Phonology
Marc van Oostendorp
16th March 2020
Contents
Contents i
1 The subject matter of phonology 1
1.1 Where in the world is phonology? . . . . . . . . . . . . . . . . . . . 1
1.2 Phonological data and methodology . . . . . . . . . . . . . . . . . . 6
1.3 Phonological theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2 The atoms of language 19
2.1 Segments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Contrast and the feature . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 A universal feature set . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 An alternative: Element Theory . . . . . . . . . . . . . . . . . . . . . 32
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Autosegmental theory 41
3.1 Tone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Autosegmental representations outside of tone . . . . . . . . . . . 50
3.3 The skeleton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4 Feature Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4 Finding patterns 69
4.1 Underlying representations . . . . . . . . . . . . . . . . . . . . . . . 69
4.2 Discovering alternations . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 The phonetic source of patterns . . . . . . . . . . . . . . . . . . . . . 70
4.4 Synchrony and diachrony . . . . . . . . . . . . . . . . . . . . . . . . 72
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
i
ii Contents
5 Syllables 77
5.1 Evidence for syllable structure . . . . . . . . . . . . . . . . . . . . . . 77
5.2 The internal structure of the syllable . . . . . . . . . . . . . . . . . . 83
5.3 Sonority . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.4 The syllable structure of English . . . . . . . . . . . . . . . . . . . . 95
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6 Phonological computation 103
6.1 Computation and representation . . . . . . . . . . . . . . . . . . . . 103
6.2 Optimality Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3 A case study: Nasal assimilation . . . . . . . . . . . . . . . . . . . . 109
6.4 Beyond place harmony . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5 A theory of constraints . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.6 Implications of the theory of computation . . . . . . . . . . . . . . 119
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.8 Sources and further reading . . . . . . . . . . . . . . . . . . . . . . . 123
7 Stress 125
7.1 Languages with stress . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.2 Metrical feet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
7.3 Syllable quantity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.4 Lexical stress: faithfulness to feet . . . . . . . . . . . . . . . . . . . . 131
7.5 The moraic theory of syllable structure . . . . . . . . . . . . . . . . 133
7.6 Stress typology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.8 Sources and further reading . . . . . . . . . . . . . . . . . . . . . . . 142
8 Prosodic structure 143
8.1 The phonological tree . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.2 The Phonological Word . . . . . . . . . . . . . . . . . . . . . . . . . . 144
8.3 The Phonological Phrase . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.4 The Intonational Phrase . . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.5 Morphological evidence for the prosodic hierarchy . . . . . . . . 155
8.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
8.7 Sources and further reading . . . . . . . . . . . . . . . . . . . . . . . 166
9 Phonology and morphosyntax 167
9.1 The difference between word phonology and phrasal phonology 167
9.2 Two layers of phonology . . . . . . . . . . . . . . . . . . . . . . . . . 168
9.3 Lexical Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.4 The life cycle of phonological processes . . . . . . . . . . . . . . . . 173
9.5 Cyclicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
10 Applying phonology 175
10.1 Phonology and spelling . . . . . . . . . . . . . . . . . . . . . . . . . . 175
10.2 Phonology and speech recognition . . . . . . . . . . . . . . . . . . . 175
10.3 Sign Language Phonology . . . . . . . . . . . . . . . . . . . . . . . . 175
10.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Bibliography 177
Chapter 1
The subject matter of phonology
1.1 Where in the world is phonology?
When a character in a comic story gets very angry, he starts shouting something
which appears in a little cloud above his head and reads as ’ A !’
Those symbols satisfy their communicative needs rather well: every reader
understands immediately what kind of intention they are supposed to reveal.
We also can easily imagine some sound — a deep grunt or growl — which
would satisfy the same goal.
It would be possible to devise a communication system which consists entirely
of sounds every human being can make, but which are difficult to notate
precisely in an alphabet writing system. Human beings have a sufficiently
large repertoire of sounds: we can whistle, hiss, snort, clap our hands, and
probably do thousands of different things. Each of those sounds we could call
a morpheme, and then combine these morphemes into words and sentences
in the way in which we know modern human language does it.
Such a language would be rather convenient for a number of purposes.
The sounds of all the word would be easily distinguishable: the chances that
two words would sound as similar as pin and pen in English would be very
small. At the same time, the words would not be too difficult to make.
Yet, as far as we know, no human language works in this way. Rather, the
words in every language seem to have much more internal structures. For
instance, all languages seem to have a finite list of vowels and consonants,
and all words consist of combinations of those — the smallest number of such
different sounds is about 11, the largest about 150.
Obviously, when phonologists talk about vowels and consonants, they talk
about sounds, not about letters, which are seen as mere graphical representations
of those sounds. Phonologically speaking, English does not have a consonant
c. The letter of that shape (<c>) is sometimes pronounced as [k] and
sometimes as [s].
Notice that I put orthographic letters in angled brackets; phonetic symbols
are put in square brackets. This is common practice. Since we cannot (yet)
incorporate real sounds in a book such as this one, we need to do with symbols
representing them. For this I use the so-called IPA system, which is explained
in section 1.2. IPA is presently used by virtually every linguist in the world. (It
is very important to remember that phonology is about sounds and not about
1
2 1.1. Where in the world is phonology?
letters. Many students confuse these two in the beginning. In an alphabetic
system, letters are a way to write sounds, but it is the sounds we study. In the
history of mankind, writing is a fairly recent invention, and still a majority of
the languages spoken today has never been written, and for those which are
written there may still be a large number of speakers who are illiterate.)
The subject matter of phonology
The discipline which studies this internal structure of language is called phonology,
in which we can recognize the Greek words φονή which means ‘sound’,phonology
and λόγος, which means ‘study’. The study of the phonology of a given language
usually starts with making an inventory of all the basic elements. These
are for instance the consonants and vowels of the language, but there can also
be other elements, which may not always be transcribed in the orthographic
system of the language in question. For instance, the following two sentences
sound different in English:
(1) a. You have eaten already.
b. You have eaten already?
The statement in (1a) sounds different from the question in (1b) because the
two sentences are pronounced at a different pitch (questions typically end in a
higher pitch). Also these tonal differences count as part of the set of primitive
elements of the language. (As we will see in chapter 3, in many languages not
just sentences but even individual words can be distinguished by the pitch at
which they are pronounced.)
If we would restrict ourselves to just making inventories, phonology would
not be a very interesting discipline — it would be a very low level bookkeeping
activity. Making such inventories however is just a very first and basic
step. Fortunately, there are more challenging and interesting questions which
immediately rear their heads. In the first place, it turns out that the inventories
of sounds are not just lists of consonants, vowels, tones, and possibly
some other things. They have much more internal structure, and furthermore
such structures are often very similar from one language to the next. Once we
observe this, we have two research questions: what is this common structure
of sound inventories in languages of the world? And what would explain that
those commonalities exist? A lot of phonological debate is about these issues,
and we will see a lot of it here.
Another type of question arises because sounds are not stable. A word or
a part of a word might have a slightly different shape in one context than in
another. Take for instance, the English plural suffix. This has three different
phonological shapes:
(2) a. -[@z]: fishes, passes
b. -[s]: cats, parks, lamps
c. -[z]: lambs, pans, beds
These shapes are dependent on the phonological make-up of the stem. It ends
in a [s], [z] or [S], it has the shape [@z] (2a). If it ends in a so-called voiceless
1.1. Where in the world is phonology? 3
consonant like p, t, k, the suffix is the equally voiceless s (2b). Otherwise, it has
the structure [z] (2c). (More about this will be said in section ??).
Again, these kinds of alternations show many similarities between languages
of the world, so that again two questions arise. First, which kinds of
alternations are available in languages of the world? And secondly, what explains
the fact that these alternations exist and that they seem to come from a
rather small repertoire of possibilities? These questions again have caused a
lot of discussion within phonology and we will discuss these here as well.
Phonology and phonetics
Now there are many more things to be heard in a word or sentence. For instance,
(1a) would sound different if it is pronounced lovingly than when it
is pronounced in anger. Yet such differences are not studied in phonology.
So which sound differences are important and which are not? Phonology is
about sounds as part of language, and that is where we typically find the answer:
we only count those sound differences which are linguistically relevant,
for instance because they correspond to a difference in linguistic meaning.
The difference between questions and statements is linguistic, the difference
between love and hate is not.
Phonology is distinguished from several neighbouring disciplines — which
partly overlaps, so that most phonologists will usually have a working knowledge
of at least some of the other disciplines.
One of these is phonetics. The differences between phonology and phonet- phonetics
ics are rather subtle and complex — some people even suggest that there is no
distinction between them at all. But usually it is assumed that phoneticians
study the physical and physiological aspects of speech sounds: they use technology
to study such sounds as part of the natural world. Phonology on the
other hand studies sounds as part of the language as a system, and the cognitive
or possibly social structures which underly this system. In terms of a
classical philosophical dichotomy, phonetics is about the body and phonology
is about the mind.
The issue is obviously very complicated, if only because it is not easy to
disentangle mind and body. If I say the word pen, at some point my mind
has to translate this word into a set of instructions to my lips and tongue and
other muscles — what is the moment in which we make a transition from my
intention to the physiology, from the mind to the body? Fortunately, in the
everyday business of linguistic analysis the question we can take a pragmatic
attitude towards this: everybody uses the tools they have to study what they
can.
There are also differences between typical phonetic and typical phonological
data. Since phonetics deals with the messiness of the physical world, a
lot of its data are gradient (they involve values which can be written as real gradient
numbers, like 1.34253... or 65.696969...: no matter how precise we are, we can
always imagine a little bit more precision), whereas phonology, which deals
with clean systems is categorical. A consonant in language is a p or an b, but categorical
not 13.4% a p. When phonology deals with numbers at all, they are therefore
usually (discrete) natural numbers: 0,1,2,....
Another difference is that phonetics is not exclusively a linguistic discipline.
This means that on the one hand, phoneticians study certain topics which
4 1.1. Where in the world is phonology?
phonologists would not consider. The difference between loving and angry
speech mentioned above is a case in point. Generally, the speech signal contains
much more pieces of information than purely linguistic ones. From the
way somebody speaks, we can draw many conclusions about their age, emotional
state and gender. We can hear how tired or agitated the speaker is, and
according to some engineers one can even hear whether somebody speaks the
truth or not. (Those engineers are involved in building polygraphs.) Such
topics will be studied by phoneticians, and not by phonologists.
Morphology and syntax
On the other hand, phonetics is obviously further removed from core linguistic
disciplines, such as morphology and syntax. Morphology is the studymorphology
syntax of the way in which words are built out of other words — how books, bookish
and bookcase are all formed on the basis of book. Syntax is the study of how
words are combined into sentences.
It is said that one key property of human language is that it has double articulation
(also sometimes called duality of patterning): every language consistsdouble articulation
of an inventory of small elements without meaning, such as consonants and
vowels. These can be combined together to form words and affixes, which
are associated to a meaning. That is the first articulation or patterning. These
smallest meaningful units can then be combined into larger units, like more
complex words or sentences. That is the second articulation.
This description suggests that the first articulation is hierarchically below
the second one (it stops below the level of the morpheme), but that is not
correct. It seems better to view the two patterns in parallel. The sounds of
language are also sensitive to the higher levels of organization. In the phonological
alternations we mentioned above — of which the different shapes
of the English plural suffix were examples — this already became apparent:
phonology should be able to somehow ‘see’ the structure that is built by the
morphology. In other languages, it can also see the syntactic structure in the
same way. However, the interaction between phonology and morphology is
in many languages much clearer than that with syntax.
Every linguistic utterance thus has the two patterns at the same time: on
the one hand, it consists of meaningless sound symbols, but at the same time
— as it were in a different dimension, it organizes the meaningful unit in a
parallel way.
Just like with phonetics, the boundaries between phonology and morphology
are not always clear. One issue is allomorphy: morphemes can have aallomorphy
different shape depending on their context. Here is an example from Kalkatungu
(an extinct Pama-Nyungan language from Australia). The genitive in
this language is expressed by -ku if the stem ends in a consonant, and by ja if
it ends in a vowel:
(3) a. t”uat-ku ‘snake’, upun-ku ‘frog’, t”untal-ku ‘moon’
b. macumpa-ja ‘moon’, ntia-ja ‘snake’, kupu-ja ‘spider’
We have seen another example of allomorphy in (2). There, we presented the
different shapes of the English plural suffix as evidence for phonology: the [z]
sound of the suffix changed its shape based on the phonological environment.
1.1. Where in the world is phonology? 5
There does not seem to be a reason to do the same for the Kalkatungu case:
the shapes ku and ja are too different from each other to be related in a sensible
way, and there also is no clear phonological reason why one would be chosen
after a consonant and the other after a vowel.
For this reason, most linguists think that the Kalkatungu alternation is not
part of phonology, but of morphology (and/or the lexicon). But quite obviously,
the boundaries are not necessarily always clear: two allomorphs might
look somewhat similar and there might be a reason why they have this shape,
but it looks a little bit far-fetched. (Notice that the Kalkatungu alternation still
refers to a phonological property of the stem, viz. whether it ends in a vowel
or a consonant.) Further, some people argue that if we already need to set up
some technology in the morphology for dealing with Kalkatungu, we do not
want to use a separate mechanism to deal with the English case.
Again, most phonologists take a pragmatic approach to these questions.
We deal with those phenomena we can account for in our theories and leave
other data to related fields which might be better equipped to deal with them.
Where in the world is phonology
If phonologists study the structure of sound systems in human language, they
must obviously assume that there is such a system to be found. The question
then arises what is the reason that languages have structured sound systems
to begin with.
We can only answer this question if we have an idea about the ontology
of language: where in the world do we locate this phenomenon? There are
approximately three possible answers.
The first is that language is an abstract, entity, which we can study independently
of its speakers, much like the way many people see mathematics.
We can call this the Platonic view, after the Greek philosopher Plato. Many lin- Platonic view
guists have taken this view — it is for instance the basis of well-known metaphors
of language as a living organism or of Portuguese being a daughter
language of Latin. Under this view, the fact that language is logically structured
is inherent in the fact that it exists: for a Platonist abstract objects will
always have a structure that is worth studying.
The second point of view sees language as a cognitive object, as something
that is ultimately represented in the brain of an individual. We can call this
the psychological or cognitive view of language. It implies that we ultimately cognitive view
believe that the structure of language is (causally) related to the structure of
human cognition. The way a phonological inventory is organized can under
such a view be understood in light of memory.
This second view has been the dominant one in phonology (as well as in
theoretical syntax and morphology) for quite a long time. It was championed
by the famous generative linguist and philosopher Noam Chomsky (1928),
but also many linguists who take very different views on the structure of language,
on linguistic methodology, etc., implicitly or explicitly see language as
an essentialy cognitive object. Since this is the dominant view, it also plays an
important role in this book.
The third view sees language primarily as a social object, as a thing that
does not belong to one individual, but always to a group of people. We can
call this the sociological view of language. It implies that the structure of lan- sociological view
6 1.2. Phonological data and methodology
guage is caused by the way in which human interaction works. Although
this way of thinking has not been as dominant as the cognitive view, it has
always been there, and is rather important too. A lot of sociolinguistic work
obviously takes this point of view, but in recent years there has also been interest
in building computer models in which it is shown that models with a
small number of computer ‘agents’ communicating with random sounds in
the course of time start converging on sound systems which look like human
language.
I will not decide here which of these points of view is the correct one. Probably
they each are correct, since they are not mutually exclusive. Language is
somehow encoded in the human brain and it is used in human interaction.
Apart from this it may also be an abstract object with mathematical properties.
In that case we will have to figure out which aspects of phonology are
best explained from which basic principles. It is however very important to
recognize that there are these different points of view and that what an individual
analyst sees as a convincing argument may depend on his or her point
of view.
There is yet another way to see the place of phonology in the world. This is
by considering the distinction between synchronic and diachronic phonology —synchronic
diachronic a distinction which is due to the Swiss linguist Ferdinand de Saussure (1857-
1913). We can see language as a static object, which exists at a given point in
time (15th Century English, Swahili as it is nowadays spoken in Kenya): in
that case we are doing synchronic linguistics. We can also study the dynamics
of language as it is (continuously) changing; that is the subject matter of
diachronic linguistics.
Again, it does not make sense to say that either of these views of language
is the correct one. It is an undeniable fact that languages (at least if seen from a
sociological view) are subject to change. But the same is of course also true for
objects in the physical world, which are constantly moving around, decaying,
etc., which does not mean that we cannot study the structure of the human
hand as if it is stable — as a matter of fact something which is really constantly
moving cannot be studied at all.
Most of phonological theory nowadays is synchronic: explanations for patterns
are sought in structures which are treated as stable, not in change. This
is definitely true for the style of theorizing which is presented here. However,
there are also many phonologists who believe that many patterns can, and
even should get a diachronic explanation. Again, it does not make sense to
stipulate that one point of view is correct and the other one is wrong, but arguments
and data will be seen in a different light from different points of view.
Images of the brain when somebody pronounces a word are not necessarily
interesting for somebody interested in diachrony.
1.2 Phonological data and methodology
Since it is concerned with finding patterns, the study of phonology requires a
level of abstraction. This emphatically does not mean that the phonologist
does not deal with emperical data. Quite to the contrary, there are many
sources of information that a modern phonologist can use to discover the ab-
1.2. Phonological data and methodology 7
stract patterns. It is therefore important to understand these different types of
data, and to be able to use different scientific methodologies to discover them.
In this section, we discuss the five most important types of data, and data
collection methods, that are used in phonology. It is important to realize that
each of them is important and shows a piece of the puzzle; but that they also
all can have problems.
Introspection and fieldwork
Classically, phonologists have derived their data about a language from asking
native speakers. Sometimes they would be their own informant, when
they were working on their own language. But in many cases, they also
worked with an informant who was not necessarily a linguist to discover the
phonology of their language. If one works with one’s own judgements, this
is called introspection; if one works with other people’s judgements, this can introspection
be called fieldwork. The advantage of fieldwork is that the judgements are not fieldwork
influenced by the speaker’s knowledge of what the theory predicts; the advantage
of introspection is that of course nobody is as patient and as willing
to ponder new examples as long as the researcher.
One simple thing one can obviously do, is try to record as many words
of the language as one can get. One problem then is how to write out the
sounds of the language. First, we need some kind of writing system. The
orthographies of existing languages, say, English, do not suffice, because they
do not faithfully record every sound distinction of the language itself, let alone
of more foreign languages.
For this reason, linguists use the so-called International Phonetic Alphabet International Phonetic Alphabet
(IPA), a convention for writing all the sounds of all (known) languages in the IPA
world. The IPA therefore consists of a large number of symbols and diacritics
which together allow the researcher to go into as much detail as is useful for
linguistic analysis. It is very important for phonologists as well as for phoneticians
to be able to read and write transcriptions in IPA. Here is an example of
a sentence transcribed in IPA:
(4) DIs Iz @n Igzæmpl
"
@v @ sent@ns trænskraIbd In aIpi:eI
See the Further Reading section at the end of this chapter for more information.
Alternatively, one can of course decide not to write out any of the sounds
at all, but make audio recordings. This may definitely be preferable while doing
the fieldwork, if only because of course you will always miss something
while transcribing. However, it is common practice to present IPA transcriptions
in analyses and publications, rather than sound files.
Transcribing random words is obviously not the most ideal way of doing
research if we are trying to find patterns. Various methods have been developed
to find the patterns we are looking for most systematically. Many of
these will be studied in the next chapters.
Data we study is of course not always from our own fieldwork or introspection.
A lot of data have been collected in this way by other researchers
over the course of the past decades, and these play an important role in the literature
since then. This does not mean that you should take such data without
criticism, because it is always possible that mistakes have been made.
8 1.2. Phonological data and methodology
It should be noted that fieldwork is often done with representative adults,
but phonologists have also worked with specific groups, for instance children
who are acquiring their language, aphasic patients who have lost part of their
language, second language learners, etc.
Corpora and databases
Another means of discovering patterns is by studying large datasets, for instance
in the form of corpora — large, typically electronic, collections of naturalcorpora
language data. These can be for instance transcriptions of conversations or of
monologues, what is usually called spontaneous speech, but also of more struc-spontaneous speech
tured interviews based on a questionnaire.
Such corpora potentially give insight in different dimensions of natural
language. For instance, they can tell us something about how frequently
words, or syntactic constructions, or individual sounds are really used in
everyday language. Also, the way in which people really speak might be
different from the way in which they think they speak — the latter is the kind
of data we obtain with the methodology described in the previous section. Finally,
corpora are also more easily verifiable than judgements — a scholar can
put the data he has used online, so that other people can verify them, or even
use them for other purposes.
A special type of corpus worth mentioning is the sociolinguistically annotated
corpus, which provides information about the background of the speakers;
typically their age and gender but also other information about their position
in society (where they were born, their occupation, the way they identify
themselves, their lifestyle). These are important in particular if we take a sociological
point of view of language, as described above and consider a language
to be a property of a human social group.
We know that any human social group of some size consists of many subgroups:
men speak slightly differently than women, etc. Given this fact, many
sociolinguists propose that the language use of an individual speaker can only
be understood as a function of the language system of the various subgroups
she belongs to. Under such assumptions, studying a corpus without the relevant
corpus information does not make a lot of sense.
On the other hand, working with corpora also has several drawbacks. One
of them is that it is very difficult to build a corpus that is really balanced and,
above all, representative. When we record people, they will almost always be
influenced by that fact and speak differently than when they speak in their
ordinary voice.
A more important problem of using corpora is that certain patterns might
just not occur even though they are possible — they are so exceedingly rare.
But it might just be these very rare patterns which show an important piece of
the puzzle. As a matter of fact, it is sometimes argued that exactly these very
rare patterns give a good indication for what is ‘really’ going on. If people
can distinguish between ‘good’ and ‘bad’ patterns which are equally rare,
they cannot have made this judgement based purely on the data they have
encountered. There must be some other factor at work — for instance, innate
knowledge in the sense of Chomsky (1968, 1986, 1995). But also other
‘external’ sources of language patterns, for instance generalization capacities
might be justified by such evidence.
1.2. Phonological data and methodology 9
A final problem with many linguistic corpora which is not inherent to the
technology is that linguistic databases often consist of transcriptions only, and
those often even in an orthographic form. It is obviously sometimes difficult
to deduce from these how the data sound, let alone what they tell us about the
phonological organization of the language in question. There are, however,
also corpora which include phonological or phonetic transcriptions; and some
of them even have sound files. (We will discuss examples in later chapters.)
Another type of electronic datasource are databases; their difference with databases
corpora is that they have more structure. A corpus consists of (transcriptions)
of texts. Extra information about individual text items such as words might
be added, as well as information about the text as a whole. Databases consist
of more structured information, for instance about whole languages.
A well-known example of a phonological (or phonetic) database is the
UCLA Phonological Segment Inventory Database, which contains tables of
all the vowels and consonants in 451 languages — it is usually estimated that
at present there are about 6,000 languages spoken in the world. Using this
database, we can investigate claims such as ‘all languages have a consonant
[t]’.
An advantage with databases is that they are relatively easy to search for
such claims. Furthermore, if we have a lot of data, we can apply statistical
techniques which can filter out individual errors. Because the problematic
aspect of databases is that they will often be eventually based on the fieldwork
methodology described in the previous subsection. This means that there may
be many errors, differences in interpretation of the data, etc.
Artificial evidence
Sometimes the kind of data we find in the world of everyday speech is not
enough. In such cases, linguists may also use artificial data, which has been
consciously created for some reason.
There are many types of such artificial data. Some of these have been created
by non-linguists. One eample of this is poetry; poets traditionally play
with the sound structure of language by using rhyme, alliteration and other
means. Especially in the case of historical data such sound patterns are often
the only indication about the phonology we have. For instance, from the fact
that two words are put in rhyming position, we may conclude that they ended
in the same or similar sounds, even though they were written differently. Similarly,
what we know about the difference between short and long vowels in
Latin is primarily derived from the fact that poets used regular alternations of
long and short syllables. The difference between a short and a long [i] was not
itself written down.
A next step on the scale of artificiality is that the researcher invents her own
data, in order to test her theories. These can be nonsense words, i.e. word-like nonsense words
sequences of consonants and/or vowels which do not occur in the lexicon. We
can then test the differences between such sequences.
One of the best known examples of this methodology is due to Morris
Halle (1923), one of the most influential phonologists ever. He observed that
there is a three-way difference between brick, blick and bnick. The first of these
obviously is a regular English word, but the other two are not. Yet English
speakers feel a difference: blick could be an English word which you just hap-
10 1.2. Phonological data and methodology
pen to not know, whereas bnick cannot. Halle therefore called blick a possible
word, and bnick an impossible word. Phonology is then not about the actualpossible word
words (whether or not a sound sequence gets assigned a meaning is seen as a
random fact) but about the set of possible words.
One can go one step further and not just invent words but whole patterns.
For instance you could wonder whether the pattern in (2) could be reversed
in some language — whether it would be possible to have a dialect of English
with plurals like the following:
(5) a. -[z]: fishz, passz
b. -[@z]: cates, parkes, lampes
c. -[s]: lambs, pans, beds
We can test this by trying to teach people this dialect. If we have found a
(synchronic) explanation for the ‘real’ pattern, it might predict that the pattern
in (5) is unlearnable, or will not survive in a community which tries to adopt
it. (The latter is of course a little bit more difficult to test experimentally.)
Experimental evidence
Another step we can take is go beyond the impressionistic recording of the
data, is to go to a laboratory and get experimental results. The advantage
of this is obviously that we can make very precise recordings, test our hypotheses
under highly controlled conditions, which might also be ‘unnatural’
but important to see how language behaves under such circumstances.
There are many different things one can do in the laboratory. I will divide
these into two types of experimentation: phonetic and psycholinguistic. The
boundaries between these two are not always clear, except that scholars tend
to identify themselves as ‘psycholinguists’ or ‘phoneticians’.
Interest in laboratory methods for phonology has grown considerably over
the past twenty years. There are special conferences and a journal on ‘Laboratory
Phonology’, but also in other conferences and in other journals there is
a rising interest in seeing phonological theories be supported or falsified by
well-designed experiments.
Phonetic measurement
Over the course of the past 100 years, phoneticians have developed a large
toolbox of instruments and techniques to study the way in which the human
body produces and perceives speech sounds as well as the way in which such
sounds are transmitted from one person to the next. Over the past few decades,
several (free) software packages have been developed which can be installed
on any (laptop) computer so that it is very easy to install a private
phonetic laboratory.
There are several advantages to phonetic data. One is that our instruments
can measure fine-grained distinctions which are not always perceivable for the
human ear. Or perhaps we should say: which we cannot make conscious. (If
people really cannot hear a difference at all, one can wonder how the distinction
can play a role in human communication.) If such patterns are system-
1.2. Phonological data and methodology 11
atic, they might be at least as important as those which we can obtain by other
means.
Another potential advantage is that phonetic measurements are performed
with computers and therefore less dependent on human interpretation. If you
study somebody’s speech by just listening to it, you might be tempted, even
subconsciously, to hear things which are not really there, just because your
theory makes you expect them to be there. Related to this is the fact that
phonetic measurements are typically more easily replicated by other researchers:
if you carefully describe the way you have performed your experiments,
another researcher will be able to do it in the same way, and arrive at the same
results.
It should be noted that these advantages do not mean that all results which
are obtained by other methods are therefore worthless or unscientific. Many
patterns in human language are quite clear and obvious to any native speaker.
The fact that the plural suffix is -z in beds but -s in cats can be easily observed
without any technological means by any native speaker of English, nor will
there be a lot of disagreement about this. For the latter reason, such data are
intersubjective: the subjective intuitions about them agree to a large extent. To intersubjective
many phonologists it would seem a waste of time and other means to try to
establish such truths also ‘objectively’, using computers.
This is then one disadvantage of phonetic measurements: they are timeconsuming,
you need to put informants into sometimes uncomfortable circumstances,
and you need special equipment (even if this is only a recorder
and a laptop), and it is not always clear that these costs are worth the result.
Furthermore, it is not always even possible to acquire phonetic data, e.g.
when we want to study the phonology of an exctinct language.
Another disadvantage is that phonetic measurements are necessarily ‘superficial’:
they can only measure things which are present outside the human
mind or human society: acoustic signals, movements of the body. Even if we
also consider brain scanning techniques — which are currently too expensive
for most linguistic researchers, but will probably be used more and more in
the not too distant future — we can see pictures of which areas of the brain
are active at some point, but not necessarily about what this means for the
human mind. If we want to know such things we have to apply techniques
from psychology, as I will discuss in the next section.
Psycholinguistic experimenting
Psycholinguistics is the field at the intersection of psychology and linguistics:
it studies the way in which humans process, produce and acquire natural language.
The results of this type of research is of course particularly relevant
if one takes the psychological point of view and sees language as primarily
something which belongs to the human mind.
A typical psycholinguistic experiment has a number of speakers of a certain
language perform a relatively easy language-related task. For instance,
they have to listen to a number of words and press a button when they recognize
that word as belonging to their language. Their performance is then
measured in various ways, e.g.: how long does it take for them to press the
button? How many mistakes do they make? And how are these factors influenced
by others, e.g. the fact that they have just heard a word which sounds
12 1.2. Phonological data and methodology
very similar in some way (starts with the same sound, has the same number
of syllables).
In this way, the psycholinguist hopes to find out how language is represented
in the brain. In our example, which dimensions of sound similarity
count as relevant for finding a word in the lexicon we as speakers all have in
our head. In turn, such knowledge about the internal structure of the lexicon
might teach the phonologist something about how sounds are organized in
language and, inversely, phonologists’ insights should ideally guide the psycholinguists’
research.
A special place has always been occupied also by the study of (first) language
acquisition. The language learning child to some extent has to face
the same task as the linguist: she has to figure out what the system of her
language is. There is one important difference, which is that the child apparently
somehow knows how to go about this task and accomplishes it within
a few years, while linguists continue to puzzle over the details. By following
children closely, we can try to learn from the way in which they apparently
acquire all these data.
But there are several other reasons why acquisition is important. One of
these is that the fact that every generation has to learn the language of their
parents is probably an important factor in language change. Small things can
and do go ‘wrong’ in that children sometimes construct a slightly different
language based on the input of their parents. If we are thus interested in the
details of diachronic phonology, it is crucial to understand how such acquisition
works. (Since language change may also be caused by adults having to
learn a foreign language, studying second language acquisition can be relevant
for the same reason.)
Formal evidence
A final type of evidence for a specific theory comes from properties of the
theory itself. Scientific theories are generally supposed to be more succesful if
they are elegant and restrictive. Since these are properties of scientific theories
in general, and not specific to phonology, we will not discuss them in detail
here.
Restrictiveness of a theory refers to the number of things which are im-Restrictiveness
possible according to the theory. The optimal theory is one which is applicable
to every real object in the world, but not to anything else. The best theory
of phonology would for instance describe and explain exactly all the sound
sytems of languages in the world, but not arbitrary collections of sounds
which can never be part of a real language. It should not just explain why
the English plural system in (2) exists, but also why the fake sytem in (5) does
not.
The arguments for formal types of evidence tend to be philosophical. An
important argument why theories should be restrictive is that in this way they
are more falsifiable. A theory which can account for everything, cannot account
for anything at all.
A second kind of formal criterion is elegance. If we have two theories of
one and the same phenomenon, we prefer the one which is more elegant.
The problem with this obviously is that this is an aesthetic criterion which
thus may be subject to personal preferences. However, there are several well-
1.3. Phonological theory 13
known and widely accepted principles of elegance. The most famous among
these no doubt is Occam’s razor, named after the fourteenth Century English Occam’s razor
scholar William of Ockham. The principle says that a theory should not contain
any unnecessary assumptions (such assumptions should be thrown out).
That in turn implies one should always go for the simplest theory, the one
with the smallest number of assumptions. It is usually very difficult to really
compare theories, because it is difficult to count assumptions, but everybody
would agree that a theory which would require a desciption of 5,000 lines is
less simple than one which can be described in 5 lines.
This is true in particular if the descriptions are written in the same language,
which preferably should be highly formalized, like a mathematical
notation, or a computer language. It has become more and more popular over
the past decades to build computer models which mimic the way language
works according to a certain theory. The goal of such computer implementations
is not only to show that a theory is indeed really elegant, but also that it
works in the first place. If you have written a computer program which does
things exactly as you say it does and of which the outcomes are furthermore
as they are in the real world, you know that the program contains the whole
theory (there are no hidden assumptions) and furthermore that it works as
required.
So which type of evidence is the best?
We have now given an overview of many types of data that are available to
the phonologist. The overview may not be comprehensive, but most data that
a phonologist deals with will be of one of the types described here.
There is an understandable human tendency to reduce such an embarassment
of riches to something more tractable and declare that only some of these
datatypes make sense, that the others are not acquired by the right methodology
and that the rest can or even should be ignored.
This is not the approach we take in this textbook. Even if you believe that
certain types of data are inherently more useful than others, all of them are
used in phonological argumentation, so a student of phonology should be able
to deal with them and evaluate them. Furthermore, I believe that evidence for
the kinds of patterns we study in phonology can be found in all of these data,
although we should always be careful: each of these type of data has their
problems and might be ‘polluted’ by other factors.
1.3 Phonological theory
We have thus seen in the preceding section that phonologists deal with a large
variety of data, trying to find patterns in them. These patterns are then described
in a theory. It is the ultimate goal of linguistics to describe what is a
possible human language and what is not. Phonology shares this goal as far
as the organization of sound patterns goes. There is a second goal, which is
to fully describe the sound patterns which we find in the existing individual
languages of the world; some phonologists work for instance exclusively on
the phonology of French.
These two activities mutually feed each other. One cannot pretend to study
what languages have in common, or what makes them different, without hav-
14 1.3. Phonological theory
ing a detailed knowledge of individual languages. On the other hand, in
studying the phonology of such a language, it is important to know which
aspects of the language are familiar also from other languages and which are
unique for the language in question.
An important aspect of any serious scholarly discipline is that it is cumulative:
we always try to build on the work of other scholars. There are manycumulative
questions, there are many mysteries. Other people will have (had) something
to say which sheds light on these issues. It is very inefficient to start all over
again every time we study a subject matter. This is also how it is with phono-
logy.
A very brief historicaly sketch
The discipline started somewhere at the beginning of the twentieth century.
A reasonable starting point is the Course on General Linguistics by the Swiss
linguist Ferdinand de Saussure (1857-1913). In this book, Saussure introduced
many concepts which turned out to be foundational for modern linguistics.
The distinction between synchronic and diachronic language study, for instance,
is due to him. But one of his key ideas was that languages can be
studied as coherent systems: that for instance the sounds of a language exist
in patterns. For the study of sounds, this implied the introduction of a
distinction between phonetics and phonology, where the latter was uniquely
occupied with the (linguistically relevant) patterns.
In terms of the discussion above, Saussure had a sociological view on these
patterns: he believed that ‘language’ only exists as a property of a community.
The speech of each individual only gives an imperfect reflection of the more
perfect abstract object which everybody shares. This was also the view of
structuralism, the most influential linguistic paradigm both in the United Statesstructuralism
and in Europe until the 1950s. Although there was also some interest in syntax
and morphology, the most succesful branch of structuralist linguistics was
phonology. Important scholars like Leonard Bloomfield (1887-1949), Nikolay
Trubetzkoy (1890-1936) and Roman Jakobson (1896-1982) produced very influential
work laying the theoretical foundations of the field. In particular,
Bloomfield’s book Language and Trubetzkoy’s Foundations of Phonology are still
important references for any (advanced) student who wants to understand
what phonology is about.
The advent of generative linguistics (of which the main phonological eventgenerative linguistics
was the appearance of The Sound Pattern of English in 1968 by Morris Halle
and Noam Chomsky) brought about several changes. The most important
one was a change in orientation: generative phonology puts the reality of
phonological patterns unequivocally in the human mind. At the same time,
Halle and Chomsky brought several technical innovations and improvements
to the theory, although there is also a lot of continuity (and cumulativity) in
technical aspects of the theory.
In the decades following the publication of the Sound Pattern, the theory
has developed in many ways, basically to the point of having become unrecognizable.
If you have studied this book, for instance, you will find it much
easier to follow contemporary literature than the work of the 1970s. A striking
aspect of the present phonological community is that it is very pluriform:
in a given phonological conference you will find many different approaches.
1.4. Exercises 15
Although most phonologists probably still take a psychological view of phonology,
the sociological view has also gained ground. Yet although there are
a lot of debates about many aspects, as is the case in any healthy discipline,
there is also a large common ground — a vocabulary in which scholars can
communicate about their ideas. My aim in this book is to bring you up to date
with that vocabulary.
Structure of this book
In this book, we will start with the smallest constituting elements of language.
As we will see in Chapter 2, these are not consonants and vowels, but even
smaller units which can be combined in certain ways to form units that are
similar to such sound units.
From there, we will start looking at larger and larger units organizing
sound in human language. Chapter 3 discusses autosegmental theory, which
explains how these ‘atoms of language’ can be combined into larger wholes,
and Chapter 4 looks into the way in which the primitives and their organization
are reflected in the phonetics.
Chapter 5 then explains the evidence for assuming that consonants and
vowels (together called segments) do not only have an internal structure, but segments
are in turn also organized into larger units, viz. syllables, in presumably all
languages of the world. If we know the segments of a word, we also know
Chapter 7 goes one step further and discusses how syllables are grouped
into metrical feet, which are used among other things for word stress. Chapter
8 then shows that there are even higher levels of phonological organization,
leading all the way up to constituents that roughly correspond to the sentence
and even the whole utterance.
The last two chapters discuss the way in which phonology relates to other
domains. Chapter 9 deals with the relation with morphology, and Chapter 10
discusses several applications: in orthography, in speech recognition and in
our understanding of Sign Language.
Every Chapter contains a section with 20 exercises. (Only this first chapter,
which is very general, has only 10.) It is very important to try to solve those
exercises; there is only one way of learning to do phonology: by doing it. Some
of the exercises can be answered by just using the material which is presented
in the book, but for others you will need to collect data, online or in other
ways. I will always explain where you can find such data. More information
can be found in the website which accompanies this book [UNAVAILABLE].
You can also find the answers to the exercises there, as well as a more extended
(and updated) version of the reference sections which are at the end of every
Chapter, and which you can use if you want to know more about a certain
topic.
1.4 Exercises
1. Would the study of the following topics belong to the field of phonology?
Discuss. If you think a certain topic is not studied by phonologists,
which other field does it belong to?
16 1.4. Exercises
a) The Hawaiian language has eight consonants: /p, k, P, h, m, n, l,
v/.
b) In French, some adjectives end in a consonant in the feminine, but
not in the masculine. (‘Small’ is [ptit] in the feminine, and [pti] in
the masculine; ‘good’ is [bon] feminine and [b˜o] masculine).
c) In some cultures, homosexual men speak differently than heterosexual
men. It can be shown in experiments that people are sensitive
to these differences and can tell above chance level what the
sexual orientation of the speaker is.
d) The English [b] sound is much more acoustically similar to the
French [p] than to the French [b].
e) There is no language in which every prime numbered syllable starts
with a p; furthermore, people cannot learn such a ‘language’.
f) Leaving out a final d or t, as happens in some dialects of English (I
kep) is considered incorrect.
g) In some dialects of English, the /r/-sound is not pronounced after
some vowels (like in car, mother, more) except if the following word
starts with a consonant.
2. For each of the topics mentioned in the previous exercise, explain which
kinds of methodology could be applied to shed light on the issues in-
volved.
3. For each of the following strings, decide whether they are existing words
of English, possible words, or impossible words: blobber, brankal, rooytkd,
trapeal, blistras, topiq. To what extent is your answer determined by
whether you take a cognitive or a sociological view of language?
4. Explain the difference between synchrony and diachrony in language.
Why does holding a Platonic or a cognitive view of language usually
imply more interest in synchronic data?
5. There seems to be an intimate relation between having a Platonic view
of language and being interested in formal types of evidence. Explain.
6. Duality of patterning is supposed to be a defining property of human language.
Explain why the following systems lack this property:
a) Programming languages like Java, Python, C++.
b) Animal communication systems, such as primate calls.
c) Western classical music.
7. Suppose we want to study the phonology of Sanskrit, an Indo-European
language that has been extinct for a long time, but of which we have a lot
of written record, like poems and prose. What kinds of methods could
we use to study this topic?
8. For each of the major types of evidence in section 1.2, give an example
of how they could be applied also to morphology or syntax.
9. ‘All phonological research should be based on data that are acquired
through phonetic experimentation.’ Comment.
10. What are the advantages of building a computer model of a theory? Can
you also think of disadvantages?
1.4. Exercises 17
Sources and further reading
Section 1.1. There are currently several other textbooks on phonology in the
market; implicitly, they all take a cognitive view on the location of phonology.
The most well-known are Gussenhoven and Jacobs (2005); Odden (2005);
Hayes (2009). They each have their merits, and in case you somehow struggle
with a topic, it can always help to see how somebody else explains it. For
more extensive information you can also refer to recent handbooks such as
van Oostendorp et al. (2011); Goldsmith (pear); de Lacy (2007,?).
A general textbook about linguistics is Freidin (2012). Some good syntactic
textbooks are Radford et al. (2009); Adger (2003), and a good book on morphology
is Lieber (2009). A nice linguistic introduction into phonetics, finally, is
Knight (2012).
The Kalkatungu data in this section are from Blake (1969).
Section 1.2. There are unfortunately no general introductions into phonological
methodology, describing all the different types of data and data collection,
but in the chapters that follow we will see individual cases. A recent
book on fieldwork are Sakel and Everett (2012).
The International Phonetic Alphabet (IPA) is maintained by the International
Phonetic Association which (among other things) has a chart of all the
IPA symbols on its website. The IPA symbols are part of the standard Unicode
character set, which implies they can be used on any reasonably modern computer
system; how to do that is explained on the website of SIL International.
A phonetic package that is often used is Praat, developed at the University
of Amsterdam in The Netherlands by Paul Boersma and David Weenink.
Praat is free of charge and available for all major platforms; it is also regularly
updated. Next to a popular module for making acoustic measurements, it
also contains modules for applying statistic calculations, modelling language
acquisition and many other tasks which a computer might want to perform
for a phonetician and/or a phonologist.
Section 1.3. Anderson (1985a) gives an authoritative overview of the history
of phonology until approximately 1980. Full references to the books mentioned
in the main text are Trubetzkoy (1939) and Chomsky and Halle (1968).
Chapter 2
The atoms of language
2.1 Segments
An important activity of linguists is the pursuit of linguistic universals — prop- linguistic universals
erties which all languages have in common. The reason for this is quite obvious:
if we know what all languages share, we can say that we have clearly
discovered something essential about human language. Furthermore, every
claim about linguistic universals can be falsified, and we have seen in the previous
section that this is a desirable property for any scientific theory.
One universal is that it is possible in all languages to divide the sound
stream into a finite set of vowels and consonants, together called segments. segments
The size of this set ranges somewhere between 15 and 150. These are actually
two different, but related universals:
(6) a. The sound structure of all languages can be divided into a finite set
of segments, S .
b. In all languages, S can be subdivided into (complementary) subsets
of vowels and consonants.
If you have been trained from an early age in a culture which uses an alphabetic
system to write its language, this may sound trivial. This is because
people tend to think, even if subconsciously, of written language as primary,
and it is obvious that every language can be written using a finite set of letter
symbols, and furthermore some are called vowels and others consonants.
We might however be misled by this cultural invention to think that there is
something real about these distinctions. For this reason, it is always good to
be critical and see what evidence we have for the reality of S .
Evidence from writing systems
On the other hand, the fact that the alphabet can work in this way, and that
also languages which use a different system can be transcribed at least in the
extended segment set of the International Phonetic Alphabet, is not a triviality,
and is itself a piece of evidence.
It would be possible to construct a language in which every morpheme
corresponds to its own unique sound: man is designated by a sneeze, woman
19
20 2.1. Segments
by clapping your hands once, and love by flapping your lips. In such a language
a man loves a woman would thus sound as SNEEZE-FLAP-CLAP. Such a
language would not necessarily be difficult to produce or to perceive. The fact
that no human language has such a lexicon is thus meaningful.
This entails that alphabetic writing is a technology that builds on a structure
which seems to be inherent in human language. We have to keep in mind,
though, that the relation between sound and letter can be very complex. One
and the same sound can be represented with many different letters and letter
combinations in a language like English, as truly, do, shoe, soon, true, lawsuit,
routine, two, screwed, jewel, manoeuvre, rendezvous, throughout en coups (all have
an [u] sound) show. The reason for this often is that spelling does not just
represent the current sound structure, but also the history of the language. In
other languages (for instance, in Italian) the relation is more one on one.
The difference between consonants and vowels becomes apparent in spelling.
There are many examples showing that a sentence without vowels can still be
read, whereas a sentence without consonants usually cannot:
(7) a. Mst ppl wll ndrstnd ths sntnc. (Most people will understand this
sentence.)
b. U iuay ooy a uea i. (But virtually nobody can understand this.)
Related to this is the finding that when speakers hear the nonsense word
[kebra], they are more likely to think of a cobra than of a zebra: the former
is one vowel ‘away’ from the word they have just heard, whereas the other is
one consonant away. Changing a vowel thus seems to matter less than changing
a consonant. (This experiment was done with speakers of Spanish and of
Dutch.) Computer-mediated communication, such as texting and Twittering
also sometimes uses this technique of leaving out vowels to shorten a message
(txtng).
One reason for this asymmetry may be that there are more consonants than
vowels (for instance: Italian has 24 consonants and 7 vowels, Arabic has 29
consonants and 3 vowels, Malay has 20 consonants and 5 vowels). Semitic
languages such as Hebrew and Arabic use this property: vowels are not usually
written in their script, except sometimes in special versions for people
learning the language or the script. There are no languages in the world which
do the opposite, i.e. which write only the vowels of words but not consonants.
This example thus not only shows that there is a distinction between consonants
and vowels, but also that this distinction involves an asymmetry between
the two sets.
Semitic templates
Semitic languages also provide another well-known type of evidence for the
existence of consonants and vowels. You can see this in the following examples
from Modern Hebrew (Semitic, Israel):
(8) gadal ‘to grow’ higdil ‘to enlarge’ gdila ‘growing’
Satak ‘to keep quiet’ hiStik ‘to quieten’ Stika ‘silence’
sagar ‘to close’ hisgir ‘to extradite’ sgira ‘closing’
2.2. Contrast and the feature 21
If you study these examples closely, you will observe that the three words in
every line have several things in common. This is true both for the meaning
and the form. As to the meaning, the first three words are about size, the
second set of three about silence, and the third set about being closed. As to the
shape, the first three words all contain the consonants g, d, l in that order, the
second triple have S, t, k, and the third triple s, g, r.
In Semitic philology, it is usually assumed that these general meanings
are indeed attached to the consonants (called consonantal roots). If you look consonantal roots
further, you will discover that the vowels also have something to add to the
meaning. The columns in the table also share certain things (the first column
contains simple verbs, the second column contains so-called causatives de- causatives
scribing how something is made larger, quieter of closer, and the third column
has gerunds. It is no accident that the vowels are also the same in each column.
We thus get the appropriate meaning of a word by combining the meaning of
the consonantal roots with those of the vocalic patterns.
Templates of these type are found in all Semitic languages, but also in unrelated
languages such as Yowlumne (Yakuts, North America) and possibly
Rotuman (Oceanic).
Psycholinguistic evidence
There is also quite some evidence that consonants and vowels are represented
in different parts of the brain. For instance, it has been shown that patients
suffering from aphasia (which is a language disorder, usually caused by head aphasia
injury or a stroke) can sometimes be affected only in the vowels — making
more errors producing them or keeping them apart — while other patients
have more problems with the consonants.
It has also been suggested that listeners tend to pay much closer attention
to variation in the production of vowels than that of consonants. The reason
for this might be that variation in vowels is more informative, for instance
telling us whether the word is emphasized, what kind of intonation it carries,
or even what the emotional state of the speaker is.
Finally, also brain scanning has shown differences between consonants
and vowels. In one experiment, changing vowels relative to consonants while
a subject in a brain scanner was reading aloud words, increased activation in a
right middle temporal area, whereas changing consonants relative to vowels
increased activation in a right middle frontal area. There thus is evidence that
consonants and vowels might even be physiologically distinct in the human
brain.
2.2 Contrast and the feature
Even though we have quite some evidence for their existence, segments are
not the ultimate primitives of phonological theory. It is usually assumed that
there are smaller building blocks — called features — which together somehow features
form the segment.
We can see these features is as instructions for the articulatory organs. Take
the French consonant [b]. In order to pronounce this sound, we have among
other things to close our lips and to vibrate our vocal folds. These are two
22 2.2. Contrast and the feature
instructions, one to the lips, and one to the vocal folds. These instructions
correspond to features, [labial] (labia is Latin for lips) en [voice].
[b] shares these features with [v], but is different from the latter sound
because it is pronounced with an explosion. This explosion also corresponds
to a feature, [stop]. The consonant [b] therefore is supposed to have roughly
the following internal structure:
(9) [b] = { [stop], [labial], [voice], ...}
I put the dots in this representation, because possibly there are other features.
However, not every possible movement we make with our lips, or anything
else we can observe about the sound counts as a feature. For instance, when
pronouncing the [b], a speaker might put her tongue in a certain position —
typically somewhere low in the mouth. However, [low] (for a low tongue
position) is not assumed to be a feature of [b], even though it is a feature of a
vowel such as [a]. The reason for this is that there are other vowels which do
not have such a low tongue position, such as [e] or [u].
How do we know which properties get a formal status in the theory as features?
We usually assume that only those traits that are linguistically relevant
are features. The primary type of linguistic relevance for features is distinctiveness.
The feature [voice] is distinctive on French [b], because there is anotherdistinctiveness
sound in the language which has exactly the same features, except that it is
not voiced. The [p] is labial and plosive, but the vocal folds do not vibrate.
On the other hand, [b] is not distinguished from any other sound only by the
position of the tongue; French does not have a consonant with closed lips and
a tongue in a high position.
The traditional test to decide that two sounds are different is the so called
minimal pair test: we construct two words with different meanings which areminimal pair test
only different in the sound in question. A relevant minimal pair in French
would be:
(10) pont ‘bridge’ [p˜O] : bon ‘good’ [b˜O]
Since the two words have different meanings, the difference between [b] and
[p] must have a relevance in French. Since they are only distinguished by
voicing, this means that French has a feature [voice]. The ultimate building
blocks of French will thus be a small set of these features. (Notice by the way
that there are more differences in the orthography, as the word for ‘bridge’
ends in a t; this is however considered irrelevant for phonology.)
In the remainder of this section, I will briefly discuss some important types
of evidence that have been adduced for the existence of the phonological fea-
ture.
Phonological activity
One important type of evidence comes from phonological alternations. Consider
the following examples from Turkish (Turkic, Turkey), paying particular
attention to the final consonant of the stem:
2.2. Contrast and the feature 23
(11) kalıp ‘mold’-NOM kalıb-a ‘mold’-DAT‘
kap ‘container’-NOM kab-a ‘container’-DAT‘
kanat ‘wing’-NOM kanad-a ‘wing’-DAT‘
tat ‘taste’-NOM tad-a ‘taste’-DAT‘
güve[tS] ‘clay pot’-NOM güve[dZ]-e ‘clay pot’-DAT‘
You can see that the final consonant is different when it appears at the end of
the word (such as in the Nominative) than when it appears before a vowel (as
in the Dative). In the former context we find { [p, t, tS,] }; in the latter { [b,
d, dZ] }. The difference between these sets is that the latter all have a feature
[voice], which the former lack:
(12) [p] [labial], [stop] [b] [labial], [stop], [voice]
[t] [coronal], [stop] [d] [coronal], [stop], [voice]
[tS] [palatal], [stop] [dZ] [palatal], [stop], [voice]
(I introduce the features [coronal] and [palatal] here. The former denotes
sounds made with the tip of the tongue at the front of the mouth, the latter
sounds which are made slightly more to the back. We will return to these features
in the next section .) We can thus draw a generalization over the data in See Section 2.3
(11) as follows:
(13) When the Dative has features F1, F2, ..., Fn, as the last segment of the
stem, the Nominative has features F1, F2, ..., Fn, except for [voice].
This is not only more elegant than simply stating for every word what the two
forms are, but also than stating that words which have [b] in the Dative, get
[p] in the Nominative, while those with a [d] get a [t].
Interestingly, consonants such as {m, n, r,...} (called sonorants) do not par- sonorants
ticipate in this kind of alternation:
(14) adam ‘man’-NOM adama-a ‘man’-DAT‘
tavan ‘ceiling’-NOM tavan-a ‘ceiling’-DAT‘
zar ‘die’-NOM zar-a ‘die’-DAT‘
From a phonetic point of view these sonorant sounds are voiced in Turkish,
just as they are in English and many other languages of the world. There are
no counterparts which are not voiced, and in particular there their voicing
they are not voiced. Given the logic just explains, this means that in the phonology
these sounds would not have a feature [voice]. This explains why they
do not participate in this asymmetry: F1...Fn is the same in the Nominative
as in the Dative.
Symmetry of inventories
Another argument in favour of seeing features as the building blocks, is that
we predict that inventories of segments tend to be symmetric, and this is indeed
what we find.
Suppose that the consonant inventory of a language consists of the features
[labial], [stop] and [nasal]. You can check that these features can be combined
24 2.2. Contrast and the feature
in 6 different ways, if we also allow a consonant without any features, but assume
that the combination of [stop] and [nasal] is not allowed. This describes
almost exactly the consonant inventory of Hawaiian:
(15) Hawaiian (Poylnesian, Hawai’i)
[labial]
p k [stop]
m n [nasal]
w l
We thus have a very regular pattern: two stops, two nasals, and two non-stop,
non-nasal consonants. And within each of these pairs, one is labial whereas
the other is not. This regularity is very nicely described with features. What is
more, it is extremely common for languages to have their consonant inventory
described in such a regular way.
Hawaiian has two further consonants ([P] and [h]) which do not fit into
this pattern. That is also a very common property of languages: there is a
tendency towards symmetry, but this is rarely absolute. There will be some
gaps in the table or some sounds which do not really seem to fit.
Any theory will have to be able to take into account both of these factors:
that there is a universal tendency towards having a symmetric inventory of
sounds, and that very few languages have an inventory that is absolutely symmetric.
One way to do it is to assume that the ‘sound grammar’ of languages
consist of a set of features plus rules about how these features can be combined
— one rule which is necessary for Hawaiian is that [stop] and [nasal]
cannot occur together in a segment, as we have seen.
Language acquisition
Features may also play a role in language acquisition: there is evidence that
children acquire features rather than segments. Young children (roughly in
their first year) acquire the individual sounds of their language. This process
is very complicated, and I cannot go into it in full detail here. But roughly,
after a first stage in which they produce all kinds of sounds and learn to say
a few words like mummy, which do not yet seem to be analysed in segments,
let alone in features, they start building up the system. And they seem to use
features to do so.
Typically, a French-speaking child, for instance may at some point have
acquired the consonants p, t and k. These are all three voiceless stops, and they
only differ from each other because the first is [labial], the second [coronal]
and the third [velar] (pronounced with the back of the tongue). These will
then be the only features which the child has acquired.
A next step might be that she acquires the feature [voice]. Our prediction
is then that she will be able to combine this new feature with the ones she
already has. In other words, that she will acquire b, d and g at more or less the
same time. Similarly, at some point she will learn m, n and N at the same time,
since these require combining a new feature [nasal] with the features that are
already in place.
Just like with the preceding evidence, feature theory alone does not make
a perfect prediction. Children each follow their individual path and are in-
2.3. A universal feature set 25
fluenced by all kinds of other factors, such as how often they hear a sound (a
child called Robin might be quicker in learning the [r] sound). But features do
play an important organizing role in the acquisitional path.
Speech errors
Speech errors also provide a possible source of evidence. While we are speaking,
we constantly make small mistakes. We usually do not really notice them,
even when we quickly correct them — and the latter does not seem to happen
very often either. By studying them carefully, however, we may learn many
things about phonological structure.
Phonologists have observed, for instance, that errors often involve exchanging
two consonants or vowels in a word or word group:
(16) a. fish grotto → frish gotto
b. fresh clear water → flesh queer water
c. brake fluid → blake fruid
d. add hoc → odd hack
e. fish and tackle → fash and tickle
In the examples in (16a)-(16c), two consonants have changed position; the
same has happened to vowels in (16d)-(16e). Interestingly, it appears very difficult,
if at all possible, to find examples where a vowel has changed positions
with a consonant. This is a new indication that vowels and consonants somehow
exist in different dimensions, at least in their representation in the human
brain.
These data also give evidence that segments can indeed sometimes be isolated:
in (16a) it isn’t the whole consonantal group gr that is pronounced in the
wrong position — although such errors do also exist — but just the [r]. In
order to make such a mistake, the speaker must thus be able to somehow separate
this sound from the others.
Also individual features can sometimes move around within a word. I
have a colleague whose child insisted that she wanted to eat skabetti. If you
analyse this and compare it to the intended spaghetti, you will observe that the
place in the mouth in which the first two plosives are pronounced have been
interchanged. However, the first obstruent is still voiceless ([k], not [g]) and
the second one voiced ([b], not [p]).
This type of child language behaviour may strictly speaking not count as a
speech error, but we find similar mistakes also in ‘real’ speech errors of adults:
(17) a. Cedars of Lebanon → Cedars of Lemadon
b. pity the new teacher → mity the due teacher
2.3 A universal feature set
We have thus seen that there is quite some evidence that segments are not
the smallest possible units of linguistic analysis, but that they can be further
decomposed in terms of features. We have also mentioned some of these features"
[stop], [nasal], [labial] and a few others.
26 2.3. A universal feature set
One could ask many questions about the status of these phonological features.
Do all languages employ the same set — or at least a selection out of
this same universal set? Or do they each make up their own features? And
what properties of sounds do they refer to? For instance, the feature label [labial],
which we have used so far, refers to a certain articulatory gesture whicharticulatory gesture
speakers make when producing the sound: they move their lips. This is the
view which we will entertain here and in most of this book.
There are logically speaking at least two alternative possibilities. The first
is that features do not refer to articulatory properties, but rather to acoustic
or perceptual ones, e.g. the effect which the sound has on the speech signal.
This view is entertained also by many phonologists. We will briefly discuss
one phonological theory in which this point of view is entertained in section
2.4 below.
The third possibility is that phonological features have no relation to the
phonetic shape of sounds at all. This would probably mean that they are
purely abstract; they would have the shape [α], [β], etc. Although this position
is also preferred by phonologists, even those people who accept prefer abstract
labels in their actual analysis usually use phonetically motivated ones.
The reason for this is that they usually do the job rather well, and furthermore
are easier to keep apart in an analysis which uses many different features.
I thus here will present a version of articulatory-based feature theory. Even
here there is a lot of variation in terminology. I have chosen the labels which
I believe are most common, but it is unavoidable that you will sometimes
encounter in papers labels which are slightly different from the ones here.
Using a dictionary or a phonetic lexicon will then usually help out.
The feature set is divided into a number of subsets which each describe
the properties of sounds. We will discuss these sets in turn on the following
pages.
Manner features
The first set of features describes the general way in which sounds are formed.
These are called manner features or also major class features. We have seen one ofmanner features
major class features those above, [stop], which denotes sounds that are produced with a temporary
obstruction at some point in the vocal tract, and a subsequent burst when
the outgoing airstream is released. Examples of stops (also called plosives) inplosives
English are [t, k, p, d, g] and [b].
Fricatives are sounds which are pronounced with slightly less obstruction.Fricatives
The result of this is that the outgoing airstreams whirls and makes a fricated
noise. Fricatives and plosives together are called obstruents. Fricatives areobstruents
different from plosives in that the former carry the feature [continuant].
All consonants which are not obstruents are called sonorants (and their fea-sonorants
ture is [sonorant]). In sonorants, the lips and tongue make even less obstruction,
so that the air can stream out of the lungs in a way that is more or less
unimpeded.
The first class of sonorants are nasals such as m and n. In these, there stillnasals
is some obstruction in the oral tract, but the airstream can escape through the
nose. (You can check this by putting a little mirror under your nose while
speaking; if you say a p, b or f, nothing happens. But when you pronounce an
2.3. A universal feature set 27
m, there will be some condensation on the mirror. This is the air that came out
of your nose. Nasals unsurprisingly have the feature [nasal].
The next class of sonorant consonants are the liquids, typically r and l (and liquids
in some langauges also some other variants). In liquids, the stream goes really
out of the mouth, but is deformed by the position of the tongue or, sometimes,
the lips or the uvula. R-like sounds are distinguished from l-like sounds in
that the former carry the feature [rhotic] and the latter [lateral].
The final class of sonorants are the glides, such as English [j, w]. These glides
are the consonants that are closest to the vowels. As a matter of fact, [j] is
pronounced as a [i], and [w] as an [u]. The main difference is that the former
are substantially shorter than the latter. It is often assumed that glides and
vowels therefore have the same feature specification: they share the feature
[vocalic] (in contrast to all the other sounds we mentioned, which are assumed
to carry [consonantal].
There are several other manners of articulation of sounds, which English
does not employ. For instance, some languages use ingressive sounds, which ingressive sounds
are produced while the air streams into the lungs rather than outside. Another
different type of producing sounds is instantiated by clicks, which we found in clicks
languages of Southern Africa (and which become known because of Miriam
Makeba’s click song. These sounds are made not by air streaming in or out
of the lungs but by sucking the tongue or lips and releasing them with a little
explosion.
As we will see in chapter 5, manner features are very useful in describing
how vowels and consonants can be arranged in a word. For instance, an
English word cannot start with a liquid followed by a plosive (*rtee, *lpate),
although these clusters are allowed in reversed order (tree, plate). Many languages
— although not all — have restrictions of these type, which often refer
to the manner of articulation.
Consonantal place features
The next set of features denote the place of articulation: the place in the vocal
tract where there is a constriction of the aistream in the case of obstruents, or
where the airstream is deformed, in the case of sonorants.
The left-right dimension in the mouth does not play a role in any phonology:
it does not matter whether you put your tongue to the left or to the right
for instance, most likely because that difference is not audible. All relevant
differences are placed in a line from the front of the mouth to the back of the
throat, usually represented in a picture like this:
28 2.3. A universal feature set
(18)
Ten different places of articulation in the world have been distinguished in
this picture. In particular within a very fine-grained phonetic analysis one
might be able to distinguish between more, but these are definitely the most
important ones. We will briefly discuss each of them in turn.
The first of these are the labial sounds, which are pronounced with (or at)labial
the lips (labia is the Latin word for lips). Examples of these in English are p, b,
m and f. The first three of these are bilabial: they are made by closing both lipsbilabial
(bi- means ‘two’); the fricative however is labiodental: it is pronounced withlabiodental
the upper teeth on the lower lip.
The second broad class of sounds are called coronal: the corona is the smallcoronal
ridge (the alveolar ridge) just behind the teeth, and these sounds are pronounced
around that area. Sounds which are made on the ridge are called
alveaolar; examples are [z, s, n] in English, but the language also has soundsalveaolar
which are pronounced with the tip of the tongue at the teeth, sich as [D, T] (the
initial sounds of these and thing respectively.
Putting the tongue a little bit further down the palate (but still on the hard
palate), one produces palatal sounds. English [S] (as in she) can be seen as anpalatal
example.
One step further down the vocal tract, we find a class of sounds which are
pronounced with the back of the tongue (the dorsum) at the soft palate (the
velum). These sounds (such as k, g) are therefore called velar or alternativelyvelar
dorsal.dorsal
At the far back of the mouth we find the uvula; some languages have
sounds which are made here, and which are logically called uvular sounds.uvular
The French pronunciation of r ([ö]) is one of them. Within the throat we then
find various sounds which are more exotic to the native speaker of English.
Many language in Northern Africa have pharyngeal consonants, like the [Q] inpharyngeal
Somali (Cushitic, Somalia). Epiglottic sounds like the plosive [Ü] are found inEpiglottic
Dahalo (Cushitic, Kenya). (It is unclear whether in the phonology we ever
need to really distinguish between these two classes of sounds.)
Lowest down in the throat (at least as the production of sounds is concerned)
is the glottis. We obviously use the vocal folds for every sound that
is voiced, but the only consonant we call glottal is the glottal stop [P], whichglottal stop
2.3. A universal feature set 29
you can hear, for instance, Cockney English (the traditional dialect of London
City) when it replaces other stops ([stOP] instead of [stOp]). Also h is sometimes
seen as a glottal sound.
In a theory of features, each of these places can be seen as a phonological
feature, so that we have [labial], [coronal], [velar], [pharyngeal], etc.
Laryngeal features
Every consonant by necessity needs to have a manner and a place. You can
therefore safely assume that every consonant has at least two features. Next
to this, a consonant can optionally have some other features.
An important class among these are the laryngeal features. We already men- laryngeal features
tioned above that there is a difference between for instance a voiced b and a
voiceless p can be expressed by assuming that the former has a feature [voice]
which is missing in the latter.
Many languages have these kinds of pairs (b-p, d-t, g-k, v-f, z-s, D-T, G-x,
etc.) Some languages also have triples. Korean (isolate, Korea) is probably the
most well-known example of this. Here are some minimal triples:
(19) bang ‘bread’ pang ‘room’ ph
ang ‘bang’
dal ‘daughter’ tal ‘moon’ th
al ‘mask’
gæta ‘to break’ kæta ‘to fold up’ kh
æta ‘to dig’
[voice] [spread glottis]
The consonants in the righthand column are aspirated, and they therefore receive
a feature [spread glottis]. (Notice that the initial plosives in English
words such as poet, taste and castle are also aspirated; but there are no minimal
pairs between aspirated and non-aspirated plosives in English.)
Vocalic features
We can now turn to vowels. Just like we can distinguish between two major
groups of consonantal features — manner and place of articulation —, we can
also also distinguish two types of vocalic features: those of aperture and those
of place.
Aperture
Aperture features describe the degree of opening of the jaw. Although there Aperture features
have been claims of languages showing four or even more degrees of opening,
it is commonly assumed that most languages have three such degrees, and
these are usually formalised with two features, [high] and [low]. We then get
the following tripartition (the following might describe the five vowel system
of Spanish or Greek):
(20) a. High vowels like [i] and [u] have the feature [high]
b. Mid vowels like [e] and [o] do not have an aperture feature
c. Low vowels like [a] have the feature [low]
30 2.3. A universal feature set
The idea is that no vowel can have the features [high] and [low] at the same
time, since these give opposite instructions to the tongue. In this way, we can
thus derive a three-way distinction with two features.
Another feature which is sometimes also considered ‘aperture’ is [ATR],vfor
Advanced Tongue Root, describing a movement for the back of the tongue. ThisAdvanced Tongue Root
feature is very often used to describe the difference between e.g. [e] and [E]
or between [o] and [O] in Bantu languages (the first of these pairs have [ATR],
but the second does not). The similar vowels in English are also sometimes
described in this way, although it is more difficult to detect actual movement
of the tongue root on these vowels.
Vocalic place
Like for consonants, the most important articulators for vowels are the lips
and the front and the back of the tongue. I adopt a tradition in which the
places of articution is expressed by the same features (with the same names)
as those of consonants. There are many other names for these features, and I
will mention some of them here as well; however, in the other chapters of this
book, I will only use the feature names mentioned first here.
The three main places of articulation for vowels, then are [labial], [coronal]
and [velar]. I will discuss these in turn.
Labial vowels are also called round (in which the feature might als be [round]).Labial vowels
One difference with consonants is that it is not unusual for vowels to carry
more than one feature. In particular, lip rounding can go together with some
movement of the tongue. For acoustic reasons, this usually is the back of
the tong: raising that back has a similar effect on the sound of the vowel as
rounding the lips. Examples of [labial, velar] vowels are [u] and [o] of which
in particular the first ([labial, velar, high]) is extremely common in languages
of the world.
The combination [labial, coronal] is much rarer. French [y] (the vowel in tu
‘you’) is an example, but English does not does have this vowel — although
some modern varieties of British English come rather close to it in their pronunciation
of [u] (as in you [jy]). The reason why this type of sound is presumably
rarer is that raising the front part of the tongue has an acoustic effect
which is very different from that of rounding the lips. While tongue and lips
thus reinforce each other when pronouncing [u], they give partially conflicting
signals when pronouncing [y]. Under these circumstances, the former vowel
has a better chance of survival than the latter.
A vowel which has only [coronal] are the so-called front vowels ([front]).front vowels
Examples of these are [i] ([high, coronal]), [e] ([coronal] without aperture specification)
and [æ] ([coronal, low]). When they are not also labial, these sounds
tend to be produced with spread lips.
Back vowels, on the other hand, have the feature [velar], which tends toBack vowels
cooccur with [labial] for the acoustic reasons just explained. Most English
back vowels are also labial (we already mentioned [u, o]). Only the low vowel
[a] is not round — which may have something to do with the fact that low
vowels are pronounced with a rather wide opening of the jaw, which makes it
difficult to round the lips at the same time. [a] can thus be specified as [vocalic,
low, velar].
2.3. A universal feature set 31
Vowels can also be placeless, i.e. missing all three of {[labial], [coronal],
[velar]}. The most well-known of these placeless vowels is schwa, the vowel in placeless vowels
schwathe second syllable of better. Its IPA symbol is [@]. There is also (at least) one
high placeless vowel, [ı], and a low placeless vowel, [2].
Vowels are sometimes also differentiated according to the pitch with which
they are pronounced. For instance, Yoruba (Niger-Congo, Nigeria/Benin/Togo)
has a minimal pair such as the following:
(21) ò.wò. ‘honour’ - ò.wó. ‘group’
The grave accent (à) denotes a relatively low tone, while the acute accent de- tone
notes a high tone. The word for honour and group are thus differentiated only
by these tones, which can be modeled by the features [high] and [low].
The final vocalic feature I want to mention is again shared with consonants:
[nasal]. For vowels, it surfaces in languages like Saraiki (Indo-European;
Pakistan, India, Afghanistan), which contrast nasal from oral vowels, as the
following (near-)minimal pair shows:
(22) a. [a:Vi] ‘forge for preparing bricks’
b. [˜a:˜V˜ı:] ‘you should come/do come’
Nasalisation is indicated by a tilde on top of the vowel. In (22b),also the approximant
consonant [V] is nasalised. Languages which have nasal vowels
also always have nasal consonants. Furthermore, the nasality of vowels is
often shared with adjacent consonants.
Monovalence and markedness
Finally, here is a word of warning. There are two interpretations of phonological
features. Here I have presented an interpretation of features as privative: privative
high vowels have the feature [high], but non-high vowels do not. When you
go on to study phonology articles, you will discover that some authors have
a different interpretation. For them, all segments have all features, but with
different values, usually denoted by + and − signs. Under such an interpreta- values
tion, high vowels have a feature [+high] and non-high vowels have a feature
[−high]. This interpretation of features is called binary for obvious reasons. binary
To a large extent, these two different notations are obviously equivalent.
But there are some differences as well. One reason why I prefer the privative
notation is that it expresses more clearly that the two ‘values’ of a feature are
not symmetric. A non-nasal vowel is in several ways simpler than a nasal one.
All languages have oral vowels, but not many have no oral ones. Furthermore,
even if a language does allow for nasal vowels, their occurrence is much more
heavily restricted than that of oral vowels. For instance, they can only occur
next to nasal consonants, as in the Saraiki example in (22b). Such properties
together are summarised by phonologists in saying that nasal vowels are more
marked than oral vowels. marked
Such asymmetries are better expressed in a theory in which nasal vowels
have an extra feature which is absent in oral vowels. The former are literally
more complex in our representations, which correspond to their being treated
as more ‘difficult’ in human language. In a theory with binary features, there
32 2.4. An alternative: Element Theory
is no formal difference: a nasal vowel is [+nasal] and an oral vowel is [−nasal],
which means that they are equally complex.
2.4 An alternative: Element Theory
The aim of this book is to introduce you to the most important ideas in the
mainstream of present-day theoretical phonology. As in any field of research,
there is however no absolute uniformity on every point; to the contrary, many
points of the theory are still very much debated, and every assumption is
questioned from time to time — as it should be.
It is therefore useful to consider an alternative approach to some of the
assumptions underlying the work of previous weeks. This is so-called Element
Theory, a theory which assumes primitive elements that are similar to, butElement Theory
not exactly the same as the features we have seen. They are also not called
‘phonological features’, but phonological elements instead.elements
An important difference between elements and features is that the latter
cannot be pronounced. A feature such as [labial] does not correspond to any
single acoustic or articulatory event in isolation; we always need to build a
segment with many other features in order to produce it. Elements on the
other hand, can be independently pronounced.
Most vowels in the world’s languages, for instance, will consist of the elements
|A|, |I| and |U| in some constellation. (We will concentrate in this
section on the phonology of vowels, because this is what a large amount of
work within this framework has been devoted to; this is not to say that the
theory has not been applied to consonants as well, however.) Elements are
usually spelled with a capital letter, and they are placed between || brackets
in order to distinguish them from features and phonological or phonetic
strings of segments. Each of these elements can be pronounced in isolation:
(23) a. |A| is pronounced as [a]
b. |I| is pronounced as [i]
c. |U| is pronounced as [u]
Other vowels can be analysed as a combination of one or more of these basic,
most primitive elements — it is of course not a coincidence that these correspond
to the three angles of the vowel triangle:
(24)
a
i uy
e o
t
t
t
t
t





