tiers? Not all
questions have answers., of course, and why-questions are
notoriously the worst in this regard. Nonetheless, we shall
pose this question, and answer it, after a fashion, with a
hypothesis.
This hypothesis harks back to the discussion at the
beginning of Chapter One that likened the act of speech to
an orchestral production, a harmonized score of several inde-
pendent musical "voices." These "voices" of the analogy
correspond to the distinctive features of phonology. At the
most superficial level, the speech signal is broken down into
a large number of independent linear parts--autosegmental
tiers--with at least as many of these tiers as there are
independent articulators. Thus there will minimally be such
a tier for the velum, for the laryngeal gesture corresponding
to pitch, and so forth. In short, the phonetic level looks
like the orchestration sketched in (3) at the beginning of
Chapter One, reproduced here as (1). As before, we shall call
the number of autosegmental tiers in a particular representa-
tion the autosegmental index of that representation. Thus the
autosegmental index of (2) is 2.
-1 . I. III I : : I 1'1 V I I I !I "f 11 1 111; 11 "
265
(1) Lips . . .Close up... .Open. ....... . ...... *..
Tongue .. .High and front... ... ..touch the palate
Velum . .Raise..............Lower...............
Larynx ..High Pitch.....Low Pitch.... ..........
(2) p i n
H L
Yet even if (1) is an adequate sketch of the phonetic
representation, it is unquestionably the wrong picture for the
more abstract--that is, psychologically real--level, where
there are atomic units like /p/, /i/, and so forth--in a word,
phonemes. The overwhelming weight of phonological work makes
it for us axiomatic that the phonological system of a language
is composed of a small number of atomic segments; each segment
is defined by several feature-specifications, but these seg-
ments act, in general, like discrete, atomic units. Thus an
/i/--a front, high syllabic segment--may be inserted epenthe-
tically; or an /n/, a coronal, non-syllablic nasal, may be
deleted. Such operations insert or delete whole units com-
prising more than one feature. The assumption that such
things may occur I shall call the Abstract Segment Hypothesis.
It is intimately connected, it can be seen, with the (Absolute)
Slicing Hypothesis of Chapter One.
In a sense, it is ironic to try to make this point here.
These atomic segments are what makes phonology possible
(especially a phonology which contains rules relating stages in
dIIN iN ,14 d l3 . 1
266
a derivation), and any alternative which denies it has a very
weighty burden of proof resting on it. But if one adopts
autosegmental theory, then although the Abstract Segment
Hypothesis is not thereby put in doubt, the question of why it
is true arises. Permit me to emphasize this: without an
autosegmental theory, the idea of segments (whether composed of
feature specifications or not) is an untestable, virtually
unquestionable part of phonology. With the introduction of
autosegmental tiers, one may view the abstract segment hypo-
thesis as a limiting case of autosegmental representation.
The Abstract Segment Hypothesis says in effect that there will
necessarily be a very small number of autosegmental tiers at
the most abstract level, and in the case where the autoseg-
mental index is 1, all the features will be on one tier. But
even if the autosegmental index is 2 or 3, the number of
features on some tier will be relatively large.
And we may ask why this is so. Why, that is, do feature-
specifications cluster together at the deepest level to form
segments incorporating many feature-specifications? This is
our question--how, and why.
Our hypothesis shall be that the peripheral devices of
the articulatory apparatus--and correspondingly, of the receiv-
ing end--are essentially parallel processing devices, as the
orchestral score image suggests.
A representation of the phonetic level, under such a con-
ception, is extremely rich in information, and correspondingly
267
2. Deautosegmentalization
The way in which the origin of autosegmental tiers would
best be explained would be if our account for the origin of
segmentation, in the more traditional sense, itself led to the
possibility of autosegmental levels. The best situation, that
is, would be to show that although autosegmental representation
was different from the traditional view, it was nonetheless a
natural consequence of the very process--in this case, of
language acquisition--necessary to arrive at the traditional
segmental view.
This, as I have indicated, is the approach we shall take.
The idea can be made most clear with a simple example.