The vowel triangle is a graphic representation of a vowel set in languages ofvowel triangle
the world. The triangle can be seen as an abstract graphic representation of
the mouth, with the lips on the right-hand side and the back of the mouth on
the left-hand side. (The vowel triangle also represents acoustic properties of
these vowels; see the section on Further Reading on page .
2.4. An alternative: Element Theory 33
Vowel systems tend to have this triangular shape. If a language has only
three vowels (such as Classical Arabic), these will be typically [i, u, a]. If it
has five, the set will be [i, u, a, e, o], and in this way the vowel triangle gets
more densely filled the more vowels we have.
This property is represented rather nicely in an element system.The three
quasi-universal vowels are the angles of the triangle, and other vowels consist
of combinations of these ‘cardinal vowels’. For instance, in a typical five vowel cardinal vowels
system we will have the following combinations.
(25) a. The combination |A|•|I| (or |I|•|A|) is pronounced as [e]
b. The combination |A|•|U| (or |U|•|A|) is pronounced as [o]
The order in which we present the combinations of features is of course irrelevant:
it does not matter whether we write |X|•|Y| or |Y|•|X|, since both
refer to the same phonological representation.
The combination |U|•|I| will be typically pronounced as [y], and the
mid front rounded vowel [ø] would consist of the combination |U|•|I|•|A|.
The system is built on the typological observation that three vowel systems
usually occupy the three corners of the vowel triangle. All other vowels are
typologically more ‘marked’: they exist in fewer languages.
This observation cannot be expressed directly in a theory which uses features.
It is not part of the formal system of ordinary feature theory that [high,front]
vowels ([i]) are much more frequent than [front,low] vowels ([æ]). This is
something which just needs to be stipulated, or derived from the phonetic
difficulty to pronounce the latter type of vowels. Within Element Theory, the
issue becomes clear from just looking at the different representations of the
two types of vowels, and the theory therefore seems more restrictive.
Notice, on the other hand, that Element Theory does not provide us with
an answer to the question why [e] and [o] are typologically much more frequent
than [y], and why the latter often behaves as a more marked vowel in
languages which have it.
A popular metaphor of elements in the phonological literature is that of
colours. We only need the three primary colours red, yellow, and blue (or red,
green and blue) to produce all other colours by mixing them in the appropriate
quantities. In the same way, we can derive (almost) all vowels from the three
elements.
The vowel system of Dutch
At first sight, Element Theory might seem obviously too restrictive. Given
three elements, we can only derive six vowels — the ones we have just mentioned.
But many languages have many more vowels. For instance, Dutch has
13 vowels (not counting three diphthongs and some vowels which only occur
in loanwords):
(26) i, y, u, e, ø, o, a, I, E, œ, O, A, @
We cannot go into all the complications of the Dutch vowel system, but we can
illustrate some of the strategies within Element Theory. In the first place, in
34 2.4. An alternative: Element Theory
our colour analogy, we mentioned that we can mix colours in the appropriate
quantities.
For phonological elements, we can express this by introducing the notion
headedness. When we combine two linguistic elements — two words in a syn-headedness
tactic phrase, two morphemes in a word, two syllables in a stressed unit, etc.
— we can always give a special status to one of them: this one is the head.
We can now extend this idea to phonological elements: if we combine
them, we can assign the head status to one of the two. This doubles our representational
possibilities. If we have a combination of two elements |X|•|Y|,
we can distinguish between |X|•|Y| and |X|•|Y|, where the underlining
denotes the headedness. Thus we get the following distinctions within the
realm of mid vowels:
(27) a. The combination |A|•|I| is pronounced as [e]
b. The combination |A|•|I| is pronounced as [E]
c. The combination |A|•|U| is pronounced as [O]
d. The combination |A|•|U| is pronounced as [o]
The phonetic interpretation of headedness might appear obvious: [E] and [O]
are lower than their counterparts, therefore they are more |A|-like, and they
have this element as their head. The head of a segment thus is the one which
has the strongest influence on the phonetic result.
The result obtained so far works very nicely for the many languages which
have a seven-vowel system: they usually have the four vowels in (27), next
to the three primary vowels of course. We are then still assuming a system
in which |I| and |U| cannot be easily combined; notice, by the way that
Swedish is a language which distinguishes between two front rounded vowels,
and thus features a headedness distinction in |I|•|U| combinations.
However, this is certainly not sufficient for Dutch, since this language still
has almost twice as many as seven vowels. A solution here comes from the
study of one of these, the schwa ([@]). This vowel is hard to describe in terms
of the elements we have seen so far: it is the central vowel, right in the center
of the vowel triangle and from an articulatory point of view it is ‘targetless’: it
does not seem to involve the active use of any specific supralaryngeal articulatory
organ.
Although we should be very careful in introducing new phonological elements
— because that would run against the spirit of the program, which
requires us to be as restrictive as possible — the special behaviour of schwa
seems to warrant the introduction of a new element, |@| (the @ sign is sometimes
used as an alternative to ‘@’ in cases where the latter is not available,
e.g. when using a computer that does not yet have the option of representing
phonetic letters).
|@| is special because it is targetless, and therefore it does not have any
effect on the realisation of the vowel.
(28) a. |A|•|@| = |A|
b. |I|•|@| = |I|
c. |I|•|A|•|@| = |I|•|A|
2.4. An alternative: Element Theory 35
d. ...
In this sense, it then behaves like 1 in multiplication or 0 in addition (the technical
term in mathematics is identity element): identity element
(29) 1x1 = 1 1+0 = 1
2x1 = 2 2+0 = 2
... ...
nx1 = 1 n +0 = n
Addition of the schwa in this way at first sight does not make our system
much more powerful, but there is one escape hatch: it may be possible to
extend the notion of headedness also to structures with a schwa. If schwa
is the head of a combination, it does have influence on the interpretation: it
centralizes it, it draws it into the inside of the vowel triangle. This seems to centralizes
give a proper description of the difference between e.g. [i] and [I] or [a] and
[A].
We can now give the following matrix of possibilities:
(30)
|A| [a] |I| [i] |U| [u]
|A|•|@| [A] |I|•|@| [I] |U|•|@| 
|A|•|I| [e] |A|•|U| [o] |I|•|U| [y]
|A|•|I| [E] |A|•|U| [O] |I|•|U| 
|A|•|I|•|@|  |A|•|U|•|@|  |I|•|U|•|@| 
|I|•|U|•|A| [ø] |I|•|U|•|A|  |I|•|U|•|A| 
|I|•|U|•|A|•|@| [œ] |@| @
The crosses  in this table denote segments which could be produced given
the combinatory rules but for which we do not have evidence for in the Dutch
system (note that we have rather arbitrarily assigned the head status in some
combinations of elements). We could try to find some reason for why certain
combinations are lacking — maybe there is some reason why |U| always
needs to be a head when it combines with |I|, or why |@| does not seem to
be the head in complex expressions (except, strikingly, the most complex one
of all: |I|•|U|•|A|•|@|).
Be this as it may, we still predict 20 possibilities by the formal system alone,
while we find only 13. However, if we compare this to a standard feature theory
of the same inventory, this fares relatively well. Within such an analysis,
we would need at least the following five features for Dutch: [back], [round],
[ATR], [high], [low]. Since all these features are binary, we have 25
= 32 logical
possibilities. The element model thus gives a thighter fit to the data.
Vowel reduction
In many languages of the world, there is an interesting difference between
stressed and unstressed positions of the word: in stressed positions we usually
find a larger number of phonological contrast, which is reduced in unstressed reduced
position. In Belorussian (Indo-European, Belorussia), we find {i, u, e, o, a} in
stressed syllables, but only {i, u, a} in unstressed position. This distributional
preference is also responsible for alternations: if an /o/ or /e/ ends up in an
unstressed position, it will be reduced to [a]:
36 2.4. An alternative: Element Theory
(31) stressed unstressed
nóGi ‘legs’ naGá ‘leg’
kól ‘pole NOM’ kalá ‘pole GEN’
vjósn1 ‘spring GEN’ vjasná ‘spring NOM’
Sépt ‘whisper’ Saptátsj ‘to whisper’
kljej ‘glue’ kljajónka ‘oil-cloth’
This type of reduction can be called ‘centrifugal’: the vowels move to thecentrifugal
corners of the vowel triangle when they do not carry stress. It can be opposed
to centripetal vowel reduction, in which all the vowels seem to move to the cen-centripetal vowel reduction
ter of the vowel triangle in exactly the same types of positions. An example
is (informal) Dutch, in which any kind of unstressed vowel can be reduced
to schwa — the difference with Belorussian is that the process is optional in
Dutch, but that is immaterial to our present discussion.
(32) formal informal
fonologie ‘phonology’ [f`onoloG´ı] [f`on@l@G´ı]
minuut ‘minute’ [min´yt] [m@nyt]
kantoor ‘office’ [kAnt´O:r] [k@t´O:r]
The distinction between ‘centrifugal’ and ‘centripetal’ is not very strong in
natural language. Some languages show both processes. For instance, Catalan
/e, a/ reduce centripetally to [@], whereas /O, o/ reduce centrifugally to [u]
(and /i, u/ do not reduce at all):
(33) stressed unstressed
sérp ‘snake’ s@rp@ntí ‘winding’
p´El ‘hair’ p@lút ‘hairy’
gát ‘cat’ g@t´Et ‘kitten’
ńúm ‘light’ ńuminós ‘luminous’
gós ‘dog’ gus´Et ‘puppy’
p´Ort ‘port’ purtuári ‘of the port’
How can we understand these processes, and the fact that together they seem
to give a fairly complete catalogue of vowel reduction phenomena in languages
of the world? Within Element Theory it is very easy to see what is
going on. All reduction processes are an instance of the following:
(34) An unstressed vowel is not allowed to carry more than one element;
delete elements if necessary.
What is more, (34) itself can be understood as a sensible restriction on phonological
structures. In an intuitive sense, unstressed syllables are less prominentprominent
than stressed ones: they are not so loud and literally less important. From perception
experiments, we know that listeners pay less attention to them. This
will be formalized — in chapter 7 — as another headedness relation: stressed
syllables are heads, unstressed syllables are not heads. Some authors equal
vowel reduction to ‘information loss’: unstressed positions can carry less information
than other positions.
2.5. Exercises 37
Centrifugal reduction now leads to one of the primary elements |I, U, A|;
centripetal reduction leads to |@|. Notice that we still need to build an asymmetry
into our system: /i, a, u/ may reduce in some languages to [@], but
/@/ will never reduce to any other vowel, not even a simple one. This should
of course have some relation to the fact that [@] is the ‘identity element’. The
fact that |F|=|F|•|@| means that in some sense |@| is part of all vowels,
but inversely, none of the primary vowel elements are part of [@]. Also in this
sense, then, |@| behaves like the mathematical zero. The fact that schwa is
the targetless vowel makes it the one which carries the smallest amount of
information — hardly any at all.
Feature theory has many other virtues, and is therefore an important topic
for any student of phonology. We will continue to use this theory, rather than
Element Theory, as the background for the following classes. It is important,
however, to realize, that fruitful alternatives exist to many of our key assump-
tions.
2.5 Exercises
1. To make phonological exercises in this and following chapters, it is important
to install a (Unicode) keyboard on your computer. On the page
of SIL you can find an explanation how to do this for the computer system
you are using. Go to this page and install a phonetic font. Then type
the following: [DIz Iz @n EgzAmp@l].
2. In the 19th Century, the French music teacher François Sudre divised a
language which had only seven distinct sounds: do, re, mi, fa, sol, la, si.
Such a language has obviously many advantages, for instance because
you can also play words on a musical instrument. Mention some disadvantages
to organizing sounds in this way.
3. The vowel [i] may have a slighly different set of vowels in a language
which only has the three vowels {i, a, u } than in a language which has
the five vowels {i, e, a, o, u }. Explain.
4. Consider the vowels of Turkish: { i, o, a, ø, u, y, e, ı }. Display these
vowels in a table which shows how you can the set in terms of features.
5. Now do the same exercise in terms of elements.
6. Consider once more the speech errors in (17). Describe exactly which
features have been moved from one consonant to the other.
7. For each of the following segments, give a full feature specification: [f,
h, J, ø, N, p, V, G, @, e]. (Look the sounds up in an IPA table if you do not
know them.)
8. For each of the following feature combinations, give a corresponding
IPA symbol.
a) [consonant, sonorant, labial, nasal]
b) [consonant, sonorant, coronal, lateral]
c) [consonant, fricative, (coronal), dental, voice]
d) [consonant, stop, (coronal), dental, voice]
e) [vowel, coronal, high]
f) [vowel]
9. For each of the following sets of sounds, describe what feature(s) they
have in common:
38 2.5. Exercises
a) [b, d, v, B, z, g]
b) [m, n, ˜a, N, M]
c) [o, u, y, w, Y]
d) [b, m, f, p, M]
e) [m, r, l, R, n, j]
f) [i, a, u, y, e]
10. The prefix in- in English has several different forms. Before a plosive it
takes the following shapes:
a) i[n]active
b) i[n]decisive
c) i[n]secure
d) i[m]popular
e) i[m]balance
f) i[N]coherent
Describe what happens in terms of features.
11. An interesting way of studying the articulation of speech sounds is the
ultrasound technique, which is relatively non-intrusive and allows us toultrasound
see the movements of the tongue. Speech scientists at the University
of Glasgow set up a YouTube Channel which represents many different
sounds. You might be able to find more clips elsewhere. Study a clip
for a dental, alveolar, palatal, velar and uvular fricative. Make a still
— a non-moving picture — at the moment in which each of these is
pronounced, and show how you can indeed see the different places of
articulation as described in this chapter.
12. Take the consonant inventory of a random language of your choice that
has not been described in this chapter (for instance, your native language,
or some other language you might be working on). Describe this
inventory in terms of the features we have seen so far. Do you find any
segments which need features that we have not yet introduced?
13. Japanese speakers who learn French sometimes replace the [y] sound of
the language, by the sequence [ju]. Explain why this might be the case.
14. In the text (p. 2.4) it is suggested that the Dutch vowel system would
be more difficult to describe with binary features. Work out such an
analysis, and show how it is more complicated than the one in terms of
elements.
15. Give an analysis of Belorussian vowel reduction in terms of binary features
rather than elements.
16. Data in some typological databases, such as UPSID, suggest that (almost)
all languages have voiceless plosives, whereas a much smaller
number have voiced plosives. Discuss how this gives evidence for privative
features.
17. What is the simplest vowel according to feature theory? What about Element
Theory? What kinds of evidence would you explore to see whether
such hypotheses are justified?
18. Lakhota (Siouan, North America) has five oral vowels ([i, e, a, o, u])
and three nasal vowels ([˜ı, ˜a, ˜u]). Explain why we would not expect a
language which would have a distribution the other way around (five
nasal vowels and three oral vowels) given markedness theory.
2.5. Exercises 39
19. The online edition of the WALS (World Atlas of Language Structures)
has, among other things, a chapter on front rounded vowels, written
by Ian Maddieson. On the map in this chapter, you can see that front
rounded vowels are very rare in languages of the world. Further, Maddieson
distinguishes between two types of front rounded vowels, mid
and low. Element Theory and Feature Theory make slightly different
predictions about which of these two should be more frequent. Do these
data provide any evidence for either of these theories?
20. The UPSID database (UCLA Phonological Segment Inventory Database)
contains overviews of the segments of several hundreds of languages.
Find in this database, languages which have:
a) [p]
b) [y]
c) [d]
d) [N]
e) [l]
Sources and further reading
Section 2.1 Many examples about the relevance of orthography to phonology
are taken from Coulmas (2003). More discussion about the difference between
consonants and vowels from a psycholinguistic point of view can be found in
Bonatti et al. (2005). The brain scanning data about the same topic is found in
Carreiras and Price (2008). Several scholars have been important for introducing
the relevance of Semitic (and other) templates into phonological theory.
Very important was McCarthy (1979)’s dissertation; Bat-El (2011) gives a good
overview of the literature on several Semitic languages, such as Arabic and
Hebrew.
Section 2.2 The role of features in acquisition has been argued for by Levelt
and van Oostendorp (2007). The examples of speech errors are from a famous
classical article by Fromkin (1973). Lahiri and Reetz (2010) give an excellent
overview of psycholinguistic evidence for phonological features. An interesting
overview and development of the concept of contrast can be found in
Dresher (2009).
Section 2.3
The Saraiki data are from Abbas (2009). Several phonologists have expressed
doubts in recent years as to the universality of the feature set; prominent
among those is Mielke (2004). (Most of these scholars would agree that
the features presented here give a good approximation for most phenomena
in most languages.)
Section 2.4
Element Theory has been worked out in particular in the framework of
Government Phonology (Kaye et al., 1985, 1990; Harris, 1994; Harris and Lindsey,
1995); related views were also found in the framework of Dependency
Phonology (Anderson and Ewen, 1987). An excellent introduction into this
subject matter is provided by Backley (2011). The Belorussian data are from
(Crosswhite, 2001; Harris, 2005). The latter author is also responsible for the
suggestion that vowel reduction is information loss.
Chapter 3
Autosegmental theory
3.1 Tone
In the preceding chapter, I have discussed the evidence that smaller the smallest
elements of phonological structure are features. The next question then is
how these features are organized into words. As an approximation of this, we
have suggested in the previous chapter that segments are bundles (or sets) of
features. This suggests that we can still divide the sound stream into discrete
time events: one segment comes after the other, although they have internal
structure.
When we have a closer look at the data, these suggest otherwise. The
features behave sometimes as independent from their segments: it sometimes
looks as if one feature is attached to more than one segment, for instance, or
even that features do not belong to any individual segment at all but are still
part of a morpheme.
It is an important question, what this organisation looks like. A very influential
view of this is autosegmental phonology. According to this theory, we autosegmental phonology
can see the organisation of speech sounds in the human mind more or less
like a musical score: every feature has its own part, which is to some extent
independent of all other parts. Their only relation is that they are all attached
to one central line, the skeleton, which keeps track of the time. The elements skeleton
of the skeleton — which resemble the notion of a segment in certain ways —
are usually depicted as x’s (we will return to this in much more detail in a few
weeks from now).
An important indication for the correctness of this view is assimilation, a assimilation
process by which two sounds become more similar to each other when they
are put in adjacent positions.
The autosegmental behaviour of features — the fact that features behave as if autosegmental behaviour of features
they are segments in their own right — was first discovered on the basis of
tone. In many languages of the world, syllables can differ from each other just
by being pronounced at a different pitch.
Tones are pronounced on vowels, and for this reason I have treated them
as vocalic features in the preceding chapter. Yet tones are demonstrably also
independent of their vocalic hosts in a straightforward way.
The following facts from Kikuyu (Bantu, Kenya) show this. The way in
which tones are distributed in the word looks rather messy at first, but they
41
42 3.1. Tone
actually come in a very simple and clear pattern.
In order to analyse the phenomenon at all, we first have to briefly consider
the morphological structure of the Kikuyu verb, which can be described by
the following template:
(35) SUBJECT (OBJECT) ROOT TENSE
to ‘we’ mo ‘him’ rOr ‘look at’
ma ‘they’ ma ‘them’ tom ‘send’ irE PAST
A verbal root such as [rOr] is preceded by prefixes which express properties
of the subject of the sentence, as well as the object when the verb is transitive.
It can also be followed by several suffixes; one of these is Tense (in this case,
expressimg that an event took place in the past).
Now if we combine these morphemes and we study the resulting patterns,
it looks at first as if (almost) any morpheme can occur both with a low tone
(marked à) and with a high tone (marked á):
(36) Subject ‘to’ Subject ‘ma’
rOr tò r`Or ìr´E má r´Or ìr´E
tò mò r`Or ìr´E má mó r`Or ìr´E
tò mà r´Or ìr´E má má r´Or ìr´E
tom tò tòm ír´E má tóm ír´E
tò mò tòm ír´E má mó tòm ír´E
tò mà tóm ír´E má má tóm ír´E
You can see for instance that the stem vowel sometimes has a high tone, and
sometimes a low tone. Which of the two is underlying, and how do we derive
the surgace tones from this underlying tone?
On closer inspection, it is not precisely true that any vowel shows any tone:
the vowel of the subject marker ‘to’ always comes with a Low tone, while the
vowel of the subject marker ‘ma’ always comes with a High tone. Furthermore,
the morpheme immediately following the subject marker always bears
exactly the same tone as the subject marker itself. In some sense, the subject
thus determines the tone of the following morpheme.
Similarly, we may observe that the final tone of the tense suffix is always
high, but that the first vowel has a varying tone: if the stem is rOr, we find
a low tone, if it is tom, it is a high tone. Thus it seems to be the stem which
determines the first tone of the suffix.
We can best understand the Kikuyu tone system if we generalise these
observations: the tone of every morpheme shows up on the following morpheme.
Every morpheme in Kikuyu thus consists of two separate parts: segmental
material on the one hand, and completely independent of that, a tone,
which is realised on the following vowel.
The underlying representations thus look as follows:
(37) to ma mo ma rOr tom irE
L H L H L H H
On the surface, every tone needs to be linked to some vowel, and none may
be left ´floating around’, due to the so-called Association Convention of autosegmental
phonology:
3.1. Tone 43
(38) Association Convention: No ‘floating’ tones are allowed on the surface, Association Convention
every tone needs to be linked to a vowel.
The Association Convention for tones is part of a more general set of requirements
on phonological structure, requiring every element in a phonological
representation to be linked to the other parts of the phonological structure, as
we will see in section ??
In many languages, the tones would be linked to the vowel in their own
morpheme, which would obviously also be the most logical option, but in
Kikuyu there apparently is a different requirement which is more important
than keeping every tone realized in its own morpheme, alignment of the tone alignment
to the end of the word:
(39) ALIGN-Tone: All tones want to be as close to the right edge of the word
as possible, given other conditions of the language.
In many tone languages of the world, we see the effect of ALIGN-Tone: tones
tend to move to the right (‘spread’). There might be a phonetic reason for
this: the realization of tones tends to be a little bit delayed. In any case, it is a
tendency we see in many tone languages.
When ALIGN-Tone would decide things on its own, regulating the realisation
of Kikuyu words as it sees fit, it would choose to have the following
representation as the best one for ma mo tom irE:
(40)
ma mo tom irE
¨¨¨

$$$$$$$$$H L H H
All tones are linked to the final vowel, and thus maximally aligned to the end
of the word. The pleasue of this comes at the enormous cost, however, of creating
a very complex tonal configuration on this final vowel, and apparently,
this is not a price which Kikuyu is willing to pay. In particular, the relation
between tones and vowels in this language is very transparantly one-to-one.
In other words, the Association Convention above can be refined to the fol-
lowing:
(41) Wellformedness Condition (WFC): Every tone in the output representa- Wellformedness Condition
tion should be linked to exactly one vowel, and vice versa.
Given strong force of the WFC in Kikuyu — it is not absolute in all languages,
as we will see later — the best we can do to maximally satisfy ALIGN-Tone is
the following:
(42)
ma mo tom irE
 
 
 
 
 
 
H L H H
Every tone is now linked as much to the right as possible, without creating
illicit ‘contour’ tones. Notice, however, that there is still one problem: the very
44 3.1. Tone
first vowel (the one of the subject marker) does not bear a tone at all. There is
no way we can solve this problem, payimg due respect to all the requirements
imposed on the Kikuyu word, and some related Bantu languages would leave
it like this in similar situations, creating a toneless syllable.
However, notice that the WFC expresses several requirements at the same
time, e.g. ‘no tone should be linked to more than one vowel’, and ‘no vowel
should be toneless’. Apparently, the former counts as a stronger violation in
Kikuyu than the latter and therefore the following repair is made:
(43)
ma mo tom irE
 
 
 
 
 
 
H L H H
Contour tones
As we have seen, Kikuyu is very strict in its requirement that vowels can be
linked to at most one tone. The idea that tones exist in their own dimension of
representation — called a tier in autosegmental theory — and that their align-tier
ment to tones can be regulated by a set of requirements — called constraintsconstraints
— is central to (modern versions of) autosegmental theory.
Another application of the idea of this theory which has proved very useful,
is the analysis of so-called contour tones. For instance, Margi (Chadic, Ni-contour tones
geria) has not two but three tones on vowels: a low tone, a high tone, and a
rising tone.
Obviously, this situation cannot be described by just two features, [High]
and [Low]. In principle, there are two ways of dealing with a situation such
as this. We can either introduce a three way featural distinction (e.g. a feature
Tone which has as values High, Low and Rising); or we can describe the
rising tone as a combination of Low followed by High. Autosegmental analysis
advises us to take the latter route, so that we can minimize the number
of primitives in our theory (there are only high and low tones, and autosegmental
association):
(44)
a a a
d
d
H L L H
For Margi, the advice that autosegmental phonology gives us turns out to
point us in the right direction. In the first place, this representation helps us
to understand what is going on with tones when morphemes combine into
complex words. Consider the following facts concerning the definite suffix
-ári. The left-hand column represents the underlying shape of the stems to
which this suffix is added (ˇa represents the rising tone):
(45) a. sál sál-árì ‘man’
kùm kùm-árì ‘meat’
b. Pímí Pímy-árí ‘water’
kú kw-árì ‘goat’
3.1. Tone 45
c. tì ty-ˇarí ‘morning’
hù hw-ˇarì ‘grave’
(45a) shows that nothing happens if the suffix is attached to a consonant-final
stem. Unlike in Kikuyu, every morpheme keeps its own home base; apparently
the tone of the suffix is high.
(45b) shows that if the stem ends in a high vowel with a high tone, this
turns into a glide. Since glides, like all consonants, cannot carry their own
tone, it looks as if the high tone disappears.
(45c) shows that something does happen if the stem ends in a high vowel
with a low tone. Again, the vowel turns into a glide, but now the tone of
the suffix changes to a rising tone. Under autosegmental assumptions, it is
very easy to understand this process: the rising tone is a combination of the
original low tone of the stem and the high tone of the suffix:
(46) a. Input:
t i a r i
L H H
b. Output:
t y a r i
  
L H H
The reason why this happens can be seen as an interaction of the impossibility
of the glide to carry the tone, and the wish of the tone to be linked to some
vowel. Notice, by the way that this is always the vowel which is closest to
the tone in some intuitive sense. In particular, we will not find the following
structure (the representation for tyárˇı):
(47)
t y a r i

L H H
The reason why we do not find this, is because there is a very hard constraint
on autosegmental representations:
(48) NoLineCrossing: Association lines may not cross NoLineCrossing
Different from all other constraints we have seen so far, NOLINECROSSING is
hard-wired into every known grammar: languages cannot fiddle with it. The
reason for this presumably has to do with the interpretation of autosegmental
representations. We are dealing in this case with two lines (traditionally called
tiers in the theory): one line on which we have the tones, and another line on
which we have the relevant vowels.
Each of those tiers represents a timeline: if element A stands before element
B on a tier, this means that the pronunciation of A precedes the pronunciation
of B. Thus, in (47), the realisation of the low tone will always precede
that of the high tones.
46 3.1. Tone
If we think about our representations in this way, it stands to reason that
association of an element X to an element Y means that the realisation of X
overlaps with that of Y in time. Thus the pronunciation of the low tone in
(46a) will happen during the pronunciation of the /i/.
But given all of this, (47) defies ordinary logic: the low tone precedes the
first high tone, but it is also realised during the pronunciation of an [i] which
follows the [a] with which the low tone is associated. In other words, the
pronunciation of the low tone will also follow the pronunciation of the high
tone. This is logically impossible: α cannot at the same time precede and
follow β (except if they overlap, but that is not the case here).
We can thus conclude that grammars can entertain all kinds of representations,
including those which are not completely well-formed (because they
display contour tones, or floating tones, or toneless vowels); but they will
never entertain possibilities which do not make any sense at all.
Another remark to be made with respect to (46), is that this raises the question
what is exactly the output representation for e.g. kwárí. We may assume
that the high tone of the stem is deleted, but it is also logically possible to
assume the following:
(49) a. Input:
k u a r i
H H H
b. Output:
k w a r i
  