Suppose that a child learning English hears the (perhaps
English) word "ha." The child, endowed with the abilities
necessary to plot distinctive feature values, hears a mess of
sound in the "ha." But this mess has some structure; first,
it begins with a distinctive breathiness arising from a
spread glottis, and then this breathing disappears. The rest
of the word "ha" is homogeneous: it consists of a vowel-type
quality we would identify as an /a/. Graphically, we have (4),
then:
(4) "ha"
breathinessbreathines , non-breathy,: I
(spread glottis) breathy r
NJouth position a
268
In a representation more suggestive of autosegmental structure,
we could write (4) as (5).
(5) [+Spr Gi] [-Spr Gi]
+Constr.
Phar.
.-High
L-Round
etc.
I take (4) and (5) to be essentially equivalent.
If the task of deautosegmentalization is the task of the
language-learner, then (5) must be converted into a non-
autosegmental representation. We can imagine at least three
ways in which the child could do this (there are, the reader
may note, others, less obvious):
(a) +Spr Gl [-Spr Glottis]
+CP
-High
-Round
The first segment here is a breathy vowel /a!; the second is
a segment specified only to bring the glottis back to its
normal position.
(b) [+Spread Glottis] Spr Glottis
+CP
-High
.Illlll I'l.1101,Iq o tA 11: 1. 1 ih 1. d
269
(b) is, of course, essentially the right solution; it
consists of a linear sequence of an /h/--a segment whose nature
is to spread the glottis but to have no influence on the
gestures made in the mouth--and then a vowel /a/, specified
as non-breathy.
(c) A third possibility is:
+Spread Glottis -Spread Glotti
+CPG+Cs
-High 
-High
-Round -Round
This analysis consists, we see, of a breathy /a/ followed by a
non-breathy (normal) /a/.
These three analyses for the underlying structure all
give rise to the same superficial form, as in (4). Why this
should be so is perhaps not obvious.
I shall assume the phonological system of an adult
speaker to be faithful enough to) reality to include, at the
superficial level, a division of the features that compose
underlying segments into the channels going to the separate
articulatory components or the feature detectors. In this
sense, even a competence grammar for English would contain a
representation essentially like (4) or, equivalently> (5).
Now if we are considering a competence grammar, it is a grammar
which is equally a grammar for production and a grammar for
comprehension. From this perspective, we may ask what the
representation on the phonetic level of (a) , (b)I and (c) would be.
ill 1 b l I 4 -!:.,I.Hm , , - I, 111:144.,
Ii I A - 4 II A 11Ph 1 41 ,i~ 'Iv ,. i A!
270
For purposes of exposition only, I shall assume that even
at the phonetic level, the features "constricted pharynx,"
"high," and "round" are on the same tier (as I already have in
(4) and (5)). The transition from representations of the form
(a), (b), and (c) to phonetic representation then consists
of autosegmentalizing the feature "Spread Glottis." (a) then
will become (6), which becomes (5) by the Well-formedness
Condition, as I have indicated.
(a) FSpread Glottis
Constr Pharynx
H<h
-R ound
(6) +Spread GlottiS
+CP
-High
-Round
\Spread Glottis
\-Spread Glot-tis]
(b) above will undergo the same change, mutatis mutandis,
producing (5) also. (c), on the other hand, will undergo the
autosegmentalization illustrated, in (7).
I 4 -. 1k III H 1: 1V li !dI !
271
(7a) +Spread Glottis -Spread Glottis
CP 4CP
-High -High
-Round -Round
(7b) [+Spread Glottis] LSpread Glottij
4CP 4CP
-High -High
-Round -Round
The question arises now whether the derived form in (7)
is to be a permissible representation at the phonetic level.
For if this level is to characterize both perception and
production (7b) is prima facie an implausible candidate for
the phonetic leve. There is no reason to say that two
occurrences of the comples +CP ] or /a/, are present.
-high
I-Round]
From the perception end of things, there is no reason to posit
other than one occurrence of it, and surely the same is true
of the production side from the phonetic point of view. In
sum, we are led back to the "obligatory contour principle" of
Leben's, discussed in Chapters One and Three, and dismissed as
false at the underlying level. We return to it, now, as a
realistic hypothesis for the phonetic level:
j I r1 4 jI, I -! I, 11,I' II I iN .. 
272
Obligatory Contour Level (Revised): At the phonetic
level, any continguous identical (auto)segments must
be collapsed into each other.
The effect of this principle, then, is to change (7b) into (5).