H H H
This would make the high vowel and low vowel stems exactly parallel. Whether
or not we accept this, seems to be a matter of taste. Scholars who like the parallelism
will readily accept this; others will point out that there is no empirical
difference between a segment linked to two tones and one linked to one tone,
and that we should therefore go for the simplest representation. The matter is
hard to decide.
We quickly look at yet another argument in favour of the representation of
rising tone as a sequence LH. We get this if we look at the underlying structure
of stems in Margi. Bisyllabic stems in Margi come in three flavours: some of
them have two low tones, some of them have two high tones, some of them
have a low tone followed by a high tone. Monosyllabic stems similarly exist
in three variants: some have a high tone, some a low tone, and some a rising
tone. Under the autosegmental assumption, we can unify these by assuming
that there are only three tonal templates in Margi: H, L, and LH:
(50) H L LH
bisyllabic ndábyá ‘touch’ g`@rhù ‘fear’ p`@zú ‘lay eggs’
t´@dú ‘fall down’ dzàPù ‘pound’ ngùrsú ‘bend’
monosyllabic tsá ‘beat’ dlà ‘fall’ hˇu ‘grow up’
sá ‘go astray’ ghà ‘reach’ vˇ@l ‘fly’
Notice that this means that, even though Margi allows (rising) contour tones,
it still only does this as a last resort: only because otherwise a tone would be
3.1. Tone 47
lost (as in the gliding cases just discussed) or because it is the only way to
express a tonal template. A bisyllabic word *g`@rhˇu is still not allowed, since it
contains an ‘unnecessary’ rising tone. We thus cannot say that the wellformedness
condition does not play a role at all; it just seems to be less stringent in
Margi.
Multiple tones vs. multiply linked tones
We thus observed that Margi has three types of disyllabic words:
1. The first syllable has a low tone, the second syllable has a high tone.
2. Both syllables have a low tone.
3. Both syllables have a high tone.
The representation of the first of these is straightforward in autosegmental
terms, but for the other two, we logically speaking have two options, which I
will illustrate on the low tone example:
(51) i.
x x
  
L ii.
x x
L L
There is one reason for assuming that the representation in (i) is the ‘real’ one:
this allows a more uniform description of disyllabic and monosyllabic forms;
recall that the latter had three tones: low, high and rising; there is no reason to
assume that low toned monosyllabic stems have two (low) tones.
This reason is not very strong, but fortunately there are other arguments;
and they point in the same direction. First, remember what happened to
monosyllabic stems when their vowel would get lost:
(52) a. Input:
t i a r i
L H H
b. Output:
t y a r i
  
L H H
And now consider the fate of bisyllabic low toned stems in the same circum-
stances:
(53) làgù làgwári ‘road’
màlà màlári ‘woman’
No rising tones are created in this case. This is much easier to understand if
we assume the representation in (51a) for stems of this type than under the
assumption of (51b). Under the latter representation, we would not expect
any difference with monosyllabic stems: if the final vowel turns into a glide,
the second low tone will go and try to find a new host on the suffix vowel,
48 3.1. Tone
creating a contour tone in the process. But under (51a) there will be only one
low tone on the stem, and this low tone does not run the risk of becoming
floating, since it can still be linked to the first vowel.
We could now wonder whether there are languages which have both (51a)
and (51b) in their inventory of phonological structures. It has been a claim
of autosegmental phonology that this is not possible; phonological structures
would be subject to the so-called Obligatory Contour Principle:
(54) Obligatory Contour Principle (OCP)Obligatory Contour Principle
OCP Adjacent identical tones are disallowed.
The OCP allows tonal tiers like ‘H’, ‘HL’, ‘L’, ‘LH’, LHL’, etc.; but it disallows
structures like *‘HH’ or *‘HLHLLH’. If we have two vowels in a row which
are pronounced at the same pitch, there is only one option: these vowels are
linked to the same tone.
Meeussen’s Rule in Bantu
In the traditional tonology of Bantu languages, an OCP related rule is called
Meeussen’s Rule (after the Belgian Bantuist Achilles Emile Meeussen, 1912-Meeussen’s Rule
1978). This rule can be illustrated by the following example, from Kirundi
(Bantu, Burundi):
(55) a. nà-rá-zì-bárììrà (I-PAST-them-to sew) ‘I was sewing them’
b. nà-rá-bàrììrà (I-PAST-to sew) ‘I was sewing’
In (55a), the high toned tense marker rá and the stem bárììrà , which also starts
with a high tone, are separated by a low tone agreement marker. Nothing
happens here; we may assume that this form represents the underlying state
of affairs quite faithfully. In (55b), on the other hand, the tense marker and the
stem are adjacent. As a result of this, the second high tone has to go.
It is quite obvious that Meeussen’s Rule describes an OCP effect: two high
tones which are adjacent are not allowed. The way to solve the OCP problem
here is to turn one of the two ‘bad’ tones into a ‘good’ tone, giving an
alternation of high and low tones.
Here is another example of the same phenomenon in a different Bantu
language (Shona; Zimbabwe/Zambia):
(56) mbwá ‘dog’ né#mbwà ‘with a dog’
hóvé ‘fish’ né#hòvè ‘with a fish’
mbúndúdzí ‘army worms’ sé#mbùndùdzì ‘like army worms’
bàdzá ‘hoe’ né#bàdzá ‘with a hoe’
bénzíbvùnzá ‘inquisitive fool’ né#bènzìbvùnzá ‘like an inquisitive fool’
Fáráì (name) nà#Fáráì ‘with Farai’
The examples show — among various other things — the following: Meeussen’s
Rule (i) applies between (certain) clitics and stems, (ii) if the clitic has a high
tone and (iii) the stem starts with a high tone. Of interest are the cases in
which the stem starts with more than one high-toned syllable. It turns out
3.1. Tone 49
that in those cases, all of those syllables become low toned, even though it
would be sufficient for Meeussen’s Rule if we would only change the first one
(witness what happens to forms such as né#bàdzá, where it also not necessary
to change the second high tone of the stem).
This behaviour of low toned words can be understood if we assume that
again in the underlying representations two adjacent syllables pronounced at
the same pitch are associated to the same tonal autosegment. If this tone has
to change, all vowels attached to it will be pronounced differently:
(57) a. Input:
n e h o v e
¨¨¨
H H
b. Output:
n e h o v e
¨¨¨
H L
Interestingly, there are certain sequences of high tones which do not change;
but there is always an extra morpheme boundary involved in those cases. For
example, we can ‘stack’ clitics in Shona, leading to sequences such as:
(58) sé#nè#hóvé ‘like with a fish’
Notice that it is only the tone of ne which changes in this case. This high tone
is not the same as the stem tone. Therefore the latter does not automatically
change with the former.
Under the assumption that Meeussen’s Rule is an instance of the OCP, the
latter principle takes two different effects in Shona:
1. It disallows sequences of the same tone in underlying forms, preferring
multiply linked tones instead.
2. It disallows sequences of high tones on the surface, solving apparent
problems not by spreading, but by changing one tone from high to low.
There is a third way in which the OCP is operative in Shona: it can also block
rules from applying. This is true in particular for a rule spreading a high tone
from the end of one word to the first syllable of the next word:
(59) zvìròngó ‘water pots’
zvìnà ‘four’
zvìròngó zvínà ‘four water pots’
Chìpó (name)
àkàbìkà ‘and then he cooked’
Chìpó ákàbìkà ‘and then Chipo cooked’
ndàkáténgá ‘I bought’
bàdzá ‘hoe’
ndàkáténgá bàdzá ‘I bought a hoe’
50 3.2. Autosegmental representations outside of tone
The last example shows that the spreading of the high tone does not occur
if the second syllable of the second word already has a high tone. Spreading
here would result, again, in a sequence of vowels linked to different high
tones, and apparently, this is disallowed.
All in all, the OCP can thus have three effects in Shona:
1. It disallows certain underlying structures (by way of a Morpheme Structure
Constraint)
2. It can trigger certain processes (H→L in clitic structures)
3. It can disallow certain processes (H spreading)
This is quite typical for phonological constraints: they can work out in different
ways, just bluntly disallowing a structure altogether, or specifying a repair
when it arises. Not every constraint will have all of these effects in every language,
but very often there is a range of ways in which languages can work
towards satisfaction of their constraints.
3.2 Autosegmental representations outside of tone
The OCP gives us a good handle on extending autosegmental ideas to areas
beyond tone. In many dialects of Dutch (Indo-European, the Netherlands/Belgium/Surinam),
the default allomorph of the diminutive suffix is -ke ([k@]). The following example
is from Bergen Dutch:
(60) vrouw ‘woman’ - vrouwke ‘woman-DIM’ [vrAuk@]
However, if the stem ends in a velar obstruent, we find the form -ske ([sk@])
instead (the second example also illustrates umlaut, which is irrelevant for
our purposes):
(61) a. vlieg ‘fly’ - vliegske ‘fly-DIM’ [vlixsk@]
b. boek ‘book’ - buukske ‘book-DIM’ [buksk@]
This can be understood as follows: bare addition of -[k@] to the stem would
result in an OCP violation on the feature [velar]:
(62)
v l i x k @
[vel] [vel]
Inserting a segment with a different place of articulation — such as coronal [s]
—, solves the problem: the two segments with the ‘bad place’ are no longer
adjacent.
A famous case of a non-tonal OCP effect is the interaction of Lyman’s LawLyman’s Law
with the Rendaku rule in a certain class of words in Japanese (Japonic, Japan).Rendaku
The latter rule turns the second element of a compound into a voiced segment;
the former expresses the condition that there is no other voiced segment elsewhere
in the word (3.2):
3.2. Autosegmental representations outside of tone 51
(63) tama ‘ball’ teppoo+dama ‘bullet’
sono ‘garden’ hara+zono ‘flower garden’
taba ‘bundle’ satsu+taba, *satsu-daba ‘wad of bills’
sode ‘sleeves’ furi+sode, *furi-zode ‘long-sleeved kimono’
Clearly, Lyman’s Law — which despite its name was first discovered by Motoori
Norinaga in the 18th century — could be stated as a specific instance of
the OCP:
(64) Lyman’s Law (OCP style): Avoid two voiced obstruents within the same
word.
The claim is thus that Lyman’s Law blocks Rendaku in Japanese in the way in
which the OCP blocks high tone spreading in Shona.
There are various interesting problems connected to this. Most important
among these is the issue that apparently vowels and sonorant consonants do
not count for the OCP; they are, as it were, invisible. The first examples in
demonstrate this clearly. The standard way of understanding this is by assuming
that these segments simply do not have a link to any [±voice] feature:
they are underspecified for that feature. The reason for this is that they do not underspecified
contrast for this feature: there are no minimal pairs of words where one has a
voiced sonorant or vowel, and the other one a voiceless one. Implicit in our
analysis of tone above was, by the way, similarly that consonants are underspecified
for tones. Again, the reason for this is that the Bantu languages we
discussed simply do not distinguish consonants from each other by tone.
Vowel Harmony
Another domain to which autosegmental analysis has been applied with considerable
success is vowel harmony, a phenomenon that can be found in many
languages of the world, albeit in different versions.
In a typical vowel harmony language, the set of vowels can be split up into vowel harmony
two (or more) disjoint subsets; all the vowels within one word are taken exclusively
from one subset. In Turkish, we can divide the set of vowels along
the round-spread dimension as well as along the front-back dimension. The
following gives a general idea of what is going on (Clements and Sezer, 1982):
(65) nom.sg. gen.sg. nom.pl. gen.pl.
‘rope’ ip ipin ipler iplerin
‘girl’ kız kızın kızlar kızların
‘face’ yüz yüzün yüzler yüzlerin
‘stamp’ pul pulun pullar pulların
‘hand’ el
“
el
“
in el
“
ler el
“
lerin
‘stalk’ sap sapın saplar sapların
We can understand this autosegmentally by assuming that the features
[±velar] and [±labial] can (and should) spread in Turkish:
52 3.2. Autosegmental representations outside of tone
(66)
[velar]
   dd
dd   
[-labial]
s a p ı n
The idea is that the phonological properties which are expressed by the harmonic
features belong to the word as a whole, and get associated to everything
within that domain; but they can ‘see’ only those things which have a harmonic
counterpart, i.e. for which the feature makes sense. Since consonants
usually do not have a harmonic sister, talking about e.g. [±round] does not
make sense, and therefore they do not participate in the harmony.
Some consonants in Turkish do have a harmonic sister, however. Exactly
those consonants can therefore participate in the harmony. I concentrate on
/k/ here, but similar things can be said aboud /g, l/:
(67) -back /k
“
/ +back /k/
k
“
ir ‘dirt’ kır ‘meadows’
k
“
el ‘bald’ kul ‘slave’
k
“
ör ‘blind’ kol ‘arm’
dik
“
‘upright’ sık ‘often’
dök
“
‘pour’ ok ‘arrow’
sak
“
in ‘calm’ sıkan ‘warning’
fak
“
ir ‘poor’ mika ‘mica’
nek
“
tar ‘nectar’ boksit ‘bauxite’
bol ‘abundant’ bol
“
‘cocktail’
kar ‘snow’ k
“
ar ‘profit’
/k, k
“
/ can also initiate harmonic behaviour themselves; to be precise on epenthetic
vowels:
(68) careful form colloquial form
‘fetters’ pranga pıranga
‘prince’ prens pirens
‘test’ prova purova
‘announcer’ spiker sipiker
‘credit’ kredi kıredi
‘cruiser’ kruvazör kuruvazör
This can also be seen in words like kulüp ‘club’, and even (given the appropriate
analysis, and in certain cases) for suffixes:
(69) nom. sg. acc. sg.
‘explosion’ infil
“
ak infil
“
ak
“
i
‘perception’ idrak idrak
“
i
‘desire’ ševk ševkı (in some dialects; older speakers)
‘confirmation’ tasdik tasdikı (in some dialects; older speakers))
3.3. The skeleton 53
3.3 The skeleton
The core of the phonological structure
Until now, we have been quite informal as to the precise structure of autosegmental
representations. We know that tones and features such as [voice] and
[velar] can behave like autosegments on independent tiers which are somehow
linked to the ‘segment’, but we have not yet developed a clear notion of
how all of these tiers are then organized into a larger structure. This is what
we will set out to do today and in the next week.
Today we discuss the central tier of autosegmental representations, the
timing tier or skeleton. Different from other autosegments, the elements on this timing tier
skeletontier do not correspond to their own independent (articulatory) instruction.
Rather, each of them is represented as a neutral symbol ‘x’; these symbolic
units are called timing slots, because their most important function is to organ- timing slots
ize all autosegments into temporal units. They are also sometimes called x
slots. (Notice that this means that the OCP should not be able to apply to this
tier, otherwise we would only be allowed to have 1 segment per word.)
Syllable structure is built on top of these timing slots, and all autosegmental
features are linked to them. An autosegmental representation for a
(hypothetical) word [pi:] with a falling tone would thus be approximately as
follows:
(70)
σ
dd  
x x x
  
b i
e
e
e
e
e
e
e
e
H L
Several remarks are in order here. In the first place, it may seem as if some of
the association lines are crossing in this representation, even though we have
argued in the first class that this is absolutely disallowed in the phonology
of natural language. The reason for this line crossing is equally trivial, however:
we are drawing a three-dimensional structure in two-dimensional space.
The line with tones is in a different dimension from the line with segmental
information; therefore the lines do not really cross. The logical problem connected
with line crossing which we discussed in the first class therefore does
not arise — this structure is perfectly legitimate.
Another thing we should note is that this representation is still overly
simple. We have conflated many autosegmental tiers by just writing /b/ and
/i/. Also the syllable structure is much too simple; we have marked it here by
one σ which is associated to three segments, but that will need to be turned
into something more sophisticated later on. In this class, however, we will
stick to these two simplifications, and we will also no longer consider the
tonal tier. Furthermore, we will in most cases stick to a further notational simplification,
which is sometimes used in the literature. We assume that syllable
positions will mark an x as being a consonant (C) or a vowel (V). In the pictures
below, we will leave out the syllable structure altogether, and note x’s as
54 3.3. The skeleton
C’s and V’s instead. All of this means that the word [bi:] gets the following
representation in our current discussion:
(71)
C V V
  
b i
If this is an autosegmental representation, we now know that we can expect
the following variants, for instance for the vowel /i/:
(72) a.
V
i b. i c.
V
dd
i E d.
V V
  
i
Indeed, all of this structures are attested: (72a) gives a regular (short) vowel;
(72b) gives a ‘floating’ vowel (we will see what floatingness means for segments
in Section 3.3 ); (72c) gives a diphthong; and (72d) gives a long vowel.See p. 57
diphthong
long vowel
We should also expect that (72a) (with a one-to-one association) is the regular
case, which every language has. This is indeed what we find. For consonants
we can set up the same set of structures:
(73) a.
x
t b. t c.
x
dd
t s d.
x C
  
t
(73a) gives a regular (short) consonant; (73b) a floating consonant; (73c) a
doubly articulated consonant; and (73d) gives a long consonant.
In what follows, we will see examples of the most important (and most
surprising) structures of these from a variety of different languages. I will
give examples of consonants and vowels each.
Long vowels in Finnish and in Germanic
The superlative suffix for nominative singular adjectives in Finnish is -in. If
we add this suffix to a vowel final stem, the stem vowel gets lost:
(74) vanha ‘old’ [vanha] vanhin [vanhin]
köyhä ‘poor’ [køyhæ] köyhin [køyhin]
The last example shows that it is only the final vowel that gets deleted. At first
sight, long vowels are exceptional: they do not get deleted, they get shortened
instead:
(75) tervee- ‘healthy’ [terve:] tervein [tervein]
rakkaa- ‘beloved’ [rak:a:] rakkain [rak:ain]
3.3. The skeleton 55
The behaviour of long vowels is hard to explain under the assumption that
they would be carrying for instance a feature [long]. Things become clearer if
we consider stems ending in two vowels. In such cases, the first vowel does
not get deleted:
(76) tärkeä ‘important’ [tærkeæ] tärkeä [tærkein]
We can unify the ‘exceptions’ in with the piece of data in if we assume that
long vowels consist of two short vowels in a row. The behaviour of tervee then
becomes completely parallel to the behaviour of tärkeä. This however would
still leave us with two options for an underlying representation:
(77)
x
e
x
e
x x
  
e
It is hard to decide on independent grounds, within Finnish, which one of
these two representations is the correct one. The language has vowel harmony,
but this affects long and short vowels, and diphtongs or vowel sequences all
alike.
In some other languages, we may find evidence that the representation
on the righthand side is the correct one. One piece evidence comes from a
phenomenon called compensatory lengthening. An instance from this comes
from the history of English. Compare the following Old English words with
their Dutch or German cognates:
(78) Old English Dutch/German
g¯os gans (Dutch)
¯oþer ander (Dutch)
s¯ofte sanft (German)
f¯ıf fünf (German)
¯us ons (Dutch)
The Old English words all have a long vowel, where the Dutch/German forms
have a short vowel followed by a nasal. There is reason to assume that the
latter is more faithful to the state of affairs in Proto-Germanic, predating all of
the Germanic languages, and that English is the language that has changed.
How can we describe what has happened? Autosegmental phonology
gives us a nice tool to provide this description: first, the nasal got lost, i.e. it
was delinked from its position on the skeletal (for some reason which we cannot
describe at this point yet). After this, the empty position was filled by
the preceding vowel. The lengthening is compensatory in the sense that the
vowel length compensates for the lost consonant:
(79) a. Input:
x x x
u n s
56 3.3. The skeleton
b. Output:
x x x
  
u n s
A few things have been noted again. In the first place, the change we are
witnessing is a priori nothing like a phonological rule in the sense we have
seen them before. We are dealing with a diachronic change, the ‘input’ in
(143) represents some stage in the history of English and the ‘output’ some
other stage; and there is as yet no specific reason to assume that any speaker
ever had both of them in his head. Still, also diachronic changes like this may
give us some insight in the mental representations of speakers.
Under this assumption, then, we actually see autosegmental phonology
at work. If a long ¯u would be nothing but a sequence of two short u’s, we
would not really understand what was going on: we would have to say that
the nasal would have turned into a full copy of the preceding vowel, which
would make the representation of this change rather complex.
Compensatory lengthening is found in many of the world’s languages.
A very well-known case can be found in Turkish. In this language, there is
actually a reason to assume that it is a synchronic process and not just the result
of language change, because, depending on sociolinguistic and pragmatic
factors, speakers can choose to delete or not delete a consonant (to be more
precise, one of /v, j, h/. When they do delete, compensatory lengthening
follows suit automatically:
(80) a. kahya ‘steward’ [kahja]-[ka:ja]
b. eylül ‘September’ [ejlyl]-[e:lyl]
c. sevmek ‘love’ [sevmek]-[se:mek]
Long consonants in Italian
Next to long vowels, we also expect to see long consonants. And indeed, there
is at least as much evidence for their autosegmental representation as there is
in the case of long vowels.
A famous case comes from Italian dialects in which we find a phenomenon
of Raddoppiamento sintattico (Syntactic doubling). In the first place, we have to
know that most Italian dialects have long consonants (or geminates as they
are usually called). For instance, there is a contrast between papa ‘father’
[papa] and pappa ‘porridge’ [pap:a], which can presumably only be described
in these terms. Yet it should be noted that on the surface, the first vowel in the
word for father lengthens, whereas the second vowel does not.
The reason for this presumably is that the following is true for Italian (as
well as for many other languages):
(81) An x-slot has to follow the stressed vowel within the syllable.
In order to understand why (81) would need to be the case, we would need
to delve deeper into the theories of syllable structure and stress; the idea is
that stress needs some space within the syllable to be expressed. This space is
3.3. The skeleton 57
already available in páppa, but it needs to be filled by the vowel in pápa (the
accents denote that stress is on the first syllable in both words):
(82)
σ
dd  
σ
  
x x x x x
  
p a p a
σ
dd  
σ
  
x x x x x
  
p a p a
There are thus two ways of satisfying (81): either by a long consonant or by a
long vowel. For some reason, the latter option is not open in the last syllable of
the word in Italian: the language does not allow words to end in a long vowel.
Therefore, what we find is a doubling of the consonant (which is dependent
on certain syntactic factors as well, hence the name):
(83) a. Cittá [s:]anta
Holy city
b. La sciammia aveva appena mangiato metá [b:]anana.
The monkey had just eaten half a banana.
c. La sciammia aveva appena mangiato quáttro [b]anane.
The monkey had just eaten four bananas.
At a sufficiently high level of abstraction, the phenomenon looks a little bit like
compensatory lengthening, except that the position to be filled is not caused
by deletion of a segment, but by stress. Again, it is hard to understand this
without autosegmental representations: why would otherwise the empty position
be filled by an exact copy of the following consonant or the preceding
vowel?
Floating consonants in French
We now turn from the doubly linked structures (long vowels and long consonants)
to the unlinked structures: the representation of floating segments. In
particular, we will have a brief look at floating consonants in the phenomenon
of French liason. French liason
In this language, final consonants of certain words are subject to a phonologically
motivated alternation: they surface before a word starting with a
vowel (noted as # V, where # is the symbol phonologists sometimes use for #
the beginning and the end of the word), but not before a word starting with a
consonant (# C) or at the end of a phrase (the double boundary sign ## indicates
this):
(84) # V # C ##
petit ami petit camerade il est petit
gros enfant gros camion il est gros
un enfant un gros enfant il y en a un
premier etage premier cas il est le premier
58 3.3. The skeleton
The nature of the consonant that surfaces before a vowel is determined by the
preceding word: petit always has [t], premier always has [r], etc. Thus, these
consonants somehow have to be present in the underlying representations of
these words.
Further, we have to distinguish the /r/ from premier from that of cher, since
the latter does not alternate, but always surfaces:
(85) # V # C ##
cher ami cher camerade ça coûte cher
The autosegmental solution is to assume that the /r/ in cher is underlyingly
linked, whereas the one in premier is floating:
(86)
σ
dd  
x x x
S E r
σ
dd  
σ
   dd
x x x x x x
p r @ m j e r
In French, like in many other languages of the world, syllables prefer to start
with a consonant rather than with a vowel: syllables are optimally CV. In
parallel to (81) we thus have (87):
(87) An x-slot has to precede the vowel in a syllable.
Because words which start with a consonant underlyingly, already satisfy (87),
nothing will happen in premier camerade: the /r/ will not find an x-slot to be
linked to, hence it will not be timed and not pronounced. Yet in premier ami,
the extra consonant projected because of this requirement, comes to the rescue
of the floating consonant.
(88) a.
σ
dd  
σ
   dd
σ
  
x x x x x x x x
  
p r @ m j e r a
b.
σ
dd  
σ
   dd
σ
  
x x x x x x x x
p r @ m j e r k a merade
3.4. Feature Geometry 59
Contour segments in Luganda
To round off our discussion of the various autosegmental possibilities of the
skeleton, we need of course also to provide evidence for the existence of structures
where more than one segment is linked to one timing slot. One piece of
such evidence we can find in Luganda, a Bantu language from Uganda. Like
many Bantu languages, Luganda has so-called prenasalized consonants such as prenasalized consonants
[mp, mb, nd, nt, Ng, Nk] and a few others. One might think of them as two
segments (a nasal and a plosive) but at the same time they behave like one
segment, for instance with respect to syllable structure (which we will not
discuss). Furthermore, they are always preceded by a long vowel:
(89) a. ku siinza ‘to worship’
b. ku toonda ‘to create’
c. mu leenzi ‘boy’
d. ku laba ‘to see’
e. ku: n daba ‘to see me’
The last example shows that the lengthening is not just another instance of a
diachronic process, but it corresponds to a productive rule of Luganda phonology.
It also shows that if we put a segment /n/ together with a stop, we
create a prenasalized consonant.
The autosegmental analysis of this is not too complicated. Apparently a
nasal will dock unto the x-slot of the following consonant, for whatever reason
(maybe because the language does not like to have two consonants linked to
independent x-slots in a row, due to some sort of OCP effect). Because of this,
the original x-slot of the nasal becomes available, and the vowel spreads, just
as in compensatory lengthening (I only draw the three relevant segments of
ku n daba:
(90) a. Input:
x x x
k u n d aba
b. Output:
x x x
     
k u n d aba
Other candidates for representations with two segments being linked to one
timing slot are affricates (e.g. [c] = /t/+/s/ linked to one slot) and doubly
articulated consonants (e.g [
>
kp]).
3.4 Feature Geometry
The Place node
If features are organized into tiers, we still have to find out how those tiers are
related to each other. Last week, we have seen that there is evidence for one
60 3.4. Feature Geometry
central timing tier, the skeleton. But this still leaves many different options.
One possibility — maybe the simplest one — is to assume that all features are
linked directly to this one central tier. This is sometimes called the bottle brush
model:
(91)
[labial]
[cor] [velar]x
[+voice]
However, there is evidence against this simple model, and pointing in the
direction of features being organized in arborescent structures; the school of
thought is called feature geometry (using a somewhat excentric definition offeature geometry
the term ‘geometry’). The most straightforward evidence here comes from the
fact that sometimes certain features group together. A well known case is place
assimilation. In many languages of the world, nasal consonants assimilate inplace assimilation
place of articulation to the following consonant. The following examples are
from Chuckchi (Palaeosiberian, Siberia), where we assume the assimilating
nasal is N underlyingly.
(92) t@N-@ì-@n ‘good’
tam-pera-k ‘to look good’
tam-vairgin ‘good being’
tam-waÈ@rÈ-@n ‘good life’
tan-tsai ‘good tea’
ten-leut ‘good head’
tan-ran ‘good house’
ten-y@ìq@t-@k ‘to sleep well’
In these cases, the nasal assimilates in the value for the features [coronal],
[velar] and [labial], but not for any other feature (e.g. it does not lose its nasality
or turn into a fricative).
We could of course assume that Chuckchi has three different phonological
rules which we could informally state as follows:
(93) a. Spread [coronal] from a consonant to a preceding nasal.
b. Spread [velar] from a consonant to a preceding nasal.
c. Spread [labial] from a consonant to a preceding nasal.
But this is very unattractive, especially because we find a similar phenomenon
in many languages of the world, and it always involves these features. But
more in general, we would want to give a uniform description of phenomena
such as this. In order to achieve this, we posit an organizing node in our
phonological representations, called a Place node. The place nodes are not
linked individually to the central skeleton, but through this organizing node:
3.4. Feature Geometry 61
(94)
x
[cor] [velar]
dd  
Place [nasal]
rrr
[labial]
We can now formulate the relevant rule in a very simple and straightforward
way:
(95) Spread the Place node from a consonant to a preceding nasal.
When we spread the place node, we spread all the relevant features at the
same time. Nasal assimilation thus gets a simple and straightforward formal-
isation.
Another type of evidence pointing in the same direction comes from debuccalization.
For instance in certain dialects of Malay, consonants in coda position debuccalization
change according to the schedule in (96) (Humbert, 1995; Botma, 2004):
(96) a. /p, t, k/ → [P]
b. /s, f, h/ → [h]
c. /m, n, N/ → [N] (a ‘placeless nasal’)
(97) a. /ikat/ → [ikaP] ‘to tie’
b. /lipas/ → [lipah]
c. /Pawan/ → [PawaN]
The traditional name for this process is ‘debuccalisation’, since all the oral articulators
become inactive. On the other hand, the manner of articulation stays
constant: a stop /t/ stays a glottal stop [P], a fricative /s/ stays a fricative /h/,
and a nasal /n/ stays a nasal, albeit a placeless one.
Again, we could formulate this in terms of three independent rules:
(98) a. Delink [coronal] at the end of the syllable.
b. Delink [labial] at the end of the syllable.
c. Delink [velar] at the end of the syllable.
This would come at a loss of generality, however, especially since again the
three processes seem often linked. For instance, the same phenomenon can be
found in London English (Lass, 1976; Gussenhoven and Jacobs, 1998).
(99) piPtS@ picture
mIiP wIlj@m meet William
nIP næPs knick-nacks
nOP næU not now
kIiP smAIlIN keep smiling
62 3.4. Feature Geometry
Introducing a Place node allows us to simplify the formalism considerably.
Both Malay and London English are subject to the following rule:
(100) Delink the Place node at the end of the syllable.
Note that this means that we assume that segments such as [P] and [h] lack
a place of articulation node. It is not the case that these segments have a
specification [-coronal, -labial, -velar]: they do not have any place features
whatsoever.
This particular assumption also makes it easier to understand why the
glottal stop very often functions as the ‘default consonant’. For instance, we
fill in this consonant in German if otherwise a situation of hiatus – two adjacent
vowels – would ensue, or if a word starts with an open syllable:
(101) Theater ‘theatre’ [teP´atK], Chaos ‘chaos’ [k´aPOs], atmen ‘to breathe’ [P´atm@n]
The reason why a consonant has to be inserted here, probably is the same as
why we have liaison in French (which we have seen last week):
(102) An x-slot has to precede the vowel in a syllable.
Different from the liaison context, there is no obvious neighbouring consonant
to fill the empty slot in cases such as in (101). Therefore the slot is filled by the
phonological rule component. We can understand why it is the glottal stop
that is inserted in contexts like this, if we assume some principle of representational
economy: if we have to insert something, we prefer to insert as little
as possible to satisfy our needs. If we need to insert a consonant, it is better to
insert one where we do not have to include a Place node (and Place features).
It is not the case, by the way, that glottal stop is the default consonant in
all languages of the world. Some languages do not allow this type of segment
at all — apparently, they disfavour Place-less consonants. In such cases, some
other consonant such as /t/ fulfills that role.
The feature tree
The next question obviously is whether the Place features are the only ones
which are organized into a seperate node. Most phonologists in the feature
geometry paradigm would agree that this is not the case, and that there is
more internal organisation to the segment. Although there is no general agreement
on this point, the following structure may be considered as fairly representative
for the mainstream:
3.4. Feature Geometry 63
(103)
± consonantal
± sonorant
[cor] [velar]
 
 
 
$$$$$$$$
e
e
rrr
rrr
rr
––––––––
Place
[±nasal]
[±continuant][±lateral]
   dd
[labial]
Laryngeal
   f
f
f[±voice]
[±aspirated]
[±high][±low]
Aperture
   e
e
Further structure is possible; for instance, Place and Aperture are often combined
into a Supralaryngeal node, combining all the instructions for organs
above the larynx. Also, the position of the features [±continuant], [±nasal]
and [±lateral] has been the topic of debate.
It needs to be observed that the claim underlying virtually all work in Feature
Geometry is that the structure in (103) — or whatever should be replacing
it — is universal: if a language has a feature [continuant], it will be organized
into the structure as indicated.
A prediction of this model is that all the organizing nodes should behave
like the Place node. There should be processes — for instance of assimilation
— which involve exactly the features that are dominated by some node
and none of the others. We will briefly review some of this evidence for the
Aperture node and the Laryngeal node.
As to the former, consider the following examples from Brazilian Portuguese
(Wetzels, 1995; Clements and Hume, 1995):
(104) 2nd person 1st person
/mOr-a-s/ [m´Oras] ‘you reside’
/mOv-e-s/ [m´Oves] ‘you move’
/sErv-i-s/ [s´Erves] ‘you serve’
/mOr-a-o/ [m´Oro] ‘you reside’
/mOv-e-o/ [m´ovo] ‘you move’
/sErv-i-o/ [s´irvo] ‘you serve’
Like in many (Romance) languages, verbs in Portuguese have a so-called
theme vowel, which behaves in some respects like a suffix, but which at the
same time is determined by the stem: the verb ‘to reside’ has -/a/- as its
theme vowel, ‘to move’ has -/e/-, and ‘to serve’ -/i/-. These theme vowel
surfaces for instance in the second person singular, which has the consonantinitial
suffix /s/, as is illustrated in the lefthand column. However, the first
person singular suffix is -/o/, and this may be a reason why the theme vowel
disappears — otherwise we would again create a hiatus.
But when the theme vowel disappears, something happens to the stem
vowel: it changes from /O/ to [o] in ‘to move’ and from /E/ to [i] in ‘to serve’.
These are changes in vocalic aperture: /O, E, a/ are low vowels ([+low,-high]),
/e,o/ are mid vowels ([-low,-high]) and /i/ is a high vowel ([+high,-low]).
What happens, then, is that the stem vowel takes over the aperture features of
the disappearing theme vowel. In autosegmental terms, we can describe this
as relinking of the Aperture node, rather than the individual relinking of the
features [±high] and [±low].
64 3.4. Feature Geometry
The argument for the Aperture node thus comes from relinking; we will
provide an argument in favour of the Laryngeal node from neutralisation.
Korean has three series of stops, traditionally called voiceless, ‘tensed’ and
aspirated (Rhee, 2002). There is no general agreement as to what exactly are
the phonetic or phonological correlates of these three dimensions, but it is
clear that they have to be described by Laryngeal features. It is also clear that
they can contrast in a position before a vowel:
(105) lenis fortis aspirated
[pal] ‘foot’ [p’allE] ‘laundry’ [phal] ‘arm’
[tal] ‘moon’ [t’al] ‘daughter’ [thal] ‘mask’
[k1n] ‘root’ [k’1n] ‘string’ [kh1n] ‘big’
However, at the end of the syllable, we only find the lenis variants:
(106) lenis fortis aspirated
[cip-to] ‘hous EMPHATIC’ *[cip’] *[ciph]
[mit-to] ‘bottom side EMPHATIC’ *[mit’] *[mith]
[pu@k-to] ‘kitchen EMPHATIC’ *[pu@k’] *[pu@kh]
This looks very similar to a process which we know from languages such as
Dutch, German, Turkish and Catalan and which is usually called final devoicing
(the example is from Dutch, in case anybody did not realize):
(107) a. Beginning of syllable:
voiced voiceless
[dAk] ‘roof’ [tAk] ‘branch’
[bAk] ‘bin’ [pAk] ‘suit’
b. End of syllable:
voiced voiceless
*[hOnd] [hOnt] ‘dog’
*[Eb] [Ep] ‘ebb’
For Dutch — as well as the other languages just mentioned — it may be assumed
that what is going on is that the feature [+voice] gets lost at the end of
the syllable; the remaining structure is then interpreted as voiceless. Korean
shows the same phenomenon, but with one difference: at least two different
features have to be lost — the ones distinguishing tensed and aspirated consonants
from lenis ones. Again, this can be profitably described if we assume
that the relevant rule is something like the following:
(108) Delink the Laryngeal node from a consonant at the end of the syllable.
This rule can even be applicable to the final devoicing languages such as
Dutch; in these languages there is only one Laryngeal feature, so it is hard
to tell a priori whether it is just this feature which is delinked, or the node
dominating it.
3.4. Feature Geometry 65
Root nodes and skeletal points
There is one more organizing node to be discussed: the root node, the node to
which all other organizing nodes, as well as individual features, are eventually
attached. This is the node in (103) which carries the features [±consonantal,
±sonorant].
The fact that the root node carries these features has an important implication
under autosegmental assumptions: we cannot spread either one of those
features independently. Whereas it is possible to spread e.g. [nasal] without
spreading any other part of the tree, spreading of e.g. [+consonantal] will always
result in total assimilation, a famous instance of which is found in the total assimilation
Lesbian and Thessalian dialects of Ancient Greek, where /s/ assimilated completely
to an adjacent sonorant segment (Clements and Hume, 1995):
(109) *gwols¯a > boll¯a ‘council’
*aws¯os > aww¯os ‘dawn’
*esmi > emmi ‘I am’
*naswos > nawwos ‘temple’
(Notice by the way that again we are not dealing with a synchronic phonological
rule in this case, but with a phonological change; which is not necessarily
the same thing.)
What is impossible, according to this model, is a change where a sonorant
would change to a stop with exactly the same place features due to assimila-
tion:
(110) amta > apta (impossible change; and impossible phonological rule).
Another implication of these assumptions, and of the analysis underlying
(109), is that the root node organizes all the features, but is still distinct from
an x-slot; for we see the process happening in (109) as spreading of the root
node with all its features from one x-slot to the next.
This assumption seems necessary also for most of the analyses we presented
last week, where it was equally the case that all the features spread together
from one skeletal point to the next.
At the same time it may be seen as a little unfortunate that we now have
two tiers which organize all the segments. Furthermore, there is an empirical
problem with this particular implementation of segmental structure in autosegmental
phonology. We know that complex segments can be for instance
affricates (sharing place features but differing on continuancy: [
>
ts,
>
pf]), pren- affricates
asalised segments (sharing place features but differing in nasality: [nd, mb]),
or doubly articulated stops (sharing all features except for place). There has
never been found any evidence for complex segments where the two parts differ
on many different dimensions (e.g. *[
>
tN], *[>pa]). This is unexpected, given
the autosegmental model.
As a methodological aside, note that an assumption underlying this criticism
is that every structure which can be generated by the formal model,
also needs to be attested in some of the world’s languages. In principle, it is
of course possible that structures such as *[>pa] do indeed exist, but only in
languages which have not yet been considered in sufficient detail: we simply
66 3.4. Feature Geometry
do not know about them yet. However, it is good practice in phonological
theorizing to assume that structures do not exist until somebody points out
that we do need them in the analysis of some language. If we would not take
this as our guideline, it would be almost impossible to compare theories: a
model which would say that ‘anything goes in natural language’ would beat
everybody else; but it would not be very interesting. In other words, we try to
make our model as restrictive as possible. The model developed so far is notrestrictive
restrictive enough from this point of view; it overgenerates.overgenerates
This problem still awaits a full formal solution at present. Somehow we
have to assume (without an explanation) that one timing slot cannot host more
than one root node. Therefore, we have to find a different representation for
complex segments.
From what we have seen so far, we can already conclude the following:
(111) Complex segments bear more than one feature (value) of a specific
type.
For instance, [
>
ts] is exactly like [t] and [s], except for one point: whereas
[t] is [+continuant] and [s] is [-continuant], [
>
ts] is both [+continuant] and [continuant].
Heavily simplifying our feature trees, we can draw the three
segments as follows:
(112) [t] [s] [
>
ts]
+ consonantal
- sonorant
[cor]
r
rrr
rrrPlace
[-cont]
+ consonantal
- sonorant
[cor]
r
rrr
rrrPlace
[+cont]
+ consonantal
- sonorant
[cor]
d
d
d
€€€€€€€€
Place
[-cont] [+cont]
In the structure for the affricate, two feature values (on the same tier) are now
linked to the same segment (in this particular case, to the same root node).
This parallels two tones being linked to one segment. A similar picture can be
drawn for prenasalised segments.1 Multiply articulated segments might be a
little bit different; the following represents [
>
kp] (again, abstracting away from
certain complexities):
1It should be noted that in the most recent literature, alternative analyses have become available
for both affricates and prenasalised segments which do not use this particular type of representation;cf.
van Oostendorp and van de Weijer (2005).
3.5. Exercises 67
(113)
+ consonantal
- sonorant
[labial] [velar]
r
rrr
rrrPlace
[-cont]
dd  
This representation is different because the two Place features are probably
not on the same tier; they are linked to the same node, but they still represent
different dimensions. Because they are not on the same tier, they are also
not temporally ordered with respect to each other; which gives the (correct)
prediction that they are realised at the same time.
3.5 Exercises
1. Give the tonal representation of ’they looked’ and ’we sent them’ in
Kikuyu, taking (43) as your model.
2. In Mende, vowels in monosyllabic words can have one of five tones:
high, low, rising, falling, or first rising and then falling. In words with
two syllables, the first syllable is always high or low, and the second syllable
is high, low, or falling; a falling tone only occurs after a low toned
first syllable. Finally in words of three syllables, all syllables have only a
high or a low tone. How can you explain this pattern autosegmentally?
(You can make up your own examples if you want to illustrate.)
3. Look at the following forms in Chizigula, in which high toned syllables
have an accent and toneless syllables do not (ku is a prefix):
(114) kudamanj ‘to do’, kudamanjiza ‘to do for’, , kudamanjizana ‘to do
for each other’, kulombéza ‘to ask’, kulombezezána ‘to ask for each
other’, kulombezéza ‘to ask for’
4. In a secret language in Thai, words are changed a little bit so that outsiders
cannot understand them: khluáì hòóm ‘banana’ is pronounced as
khlóòm huàí and té`n ram ‘dance’ as tá `m ren (acute accent denotes a high
tone, grave accent a low tone, no accent on a vowel is a mid tone). How
can this be constructed as an argument for autosegmental phonology?
5. In some varieties of Latin American Spanish something remarkable happened
to the [s] at the end of a syllable. Give a precise description of what has
happened (the circle under a vowel indicates that the vowel is voiceless):
(115) European Spanish L.A. Spanish
mismo mii
˚
mo of mim
˚
mo ‘same’
fosforo foo
˚
foro ‘match’
6. In an innovative variety of the dialect of Shanghai we find an interesting
tonal pattern. Consider first the following underlying representations
68 3.5. Exercises
for several morphemes (tonal specifications are added in parentheses;
M denotes a mid tone, a third tonal level in some languages):
• çi ‘fresh’ (HL); wa ‘yellow’ (LH); du ‘big’ (LH) N ‘dish’ (LH); ço
‘small’ (MH)
Now consider the tones of following phrases:
• ço+N → ço (M) N (H) ‘small fish’
• çi+N → çi (H) N (L) ‘fresh fish’
• wa+N → çi (L) N (H) ‘yellow fish’
• ço+wa+N → ço (M) wa (H) N (L) ‘small yellow fish’
• çi+wa+N → çi (H) wa (L) N (L) ‘fresh fish’
Describe what is going on here in autosegmental terms, and give the
tonal pattern for ‘big yellow fish’.
7. Consider the following forms in Tiberian Hebrew.
(116) a. seefer ‘book’
b. geSem ‘rain’
c. Pi:S ‘man’
d. ha:r ‘mountain’
Now look at the following forms, to which we added the definite determiner
ha:
(117) a. has:eefer ‘book’
b. hag:eSem ‘rain’
c. ha:Pi:S ‘man’
d. ha:ha:r ‘mountain’
Assume laryngeals and pharuungeals cannot geminate. Give a description
of what happens.
Discussion and further reading
The Kikuyu data are wonderfull described and analysed in Goldsmith (1990);so
are the Luganda data later on in this chapter.
The Margi data are from (Hoffmann, 1963; Williams, 1976; Kenstowicz,
1994)
Since the work of Odden (1986), it is no longer assumed that the OCP is
a universal principle, but it can still be seen at work as a tendency in some
languages.
The Shona data are discussed in Odden (1980); Myers (1987); Kenstowicz
(1994)
The Bergen Dutch data are from van Oostendorp (1998)
The interaction between Rendaku and Lyman’s Law is the topic of extensive
literature. See for example Itô and Mester (2003).
The discussion of Finnish follows Keyser and Kiparsky (1984); Gussman
(2002). Compensatory lengthening in Old English was described in (Ewen
3.5. Exercises 69
and van der Hulst, 2001; Gussman, 2002); the Turkish data have also been
amply discussed, for instance in Sezer (1986); Goldsmith (1990); Kenstowicz
(1994); Gussman (2002). Italian Raddioppiamento Sintattico has been described
by (Nespor and Vogel, 1986).
The Chuckchi data are in Odden (1987); Clements and Hume (1995).
Chapter 5
Syllables
5.1 Evidence for syllable structure
Where in the world can we find syllables? The best place to look might be in
poetry. In many poetic traditions, every line in a poem has a fixed number
of syllables. Very famous in this respect is the Southern Slavic epic tradition:
many poems in Serbian and Croatian poetry consist of lines of exactly ten
syllables each (decasyllables; deka is the Greek word for ten): decasyllables
(123) Što se bili u gori zelenoj?
What itself be-white in mountain green
al su snizi, al su labutovi?
or is snow, or is swans
da su snizi, ve´c bi okopnili,
if were snow, already it melted
labutovi ve´c bi poletili
swans already fly-away
“What is white on the green mountain? Snow or swans? If it were
snow, it would already have melted away; and swans would already
have flown.”
These are the first lines of ‘The Mourning Song of the Noble Wife of the Asan
Aga’ (Asanaginica), a folk ballad from 1646-49 and a region which currently
belongs to Croatia. Although decasyllables are still occasionally written by
modern authors in the region, their origin lies in a medieval oral tradition;
they were composed by poets who did not necessarily write their poems. This,
and the fact that there are other traditions around the world which ‘count
syllables’ in poetry, shows that the syllable is an intuitive concept for human
beings, and does not necessarily depend on their literacy.
Similarly, we can observe that people tend to find it easier to count the
number of syllables in the word syllabicity than the number of segments in
sounds. There is evidence that children know how to syllabify words before
they can divide them into segments and before they can write.
At the same time, it turns out rather difficult to give a precise definition of
the syllable that covers all the intuitions. In addition, syllables so far have n
not been straightforwardly detectable in the acoustic signal, possibly even less
so than individual segments. Even so, the phonological evidence in favour
77
78 5.1. Evidence for syllable structure
of the syllable is manifold. Many different observations can be phrased in a
much more elegant way if we accept the notion of a syllable. Furthermore, the
syllable comes very close to being universal: most, if not all, languages give
evidence for it.
We will now review the most important types of evidence that have been
put forward for recognizing the syllable as a formal unit of phonological analysis.
We will see that this evidence comes from many different angles, and
converges on two properties that syllables have:
• They are constituents, i.e. groups of smaller units (segments, in this case)constituents
which behave as a unit in some ways.
• They are headed, i.e. one of the smaller units is more prominent than theheaded
others and determines the properties of the constituent as a whole. We
call this prominent unit the head; the other units are called the dependents.head
dependents In syllable structure, the head is typically a vowel.
We will discuss this terminology in more detail in section (5.2). For now,
we should note that headed constituents are also known in other branches
of grammatical description, such as syntax.
In order to represent these two properties, phonologists draw diagrams
like the following (for the English word drop):
(124)
d r p
σ
The Greek letter σ (sigma) denotes the syllable node, and the lines — which
are sometimes thought of as association lines — the fact that segments belongSee Section ??
to the constituent. The vertical line between σ and O is special and denotes
the relationship between the head and the constituent. A more refined view
of this structure will be given in section 5.2.
Stress
The most important argument in favour of syllables is consistent asymmetries
between vowels (V) and consonants (C) in human language. There are many
phenomena where the vowels can ‘see’ each other at a distance across consonants,
whereas there are very few cases where consonants could see each other
across vowels.
One such phenomenon is word stress . Many (albeit not all) languages haveword stress
See Chapter 7 one vowel in the word standing out for being more prominent than the others:
it has a higher pitch and/or it is longer and/or it is louder. A simple example
is Southern Peruvian Quechua. Stress in this language is on the penultimate
vowel (with some exceptions):
(125) a. "wassi “house”
b. was"sipi “in (the) house”
c. wassi"kuna “houses”
d. wassiku"napi “in (the) houses”
5.1. Evidence for syllable structure 79
In order to determine where the stress falls in these words, one needs to
count syllables. The absolute number of segments is irrelevant: in (125a) the
stressed vowel is the fourth segment counted from the end, whereas in the
other words, it is the third segment. In order to determine the exact position
of the stress, we thus should not count segments, but we should count vowels
from the end and just disregard the consonants.
Importantly, there are no similar known processes which apply to the second
consonant from the end and disregard all the vowels. We can account for this
in the following way: we assume that stress is not so much a property of
vowels as it is of syllables. We thus assign stress to the second syllable from
the right. However, we also assume that this property is first and foremost
expressed on the head of the syllable, which is typically the vowel. Since consonants
are not typically heads, but dependents of syllables, it is not possible
to apply a similar technique to them to get a system to count consonants and
not vowels.
An even stronger argument comes from those languages in which not just
the distance from the edge plays a role in terms of stress but also the structure
of the individual syllables. Many dialects of Arabic (Afroasiatic) are of this
type. Let us briefly consider Palestinian Arabic as an example. In this language,
stress falls preferably on the rightmost heavy syllable (i.e. a syllable that heavy syllable
is closed by a consonant or that contains a long vowel; as opposed to a light
syllable, which ends in a short vowel). If there is no such syllable, then stress light syllable
falls on the first syllable of the word (the final syllable does not count):
(126) a. Words with a heavy syllable:
i. [ba"Su:fiS] ‘I don’t see’ ([Su:] has a long vowel, and hence is
heavy)
ii. [ka"tabti] ‘you FEM SG wrote’ ([tab] is closed and hence heavy)
b. Words without a heavy syllable
i. ["darabu] ‘they hit’
ii. ["zalama] ‘man’
It would be difficult to describe this system without referring to syllables and
their structure, but just considering vowels and consonants, as you can try out
for yourself. Furthermore, there are many languages in the world that make
this distinction between heavy and light syllables, as we will see in Chapter
7. This therefore gives us a piece of evidence that is not direct (we establish
the existence of the syllable not by directly observing that it is there) but still
very forceful, in particular because it turns out that we need the concept to
understand many other, at first sight unrelated phenomena.
Reduplication
Another phenomenon where we find evidence for the syllable is in reduplication,
a morphological process in which a portion of a stem is copied to express reduplication
a certain meaning. In almost all known cases, the copied part corresponds to
a phonological constituent; in several languages this is a syllable.
80 5.1. Evidence for syllable structure
Yaqui (Uto-Aztecan), for instance, reduplicates the first syllable of the word
(there is a lot of discussion in the literature as to what the semantics of this reduplication
really is; it could be some intenser reading of the verb stem):
(127) a. vu.sa vu.vu.sa ‘awaken’
b. chi.ke chi.chi.ke ‘comb one’s hair’
c. he.wi.te he.he.wi.te ‘agree’
d. ko.a.rek ko.ko.a.rek ‘wear a skirt’
e. vam.se vam.vam.se ‘hurry’
f. chep.ta chep.chep.ta ‘jump over’
g. chuk.ta chuk.chuk.ta ‘cut with a knife or saw’
h. bwal.ko.te bwal.bwal.ko.te ‘soften, smooth’
The dot between two segments denotes the syllable boundary. The examples
in (127a-d) show that if the stem starts with an open syllable, a sequence of
a consonant followed by a vowel (‘CV’) is copied, whereas (127e-h) show examples
of a stem starting with a closed syllable, which is also faithfully copied
as a consonant-vowel-consonant sequence (‘CVC’). Describing such a morphological
process requires referring to the syllable, and in this sense, then,
reduplication provides another piece of indirect evidence for its existence.
Language games
Language games also sometimes refer to the concept of the syllable. A case
in point is Vesre, a secret language originating in the underworld of Argentina
(around Buenos Aires) and Uruguay, and sometimes used in tango lyrics (for
instance in the famous tango song “¿Qué querés con ese loro?”). Vesre takes a
Spanish word and puts the syllables in the opposite order:
(128) Spanish Vesre gloss
a. pizza zapi ‘pizza’
b. caballo llobaca ‘horse’
c. réves vesre ‘inverse’
Other languages have similar procedures in language games (e.g. French Verlan
and Tagalog Binaliktad).
Another common type of language game is one in which a syllable is added
before or after every syllable of the original word, as in the following
example from Hausa (Chadic) in which da is prefixed to every syllable of the
word:
(129) tsíntsíyáa → dá-tsín-dà-tsìi-dà-yáa ‘broom’
(Notice that interesting things also happen to the tones, but we will ignore this
here.) In order to be able to play this game, one needs to be able to divide the
word into syllables in the first place; further, it probably is no coincidence that
the prefixed sequence is also a syllable. (At this point we obviously do not yet
have a precise definition of where the syllable boundaries are; we will return
to this later in this chapter. )See Section 5.2
5.1. Evidence for syllable structure 81
The crucial aspect of such language games is that they play a role in a context
that is playful and oral. Just like in the case of the South Slavic poetry at
the beginning of this chapter, this demonstrates that ordinary language users
do not find it difficult and do not need a high level of literacy for ordinary
language users to acquire such a system, or to understand it. This is not true
for sequences that form phonological non-constituents. There are apparently
no language games in which one would take random segment sequences and
revert them or prefix them by other material.
We should take approach from language games with some caution, however.
There is obviously always something ‘artificial’ about them: the rules
of these games have been made up with the purpose of playfulness or deception
and there is no ‘real’ morphological process which shifts around syllables
in this way. Nevertheless, it is suggestive that human creativity in inventing
succesful language games seem to have boundaries, and these boundaries are
among other things defined by constituency.
Psycholinguistic evidence
Another type of evidence concerns several type of data from psycholinguistic
research. One type concerns speech errors. For many decades, psycholinguists
have been studying the mistakes which people occasionally make from
a phonological point of view. One of those is so-called blending, in which two blending
words get inadvertently mixed. An example is yout, which blends the beginning
of yell and the end shout.
We can call the point where one word turns into the other (in this case,
between y and out) the break. Where do these breaks occur exactly? Already in break
1972, the psychologist D.G. McKay studied this question, based on a corpus
of mistakes made by university professors in Vienna at the end of the 19th
Century. McKay did something simple: he looked whether breaks occurred
within syllables (such as in the case of yout) or at the boundary between two
syllables (e.g. in war.der from warmer and col.der). Here are his results:
(130) Breaks within syllables Breaks between syllables
Data% 40 60
Chance % 64 36
The ‘chance’ level means that a clear majority of 64% of all boundaries between
segments are within syllables (which means that syllables span more than
two segments on average). Still, breaks are much more likely to occur at the
boundaries between two syllables than between any two segments within a
syllable. This shows then that somehow syllables play an important role in
language planning, and this has been confirmed by many studies since, also
for speakers who were not Viennese university professors.
There is also evidence that the syllable plays a role in the perception of
speech. In a famous experiment, a group of French speakers listened to long
lists of individual words. They were asked to find a specific sequence in these
words, for instance pa or pal, and press a button if they heard this sequence.
The trick was that the word list would contain items that started with the
same three segments (in our example p-a-l), but which were syllabified in a
different way; in our case these would be palace ‘palace’ which has an initial
82 5.1. Evidence for syllable structure
syllable pa, and palmier ‘palm’, which has an initial syllable pal. (Other pairs
were carotte-carton, tarif-tartine, garage-gardien, balance-balcon.)
It turned out that people were much faster in detecting pa in palace than
pal, and much faster detecting pal in palmier than pa, as the following graph
illustrates (the horizontal axis denotes the type of sequence the subjects were
looking for, the vertical axis how much time it took them to press the button):
(131)
These data show that the French speakers divided the word into syllables
while they were listening to them, and found it more difficult to hear sequences
that were smaller or bigger than a syllable, although they managed
to do that as well.
Later experiments showed that the same effect does not hold for all languages.
In particular, English speakers did not show the effect of syllable
structure at all, and found it just as easy to find pa as pal in any word which
contained those sequences. The reason may be that syllable boundaries in
English are relatively blurry as compared to e.g. German and French: it is difficult
to find agreement about where exactly the boundary is in words like
bitter. It might therefore be less helpful to people listening to English to pay
close attention to syllables than it is for speakers of other languages.
Syllabaries
The final type of evidence for syllables we want to discuss are writing systems.
You are probably familiar with alphabetic systems, in which every letteralphabetic systems
corresponds (roughly) to a segment — the Roman alphabet is such a system,
and this book is written in it, although the correspondence between sound and
letter in the case of English is very rough — and logographic systems, in whichSee Section 10.1
logographic systems every symbol stands for a meaning unit, such as a morpheme: Chinese is often
cited as an example of the latter (although matters are more complicated,
especially in Modern Chinese).
However, another quite widespread system is syllabic writing, in which thesyllabic writing
syllable is the smallest unit of writing. One example is (or was) the Akkadian
cuneiform, a system that was invented by the Sumerians around 4000 BCE andcuneiform
written on clay tablets. The system had several logograms as well, but at its
core it was syllabic. Here are some examples:
5.2. The internal structure of the syllable 83
(132)
Other examples of languages which have, or used to have, a writing system
based on syllables are Ge’ez (Afroasiatic), Mayan (Mayan), Mycenaean Greek
(Indo-European) and Cherokee (Iroquoian). Many alphabetic systems seem to
have developed out of a syllabic system at some point: when humans invented
the art of writing sounds, the first building blocks that apparently came
to mind were syllables. In many cases, these building blocks were only very
simple syllables, consisting of a consonant followed by a vowel. In the following
section, in which we examine the internal structure of the syllable, we will
discover why.
5.2 The internal structure of the syllable
Even though we are thus unable to directly observe the phonological syllable,
there are plenty of reasons to assume that it has an organizing role in the
sound systems of many, if not all, languages of the world. There is also evidence
that the structure in (124), of one vocalic head and a few consonants
preceding and following the vowel, is too simplistic.
You may have already observed in the preceding section that we sometimes
took decisions that require more motivation than we actually gave. For
instance, in the French perception experiment we assumed that pal is a syllable
in palmier, but not in palace. Similarly, in the Yaqui example in (127), we were
tacitly assuming that some consonants belong to the same syllable as the preceding
vowel, whereas others do not. This may have seemed quite plausible,
because it is the same as one would do in English, but how do we motivate
these divisions further?
In this section, we will study the internal structure of the syllable in somewhat
more detail. This structure looks as follows (again for the English word
drop):
(133)
σ
$$$$
O R
d
d
e
e
e
e
e
N C
d r O p
The capital letters in this graph denote an onset (O), a rhyme (R), a nucleus (N) onset
rhyme
nucleus
and a coda (C). These are all headed constituents: in drop, d is the head of the
coda
onset, p the coda and O the head of the nucleus, the rhyme and the syllable.
Onset
There is one property of headedness in constituents that we have not mentioned
yet: heads are obligatory, whereas dependents are not. This means that
84 5.2. The internal structure of the syllable
every syllable in every language always will be expected to have a nucleus
and a rhyme, whereas not every syllable necessarily has an onset.
An example is Indonesian (Austronesian). Here is how you count from
one to ten in this language:
(134) satu, dua, tiga, empat, lima, enma, tujoh, delapan, sembilan, sepuluh
Most numerals in this language start with a consonant, but the words for four
and six start with a vowel. Since we assume that all words are divided into syllables,
this means that at least these words start with a syllable with a rhyme,
but no onset. Similarly, the word for ‘two’ has two syllables, of which the
second does not have an onset (a configuration in which two vowels occur
next to each other without an intervening consonant is called hiatus).hiatus
Obviously English is like Indonesian in that it has syllables without an onset
— one only has to look at a few words like eight and eleven to get convinced
of that. Many other languages also allow onsetless syllables.
The requirement that syllables have onsets is on the other hand very strong
in certain languages. An example of this is Axininca (Maipuran). Whenever
the concatenation of morphemes would result in an onsetless syllable, an
epenthetic [t] is inserted in this language, as the following examples demon-
strate:
(135) a. /no-N-koma-i/ [noNkomati] ‘he will paddle’
b. /no-N-koma-aa-i/ [noNkomataati] ‘he will paddle again’
c. /no-N-koma-ako-i/ [noNkomatakoti] ‘he will paddle for’
d. /no-N-koma-ako-aa-i/ [noNkomatakotaati] ‘he will paddle for it again’
(135) shows several affixes which are added to the base form koma. Whenever
such a suffix starts with a vowel and the preceding stem or suffix ends with
a vowel, a consonant t is inserted. This t provides the following syllable with
an onset. (Such insertion of phonological segments is called epenthesis.)epenthesis
There are thus languages in which onsets seem obligatory, like Axininca,
and languages in which they are optional, like Indonesian. There do not seem
to be languages in which onsets are consistently disallowed. Since rhymes
and nuclei are also allowed (because they are heads), we have the following
universal:
(136) All languages allow CV syllables.
Syllables with one consonant followed by one vowel are also called core syllables.
There are no other syllable types which can make this claim to uni-core syllables
versality. Core syllables are typically the first that are acquired by children,
even in languages that allow other syllable types. Indeed, it has even been
demonstrated that at a very early stage of acquisition, children replace even
the smaller V syllables with CV structures. The following examples are from
Dutch (Indo-European):
(137) adult form child’s output
/oto/ [toto] ‘car’
/api/ [tapi] ‘monkey’
5.2. The internal structure of the syllable 85
In these examples, the first syllable of the word has no onset in the adult language.
It is filled with a t in the child language, just like in Axininca. (The fact
that it is a t in both cases is a coincidence.)
Certain languages also allow for more complex onsets. In many languages, complex onsets
these will consist of two consonants of which the first is an obstruent and the
second a sonorant, such as in Sanskrit (Indo-Aryan):
(138) prath ‘spread’, svaé ‘embrace’, mjaks ‘glitter’, mna: ‘note’, grabh ‘seize’
Sanskrit v is probably a glide, and thus a sonorant, as it can alternate with the
vowel u.
The fact that two consonants can occur together at the beginning of a word
does not in itself provide sufficient indication that they form a constituent.
However, just as there is strong phonological evidence for the syllable , there See Section 5.1
are also good arguments for taking this position.
As a matter of fact, one example briefly discussed above already sheds
some light on the issue: the blended form yout for shout and yell. We argued
that most blends occur at the boundary between two syllables, but there also See p. 81
are examples where blending happens within a syllable. In many cases, like
in yout, it then occurs at the boundary between onset and rhyme.
Blending is not always a speech error in English. It can sometimes also be
a way in which speakers actively create new words out of two existing ones:
(139) smog (smoke and fog), brunch (breakfast and lunch), motel (motor and hotel),
infotainment (information and entertainment)
Interestingly, also in these cases, the blending occurs exactly at the boundary
between two constituents. For instance, in brunch, it occurs exactly between
the onset and the rhyme; there are no blends of the shape breanch.
As to the headedness of onsets, the most popular type of evidence comes
from reduction processes: if, for some reason, only one of the two consonants
is pronounced, it is typically the first one, the head. (Recall that headedness
involves the claim that the head is necessary but the dependent is less so.)
One piece of evidence comes from language acquisition. In a study on the
acquisition of children with a cochlear implant, i.e. a permanent hearing aid cochlear implant
that is implanted close to the ear, it was shown that these children tend to
reduce clusters at some stage of their acquisition:
(140) adult word child form
‘clocky’ [khakki]
‘frog’ [fO˚g]
‘brush’ [bUtS]
There was a handful of cases in which it was the second consonant that survived
[lhak] for ‘clock’, but the cases where it was the first segment were an
overwhelming majority. Very similar results were obtained with normally
hearing children in many experiments as well.
Furthermore, adult languages also sometimes show signs of the same phenomenon.
An example is reduplication in Sanskrit. Recall that the morphological
process of reduplication can copy a syllable. In Sanskrit, the copied See Section 5.1 on page 79
86 5.2. The internal structure of the syllable
syllable is simplified: one of the simplifications is in the onset, which consists
of only one consonant. (Other simplications involve the nature of the
vowel, which is of no concern to us now.) This is always the head. In the
examples, the reduplication adds perfective meaning to the stem; I added a
dash between reduplicant and base for clarification:
(141) base reduplicant
prath ‘spread’ pa-pratha
svaé ‘embrace’ sa-svaéa
mjaks ‘glitter’ mi-mjakùa
mna: ‘note’ ma-mnur
The fact that in processes such as this, it is uniformly the first consonant that
is preserved, can be described nicely by claiming that it is a head — since that
is the only obligatory segment by definition.
Rhyme
As in the case of the onset, the evidence for the rhyme comes from several different
sources; we will discuss only a few. One we have actually already seen
above, in our discussion of Arabic stress : languages distinguish between lightSee p. (126)
and heavy syllables in the assignment of stress. This distinction can (almost)
always be described as a difference between a branching rhyme (with more thanbranching rhyme
one segment, which is heavy and attracts stress) and a non-branching rhyme
(with one segment, which is light). In the case of Arabic, tab in katabti ‘you
wrote’ has stress because it is heavy, but ra in darabu is light and therefore
unstressed.
Another piece of evidence comes from a phenomenon called compensatory
lengthening. An instance of this can be found in the history of English. Com-compensatory lengthening
pare the following Old English words with their German cognates:
(142) Old English German
go:s Gans
o:þer Ander
so:fte sanft
fi:f fünf
u:s uns
The Old English words all have a long vowel, where the German forms have a
short vowel followed by a nasal. There is reason to assume that the latter language
is more faithful to the state of affairs in Proto-Germanic, predating all
of the Germanic languages, and that English is the language that has changed.
How can we describe what has happened? The assumption of the rhyme
gives us a nice tool to formulate this description: the nasal was lost, i.e. it was
delinked from its position on the skeletal tier in cases where it was followed
by a fricative. After this, the empty position in the rhyme was filled by the
preceding vowel. The lengthening is compensatory in the sense that the vowel
length compensates for the lost consonant (we disregard the position of the s
for now 5.4 ):See Section 5.4
5.2. The internal structure of the syllable 87
(143) a. Input:
R
dd
x x x
u n s
b. Output:
R
dd
x x x
  