In so doing, we find that (a), (b) and (c) at the phonological
level all correspond to (5) at the phonetic level.
What more can we say about the process of deautosegmen-
talization that takes (5) into one of (a), (b) or (c)? Certain
constraints are clearly going to be necessary in aiding the
language-learner in this task. First of all, notice what
results if we "limit" the use of the revised Obligatory Con-
tour Principle. When we viewed it from the point of view of
acquisition, we see that for the learner to invoke the
Obligatory Contour Principle is for him or her to reduplicate
the autosegment, as in (7b), before merging the two levels
together to form one sequence of segments. If we postulate
this as a marked or psychologically more complex task than the
alternatives (yielding, here (a) or (b)), we derive, as a
consequence, a form of Leben's original Obligatory Contour
Principle. That is, the contour's being obligatory at the
phonetic level carries through to being the unmarked or expec-
ted situation at the deeper level. Thus it is no accident that
Leben's formulation had a lot going for it; it described the
linguistically normal situation. It was, however, due to a
condition on simple grammars to learn, not on (as Leben
suggested) knowable grammars.
273
Another consequence of this is that there will be, in
the unmarked case, segments unspecified for some features. In
(b) above, .for example, the segment /h/ is specified only for
the feature Spread Glottis, and unspecified for all other
features.
This conclusion is only a step (though a real one) past
the conclusion that we had reached already in Chapter One.
There it was shown that morphemes could be unspecified for
tonal features., This could happen if tone was autosegmental,
and the morpheme had only a syllabic (non-tonal) entry in the
lexicon. Since a morpheme could also be specified with a tonal
autosegment, this left three categories for a morpheme to fall
into with respect to the tonal features: marked +, marked -,
or simply unmarked. We are drawing a different conclusion
now, formally speaking, but it varies from the same conditions
of autosegmental representation. The process of de-autoseg-
mentalization, along with our hypothesis of the obligatory
contour principle application being a marked state, leads
directly to non-fully specified segments.
The process of deautosegmentalization, as a task for the
language-learner, then, consists of learning which sets of
feature-specifications on separate tiers may be merged together
to form an acceptable segment in that language. The child
learning English will eventually learn that there are no
breathy vowels, but this is a fact about English, not language
in general.
I I 4,
Should the child not find any consistent system by which
to merge some autosegmental tier with the other sets of
features, or should the separate tier of segments warrant
status as an entry in the lexicon, this leaves the grammar with
more than one linear sequence of segments at the underlying
level. And in such a way, autosegmental phonology, as we have
seen it in the first three chapters of this thesis, would
arise. The separate linear sequences of segments would then
each constitute .items that would have a place in the lexicon
of that language; and the phonology would develop melodic
association rules to link these two (or more) lexical entries
together, as we have seen in Chapter Three.
3. Conclusion
Throughout this thesis has run one tension: that between
the analysis of phonological phenomena in terms of distinctive
features, and seeing phonological representations as composed
of atomic, indivisible units. The motive force has been the
belief that both of these are correct, and therefore ultimately
reconcilable. But such occurrences as contour tones on short
vowels seemed to threaten these assumptions; and the auto-
segmental hypothesis arose as a way to save these two basic
principles and also save the appearances.
There are some conclusions to be drawn from the discussion
in the first three chapters. It is an interesting realization
that the formalism of generative phonology is insufficient,
4 I 14 1.1! I-] I l l 1,
275
and that a multi-linear geometry is needed to deal with what
have traditionally be called "suprasegmentals." Still, it is
more the tradition--intellectually speaking--than any meta-
physical consideration that has kept generative phonologists'
representations on a straight line.
The conclusions that may carry some real interest are
those that lead to an understanding of the ontogeny and
phylogeny of language; and I have suggested in this last
chapter how we may interpret autosegmental phenomena as very
particular evidence for a human capacity--indeed, drive--to
restructure sensations to fit a linear geometry of perception.
This condensation, which I called "deautosegmentalization,"
is motivated not out of some physical or psychological in-
ability, we may speculate, but rather in order to restructure
phonetics, on both the perception and the production side, to
fit in with the extraordinary human invention of language.