u n s
In this example, the skeletal tier (which we have disregarded so far) has been
added explicitly in order to see more clearly what is going on. The crucial step
in the argument is that compensatory lengthening only involves consonants
that get lost in the rhyme. There are basically no known cases in which compensatory
lengthening involves an onset consonant that gets lost and results
in lengthening of the vowel. CL is thus restricted to the domain of the rhyme.
The change we are witnessing in (143) is a priori different from the phonological
processes we have seen before. We are dealing here with a diachronic
change, the ‘input’ in (143) represents some stage in the history of English and
the ‘output’ some other stage; and there is as yet no specific reason to assume
that any speaker ever had both of them in his head. Still, diachronic changes
like this, too, may give us some insight in the mental representations of speakers.
After all, there must be a reason why it is only deletion of consonants in
‘heavy’ syllables that results in a long vowel.
On this assumption, then, we see autosegmental phonology at work. If a See Chapter 3
long u: would be nothing but a sequence of two short u’s, we would not really
understand what was going on: we would have to say that the nasal would
have turned into a full copy of the preceding vowel, which would make the
representation of this change rather complicated.
Compensatory lengthening is found in many of the world’s languages. A
well-known case can be found in Turkish (Turkic). In this language, there
is good reason to assume that the process is synchronic, because, depending
on sociolinguistic and pragmatic factors, speakers display optional deletion
of a consonant (to be more precise, any of /v, j, h/. When deletion occurs,
compensatory lengthening follows suit automatically:
(144) a. kahya ‘steward’ [kahja]-[ka:ja]
b. eylül ‘September’ [ejlyl]-[e:lyl]
c. sevmek ‘love’ [sevmek]-[se:mek]
Alternatively, one could assume that words are just stored in two different
ways in the Turkish lexicon and that one is historically derived from the other.
However, such an analysis would disregard the fact that the relation between
not having a consonant and having a long vowel is systematic.
88 5.2. The internal structure of the syllable
Nucleus
Within the rhyme, we sometimes distinguish a nucleus and a coda. The latter
(like the onset) is the exclusive domain of consonants, whereas the former
is assumed to contain vocalic material, for instance diphthongs, sequences ofdiphthongs
two vowels that occur in the same syllable (with a technical term: that are
tautosyllabic).tautosyllabic
A well-known example of a complex nucleus comes from French (IndoEuropean).
This language has words such as the following:
(145) trois ‘three’ [trwa], croix ‘cross’ [krwa], pluie ‘rain’ [pl¨wi], truite ‘trout’
[tr¨wit]
These words all start with three consonants: a sequence that we can recognize
as a regular onset also in English, followed by a glide [w]. Since there is no
place for this [w] in the onset, it has to be postulated somewhere else, and the
nucleus seems an obvious place to do so.
Interestingly, there is independent evidence that indeed [w] can function
in a nucleus in French. In order to see this, we have to look briefly at the
definite determiner (the word meaning ‘the’). This small word is sensitive todefinite determiner
whether the following word starts with an onset, as the following examples
demonstrate:
(146) camerade ‘friend’ [kamrad] ami ‘friend’ [ami]
le camerade ‘the friend’ [l@ kamrad] l’ami ‘the friend’ [lami]
les camerades ‘the friends’ [le kamrad] les amis ‘the friends’ [lez ami]
If the following word starts with a consonant, the singular determiner has a
schwa and the plural determiner has [e]. If the following word starts with
a vowel, on the other hand, the singular determiner consists of a consonant
only, whereas the plural determiner has a [z] after the [e]. (Notice that the
allomorphy is reflected in French orthography for the singular, but not for the
plural.)
We can already understand what triggers this allomorphy: French wants
to create an onset in the first syllable of ami. In the singular we do this by
dropping the schwa at the end of [l@], whereas in the plural we do it by pronouncing
an extra [z] after [le].
All of this means that the definite determiner gives us a test to see whether
a noun starts with an onset or not. We can now apply this test to words starting
with our glide [w]:
(147) watt ‘watt’ [wat] oiseau ‘bird’ [wazo]
le watt ‘the watt’ [l@ wat] l’oiseau ‘the bird’ [lwazo]
les watts ‘the watts’ [le wat] les oiseaux ‘the birds’ [lez wazo]
It turns out that there are two types of words in French. Some words, such as
watt, behave as if they start with an onset, and we should therefore conclude
that the [w] is in this onset. Other words, such as oiseau behave like ami. In
these words, then, the [w] can only be part of the nucleus.
5.2. The internal structure of the syllable 89
It is important that allomorph selection is consistent for the singular and
the plural of the definite determiner. There are no words which take e.g. [l@] in
the singular and [le] in the plural. This makes it unlikely that speakers just remember
what the singular and plural determiners are for every word. Rather,
speakers have a system in their minds in which one form of the determiner
goes with onset-initial words and another form with non-onset initial words.
It is this system that can be neatly described under the assumption of subsyllabic
constituency.
Coda
Whereas every syllable has a nucleus and a rhyme, and onsets are at least allowed
in all languages, and often preferred, the coda position is restricted, and
dispreferred in many different languages. As a matter of fact, there is an array
of languages which do not have coda consonants at all. Examples of these are
Fijian (Malayo-Polynesian), Mazateco (Mesoamerican) and Cayuvava (isolate).
As a matter of fact, the following implicational universal seems to hold:
(148) If a language has closed syllables, then it also has open syllables.
In other words, all languages have open syllables (syllables without a coda),
but only a subset has closed syllables.
As we mentioned, Boumaa Fijian is an example of a language without
closed syllables. In order to repair potential violations of this generalisation,
the language employs vowel epenthesis, the insertion of a vowel. We can see this vowel epenthesis
at work in loanword adaptation. If a word with a closed syllable is borrowed
(from English), a vowel is epenthesized to make the word more well-formed:
(149) Vowel epenthesis in Boumaa Fijian
a. kaloko ‘clock’
b. aapolo ‘apple’
c. tSone ‘John’
As you can see, various other changes are also performed on these words. For
instance, in the first word, the complex onset [kl] is broken up by a vowel,
because Fijian does not allow complex onsets either. Similarly, Boumaa Fijian
disprefers /dZ/, and the first consonant of John has therefore turned into the
affricate [tS]. Another thing you may observe that there are several potential
vowels which can be epenthesized. This will not concern us here, either. The
important fact is that final ‘coda’ consonants are not allowed, and vowels are
epenthesized to ensure this. In this way, the consonant can be saved by being
pronounced as an onset. Here are the syllable structures for the two lan-
guages:
90 5.3. Sonority
(150) English → Fijian
N C
RO
σ
dZ 6 n
N N
R RO O
σ σ
tS O n e
The arrow in this example indicates that the English form is in some way
‘underlying’ to the Fijian form. At some point, speakers of Fijian must have
adapted the English word to the phonological system of their language, which
does not have codas.
Even in languages that do allow coda’s, the coda position is often quite
restricted. For instance, Japanese (Japanese-Ryukyuan) only allows coda consonants
if they share their place of articulation with the immediately followingSee Section x
consonant. We thus find words such as those in (225a), whereas the forms in
(225b) are not allowed.
(151) a. kap.pa ‘a legendary being’, kit.te ‘stamp’, gak.koo ‘school’, tom.bo
‘dragonfly’, non.do ‘tranquil’, kaN.gae ‘thought’
b. *kap.ta, *tog.ba, *pa.kap, etc.
This is not exclusive for Japanese; we also find it in an unrelated language
such as Ponapean (Micronesian). In this language, we can see that this restriction
takes a phonological effect: it causes vowel epenthesis, as the following
examples demonstrate:
(152) /ak-dei/ a.ke.dei *ak.dei ‘a throwing contest’
/kitik-men/ ki.ti.ki.men *ki.tik.men ‘rat INDEF’
/naNkep/ *na.Ni.kep naN.kep ‘inlet’
Another way in which the restriction on codas can be satisfied is by deletion
of the offending consonant. Also this is attested in some of the world’s languages,
e.g. in Jola-Fonyi (isolate):
(153) /let-ku-jaw/ le.ku.jaw *let.ku.jaw ‘they won’t go’
/jaw-bu-Nar/ ja.bu.Nar *jaw.bu.Nar ‘voyager’
(We leave it as an open question why the final consonant of the word does not
need to be deleted; word-final consonants tend to show a slightly different
behaviour in many languages of the world.)
5.3 Sonority
The study of the way in which consonants and vowels are arranged in a word
is called phonotactics. Obviously, dividing the word into syllables, and sub-phonotactics
dividing syllables in further subconstituents sets one step in the direction of a
phonotactic theory.
5.3. Sonority 91
Yet, so far we have been mostly implicit about another necessary step: we
need to determine which segments can go into which positions in the onset,
nucleus and rhyme. We have briefly and informally mentioned several
such restrictions, e.g. that nuclear positions are (typically) vowels, and that
languages allow only a small subset of consonants to appear in coda position.
Further, we have observed that a typical complex onset consists of an
obstruent as the head and a sonorant as the dependent.
In order to make more sense of these observations, phonologists often invoke
the notion of the sonority scale, which looks roughly as follows (there sonority scale
are much more refined versions of the scale, taking into account many more
categories):
(154) Sonority scale
obstruents < nasals < liquids < glides < vowels
1 2 3 4 5
The notion of sonority was introduced already in 1881, by the German linguist
Eduard Sievers (1850-1932), but there is no absolute consensus on the
precise phonological or phonetic definition of sonority; as a matter of fact,
many different definitions are used. For instance, it has been equated to the
openness of the vocal tract and to amplitude (relative loudness). We will ignore
the issue here; intuitively, the notion seems to correspond to ‘similarity
to a vowel’. Obstruents are the absolute anti-vowels. They have a complete
or almost complete constriction, they can be voiceless, etc. Nasals are already
more vowel-like, for instance because they are inherently voiced, although,
like plosives, they also involve complete closure of the oral cavity. The constriction
of liquids is less constricted, and glides are obviously the consonants
that come closest to the vowels.
If we use the numbers in (154) and transfer them into columns of asterisks
(so obstruents get a column of height 1, nasals of height 2, etc.), we can
represent the syllable structure of the English word trim as follows:
(155) *
*
* *
* * *
* * * *
t r I m
1 3 5 2
This mountain-like structure is typical of human language syllables. In particular,
the vowel is always the highest element (not surprising, given our
informal definition; the nucleus is therefore also sometimes called the peak). peak
Furthermore, the segments before the peak (thus, those in the onset) gradually
rise in sonority, whereas those following the peak, fall.
There are several points where languages may differ. One is in the required
steepness of the rise before and after the peak. Generally, languages prefer to
have a rather steep rise in the onset. For instance, in English, [kr] (creek) and
[kl] are fine clusters, rising from the 1 of [k] to the 3 of [l], but [kn] is not, and
neither is any other cluster of an obstruent and a following nasal (disregarding
92 5.3. Sonority
sn and sn clusters, to which we will return ). The reason for this is that theSee Section 5.4
dispersion — the difference in sonority — between an obstruent and a nasal isdispersion
not large enough. Languages can differ on this point: in German, the word
Knie ‘knee’ is still pronounced with the initial [kn] cluster, which was lost in
English. We can thus say that the minimal dispersion in the English onset is
3 − 1 = 2, whereas in German it can be 3 − 2 = 1. (German does allow onsets
with greater steepness such as [kr] and [kl] as well.)
The fall after the peak, on the other hand, tends to be less steep. In many
languages, obstruents are not allowed in the coda although nasals and liquids
are. We will return to this below .See p. 94
The irrelevance of word-edges
We will now take a further look at the structure of the onset. The following
words are from Attic Greek, an ancient dialect of Greek (Indo-European) in
which the classical playwrights Aeschylus, Sophocles, Euripides and Aristophanes
wrote their major works (in the 5th Century BCE). Consider the forms
in (156), where I have denoted the syllable boundary in each case:
(156) mi.kron ‘small’ ok.to: ‘eight’
pa.tri ‘father’ des.mos ‘fitting’
(oi)a.gros name hag.nos ‘holy’
How can we know where these syllable boundaries are correct? Ancient
scribes would not even write spaces between words, let alone that they would
explicitly mark the boundaries between syllables. Still we have good reason
to assume that the syllable boundaries are indeed in this position.
This reason is that the playwrights would write their works in verse, which
consisted of regular patterns of light and heavy syllables (in a typical pattern,
every line would consist of a number of dactyls: either one heavy syllable fol-dactyls
lowed by two light, or two heavy syllables in a row). The first syllables of
the words on the left systematically appear in the position of light syllables,
whereas those on the right systematically appear in a heavy position. From
this we can conclude that the first consonants in the clusters on the left belonged
to the onset whereas those in the clusters on the right belonged to the
coda.
You will observe that this fits nicely with the theory of the sonority profile
lined out above. The complex onset clusters start with an obstruent and are
followed by an r. On the other hand, the clusters in the column on the right
consist either of two plosives or of a plosive followed by a nasal. If Attic
Greek had the same dispersion profile as English, then the reason for this is
easy to see: a consonant cluster forms a complex onset only if it has the right
sonority profile; if not the consonants form a coda-onset sequence (i.e., they
are heterosyllabic).heterosyllabic
Although this argumentation seems solid, it runs into trouble once we consider
the following words:
(157) phthero ‘I destroy’, ksenos ‘stranger’, skapto: ‘I dig’
Exactly the clusters that were avoided word-internally as complex onsets seem
to appear here at the beginning of a word. Since there is nowhere else for these
5.3. Sonority 93
consonants to go, the conclusion seems unavoidable that Ancient Greek did
have remarkable onset clusters after all.
You may have noticed however, that also English has a number of clusters
which do not fit the template that we have described so far. These are words
such as skate, spy, steam, spray, splash, stream and scream. Not only do these
words all start with two obstruents, in spite of the high demands on dispersion
which English otherwise displays, but words like spray and splash even start
with no fewer than three consonants.
In particular the latter facts give us an indication of what is going on here.
All these clusters look like they have an s followed by what is otherwise a
normal complex onset: an obstruent followed by a liquid. It looks as if in
English, a word can be preceded by an s that does not belong to the syllable
structure proper, so that we have the following structure for a word like splash:
(158)
N C
RO
σ
s p l æ S
The reason why s possesses this mysterious property in English is unclear (one
phonological article about the issue is called “Do you believe in magic?”), but
this property of being extrasyllabic, i.e. not belonging to the syllable structure extrasyllabic
might be a property of a number of consonants in Greek. Apparently, the
following is true for some languages:
(159) At the beginning of the word, at most one consonant may be extrasyl-
labic.
In English and some other languages, extrasyllabicity is restricted to /s/; in
Greek, other consonants can have it as well. One might object to this move
that it is a trick to save the original hypothesis. This criticism is justified to the
extent that we do not really understand what is going on here, but notice that
we have at least restricted the ‘trick’ to one position in the word.
Furthermore, there is evidence that the ‘extrasyllabic’ s indeed behaves
the same as the Greek consonants. This evidence comes from Italian (IndoEuropean),
a language that has onsets very similar to the English ones: normally
an onset consists of at most two consonants of which the first is an
obstruent, but str, sp, etc. are also allowed, in other words, /s/ can be extrasyllabic.
In this language, the one syllable in the word which has stress, needs
to be heavy . This can be seen in the following facts: See Chapter 7
(160) a. fato ‘fate’ ["fa:to]
b. capra ‘goat’ ["ka:pra]
c. parco ‘park’ ["parko]
d. pasta ‘pasta’ ["pasta]
94 5.3. Sonority
Examples (160a) and (160b) show that the vowel is usually lengthened to satisfy
this condition, regardless whether one or two consonants follow, as long
as the consonants form a well-formed onset. In (160c), you see that this does
not happen before a cluster [rk], which has the wrong sonority profile for being
an onset. Finally, (160d) shows that /st/ behaves as the non-onset [rk]
rather than the onset [pr], in spite of the fact that it can occur at the beginning
of the word (as in stella ‘star’).
Notice that this is exactly the same pattern as we saw for Attic Greek before:
clusters which are possible at the beginning of the word do not behave as
onsets in the middle of the word, but one of the consonants becomes a coda,
making the previous syllable heavy.
Interestingly, there is even evidence for a special position of the /s/ at the
beginning of a word. Italian has two forms of the masculine definite determiner
(lo and il), which are distributed in a way that looks very similar to what
we saw in French :See p. 88
(161) l’est ‘the East’ [lest] (/lo est/) il burro ‘the butter’ [il bur:o]
il clima ‘the climate’ [il kli:ma]
We find il if the word starts with a simple or complex onset, but lo if it does
not. This gives us the perfect test for s clusters’ as (162a) demonstrates, these
show the expected behaviour, viz. they do not start with an onset:
(162) a. lo studente ‘the student’ [lo studente]
b. cf. il senatore ‘the senator’ [il senatore]
Clearly, the words do not behave exactly the same as those starting with a
vowel (in the latter case the form of the determiner is l not lo), but the similarity
is striking enough to count as evidence that our analysis of the special
status of /s/ is right.
The syllable structure of studente thus can be drawn as follows, with the s
outside of syllable structure:
(163)
N
RO
σ
N C
RO
σ
N
RO
σ
s t u d e n t e
Syllable contact
Sonority does not only play a role within the syllable, but also across syllable
boundaries. In particular, many languages require codas to have a lower sonority
than the following onset consonant. For instance, in French, although
both [pat] (pâte ‘pastry’) and [ri] (que je rie ‘that I laugh (SUBJUNCTIVE)’ are
well-formed syllables, the combination *[pat.ri] is not well-formed; we find
[pa.tri] (patrie ‘fatherland’) instead.
We call this requirement the Syllable Contact Law:
5.4. The syllable structure of English 95
(164) Syllable Contact Law (SCL) Syllable Contact Law
If C1 is in the coda, and C2 is the head of the onset of the following
syllable, the sonority of C1 shouldnot be smaller than that of C2
We already implicitly used the SCL above, in our discussions of Attic Greek
and Italian. The SCL can also be seen at work in Korean (isolate). In this language,
whenever a sequence arises that would violate the SCL, a phonological
process applies to repair this unwanted configuration:
(165) Input Output Gloss Related form
Nasalisation: /sip-ny@n/ [sim.ny@n] ‘ten years’ [sip-1l] ‘ten-ACC’
/kam-li/ [kam.ni] ‘supervision’ [to-li] ‘ethics’
Lateralization /non-li/ [nol.li] ‘logic’ [non-mun] ‘research paper’
/tik1t-li1l/ [ti.k1l.li.1l] ‘t and l’ [ni1n-tik1t] ‘n and t
Nasalisation (of plosives and laterals) and lateralisation (of plosives and nasals)
apply either to the first or to the second segment in the sequence, depending
on certain complicated factors which will not concern us here. What is
important for us is that these processes always apply to clusters that violate
the SCL and results in clusters that no longer do: the SCL is the trigger of the trigger
process.
Neither of these processes occurs if the underlying cluster satisfies the SCL:
(166) a. /kun-tæ/ - [kun-tæ] ‘army’
b. /kal-ku/ - [kal.ku] ‘desire’
c. /kal-maN/ - [kal-maN] ‘desire’
This is a strong indication that what is at work in these cases is the SCL. Sonority
restrictions of this type show up in many languages. Importantly, the
reverse effects are never found; these would be languages which have e.g. rt
onsets, but no tr; or which allow ak.la syllable contacts, but not al.ka.
5.4 The syllable structure of English
To conclude this chapter, we briefly consider the syllable structure of one language.
English is a good choice, given that you will be familiar with it when
you can read this book, but also because it has a reasonably complex syllable
structure. (There are languages which only have core syllables, hence
are much simpler, but also languages like Polish, Georgian and Berber, which
are much more complex). We will see that the theory developed in the preceding
sections gives a good frame for understanding the phonotactics of English,
although several details require further elaboration. I concentrate on the consonantal
positions here (i.e. onset and rhyme); English vowels are a rather
complicated area.
Let us first consider the onset. The constituent is obviously not obligatory
in English, given words like English and onset, which start with a vowel, yet
are perfectly acceptable to the English speaker.
Simple onsets can be filled by any consonant, with one notable exception:
the [N] can occur at the end of the word (sing), but not at the beginning (*ngis),
96 5.4. The syllable structure of English
and, in most varieties of English, not in the onset of a word-internal syllable
either. A word such as finger is pronounced as [fINg@~], not as *[fIN@~]. (There
are a few marginal cases such as dinghy) It is true that singer is pronounced
as [sIN@~], but this is only possible because there is a morpheme boundary
between the [N] and [@]. (More on the interaction with morphology in Chapter
9 .)See Chapter 9
As to complex onsets, we can draw the following table (+ denotes that a
combination exists, − that it does not or is very marginal):
(167) C2 l r w
C1
p + + −
t − + +
k + + +
b + + −
d − + +
g + + +
f + + −
T − + +
s + − +
The first segment of a cluster is always a ‘simple’ obstruent. Affricates such asSee Section x
[tS, dZ] do not occur in clusters at all (*Jrohn, *chleaf). The second consonant in
a cluster is either one of the liquids [l, r] or the glide [w].
You can see that not all cells are filled in the table. In particular, the combinations
[tl, dl, Tl] and [pw, bw, fw] are missing. If you observe these triples
closely, you will discover that the first three all involve a coronal obstruent
followed by a coronal liquid, and the second triple all a labial obstruent followed
by a labial glide. In other words, English onsets satisfy the following
criterion:
(168) The two segments in the onset cannot have the same place of articula-
tion.
The odd one out is of course coronal s which can be followed by a coronal l,
but as we have already seen , s can occupy a position outside of the syllable.See p. 93
The reason why sr does not occur in English is a mistery; it seems that in
onsets with r, English prefers [S] (shrill, shrimp).
We now turn to the coda position in English. We can observe that every
consonant can occur in this position, except [h]: sip, sick, sit, sin, ill, in, rush,
rib, kid, etc. Furthermore, we also find consonant clusters:
(169) harp, help, lamp, walk, dark, rank, old, word, wound, etc.
We can take these as evidence for English having a complex coda, which we
have not seen so far:
5.5. Exercises 97
(170)
N C
RO
σ
h a r p
The sonority profile of the codas in (169) is always the same: a liquid or nasal
followed by an obstruent. A coda is therefore almost the mirror image of
an onset, except that nasal+obstruent sequences are viable codas (lamp, rank,
hand) but the opposite is not a well-formed onset in English (*pmal), although
it is in other languages.
However, this is not the whole story. In the first place, complex codas are
(almost) completely restricted to the word-final syllable in English: a word
like *lampbroo does not seem well-formed. In the second place, there are quite
a number of apparent counterexamples to the generalization that complex codas
have a falling sonority profile:
(171) act, lapse, past, apt, etc.
The list becomes even longer if we consider morphologically complex forms
(length, depth, barred, etc.) Notice that in all these cases, the final, ‘offending’
consonant is a coronal obstruent. It seems that just like the voiceless
coronal fricative s can be an exception at the beginning of the word, all voiceless
coronal plosives can be exceptional at the end of the word.
5.5 Exercises
1. Draw the syllable structure of the following English words: stream, black,
each, blister.
2. The American poet Adelaide Crapsey (1878-1914) became known for developing
a verse form she called the cinquain, which was supposed to be
an American analogue of the haiku. Below are two examples from her
work:
(172) a. Triad
These be
Three silent things
The falling snow... the hour
Before the dawn... the mouth of one
Just dead.
b. Amaze
I know
Not these my hands
And yet I think there was
A woman like me once had hands
Like these.
98 5.5. Exercises
Describe the form of the cinquain. Can you give a reason why the existence
of such a form (created by a 19th Century American author) is less
compelling an argument for the syllable than the South-Slavic decasyl-
lable?
3. A spoonerism is an error in which sounds in two different words getspoonerism
transposed. Here are some examples of spoonerisms by English speak-
ers
(173) a. three cheers for our queer old dean ‘dear old queen’
b. is it kisstomary to cuss the bride ‘customary to kiss the bride’
c. a blushing crow ‘a crushing blow’
d. a well-boiled icicle ‘well-oiled bicicle’
e. is the bean dizzy ‘dean busy’
f. frish gotto ‘fish grotto’
g. flake bruid ‘brake fluid’
h. spicky toint ‘sticky point’
i. The Shaming of the True ‘taming of the shrew’ (title of a rock
opera)
In what way do such errors provide evidence for subsyllabic constituency?
What about speech errors such as frake bluid?
4. Italian, like French, has a number of rising diphthongs (e.g. [je, ja, we,
wa]. Considering the fact that these rising diphtongs can occur after a
single consonant, but never after two in Italian, what is the difference in
syllabification between the two languages?
(174) a. pieno ‘full’ [pje:no]
b. chiave ‘key’ [kja:ve]
c. quello ‘that’ [kwel:o]
d. guado ‘ford’ [gwa:do]
5. What explains the epenthesis of 1 in the following words in Lénakel
(Tanna)?
(175) Underlying Epenthesis
a. /t-n-ak-ol/ [t1.na.gOl] ‘you will do it’
*[tna.gOl]
b. /ark-ark/ [ar.ga.r1k] ‘to growl’
*[ar.gark]
c. /kam-n- ¯man-n/ [kam.n1. ¯ma.n1n] ‘for her brother’
*[kam.n ¯mann], *[kamn. ¯mann]
6. Here are some verbs in present and future tense in Tagalog (Austrone-
sian):
5.5. Exercises 99
(176) Present Future
bili ‘to buy’ bibili ‘will buy’
talon ‘to jump’ tatalon ‘will jump’
alis ‘to leave’ aalis ‘will leave’
kain ‘to eat’ kakain ‘will eat’
matulog ‘to sleep’ matutulog ‘will sleep’ (you can consider ma a prefix)
maligo ‘shower’ maliligo ‘will shower’
How is the plural form derived from the singular? How can this type of
morphology be described?
7. In the history of Spanish clusters of consonants sometimes reversed their
order. Look at the following examples.
(177) Latin Old Spanish Middle Spanish
spatula espadla espalda ‘blade’
retina riedna rienda ‘rein’
titulo tidle tilde ‘tilde’
The relevant clusters have been underlined. What principle of syllabification
could explain this change?
8. Voiced obstruents are disallowed in certain positions of the word in German
(Indo-European). Consider the following arguments and explain
why it can be taken as an argument for the syllable (the examples in the
lefthand column are real words of German, the examples on the right
could never be).
(178) a. [di:p] ‘thief’ *[di:b]
b. [ra:t] ‘wheel’ *[ra:d]
c. [bErk] ‘mountain’ *[bErg]
d. [gla:s] ‘glas’ *[gla:z]
e. [moti:v] ‘motive’ *[moti:f]
f. [vOtka] ‘vodka’ *[vOdka]
g. [vIkvam] ‘wigwam’ *[vigvam]
h. [bEtmInt@n] ‘badminton’ *[bEdmInt@n]
9. It could be argued that open syllables satisfy the Syllable Contact Law
in the best possible way. Comment.
10. The oldest known form of Greek (called Mycenaean Greek) was written
in a syllabic writing system called Linear B. Some of the properties of
this system are that l, r, m and n are not written at the end of the word or
before another consonant: for instance, one wrote pa-ta instead of panta.
Another property was that certain consonant clusters were written
as more than one syllable (po-to-li-ne for ptolin). A third property was
that word-initial s was omitted before a consonant (stathmos became tato-mo).
Comment on each of these three observations from the point of
view of syllable theory.
11. English words are nowadays being borrowed into many languages. Sometimes
there phonological shape is adapted to the borrowing language.
Consider the following words in Brazilian Portuguese (very similar things
happen e.g. in Arabic):
100 5.5. Exercises
(179) English → Portuguese
[sl]ide [isl]ide
[sn]ob [isn]ob
[st]and [ist]and
Comment on the reason for this epenthesis from the point of view of
syllable theory.
12. In many varieties of English a [j] sound is inserted before an [u] in words
like pew, cue and hue. Give evidence to demonstrate whether the [j] is
inserted in the onset or in the nucleus.
13. The online edition of the WALS (World Atlas of Language Structures)
has, among other things, a chapter on syllable structure, written by Ian
Maddison. The map shows three types of languages: with simple syllable
structure, with moderately complex syllable structure and with
complex syllable structure. Give the templates that are used for each
of them, and draw them as in the form of a syllable bracket. (You are
allowed to put segments in parentheses.)
14. Look at the following data from Kazakh (Turkic) and explain how they
are evidence for the Syllable Contact Law.
(180) /kol-lar/ [kol.dar] ‘hands’ cf. [al.ma.lar] ‘apples’
/murin-ma/ [mu.rin.ba] ‘nose-INT cf. [kol.ma] ‘hand-INT’
/koNWz-ma/ [ko.NWz.ba] ‘bugs’ cf. [ki.jar.ma] ‘cucumber’
15. (If you are a fieldworker.) Working with an informant, give a complete
overview of the syllable structure of some language (other than English)
along the lines of section 5.4. (Since we have not covered the whole
body of phonological knowledge on syllable structure in this chapter, it
may well be that you encounter things that do not fit the model of this
chapter. Describe those too, and explain what the problem is.)
16. (If you are a computer programmer.) The CELEX database is a database
of English (as well as Dutch, German and Tuvan) words, which
you can access among other things in syllabified form. Build a program
that makes an inventory of all the syllables of English. Do you find any
syllables which do not fit the templates described in this chapter?
Sources and further reading
Section 5.1. The text of the Asanaginica can be found on the internet at Wikisources.
The data on syllable structure in Quechua (Quechua) are taken from
O’Roerke (2008). Watson (2011) gives an overview of stress in (Palestinian) Arabic.
The reduplication data from Yaqui are discussed in Haugen (2003). More
on the Tagalog language game: Conklin (1956); the Hausa language game is
documented in Alidou (1997). A classic study on the role of syllable structure
in speech errors is MacKay (1972); his data in the study reported here were
from Meringer and Mayer (1895). A classical collection of papers on the relevance
of speech errors for phonology is Fromkin (1973). The classical article
on syllabification in children is Liberman et al. (1974). The experiment of perception
in French was originally reported on in Cutler et al. (1986). More on
writing systems and their linguistic analysis can be found in Coulmas (2003).
5.5. Exercises 101
Section 5.2. The relevance of the Axininca Campa data for phonological theory
was pointed out by McCarthy and Prince (1993a). The acquisition data
of Dutch children are from Fikkert (1994). The facts about reduplication in
Sanskrit are discussed in more detail in Kennedy (2011). The study on complex
onsets in children with a cochlear implant is Chin (2006). Old English
Compensatory Lengthening has been described in Campbell (1959) and
Hogg (1992) and further discussed in Ewen and van der Hulst (2001); Gussman
(2002). The Turkish phenomenon is discussed in Sezer (1986); Goldsmith
(1990); Kenstowicz (1994); Gussman (2002). A standard reference to the modern
analysis of Compensatory Lengthening is Hayes (1989).
The French examples are discussed in Kaye (1989), who gives them a slightly
different interpretation. The Fijian case is analysed in much more detail in
Kenstowicz (2007). One study on Japanese (and other) syllable structure is Itô
(1986).
Section 5.3 The notion of sonority is discussed in Parker (2002, 2008). It was
introduced into phonology by Sievers (1881); the particular scale discussed
here is from Clements (1990). The data from Attic Greek are discussed among
others in Steriade (1988); Kenstowicz (1994); Kiparsky (2003). The article “Do
you believe in magic?” is written by Kaye (1992). The Italian data are discussed
in that article as well as in Chierchia (1986) and Davis (1990). Krämer
(2009) summarizes the literature on Italian phonotactics. Work on the Syllable
Contact Law includes Vennemann (1988); Clements (1990); Gouskova
(2003).The Korean data are from Davis and Shin (1999), and the data from Old
Spanish are from Holt (2004).
Section 5.4. Some descriptions of English phonotactics are Lass (1976); Harris
(1994); Hammond (1999).
General A recent book on syllable structure (presenting a new view) is Duanmu
(2009). Hyman (2008) lists and explain all known universals in phonology, including
those on syllable structure.
Chapter 6
Phonological computation
6.1 Computation and representation
We have so far concentrated on the internal structure of phonological forms:
the way phrases and words are organized into smaller units of sound structure.
Theories about this aspect are often referred to as theories of phonological
representation: how are the concrete physical events corresponding to representation
speaking represented in the language system, and in the mind.
Next to this, any phonological theory needs to talk about computation as computation
well. Theories of computation talk about how phonological forms are related
to each other. For instance, it is not unreasonable to assume that the following
words start with something which is the same thing (the same morpheme).Here morpheme
is an example. Turkish is a language which displays the process of vowel harmony,1
which means, roughly, that suffixes take a different form based on the
phonological shape of the stem to which they are attached:
(181) a. ip ‘rope’, ip-ler ‘ropes’, *ip-lar, *up-lar
b. pul ‘stamp’, pul-lar ‘stamps’, *pul-ler, *pil-ler
We want to say that the thing which expresses the plural is always the same
morpheme (LAR), which sometimes shows up as lar and sometimes as ler.
The relation between ler and lar is the topic of computational phonological
theories. Such theories usually assume that one of the two — say, lar — is
underlying and that the other derived from it (for instance by spreading a feature underlying
derivedfrom the stem to the base. We tacitly already used such derivational terminology
already below.
Seen in this way — which seems to be the standard view and is the view
we will also adopt here — phonological computation thus takes an input, an
underlying form, as an input and some more concrete form, a surface representation,
as the output. Phonology thus works as a little computer which
transforms things which are there in our mental lexicon, where the underlying
representations are thought to reside, to things which are closer to the
phonetic reality. The latter are derived from the former; since this is the most
common view, theories of computation are often also called theories of deriv-
ation.
1Kornfilt (1977); Clements and Sezer (1982); Hulst and Weijer (1995); Bakovic (2000).
103
104 6.1. Computation and representation
Theories of representation and theories of computation are largely independent
of each other, since they are designed to explain different kinds of
phenomena. In (181) above, it is a fact about phonological representations in
Turkish that all vowels in a word are either front vowels or back vowels. It is
a fact about derivations that it will be the suffixes which change in stem-suffix
combinations, and not the stems.
The theory of representations will tell us what are the objects that a phonological
theory can talk about — for instance, phonological features, phonological
segments, autosegmental tiers, syllables, etc., as well as what kinds
of relations these will entertain with each other: it gives us a measure for the
well-formedness of an individual representation. The theory of computations,
on the other hand, tells us how different representations can be related to each
other: when one morpheme takes a different shape in different phonological
contexts, how these shapes are related to each other; what are the possible
changes a phonological representation can undergo.
Underlying representations
There is a difference between printed letters and hand-written letters: the
former are all independent from each other on the paper — except in the case
of so-called ligatures like ‘fi’ — but the latter are usually connected to each
others by all kinds of small lines. A hand-written <n> looks a little different
after a <n> (line is coming from below) than after a <o>.
Sounds tend to adapt themselves even more to each other than handwritten
letters, both phonologically and phonetically. For instance, the last
consonant of the English prefix in- changes according to the context, witness
words such as:
(182) impopular, inactive, impopular, i[N]consistent, irregular, illegal.
We have of course already seen that this is the result of spreading a Place
of Articulation feature from the base to the nasal consonant. In order to do
make explicit how assimilation works as a computation, we have to add a
dynamic element to our theory, which can explain why certain things change
and others do not. Such a theory should also describe why certain changes are
possible in some languages but not in others, even though they have to deal
with similar problems: other languages may react to ‘problematic’ sequences
such as /np/ in a different way, for instance by deleting the nasal.
In order to do set up a theory of computation, we distinguish two types
of phonological representations for every form: an underlying representation —underlying representation
the form as we assume it is stored in the mental lexicon of the speaker — and
a surface representation — the phonological form which is closest to the actualsurface representation
speech sounds and/or articulatory instructions. One underlying representation
(/In/ for the prefix) can correspond to more than one surface representation
([In], [Im], [IN], etc.) and it is not the case that underlying representation
and surface representation are necessarily distinct: underlying /In/ corresponds
to surface [In]. We will assume that underlying representations and
surface representations have the same formal structure: they consist of features,
autosegmental tiers, syllables, etc., in exactly the same way.
Let us concentrate on one fairly simple process, final devoicing in German,final devoicing
6.1. Computation and representation 105
illustrated in (183):
(183) a. Hund ‘dog (sg.)’ [hunt], Hunde ‘dogs (pl.)’ [hund@]
The first observation we have to make is that we know that [@] is a plural suffix
in German. We also find it for instance in Schuh ‘shoe (sg.)’ [Su] - Schuhe ‘shoes
(pl.)’ [Su@]. We thus have a situation where we have to assume that the forms
hunt and hund are derived from the same underlying form.
As a rule of thumb, we can safely assume that the underlying form always
is a surface form somewhere, so that we have two possibilities: either /hund/
is the underlying form or /hunt/ is. Notice a notational point here: we write
underlying forms between dashes //, whereas output forms are written in
square brackets []. A further notational point is that we write a derivational
relation between two forms with an arrow, pointing from the underlying form
to the phonetic form, as follows:
(184) /A/ → [B]
So which of the two /hund/ or /hunt/ is the underlying representation for the
German word for ‘dog’? In principle, both are possible and the isolated fact in
(183) does not provide us enough information to decide. We have to consider
the overall system of German. Notice that we have to decide between the
following two derivations:
(185) a. /hund/ → [hunt] (at the end of the syllable)
b. /hunt/ → [hund] (before @)
How do we choose? One observation we can make is that there actually are
no German words ending in a [d] (or any other voiced obstruent, for that
matter). This is captured by the derivation in (185a): even if a word starts its
life in the lexicon having an /d/, it will never make it to the surface as a [d],
but will become a voiceless consonant, in something like the derivation (185a).
Derivation (6.1), however, has nothing to say about this fact.
On the other hand, cannot easily be extended to other kinds of facts. For
instance, there are many German words in which the /t/ does not change
before a schwa. For instance, the adjective for ‘colourful’ is bunt. When we inflect
this adjective, it can get a schwa, but the /t/ remains unchanged: bun[t@].
Although descriptively for this one individual fact both analyses are equivalent
in linguistics, and in science more generally, we aim to choose the analysis
that is most easy to generalize to other facts. Since it is easy to generalize
(185a) (viz. to (186)), but not (6.1)), we assume that the former holds, and
hence /hund/ is the underlying form:
(186) /. . . [Voice]/ → [. . . ] (at the end of the syllable)
Notice that the condition ‘at the end of the syllable’ is also a familiar one: it is
an effect of the special status of codas.
106 6.2. Optimality Theory
6.2 Optimality Theory
In this chapter, we will look at one specific theory of phonological computation,
Optimality Theory (OT), which for the past few decades has been dom-Optimality Theory (OT)
inant in this area, albeit in several varieties (we concentrate here on the most
classical, standard variety of the theory).
Operations
Like other derivational theories, Optimality Theory involves derivation from
an input form for every morpheme to an output form. We can get from one to
the other by applying a number of minimal operations, each defined in terms ofminimal operations
the representations we use; so autosegmental structures, syllable structures,
etc.:
• We can add or delete a feature – and in this way we may turn /hund/ into
[hunt] by deleting the feature [voice] at the end
• We can spread features such as tone, or vocalic features such as we have
seen for Turkish
• We can shorten or lengthen consonants and vowels — we have seen an
instance of this where we derived cittá [s:]anta from a form with an underlyingly
short consonant in ItalianKlopt dit?
• We can epenthesize (insert) consonants and vowels — such as has happenedepenthesize
for instance in the Axininca Campa word [noNkomati] which is derived
from /no-N-koma-i/ (‘he will paddle’).Klopt dit?
• We may change syllable structure as exemplified e.g. in the derivation
between /hund/ (where /d/ is in the coda of the first syllable) and
[hun.d@] (where it occurs in the onset of the second syllable).
• We may change the stress structure of the word.Klopt dit?
We could apply each and every one of these operations on any underlying
form. In actual practice, we only take action however if this improves the
form in some sense; if the output form becomes better than the input form.
As we have seen, the reason to delete the feature [voice] in /hund/, is that
in this way we can satisfy a requirement on syllable coda’s. We can write such
a requirement (in OT, these are usually called constraints) as follows:
(187) DEVOICE: Consonants at the end of the syllable should not have the
feature [voice].
Notice that this formulation of the constraint presupposes certain representational
assumptions, for instance that we can distinguish consonants, that there
is syllable structure and consonants can occur at the end of syllables, and that
we have a feature [voice]. (184) is probably an instance of a more general principle
requiring coda consonants to have as few consonantal features as they
can get.
An important step in OT thinking is that we only delete [voice] in /hund/
because DEVOICE asks for it. We would not delete [voice] in e.g. the word
denn ‘then’ [dEn], because there is no constraint which requires deletion in this
case. To the contrary, there is a general principle of economy or faithfulness —faithfulness
as it is called in OT — which states the following:
6.2. Optimality Theory 107
(188) FAIHFULNESS: The surface representation should be as close to the underlying
representation as possible; do nothing.
If DEVOICE is not operative in German, FAIHFULNESS is; therefore *[tAn] is a
bad output form for /dAn/; it has violated FAIHFULNESS without necessity.
Two functions
So how do we decide which operations can be applied and which cannot? The
(phonological) grammar consists of two functions, called Gen (Generator) and Gen (Generator)
Eval (Evaluator). Gen takes an input form and blindly applies phonological Eval (Evaluator)
operations to it in any conceivable combination. In this way it creates a very
large number of possible output forms, called candidates. candidates
As a matter of fact, this number will be infinite in classical OT. If we take
the input form /hund/, we can change all the features of all four segments,
but we can also go on adding consonants and vowels to this structure indefinitely.
Adding 5,000,000 consonants to this form will probably not improve
the structure for any kind of constraint, but Gen is assumed to be blind to this.
The idea is not to be psychologically real (nobody assumes that every time
you utter a word, you go through all logically possible things you could do
to that word), but to offer a precise model of what the best possible form in a
language is.
The output of Gen thus is a very large set of candidates. The function Eval
takes this set as its input and determines which single one of these best satisfies
all grammatical principles, including the two we have just introduced,
DEVOICE and FAIHFULNESS. The form is thus a kind of compromise: something
has changed, but only in order to satisfy the needs of the language.
Schematically, the derivation can now be drawn as follows:
(189)
/hund/ underlying representation
¨
¨¨¨
¨¨¨
r
rrr
rrr
 