This process of deautosegmentalization permits the human
speaker to develop a lexicon built out of items that are
linear sequences of elements from small inventory. Such a
task is not difficult for the human, and so the endeavor of
learning and using the other components of the grammar may
proceed smoothly, each component--morphology, syntax, and
semantics--free to develop the kind of geometry uniquely ap-
propriate to it, unconstrained by the particular structure of
the vocal apparatus, the very organs through which these com-
ponents, and by them, thought, ultimately find expression.
276
BIBLIOGPAPHY
Arnott, D.W. 1964. "Downstep in the Tiv Verbal System". African
Language Studies 5, 34-51.
Bird, Charles. 1966. "Determination in Bambara", Journal of West
African Languae 3, 5-11.
Bloch, Bernard. 1948. "A Set of Postulates for Phonemic Analysis",
Language 24, 3-46.
Bloch, Bernard and George L. Trager. 1942. Outline of Linguistic
Analysis. Linguistic Society of America Special Publication.
Bolinger, Dwight. 1955. "Intersection of stress and intonation."
Word 11, 195-203.
Bolinger, Dwight. 1958. "A Theory of Pitch Accent in English", in
Bolinger (1965).
Bolinger, Dwight. 1965. Forms of Eng1ish; Accent, Morpheme,, Order.
Harvard University Press, Cambridge, Mass.
Bolinger, Dwight. 1972. "Accent is Predictable -- if you're a
Mind Reader", Language 48, 633-645.
Bolinger, D. and Louis Gerstman. 1957. "Disjuncture as a Cue to
Constructs", in Bolinger (1965).
Bresnan, Joan, 1972a. "On Sentence Stress and Syntactic Transforma-
tions", in Brame, ed., Contributions to Generative Phonology,
University of Texas Press.
Bresnan, Joan. 1972b. Theory of Complementation in English Syntax.
MIT PhD dissertation.
Bresnan, Joan. 1976. "Towards a Realistic Model of Transformational
Grammar", presented at 1976 MIT-AT&T session on Communication,
MIT.
Chomsky, Noam. 1965. Aspects of the Theor of Syntax. MIT Press,
Cambridge, Mass.
Chomsky, Noam. 1970. "Remarks on Nominalizations", reprinted in
Chomsky, Studies on Semantics in Generative Grammar, Mouton.
Chomsky, Noam and Morris Halle. 1968. The Sound Pattern of English.
Harper and Row, New York.
Clark, Mary. 1975. "Evidence from Igbo for a Theory of Tone."
Unpublished paper, University of Massachusetts at Amherst.
N II il l l i v I E -
277
Clements, G.N. 1976. "Vowel Harmony in Non-linear Generatie
Phonology: the Autosegmental Theory (I)". Unpublished
paper, Harvard University.
Cohen. , Michael. 1974. "Topics in Tonga Tone". Unpublished
paper, MIT.
Crystal, David. 1969. Prosodic Systems and Intonation in English.
Cambridge University Press.
Dubuisson, Colette and John Goldsmith. (in press) "A Propos de
l'Inversion du Sujet Clitique en Francais" in Montreal
Workin Peers in Linguistics vol.
Ednundson, T. and J.T.Bendor-Samuel, 1966. "Tone Pattern of Etung",
Journal of African Languages 5, 1-6.
Emonds, Joseph. 1970. Root and Structure-Preserving Transformations.
MIT PhD dissertation. (forthcoming from Acadeftic Press).
Goldsmith, John. 1974. "English as a Tone Language". MIT ditto.
To appear in Communication and Cognition 1977.
Goldsmith, John. 1975a. "An Autosegmental Typoloqy of Tone". In
Proceedijs from the Fifth North Eastern Linguistics
Society, ed. Ellen Kaisse, Harvard University.
Goldsmith, John. 1975b. "Tone Melodies and the Autosegment". In
Proceedings from the Sixth Conference on African Linguistics,
in Ohio State Working Papers in Linguistics 20, ea. R.Herbert.
Goldsmith, John. In press. "An Overview of Autosegmental Phonology".
inguistc Analysis Vol 2, No. 1, 1976.
Green, M.M. and G.E. Igwe. 1963. A Descriptive Griamar of Igbo.
Oxford University Press.
Halle, Morris. 1973. "Stress Rules in English: A New Version".
Linguistic Inquiry IV,4.
Halle, Morris. 1975. "On Russian Accentuation". Slavic and
East Eur2ean Journal 19, No. 1.
Haraguchi, Shosuke. 1975. The Tone Pattern of Jaanese: An Auto-
segmental Theory of Tonology. MIT PhD dissertation.
Harris, Zellig. 1941. Review of Trubetzkoy, Grundzge de Phonologie.
Lang 17, 345-349.
Harris, Zellig. 1944. "Simultaneous Components in Phonology". Language
20, 181-205.
Haugen, Einar. 1949. "Phoneme or Prosodeme". Language 25, no.3.
I' 1,1111iii-MIM, 41-1
278
Higgins, F.R. 1974. "On the Use of Idioms as Evidence for Movement:
A Cautionary Note". Presented at the Winter 1974 LSA Conference.
Kahn, Dan. Forthcoming. The Syllable in Phonology (tentative title).
MIT PhD thesis.
Kiparsky, Paul. 1973. "The Inflectional Accent in Indo-European".
Language 49, No. 4.
Leben, Willi -. 1973. Suprasegmental Phonology. Unpublished MIT
PhD dissertation,
Liberman, Mark. 1975. The Intonational System of Eng MIT PhD
dissertation.
Lovins, Julie. 1971a. "Nelodic Conspiracies in Lomongo Tonology".
PaLs._ from the Seventh Regional Meeting of the Chicago
Lin1u4stics Socie.
Lovins, Julie. 1971b. "The tonology of Lomongo Reduplication".
Studies in African Lgruistics 2, No.3, 257-270.
Lunt, Horace. 1973. "Remarks on Nasality: The Case of Guarani".
In Anderson and Kiparsky, eds., A Festschrift for Morris
Halle, 1973, Holt, Rinehart and Winston, New York.
May, Robert and John Goldsmith. 1975. "Tone Sandhi in Sanskrit".
Unpublished paper, MIT.
McCawley, James. 1970. "A Note on Tiv Conjugation". Studies in
African Linguistics, 1, 123-239.
McCawley, James. 1973. "Some Tonga Tone Rules", in Festschrift for
Moris Halle.
Mc Cawley, James. 1974. "Global Rules and Bangubangu Tone", in
C.Kisseberth, ed., Studies in Generative Phonology.
Palmer, F.R., ed. 1970. Prosodic jAnjlysis. Oxford University Press.
Rivas. Alberto. 1974. "Nasalization in Guarani" , Unpublished MIT
paper.
Rivas, Alberto. 1975. "Nasalization in Guarani", in E.Kaisse, ed.,
Paper Efrom the Fifth North Eastern Linguistics Society Meeting.
Selkirk, Elizabeth. 1972. The Phrase Phonolog of English and French.
MIT PhD dissertation.
Spa, Jaap. 1973. Traits et Tons en Enya, Phonologie Generative d'une
Lagi antoue. Mus7
lgusee Royale de L'Afridue Centrale, Tervuren,
Belgium.
- il1 I 4 I l I I 1 1.1
279
Trager, G.L. and H.L.Smith. 1951. An Outline of English Structure.
Washington, American Council of Learned Societies, Studies
in Linguistics Occasional Papers #3.
Troubetzkoy, N.S. 1939. Princies de Phonologie. Editions Klincksieck,
Paris.
Vanvik, Arne. 1961. Stress inPresent Day English. Norwegian Universities
Press. Bergen, Oslo.
Voorhoeve, J., A.E.Meussen, K.F. DeBlois. 1969. "New Proposals
for the Description of Tone Sequences in the Igbo Completive
Phrase". Journal of West African Languages 6, 79-84.
Williams, Edwin. 1971. "Underlying Tone in Margi and Igbo". Unpub-
lished MIT paper. To appear in Linguistic Inquiry, 1976.
Woo, Nancy. 1969. Prosody and Phonology. Indiana University
Linguistics Club publication; MIT PhD dissertation.
1 i I4iJ~
280
Biographical Note
The author was born in Forest Hills, New York, in 1951, and
attended Forest Hills High School until 1968. He received
an A.B. from Swarthmore College in 1972 for studying philosophy,
mathematics, and economics, mostly. He attended MIT from
1972 to 1976 where, except for this present dissertation, he
worked primarily on semantics and syntax. He shall be moving
shortly to Bloomington, Indiana, there to join the Linguistics
Department of Indiana University.