 
 
 
d
d
d
d
€€€€€€€€€€
ˆˆˆˆˆˆˆˆˆˆˆˆˆˆˆˆ
Generator
/hund/ /hund@/ /huda/ /hAnt/ /und/ /hunt/ ... candidates
Evaluator
surface representation
rr
rrr
rr
¨¨
¨¨¨
¨¨
d
d
d
d
 
 
 
 

$$$$$$$$$$$$$$$$
/hunt/
In this case, Eval will choose [hunt] as the definitive surface structure for German
because it satisfies DEVOICE without making all kinds of unnecessary
changes. For the principle of FAITHFULNESS, however, this is not the best
possible form; that would have been [hunt].
This observation has a few implications; in the first place, the actual surface
form is not perfect in the sense that it satisfies all possible constraints. It is
not possible to be perfect in this sense, since constraints can impose conflicting
demands. This is why the theory is called Optimality Theory: the ‘winner’ of
the evaluation is not necessarily impeccable, but it is optimal; the best one can
do.
108 6.2. Optimality Theory
Secondly we may observe that apparently DEVOICE has more weight in
the grammar of German than FAITHFULNESS. We can write this down as fol-
lows:
(190) DEVOICE FAITHFULNESS
(pronounce: ‘DEVOICE dominates FAITHFULNESS’)
There are also languages in which the order of these constraints is reversed.
There is no devoicing in the Yiddish word [hund] and we may assume that
the reason for this is that in this language we prefer to be faithful rather than
satisfy this particular requirement on syllable structure wellformedness. The
crucial difference is that German has the order in (190), whereas Yiddish has
the ordering in (191).
(191) FAITHFULNESS DEVOICE
An interesting assumption of (the classical version of) OT is that all constraints
are universal; languages differ only in the relative ordering of the constraints.
Metaphorically speaking, phonology in all languages consist of a number of
forces, and these forces are always the same. The only difference between
languages is how powerful each and every one of these forces is. Constraint
ranking is the only possible difference between two languages; in this sense,
OT is a strong theory of language variation: it claims that systematic differences
between languages always can be described with an ordering of universal
constraints.
Faithfulness is not one thing
Our analysis of final devoicing in German is not completed yet. It is true that
[hunt] satisfies DEVOICE, but this is true also for e.g. [hAnt] and [hund@]. So
why is the former the winner? The answer is relatively easy to give for [hAnt].
Like [hunt], this form violates FAITHFULNESS, in that it has deleted a feature
[voice], but it has done even more: it has also changed the specification for
[round] on the vowel, and this is an unnecessary extra violation of FAITHFULNESS.
Apparently, we do not just count whether or not a constraint is violated,
but also how often this is the case.
Matters are more difficult for the comparison with [hund@]. In order to
get there from our underlying form, we arguable need to take only one step:
insert an empty vocalic position. So why does this form lose from our winner
[hunt]? We will have to split up our cover constraint FAITHFULNESS into a
more fine-grained structure of constraints which are all ordered. In particular,
we will need at least the following two faithfulness constraints:
(192) a. KEEP-FEATURE: All features in the underlying representation must
be present in the surface representation
b. *FEATURE: All features in the surface representation must be present
in the underlying representation
c. (Dutch): *FEATURE KEEP-FEATURE
We can now consider [hunt] as a better surface structure than [hund@], because
the former violates a lower-ranked constraint than the latter.
6.3. A case study: Nasal assimilation 109
Tableau
It is common practice to draw the evaluation of surface candidates in a socalled
tableau; in the case at hand, this tableau looks like this: tableau
(193)
/hund/ DEVOICE *FEATURE KEEP-FEATURE
hund *!
hunt *
hund@ *!
hAnt **!
This should be read in the following way. In the left-hand column you see
the underlying representation on top. Immediately below it you see some of
the more interesting output candidates. Given that there are infinitely many
things we can do, it is impossible to draw all of them, but you do not have to
worry about this: it is typically possible to determine which forms are relevant
for a discussion and which are not.
From left to right, you see the names of the relevant constraints, in the order
in which the grammar (of German, in this case) has ordered them. An
asterisk in a cell indicates that the form in question violates a constraint, and
two asterisks indicate that it violates the constraint twice. An exclamation
mark behind an asterisk indicates that this violation is ‘fatal’ for the form in
question; it is the reason why this form is not the ultimate winner. The pointing
finger directs the reader’s attention to the form which has no fatal violations
and is therefore the optimal form and the actual surface structure.
6.3 A case study: Nasal assimilation
Let us now turn to an example which is slightly more complicated, viz. the
behaviour of the English prefix in-, which displays nasal assimilation. Nasal
assimilation is a phenomenon which is much more wide-spread in languages
of the world, and has been analysed in terms of autosegmental representa-
tions.
How are we going to integrate such an analysis into an OT framework?
We obviously need to have the right constraints, which is a craft in its own
right. This section will be an exercise in formulating one such constraint.
Consider the input representation (194a) and the candidate outputs in (194b).
(194) a. /in+polite/
b. { [inpolite], [iNpolite], [impolite] }
The fact that [impolite] is the winner means that we prefer a structure in which
the nasal and the following consonant share their Place node. Let us formulate
the relevant constraint as follows:
(195) PLACEHARMONY (first version): A nasal has to bear the same Place
node as a neighbouring consonant.
Clearly, impolite is the only form among our set of candidates satisfying this
constraint, if we assume that it has the following structure:
110 6.3. A case study: Nasal assimilation
(196)
I m p o l a j t
  
Place
However, as soon as we make our set of candidates just a little bigger, we
will see that there are more possibilities. Take for instance the form intolite, in
which the stem consonant has adapted to the prefix, rather than the other way
around:
(197)
I n t o l a j t
  
Place
Presumably, the nasal in the input /in+polite/ is more sensitive to harmony
than the plosive is. There are two possible explanations for this difference. It
could be that there is some internal difference in the structure of nasals and
segments which causes the asymmetry; alternatively, this could point to a difference
between affix segments and stem segments. Although the second explanation
has something to say for it (as we will see later in chapter ), there are***
also clear indications that there is something right about the first explanation.
In English place assimilation, for instance, there is always a nasal involved:
there is no assimilation in words such as actor (*[ækkOr], *[ættOr]). On the
other hand, nasals are always homorganic to a following plosive also inside a
word (land, antenna, camp, lampoon, bank, banquette).
Let us assume than for the moment that the second analysis is the correct
one. We can build this restriction into our theory in various ways. We could
make it a matter of a faithfulness constraint, which would somehow say that
nasals tend to be less faithful to their underlying representation than other
consonants. Alternatively, we may build the restriction into our definition of
the constraint PLACEHARMONY, which we will do here:
(198) PLACEHARMONY (second version): A nasal at the surface structure
has to bear the underlying Place node of a neighbouring consonant.
This now explains why /in+polite/ is not rendered as [intolite]. But we are
still not completely satisfied. This second version suggests that every nasal
will borrow the place of its neighbor, either on its left or on its right. This is not
true for English, witness words such as techno in which there is no assimilation
at all.
Again, there are several possibilities. We could assume, for instance, that
the difference with the previous case is that nasal occurs on the righthand side
of the neighbouring consonant rather than on its left. We could now revise
PLACEHARMONY in the following way:
(199) PLACEHARMONY (third version): A nasal at the surface structure has
to bear the underlying Place node of a neighbouring consonant on its
righthand side.
6.3. A case study: Nasal assimilation 111
This version will work sufficiently well for English, even though the question
remains open what is so special about the righthand side of the nasal. Yet if
we consider other languages, we discover soon enough that ‘righthand side’
and ‘lefthand side’ are not the right concepts to be used. Since OT assumes
that constraints are universal, we are however on the quest for a constraint
which can explain the facts of as many languages as possible.
We will have a brief look at Dutch dialects. Many of these dialects — we
take the dialect of Hellendoorn, a small town in the north east of the Netherlands,
as an example — show syllabic nasals, for instance as the infinitival
ending: eten ‘to eat’ [Etn
"
]. The nasal forms the nucleus of the syllable on its
own in cases such as these. Interestingly enough, also this nasal is sensitive to
place assimilation, and shows up with the same place as the preceding con-
sonant:
(200) roe[pm
"
] ‘call’
wer[kN
"
] ‘work’
po[fM
"
] ‘roast’
Syllabic nasals can also borrow their place from their neighbour on their righthand
side, for instance if they function as indefinite determiner clitics:
(201) [n
"
] doeve ‘a pigeon’
[M
"
] fietse ‘a bicycle’
[m
"
] bal ‘a ball’
[N
"
] keer ‘once (a time)’
Apparently, left and right are not the relevant categories, at least not in Hellendoorn
Dutch. Still, also in words like opnemen ‘take on’ or pneumatischpneumatic
we would not find assimilation in this dialect, showing that not every nasal assimilates
in place. The correct definition of the (universal) constraint on Place
assimilation is not sensitive to these categories, but instead of this to syllable
structure. From your introductory class to phonology you may recall that it
is usually assumed that syllables form constituents of the following type (disregarding
various details):
(202)
σ
$$$$$
O R
r
rr
r
rr
t r I k
The generalisation now seems to be that nasals within the coda assimilate, but
nasals within the onset do not. An improved version of PLACEHARMONY will
therefore say:
(203) PLACEHARMONY (fourth version): A nasal in the coda has to bear the
underlying Place node of a neighbouring consonant.
112 6.4. Beyond place harmony
We could speculate why a nasal in the rhyme has this peculiar property. It
seems reasonable to relate it to the effect of DEVOICE and other constraints
we have seen, which all state that independent consonantal features in the
rhyme are undesirable. In some sense, rhymes are the domain of vowels and
vocalic material and consonants are aliens in that domain; consonants belong
to the onset, where they are much less restricted: they do not have to devoice
and they do not have to assimilate. We could now go on to find a general
constraint of this sort which will give us all the right results, but that is not
something we can go into here.
It is still (intentionally) unspecified in our constraint which of the two
neighbours is going to lend its place in case of a choice. Hellendoorn facts
shed light on this issue as well:
(204) a. loop [N] keer ‘walk one time’
b. (ik heb) de kat [m] bettien (gevoerd) ‘(i feeded) the cat a little bit’
c. (ik heb het) rek [m] verfien (gegeven) ‘I painted the rack (I gave the
rack a little paint)’
In these cases there seems to be a preference for the consonant on the righthand
side. Does this mean we will have to build the notions ‘left’ and ‘right’ into our
theory after all? An important observation is that in these cases we are considering
a determiner which entertains an intimate relationship with the noun on
its righthand side and a much less intimate relationship with the word (verb
or noun) on its lefthand side. The former is within the same syntactic phrase,
but the latter is not. It thus is not necessary to distinguish between left and
right; we just have to understand that the nasal attracts place from its closest
neighbour in terms of syntactic structure. That could be built into the ultimate
version of the constraint as well, but we will refrain from doing that here.
6.4 Beyond place harmony
In our discussion of nasal harmony, the notion of coda is important, which we
know from chapter 5 (more precisely, section 5.2). We have already seen there
that coda consonants are weak, and dispreferred in many different languages.
As a matter of fact, there is quite an array of languages which do not have
coda consonants at all. Examples of these are Fijian, Mazateco and Cayuvava,
and the following implicational universal seems to hold over known phonological
systems:
(205) If a language has closed syllables, then it also has open syllables.
All languages have open syllables (syllables without a coda), but only a subset
also has closed syllables.
In section 5.2 we have seen that Fijian is an example of a language without
closed syllables. In order to repair potential violations of this generalisation,
the Boumaa dialect employs vowel epenthesis, the insertion of a vowel. If avowel epenthesis
word with a closed syllable is borrowed, a vowel is inserted to satisfy the
constraint against closed syllables (the following is repeated from example
(149) in chapter 5):
6.4. Beyond place harmony 113
(206) Vowel epenthesis in Boumaa Fijian
a. kaloko ‘clock’
b. aapolo ‘apple’
c. tSone ‘John’
In order to capture this effect, we can posit a constraint NOCODA:
(207) NOCODA: Syllables should not have a coda, *C ]σ
Like all constraints, the constraint NOCODA should be assumed to be universal,
it is present in all grammars. The difference between English, allowing
codas and Fijian, disallowing them, is one in constraint ranking with respect
to a faithfulness constraint (ignoring a few segmental differences between the
languages):
(208) a. NOEPENTHESIS: (= a subtype of faithfulness) Do not insert vowels
b. English grammar: NOEPENTHESIS NOCODA
c. Fijian grammar: NOCODA NOEPENTHESIS
(209) a. English
/ÃOn/ NOEPENTHESIS NOCODA
ÃOn *
ÃOne *!
b. Fijian
/ÃOn/ NOCODA NOEPENTHESIS
ÃOn *!
ÃOne *
If we now look at the other side of the syllable template, the onset constituent.
The typological behaviour here is quite different. We can posit an implicational
universa here as well, but it runs in the opposite direction:
(210) If a language has syllables that lack an onset, then it also has syllables
that have an onset.
In other words, all languages have so-called CV syllables, but not all languages
have syllables that consist of only a V; we have seen an instance of
such a language last week: Axininca Campa. We have expressed this informally
before in the observation that the onset is the consonantal domain and the
rhyme the domain of the vowels.
In order to describe this situation, we need a constraint of the following
type:
(211) ONSET: Every syllable should start with a consonant.
Note that this constraint is almost exactly the mirror image of NOCODA; together
they describe the ideal syllable template CV, which all languages have.
Formally, the reason for this state of affairs is that no matter how high or how
low we rank the faithfulness constraints with respect to these two constraints,
114 6.4. Beyond place harmony
CV syllables will always surface, viz. when they are underlying: nothing will
have to change to them in order to get to the surface.
Another universal follows from these two constraints:
(212) An underlying (monomorphemic) sequence VCV will be syllabified in
all languages as V.CV
(212) is not completely self-evident. It is not hard to imagine a world in which
it would not be true. For instance, in French pat in pâte ‘pastry’ [pat] and e in
é(gale) ‘equal’ [e (gal)] are both well-formed syllables, so why do we syllabify
pâté ‘paste’ [pate] as [pa.te] rather than *[pat.e]? The answer is that these two
constraints conspire to this result:
(213) a.
/pate/ ONSET NOCODA
pa.te
pat.e *! *
b.
/pate/ NOCODA ONSET
pa.te
pat.e *! *
In other words, every language which has pâté will syllabify it in the French
way. Faithfulness constraints are irrelevant, at least as long as we assume that
there is no syllabification in underlying representation (a standard assumption
although it is sometimes contested).
Epenthesis
We have seen that both Fijian and Axininca Campa solve their problems with
syllable structure by way of vowel epenthesis. We will now go into this a little
deeper for the latter language.
The ONSET constraint is very strong in Axininca. Whenever the concatenation
of morphemes would result in an onsetless syllable, an epenthetic [t] is
inserted:
(214) a. /no-N-koma-i/ [noNkomati] ‘he will paddle’
b. /no-N-koma-aa-i/ [noNkomataati] ‘he will paddle again’
c. /no-N-koma-ako-i/ [noNkomatakoti] ‘he will paddle for’
d. /no-N-koma-ako-aa-i/ [noNkomatakotaati] ‘he will paddle for it again’
Once we introduce a specific faithfulness constraint against epenthesis, we
have all the constraints set in place to describe this behaviour:
(215) NOEPENTHESIS: Segments in the output should also be present in the
input.
(216)
/no-N-koma-i/ ONSET NOEPENTHESIS
noN.ko.ma.ti *
noN.ko.ma.i *!
6.4. Beyond place harmony 115
However, given the properties of the Gen function, we should also take into
account numerous other candidates. Most interesting among these are those
forms which satisfy both ONSET and NOEPENTHESIS. This is certainly possible;
for an input /no-N-koma-i/ there is an output candidate [noN.ko.ma] in
which nothing is epenthesized, but there is also no ONSET violation.
The point here is that here a different type of faithfulness constraint is violated,
viz. one against deletion:
(217) NODELETION: Underlying segments (vowels) must be preserved in
the output.
Apparently, this constraint dominates NOEPENTHESIS in Axininca:
(218)
/no-N-koma-i/ ONSET NODELETION NOEPENTHESIS
noN.ko.ma.ti *
noN.ko.ma.i *!
noN.ko.ma.i *!
We have ordered ONSET NODELETION, but it is not very hard to see that
we would have got the same result if we would have ordered these constraints
in the opposite order, given the fact that NOEPENTHESIS is low ranking (you
will be asked to show this in exercise 10). In cases like this, we say that the
ordering is irrelevant, which we write down as ONSET, NODELETION, so with
a comma instead of the -sign.
More generally, we can order three constraints in 3x2x1 = 6 different logically
possible ways: all three constraints can be in the first position, but once
we have chosen one, only the two remaining ones can be put in the second
position, and if they are fixed, only the one remaining constraint can be put
in the final position. Similarly, if we have four constraints, the number of
orderings is 4x3x2x1=24, and the number of orderings for five constraints is
5x4x3x2x1=120. These numbers are also written as 3!, 4! and 5!, respectively,
in mathematics, which are pronounced ‘the factorial of 3, 4, 5’ respectively.
If we write down all possible orders for a given set of constraints, we get a
factorial typology. The prediction is that every individual grammar should de- factorial typology
scribe some (possible) human language.
We thus have six different possible constraint rankings for our three constraints.
Yet some of these grammars produce exactly the same result no matter
what the input is:
(219) Factorial typology for { ONSET, NODELETION, NOEPENTHESIS }:
1. ONSET, NODELETION NOEPENTHESIS: Consonant epenthesis
to create onset (e.g. Axininca)
2. ONSET, NOEPENTHESIS NODELETION: Vowel deletion to create
onset (e.g. Modern Greek)
3. NOEPENTHESIS, NODELETION ONSET: Onsetless syllables freely
allowed (e.g. English)
We this have three possible different languages, according to this miniature
typology. Every language should fit into one of these three categories.
116 6.4. Beyond place harmony
Harmonic bounding
It is one prediction of Optimality Theory that changes never happen without
a cause. If we delete something, we violate NODELETION; if we insert something,
we violate NOEPENTHESIS. Such violations will only be allowed if they
help us satisfy a higher-ranked constraint. Violation of constraints is always
minimal, because there will always be a competing candidate which has lessminimal
violations, and unnecessary violation of constraints will not help a candidate
in the struggle for life.
In order to see this, consider the following example from Lenakel. The
relevant syllable structure constraint in this language is slightly little different
from what we have seen so far, although it is clearly related:
(220) *COMPLEX: Onsets and codas should not contain more than one con-
sonant.
This constraint is responsible for the fact that consonant clusters are broken
up by an epenthetic vowel [1] if they would result in syllables with complex
marginal clusters:
(221) a. /t-n-ak-ol/ [t1.na.gOl] ‘you will do it’
*[tna.gOl]
b. /ark-ark/ [ar.ga.r1k] ‘to growl’
*[ar.gark]
c. /kam-n- ¯man-n/ [kam.n1. ¯ma.n1n] ‘for her brother’
*[kam.n ¯mann], *[kamn. ¯mann]
This can be described by assuming the ranking *COMPLEX(, NOINSERTION) NOEPENTHESIS
for Lenakel. Now study the following alternative candidates for these forms:
(222) a. *[a.r1.ga.r1.k1]
b. *[t1.na.gO.l1]
c. *[ka.m1.n1. ¯ma.n1.n1]
Like the real winners, all these candidates satisfy *COMPLEX and violate NOEPENTHESIS.
The problem is, however, that they violate this constraint more
than necessary.
It would be necessary to violate NOEPENTHESIS as often as these forms do
it, if Lenakel would have a high-ranking NOCODA, but apparently this is not
the case: Lenakel allows closed syllables, so that the language can be assumed
to have the following constraint ranking:
(223) *COMPLEX NOEPENTHESIS NOCODA
Yet even in a language which disallows closed syllables, a candidate such as
the following would never win:
(224) *[a.r1.ga.r1.1.k1]
6.5. A theory of constraints 117
It is safe to assume that this particular form would never win in any language,
given the input we studied. It contains an epenthesis which does not improve
anything, and it is harmonically bound by other forms which do not violate thisically bound
constraint.
Another sense in which vowel epenthesis in Lenakel is minimal is in its
choice of the central vowel [1] as the epenthetic vowel. This vowel (as well
as its non-high counterpart [@]) very often serve as the epenthetic vowel. We
know why this is: these vowels are quite empty, since they do not contain
place features. By inserting them rather than place-bearing vowels, we epenthesize
as little as possible into our phonological structure.
6.5 A theory of constraints
You may have noticed that many of the constraints which have been presented
here talk about codas in one way or another. Codas are marked positions
for consonants. In some languages, they are disallowed altogether, but even in
languages which do have them, they are restricted. French word-final floating
consonants only show up if there is an onset position created for them,
rather than a coda position. Nasals in the rhyme borrow their place features
from their neighbour. Obstruents undergo final devoicing in the coda in many
languages. We will study a few more examples in this chapter.
Here is the first such an example. Japanese only allows coda consonants if
they share a place of articulation with the immediately following consonant.
We thus find words such as those in (225a), whereas the forms in (225b) are
not allowed.
(225) a. kap.pa ‘a legendary being’, kit.te ‘stamp’, gak.koo ‘school’, tom.bo
‘dragonfly’, non.do ‘tranquil’, kaN.gae ‘thought’
b. *kap.ta, *tog.ba, *pa.kap, etc.
The constraint which is responsible for this is the so-called Coda Condition,
well-known from the study of Japanese phonotactics:
(226) CODA-COND: Consonantal place features should occur in a position
outside the coda.
Note that the constraint is satisfied by the forms in (225a) under autosegmental
assumptions: the place features are all in an onset postion; CODACOND
does not care that they are also in a coda. The only structure it militates
against is one where place features occur in a coda position to the exclusion of
other positions.
The CODA-COND is not idiosyncratic to Japanese; we also find it in an unrelated
language such as Ponapean. In this language, we can see that it takes
a phonological effect: it causes vowel epenthesis, as the following examples
demonstrate:
(227) /ak-dei/ a.ke.dei *ak.dei ‘a throwing contest’
/kitik-men/ ki.ti.ki.men *ki.tik.men ‘rat INDEF’
/naNkep/ *na.Ni.kep naN.kep ‘inlet’
118 6.5. A theory of constraints
Another way in which CODA-COND can be satisfied is by deletion of the offending
consonant. Also this is attested in some of the world’s languages,
e.g. in Diola Fogny:
(228) /let-ku-jaw/ le.ku.jaw *let.ku.jaw ‘they won’t go’
/jaw-bu-Nar/ ja.bu.Nar *jaw.bu.Nar ‘voyager’
/jaw-bu-Nar/ *ja.bu.Na ja.bu.Nar ‘voyager’
(We leave it as an open question why it is the first consonant which is deleted
rather than the second one.)
We can now see CODA-COND as one member of a ‘family’ of constraints,
all of them having parallel definitions:
(229) a. CODA-COND: Consonantal place features should occur in a position
outside the coda.
b. FINALDEVOICING: Consonantal [voice] should occur in a position
outside the coda.
c. NASALHARMONY: Nasal place features should occur in a position
outside the coda.
d. NOCODA: Consonantal features should occur in a position outside
the coda.
In a theory of phonological computation which is based on constraints, such
as OT, one should obviously have a theory about what is a possible constraint.
If we are allowed to freely formulate new ‘universal’ constraints all the time,
we cannot say that we have much of a theory. We do not make any specific
predictions about what is and is not possible in human language, since we
can always change the structure of the theory once we encounter a new phe-
nomenon.
Within OT, we posit that all constraints are universal; that is already a
restriction of some sort, since we at least need to show how a constraint which
we posit for one language plays a role in (all) other languages of the world.
But if we can freely invent constraints, then we can have a constraint X and
a different constraint ¬X which says exactly the opposite, and which would
‘explain’ why we do not see the effect of X in all languages: because many of
them would happen to have ¬X X, and ¬X would just make X ineffective
whenever its occurrence would be unpleasant to us.
Organizing constraints into families such as we have done in (229) is a
first step towards building a better theory of constraints. We could build one
schematic constraint from which the various concrete instances in (229) can be
derived by instantiating the variable F in different ways:
(230) CODA-COND(F): Consonantal feature F should occur in a position outside
the coda.
We could now say that the universal set of constraints consist only of concrete
instances of a small set of constraint schemes (or even that an individual language
chooses one or more instances of the scheme in its actual grammmar.)
It is then unexpected that constraints are needed in the analysis of a language
which does not fit into some general schema.
6.6. Implications of the theory of computation 119
6.6 Implications of the theory of computation
Psychological reality
One can see the kind of computation here as reflecting the mapping which
speakers probably do while speaking: one retrieves a form from the lexicon,
and transforms this in certain ways in order to get to something that can actually
be pronounced. The underlying representation stands for the former, the
output representation for the latter. Similarly, there must be a mapping going
on while listening: the sound waves we here must somehow be mapped onto
the structure of words as we remember them.
Phonological computation models some part of this process: it does not
deal with actual sound waves, or instructions to the articulatory organs, nor
does it deal with configurations of neurons, but it does represent the mapping
in some way.
The model of Optimality Theory also abstracts from what is presumably
the psychological reality in some other way. We most probably do not entertain
an infinite number of possibilities every time we utter a word. In actual
practice, the generator function will thus be restricted in some way, and
in any case, the theory here seems to be able to describe more why a certain
input-output mapping is made than how it is made.
Altogether, an OT mapping thus gives a fairly abstract account of what is
going on; but the claim by its practitioners is that it is also the best, or the
most precise account we have. There are various alternatives as well, for instance,
people who claim that there is no mapping at all, and that all forms are
stored. So German speakers remember both Hund [hunt] and Hunde [hund@]
separately. The fact that one ends in a [t] and the other has a [d], and furthermore
that we systematically find no [d] or other voiced obstruent at the end of
a German word, is then seen as the result of one or more historical processes,
not as something that needs to be represented in the grammar. Under this
view, there is no phonological grammar, there is just a collection of words,
each of them the result of some path through history.
An argument against this is that people show that they have knowledge
of patterns such as the devoicing of obstruents in a coda. German speakers
display this knowledge in various ways. For instance, it is a mark of a German
accent when speaking English to also devoice consonants in that language;
but German speakers also do it when borrowing words from English or other
languages which do not show final devoicing. Furthermore, in laboratory
experiments, Germans will not accept words ending in a voiced obstruents as
plausible German words.
All of this implies that even if speakers store both the singular and the
plural form for words such as Hund ‘dog’, they also have some way of accessing
the regularity of the sound correspondences between these words. From
experimental work, we know that speakers do not know about all statistically
significant patterns which linguists can detect in a language, but only about
the ones which somehow make phonological sense, such as the devoicing pattern.
When confronted with loanwords, they will only adapt them to such
patterns, not to patterns which seem completely random from a phonological
point of view. In our terms, they can see patterns which can be expressed by
the machinery of phonological computation, but not random other patterns.
120 6.6. Implications of the theory of computation
This of course implies that we take the cognitive view on phonology seriously.
It seems reasonable to say that there is phonological computation, and
that it may even be ‘optimizing’, although the precise way in which it is implemented
in the brain may be different from the tableaux we draw in an OT
analysis.
Typology
The computational theory of Optimality Theory furthermore provides us with
an interesting view on linguistic typology. Remember that the claim is that the
only differences between languages are in the ranking of constraints, while
these constraints themselves, as well the representations about which they are
computed are universal. All languages have coda constituents in some sense,
but in some languages the constraint against them are disallowed.
Obviously, when we say that the only difference in languages is in their
constraint ranking, we mean the only systematic differences in their sound
system. The fact that the French word for ‘tree’, arbre, sounds very different
from the English word, is not the result of the constraint ranking, but from
an arbitrary, and hence non-systematic, fact about the French and English lex-
icons.
Still, the claim that all systematic differences between languages are describable
in terms of constraint rankings is a fairly strong one. As we have
already pointed out, it means that we make a claim also about universals:
every constraint should be present in all languages, even though its effect
might be covered by other constraints in some languages. Those constraints
should then actually be identifiable, and themselves also universal.
Furthermore, the claim is that every permutation of constraints gives at
least a possible human language. This hypothesis can then also be tested. For
instance, we can check whether we find an attested human language which
actually behaves according to the constraint ranking we have established.
If we do find such a language, we find a confirmation of our theory. It is
well-known, however, that part of the scientific methodology is to look not for
confirmation (or verification), but for falsification of the theory. Unfortunately,
it is rather difficult to find such a falsification. The fact that we do not find
a language which confirms the expected pattern in itself, could be due for
instance to the fact that it is impossible to check all existing languages. But
what is worse, the languages we now have in the world almost certainly do
not show all possible human languages: some of the latter may simply already
be extinct without leaving a trace, or yet to arrive in the world, or may even
never actualize for extra-linguistic reasons (the people who would speak such
a language just give up their language altogether before it could change into
the required pattern).
It is important to see, then, that the claim is not about existing or nonexisting
languages. The claims of cognitively inspired theories of language
are ultimately theories about human beings. The claim of Optimality Theory
is that humans can compute some languages, but not others. The ultimate
test would therefore be to take a given constraint ranking, apply it to some
complete lexicon of words, and see whether we can raise a population using
this language.
6.7. Exercises 121
Such an experiment is of course not feasible, but linguistics sometimes
use approximations to it, for instance teaching artificial (miniature) languages
to (adult) speakers, displaying the required pattern and compare the way in
which those speakers acquire patterns that could not be generated by any
ranking of the constraints. (More on artificial language learning was said in
section 1.2.)
6.7 Exercises
1. Consider the following forms in Yoruba, and provide an analysis in
terms of the constraints given in the chapter. Give a constraint ranking
and a tableau for the first form, with some of the reasonable candidates.
For the purposes of this exercise, you may ignore what happens to tones.
/bu ata/ [bata] ‘pour ground pepper’
/g´e ol´u/ [g´ol´u] ‘cut mushrooms’
/ta epo/ [tepo] ‘sell palm oil’
2. Consider the following forms in Diola Fogny, and provide an analysis in
terms of the constraints given in this chapter. Give a constraint ranking
and a tableau for the first form, with some of the reasonable candidates.
/let ku jaw/ [lekujaw] ‘they won’t go’
/ujuk ja/ [ujuja] ‘if you see’
/kobkoben/ [kokoben] ‘yearn’
3. Consider the following forms in Lebanese Arabic, and provide an analysis
in terms of the constraints given in this chapter. Give a constraint
ranking and a tableau for the first form, with some of the reasonable
candidates.
/Pism/ [Pisim] ‘name’
/Pibn/ [Pibin] ‘son’
/Sigl/ [Sigil] ‘work’
4. Consider the following forms in Samoan, and provide an analysis in
terms of the constraints given in this chapter. Give a constraint ranking
and a tableau for the first two forms, with some of the reasonable
candidates.
[olo] ‘rub’ [oloia] ‘rub (perfective)’
[sopo] ‘go across’ [sopoPia] ‘go across (perfective)’
[aNa] ‘face’ [aNaia] ‘face (perfective)’
[asu] ‘smoke’ [asuNia] ‘smoke (perfective)’
[tau] ‘repay’ [tauia] ‘repay (perfective)’
[tau] ‘cost’ [taulia] ‘cost (perfective)’
5. Consider the following forms in Turkish, and provide an analysis of the
consonant alternation in terms of constraints. You may have to form
a new constraint, modeled on the constraints you have seen. Give a
constraint ranking and a tableau for the first four forms, with some of
the reasonable candidates.
122 6.7. Exercises
Nominative Dative
[sap] [sapa] ‘stalk’
[elmas] [elmasa] ‘diamond’
[ev] [eve] ‘house’
[tat] [tada] ‘taste’
[at] [ata] ‘horse’
[deniz] [denize] ‘sea’
[kap] [kaba] ‘container’
[masraf] [masrafa] ‘expense’
6. Draw tableaux for the dataset in (221), taking also the hypothetical forms
in (222) into the datasets.
7. Draw a tableau for a hypothetical input /kapta/ in Japanese, assuming
it comes out as [kappa] (cf. the dataset in (225)).
8. How many different constraint rankings do we get with 7 constraints? If
we call them A, B, C, D, E and F and we assume that (only) the relative
ranking of A and B does not matter, in the sense that A B always gives
the same language as B A, how many different languages are produced
by these different rankings?
9. There are approximately 7,000 languages in the world. With 8 constraints
we can generate more than 40,000 different rankings. Already
within this chapter we introduced more than 8 constraints, which however
do not suffice to describe all phonological phenomena in languages
(let alone all linguist phenomena). One might therefore claim that the
theory predicts too many different languages. Discuss.
10. On page 115, it is claimed that it does not make a difference for Axininca
Campa whether we assume that ONSET NOEPENTHESIS NODELETION
or NOEPENTHESIS ONSET NODELETION. Show that this is correct,
by showing the tableaux for some relevant examples.
11. Korean has a both plain and aspirated stops. Consider the following
table, and use an (adapted version of) a constraint from the main text to
give an analysis, as well as a typology which includes at least German
and English as well.
pat ‘field’ path
e ‘on the field’
tat ‘close’ tath
@ ‘to work’
pu@k ‘kitchen’ pu@kh
i ‘in the kitchen’
sak ‘old’ saka ‘to be old’
Are any other languages also predicted by your typology? If yes,
what kind of patterns would you find in such languages? (It would be
even better if you could name a concrete language, of course.)
12. It has been claimed in this chapter that V CV sequences in all languages
tend to be syllabified as V.CV . However, in English a word like inadequate
gets syllabified as in-a-de-quate. What could the reason be? Invent
a constraint which could do the right job. Place it in a constraint hierarchy
and draw a tableau to show how your analysis would work.
13. Given an analysis of Belorussian centropetal reduction (in section 2.4 on
page 35) in terms of Optimality Theory. You will have to invent your
own markedness and faithfulness constraints.
14. It has been claimed in the literature that there are languages which have
syllable-final devoicing (as we have seen), but no languages with syllableinitial
devoicing. How is this related to our observations on ONSET and
NOCODA in this chapter?
6.8. Sources and further reading 123
15. Suppose we want to test the claim of the previous exercise that there
are languages which have syllable-final devoicing (as we have seen),
but no languages with syllable-initial devoicing. Set up a small artificial
language experiment to study this claim: create two small artificial
languages showing the two patterns (but otherwise the same).
(You can now also test these two languages if you have sufficiently
large groups of, say, 5 people each, willing to try to learn them. Do they
succeed applying the rule on new forms?)
6.8 Sources and further reading
Section 6.1. Anderson (1985b) shows how phonological theory throughout its
history has been involved with the study of representations as well as with the
study of computation. The subdivision between underlying and surface representations
has been introduced by early generative phonology, in particular
Chomsky and Halle (1968).
Section 6.2. Prince and Smolensky (1993) is the classic text on Optimality
Theory; Kager (1999a) gave the first introductory text, nicely summarizing
the classical version of the theory up until the moment of the publication of
that book for beginning students. Another nice textbook — also talking more
generally about doing advanced phonological research — is McCarthy (2008).
Final Devoicing has been the topic of debate for many languages. Iverson
and Salmons (2011) gives a nice overview.
Section 6.3. Nasal place assimilation is a type of local assimilation, and as
such it has been described by Zsiga (2011). The data from the Hellendoorn
dialect are taken from Nijen Twilhaar (1990).
Section 6.4. The Fijian data are from Kenstowicz (2007). The observation
that VCV is syllabified as V.CV in all languages is made in various places, for
instance in Charette (1991). McCarthy and Prince (1993a) also discuss this, as
well as the data on Axininca Campa. The standard source on Lenakel data is
Lynch (1974).
Section 6.5. The Japanese data are from Ito (1986), the Ponapean data from
Rehg and Sohl (1981), and the Diola Fogny data from Sapir (1965). The idea
of constraint families originates with McCarthy and Prince (1993b); I do not
think that the particular family organisation here has been proposed in the
literature, but the observation that the various constraints resemble each other
has been made more often.
Chapter 7
Stress
7.1 Languages with stress
So far, we have seen that languages can organize features into segments and
segments into syllables. Although it has sometimes been claimed that the arguments
for syllables are weaker in some languages than in others, it seems
fairly uncontroversial to assume that all languages share this kind of organ-
isation.
We could now wonder whether there is also any kind of higher order organisation.
Does a speaker of a human language simply utter a string of syllables,
one after the other, or are these syllables in turn organized into higherorder
units?
At least certain languages seem to give evidence for this. These are languages,
such as English, which have stress: one syllable stands out among the stress
other syllables as being particularly prominent. It is not always clear what the
exact phonetic correlates of this ‘prominence’ are — it is a partly languagespecific
mixture of higher pitch, longer duration and longer intensity — but
speakers will agree which syllable in a word is more prominent. This syllable
can be called the head of the word. head
A proper subset of these languages also has more organisation of the word,
in the sense that some syllables in longer words have secondary stress, i.e. they secondary stress
are not as prominent as the head syllable, but more prominent than other
syllables in the word. This secondary stress is often rhythmic: stressed and
unstressed syllables tend to alternate each other.
Not all languages show evidence for stress, and of those which do, not all
languages show also secondary stress. Such languages typically have other
ways of organizing the word, e.g. by certain autosegmental tonal patterns.
In any case, there seems little doubt that something like a word plays a role word
in the phonological organisation; this word may not always be exactly what
people write in between spaces, or as separate characters, but is is remarkably
often something coming close to that. In stress languages (on which we will
concentrate in this chapter), there furthermore is evidence that such words
have a hierarchical internal structure: like a syllable has a nucleus, a word has a hierarchical internal structure
head syllable. head syllable
Secondary stress furthermore gives evidence for a further level of organisation,
in between the syllable and the word: that of the foot. The heads
125
126 7.2. Metrical feet
of feet surface as having secondary stress, and the feets themselves have one
head, the head foot. The syllable which is the head of this head foot has the
primary stress. Thus in the English word encyclopedia, pe is the head of the
main foot, and en is the head of a secondary foot, whereas all other syllables
are unstressed.
7.2 Metrical feet
The notion of a foot is derived from the study of classical metrics, the study offoot
rhythm in verse; it has been extended to the study of the rhythmic grouping
of syllables within the word. In English (as well as Dutch) poetry, poetic feet
are usually bisyllabic, they consist of two syllables. One of these two is more
prominent than the other, and this gives us two options:
In the first option, the first syllable is the most prominent one; we then
have a trochee:trochee
(231) s w s w s w s w
(Ón thˇe) (shóre stˇood) (Hí- ˇa)- (wá- thˇa)
s w s w s w s w
(Túrned ˇand) (wáved hˇıs) (hánd ˇat) (pár- tˇıng)
(Henry Wadsworth Longfellow, Hiawatha)
In this example we see some useful notation illustrated. Accented syllables
are denoted by an accent (á), unaccented ones by a breve symbol (ˇa). Furthermore
we put an s (for strong above the accented syllable and a w (for weak)
above an unaccented one. The brackets indicate that the syllables are grouped
in a foot.
The other possibility is that the second syllable is the most prominent one;
we then have an iamb:iamb
(232) w s w s w s w s w s
Now is the win- ter of our dis- con- tent
w s w s w s w s w s
Made glo- rious sum- mer by this sun of York;
(William Shakespeare, Richard III)
Iambic and trochaic feet are the most important building blocks in the stress
systems of most (stress) languages as well as in poetry. As we have already
indicated, feet are different from all other levels of phonological organization
(segments, syllables, words) in one important way. Although it is hardly
ever contested that all languages have features, segments and syllables, there
is quite a number of languages for which there is no evidence for metrical
feet; for these languages it cannot be said that some syllable is systematically
stronger than its phonological neighbours.
Languages which do have feet, however, very often choose to have either
iambic or trochaic feet. This seems furthermore to be a choice which is made
within a language once and for all; there might be no languages in which the
two types of feet are mixed.
Pintupi, a Pama-Nyungan language of Australia is a typical example of
a language with trochaic feet. In our linguistic transcription, we only note
7.2. Metrical feet 127
stress, by an accent marker " placed in front of the syllable with primary stress,
and placed in front of a syllable with secondary stress:
(233) a. "σσ "païa ‘earth’
b. "σσσ "tj
uúaya ‘many’
c. "σσ σσ "maía wana ‘through from behind’
d. "σσ`σσσ "puíiN kalatj
u ‘we (sat) on the hill’
e. "σσ σσ σσ "tj
amu lımpa tj
uNku ‘our relation’
The notation which we use in these examples is convenient because it is compact.
However, many phonologists really think of these structures in terms of
trees. The form in (233d), for instance, can be pictured as follows: trees
(234)
ω
ˆˆˆˆˆ
F F
rrr
rrr
σ σ σ σ σ
"pu íiN ka la tj
u
The straight lines here represent ‘heads’ — the most prominent members in a head
constituent — whereas slanted lines represent ‘dependents’ — less prominent dependent
members. Furthermore, F abrreviates ‘Foot’, and ω is often used in the literature
for the phonological word. Thus pu is the head of the (trochaic) foot
puíiN, and this foot is in turn the head of the whole word. For this reason, pu
gets most stress in the word (primary stress), whereas "ka (the head of a foot
which is not the head of the word, viz. the foot kala) gets less stress (secondary
stress) and the other syllables get no stress at all.
In a word with an odd number of syllables, such as the one in (234), there
will be one syllable which does not participate in the foot structure at all: it
is unfooted. Languages can choose where they leave their unfooted syllable, unfooted
but usually this will be at one of the two edges of the word: in Pintupi, this
is the righthand edge of the word. In other languages, such as MalakMalak
(another Australian language, spoken in Western Arnihem), it is the left edge
of the word:
(235) a. "wuru ‘arm, rivulet’
b. mel"papu ‘father (emphatic)’
c. "munan kara ‘beautiful’
d. ar"kini yaNka ‘we are all going to stand’
e. "nönö rönö yunka ‘you (pl) will lie down’
From studying the words with an even number of syllables, it is again easy to
see that the language has trochaic feet: stress is always on the first syllable
in such words, and then alternates. Furthermore also in MalakMalak, the
head foot is the first syllable. However, in words with an odd number of
128 7.2. Metrical feet
syllables, the languages differ. One syllable is left out of the template, and
in MalakMalak this is the first, whereas in Pintupi it is the last. As far as
we know, these are the main options; there are no languages which leave for
instance the syllable right in the middle of the word unparsed.
We thus have distinguished two axes along which languages may vary:
(236) a. iambic feet vs. trochaic feet
b. first syllable vs. last syllable unfooted in words with an odd number
of syllables
This gives us a miniature typology of four different kinds of languages: we
expect two types of iambic languages as well as the two types of trochaic
languages we have seen. Creek is a famous example in the literature of a
language with iambs.
(237) a. co"ko ‘house’
b. a"mifa ‘my dog’
c. apata"ka ‘pancake’
d. anoki"cita ‘to love’
e. isimahici"ta ‘one to sight at one’
Again, that we are dealing with an iambic system rather than with a trochaic
one, is something we can most easily see in words with an even number of
syllables; a word with two syllables simply has an iambic pattern, which is
most easily explained if we assume that it consists of one iambic foot. Similarly,
also words of four syllables have stress on the final syllable, which makes
us assume that there must be two feet there, even though the secondary stress
on the first foot is not noted on the data we have.
From the odd-numbered syllable words we can furthermore learn that it
is the last syllable of the word which is unfooted. For some reason, this seems
to be the option which is chosen by most iambic languages; as a matter of
fact, some scholars believe that all iambic languages choose to leave the final
syllable unfooted rather than the initial one.
One potential example of an iambic language leaving the first syllable unparsed
is Weri, but the data for this language are rather sketchy:
(238) a. Nin"tıp ‘bee’
b. kuli"pu ‘hair of arm’
c. u lua"mit ‘mist’
d. aku nete"pal ‘times’
Again, the words with an even number of syllables give us a good indication
of the foot structure. They have the same shape as Creek words, although
in this case we do have evidence for secondary stress as well. It is the oddnumbered
words which show a different pattern, and one which makes us
believe that the first syllable might stay unfooted.
It is not clear at present why there are no well attested examples of this
type of language; given the typology suggested above, we would expect four
7.3. Syllable quantity 129
kinds of languages, but in actual practice only three seem to be attested so
far. Languages seem to prefer to have at least some stress at one of the first
two syllables of every word; maybe this is because stress is often used as a
demarcation of the edges of words. Because the speech signal is uninterrupted, demarcation
the listener needs cues as to where the word boundaries are in order to be
able to make sense of what she hears. Stress can be one such cue that a new
word has begun. Allowing words to start with two unstressed syllables might
makethis task too complex.
Leaving this problem aside for a moment, we now have to translate our
different options into OT constraints. One way of doing this is the following:
(239) a. i. ALIGN(Foot, Left, σ, Left): The left edge of a foot should be
aligned with the left edge of the head syllable (so: the heads
are on the lefthand side, feet are trochaic).
ii. ALIGN(Foot, Right, σ, Right): The right edge of a foot should
be aligned with the right edge of a syllable (so: the heads are
on the righthand side, feet are iambic).
b. i. ALIGN(Word, Left, Foot, Left): The left edge of a word should
be aligned with the left edge of a foot (so: no unfooted syllables
at left edge).
ii. ALIGN(Word, Right, Foot, Right): The right edge of a word
should be aligned with the right edge of a foot (so: no unfooted
syllables at right edge).
These constraints are instances of some more general family of constraints,
aligning phonological and morphological edges to each other. You can see
that there would be a general template which these constraints satisfy:
(240) ALIGN(X, Left/Right, Y, Left/Right)
It will be left as an exercise to formulate more instances of this particular constraint
schema.
7.3 Syllable quantity
In the languages we have considered so far, all syllables are treated equally.
This is a pattern that we find quite often among stress languages. However, in
a substantial number of the world’s languages, stress is quantity sensitive: the quantity sensitive
stress system looks at the structure of syllables and distinguishes between (at
least) two types of them: heavy and light syllables. The distinction is usually heavy
lightconnected to the structure of the rhyme in the following way:
(241) In heavy syllables, there are (at least) two positions in the syllable
rhyme; in light syllables, there is only one position.
Having two positions in the rhyme means having a long vowel, a diphthong
or being a closed syllable. Variations on this theme are also possible. For
instance in certain languages, syllables are heavy iff they are closed by a consonant
of a certain type, and light otherwise.
130 7.3. Syllable quantity
A famous example of a language with a quantity-sensitive system is the
Uto-Aztecan language Tübatulabal. In this language, the distinction between
light and heavy syllables is made in the following way:
(242) In heavy syllables, vowels are long; in light syllables, vowels are short.
You may check for yourself how this distinction can be seen as a special case
of (241) if we base ourselves on autosegmentalist assumptions on the skeleton.
Consider the following data (we do not distinguish between primary and
secondary stress in these examples because we are only interested in foot
structure):
(243) a. i ponih win ‘of his own skunk’
b. wi taNha tal ‘the Tejon Indians’
c. wıtaN hata la:ba cu ‘away from the Tejon Indians’
d. yu: du: yu: dat ‘the fruit is mashing’
e. ta: hawi la:p ‘in the summer’
f. wa ša: gaha ja ‘it might flame up’
g. ana Ni: nin1 mut ‘he is crying wherever he goed (distr.)’
h. p1 t1p1 t1:di nat ‘he is turning it over repeatedly’
One thing which strikes us if we study these examples, is that all long vowels
are stressed. This is the reason why we say that stress is quantity-sensitive in
this language: the ‘normal assignment’ of feet gets interrupted by the requirement
that heavy syllables want to be stressed.
In grammatical terms this can be seen as a result of a constraint which is
usually called WEIGHTTOSTRESS:
(244) WEIGHTTOSTRESS: Heavy syllables should be stressed.
This constraint has a very high ranking in Tübatulabal grammar — it is never
violated. Another observation we can make is that a light syllable before a
heavy syllable stays always stressless, whereas light syllables following them
are sometimes stressed. This is an indication that we are dealing here with
an iambic system: light syllables tend to go into feet with a head on their
righthand side. A similar conclusion may be drawn from the first two words,
in which there is no heavy syllable at all, and in which the stress pattern is
weak strong weak strong (abbreviated as wsws).
In other words, the language seems to have a basic iambic pattern, and we
may assume that also the constraint in (239a-ii), repeated here, is operative:
(245) ALIGN(Foot, Right, ´σ, Right) (henceforth abbreviated as IAMB)
Together, these two constraints will give analyses such as the following, in
which we placed feet in between parentheses:
(246) a. (i po)(nih w1n)
b. (wi taN)(ha tal)
7.4. Lexical stress: faithfulness to feet 131
c. (ha ni:)la
d. wi(taN ha)(ta la:)(ba cu)
The last two examples are not in accordance with the facts of (243), as you can
check for yourself: we predict the last syllable in (246c) and the first one in
(246d) to be stressless, but this is not the case. They are stressed.
The difference between Tübatulabal and the languages we have seen so far
is that in the latter all feet need to be binary: they need to have both a head and
a dependent. In words with an odd number of syllables, the one ‘remaining’
syllable stays outside of the foot structure and is unstressed.
For Tübatulabal,a foot can also only have a head and no dependent. Formally,
Creek and the other languages have a high-ranking constraint on foot binarity:
foot binarity
(247) FOOTBIN: A foot needs a dependent.
In Tübatulabal, this constraint is dominated by another constraint, which is
violated in the other languages:
(248) PARSE-σ: Every syllable needs to be parsed into a foot.
We thus get the following typology for iambic languages (something similar
could be done for trochaic languages):
(249) a. Tübatulabal: PARSE-σ FOOTBIN
b. Creek (and Weri): FOOTBIN PARSE-σ
The difference between the two types of languages wil only be seen in words
with an odd number of syllables. In Tübatulabal, the remaining syllable has
to be put in a foot, even if that foot is less than perfect as a result. In Creek,
one prefers to keep all feet binary, even if that leads to the one syllable being
left out of foot structure altogether.
Notice that we can see from examples such as (243d) that FOOTBIN is indeed
lowly ranked in Tübatulabal: this word consists exclusively of feet which
have only one syllable. The reason for this is of course that all syllables (but
the last one) are heavy. In other words, this piece of data provides us with
evidence that WEIGHTTOSTRESS FOOTBIN, but also that IAMB FOOTBIN
(because otherwise we could have solved our problem by making the last two
syllables of (243d) into one foot). All in all, we thus have established the following
miniature constraint ranking for Tübatulabal stress:
(250) PARSE-σ,WEIGHTTOSTRESS,IAMB FOOTBIN
7.4 Lexical stress: faithfulness to feet
Word stress in Modern Greek is quite puzzling at first. We may observe that
stress can be on many different syllables of the word:
(251) a. last syllable: ura"nos ‘sky’
132 7.4. Lexical stress: faithfulness to feet
b. penultimate syllable: ku"baros ‘godfather’
c. antepenultimate syllable: "anTropos ‘man’
How are we going to account for this lexical variation? An obvious answer to
this is: apparently Greek has feet already present in the underlying representation,
and a strong faithfulness requirement on underlying foot structure:
(252) FAITHFOOT: Do not delete underlying feet.
Suppose there are reasons to assume that Greek feet are trochees, and furthermore
that pyjama (penultimate stress) represents the default. These reasons
are manifold; one of them is language acquisition, in which children tend to
regularize the other patterns to this one. This gives us the following ranking:
(253) TROCHEE FAITHFOOT ALIGN(Word, Right, Foot, Right), FOOTBIN
We get the following tableaux for our three example words (leaving out candidates
without trochees):
(254) a.
ura("no)s FAITHFOOT ALIGN FOOTBIN
("ura)nos *! *
u("ranos) *!
ura("nos) *
b.
kubaros FAITHFOOT ALIGN FOOTBIN
("kuba)ros *!
ku("baros)
kuba("ros) *!
c.
("anTro)pos FAITHFOOT ALIGN FOOTBIN
("anTro)pos *
an("Tropos) *!
anTro("pos) *! *
Note that it is not necessary to posit an underlying foot for the default stress
structure kubáros. This is what it means to be default: the grammar will assign
the appropriate structure without instructions from the underlying form. (But
note that it would do no harm to assign underlying structure either.)
The constraint TROCHEE is ranked most highly since there is no evidence
that there is ever an iambic structure in Greek. Even words such as uranós or
are analysed as (ùra)(nós).
Still, not everything is possible. One observation to be made is that Greek
— like many other languages — displays the effects of a so-called three-syllable
window: stress is on one of the last three syllables of the word, but never outside
it. In other words, (monomorphemic) forms of the following type are
unattested in Modern Greek:
(255) *"makaroni
7.5. The moraic theory of syllable structure 133
The reason for this is straightforward. If we posit an underlying structure
(máka)roni, the last two syllables are still unfooted. We can then parse these
two into a new foot, which will receive primary stress, because this is always
on the last foot of the word in Greek. This makes (máka)roni different from
(ánTro)pos, where there is no room to build an extra binary foot.
7.5 The moraic theory of syllable structure
There is a popular alternative to the representations of syllable structure that
we have seen so far. Under this conception, the syllable does not consist of an
onset and a rhyme, but of two mora’s (from the Latin word meaning ‘a short mora’s
period of time’ or ‘delay’). The main generalisation underlying this theory is
the following:
(256) a. Heavy syllables consist of two mora’s
b. Light syllables consist of one mora
In other words, if both long vowels and coda consonants count, mora’s are the
same as positions in the rhyme. However, we can also model other kinds of
languages using mora’s.
Suppose we are dealing with a language in which closed syllables and
syllables with a long vowel are heavy, whereas other syllables are light. We
can represent syllable structure in this language in the following way:
(257) a. light b. heavy c. heavy
σ



µ
C V
σ



dd
µ µ
C V C
σ



dd
µ µ
  
C V
In a language in which only long vowels count as heavy, on the other hand,
we get the following structures:
(258) a. light b. light c. heavy
σ



µ
C V
σ



µ
dd
C V C
σ



dd
µ µ
  
C V
It is usually assumed that the mora’s take the position of skeletal points; the
C’s and V’s in this figure represent root nodes. This means that the phonological
timing in this model is slightly different from that in the theory we have
developed in earlier chapters, the one based on a skeleton with x-slots: onset
consonants do not count for timing, for instance.
134 7.5. The moraic theory of syllable structure
One language to which moraic analysis has been applied quite succesfully,
is Japanese. As a matter of fact, the mora, called haku in Japanese, plays an important
role in traditional Japanese linguistics. For instance, in the ‘phonetic’
part of Japanese spelling, heavy syllables are represented by two symbols,
whereas light syllables are represented by one. Traditional Japanese poetry
(like haiku) is also based on counting 5+7+5=17 mora’s (rather than syllables,
as in Western renditions of haiku).
It is generally considered true that moraic theory solves two problems of
traditional syllable structure/stress analysis. In the first place, weight usually
refers to coda consonants and not to onset consonants (as in our discussion of
Tübatulabal stress).
In the second place, compensatory lengthening of vowels is claimed to
be always the result of the deletion of coda consonants, and never of onset
consonants. So in in our discussion of the history of Germanic in section 5.2
on page 86, we saw that e.g. gans can correspond to ga:s: the n gets deleted
and the a takes its place. The inverse does not happen: when a consonant in
the onset gets deleted in some language, vowel lengthening is not the result
(so there are no languages where onset deletion turns gans into a:ns.
Moraic theory provides us with a formal language which can link these
observations and express them in a uniform way. As to weight, we can posit,
that some languages build stress feet on syllables, whereas others build them
on morae (more on this below). Compensatory lengthening can now be described
in terms of mora preservation: if a coda consonant is lost, it may leave
a mora behind, which will then be filled by the vowel. However, if an onset
consonant is lost, there is no resulting mora, and hence no possibility for onset
loss.
It should be noted, that there is some discussion in the literature on the
validity of both of these claims. For instance, there are a few languages for
which it seems to be true that onsets count for weight. One famous instance
of this is Pirahã, which has the following stress rule:
(259) Stress the rightmost heaviest syllable of the last three syllables of the
word.
Like many other languages, Pirahã thus displays a three-syllable window at
the end of the word: it is as if the stress assignment process only look at those
last syllables, Within, this window we choose the heaviest syllable, where the
notion of ’heaviness’ is defined according to the following hierarchy:
(260) PVV > BVV > VV > PV > BV (> V)
(P = a voiceless plosive, B = a voiced plosive; a > b means a is heavier
than b)
The notion of weight is thus fairly complex in Pirahã, but it can be decomposed
into the following:
(261) a. long vowels are heavier than short vowels
b. syllables with an onset are heavier than syllables without an onset
c. syllables with a voiceless onset are heavier than syllables with a
voiced onset
7.5. The moraic theory of syllable structure 135
Here are a few examples illustrating these effects (I leave out tone markings
and note stress with an acute accent):
(262) a. "kao.ba.bai ‘almost fell’
b. "kaa.gai ‘word’
c. "bii.ao.ii ‘tired’
d. pia.hao.gi.so."ai.pi ‘’
e. "Pa.ba.gi ‘toucan’
f. Pa.ba."pa ‘Amapá’
g. ho."ao.ii ‘shotgun’
h. pao.hoa."hai ‘anaconda’
i. ti."po.gi ‘species of bird’
(261b) is relevant to our present discussion in particular; it shows that at least
in some languages onsets do seem to be relevant to the calculation of syllable
weight — there is a handful of languages for which a similar claim has been
made.
Compensatory lengthening may also be attested, albeit again in marginal
cases. A rather well-known example is the Samothraki dialect of Greek, where
deletion of an onset /r/ may result in lengthening of the preceding vowel:
(263) /roGa/ → [o:Ga], /riz/ → [i:z], /rema/ → [e:ma], /roGa/ → [o:Ga],
/ruxa/ → [u:xa], /rafts/ → [a:fts]
(In spite of the phonological notation, it may not be clear that we are dealing
with a synchronic process in this case; the underlying representations here as
a matter of fact represent Standard Greek and other dialects, but we have no
a priori evidence that these are also the underlying representations for Samo-
thraki.)
What are we going to do with this type of evidence? A reasonable first
approach might be to be very sceptical about it: if our theory forbids it, and
the data are so rare, maybe there is something wrong with the sources we
have.
However, in this case, this line of attack will not work. In the first place,
the Pirahã data stem from afield worker who has spent a large amount of time
on his work on this particular language. The Greek dialectological data might
be a bit more shaky, but they have been confirmed by some other speaker. In
the second place, it is not really true that our theory ‘forbids’ these facts; there
is nothing very deep inherent to any of the theories presented thus far which
would disallow onsets to carry morae.
But if this is the case, we are dealing with a typological puzzle: why are
data of the Pirahã/Samothraki Greek type so rare as compared to similar effects
with coda’s? The answer to this might fall outside of the domain of
formal linguistics proper: it might have something to do with the phonetic
perceptability of codas vs. onsets, for instance.
But given all this, it becomes less clear that mora theory is really superior
to the more traditional theories we have seen. If the two reasons why it
is introduced in the first place do not really seem to fall within the realm of
136 7.6. Stress typology
formal phonological analysis proper, mora theory mainly becomes a convenient
notation to talk about the interaction between syllable structure and stress.
As an example of such a notational property, consider the following. The
representation of short vs. long vowels will be as in (264a) within mora theory;
the representation of short vs. long consonants will be as in (264b):
(264) a. short long
µ
V
µ µ
  
V
b. short long
C
µ
C
One observation which is nicely represented by these pictures is that geminate
(long) consonants do not occur in onsets — although, again, there seem to be
a few exceptions. Typically, a long consonant will be attached to the coda of
one syllable, and the onset of the next one:
(265) [at:a]
σ σ
dd 



µ µ µ
a t a
Most languages allow for only monomoraic or bimoraic syllables; syllables withmonomoraic or bimoraic syllables
one or two morae. This means that long vowels could not be followed by long
consonants. The following example is from Koya:
(266) /ke: t: o:ïãa/ [ket:o:ïãa] ‘he told’
/o:t:o:ïãu/ [ot:o:ïãu] ‘he brought’
These facts can be understood under assumption of the representations in
(265), plus a requirement that Koya syllables have at most two morae, and
the idea that nongeminate consonants are never moraic in Koya — this explains
why the vowel before the cluster [ïã] does not have to be shortened.
Note that especially the latter fact is more difficult to express in a nonmoraic
framework.
7.6 Stress typology
Determining where a language fits in our typology
In this chapter, we have introduced a simple model of stress typology. The
core of the foot typology are a number of binary distinctions:
7.6. Stress typology 137
(267) a. Feet are left-headed (trochaic) or right-headed (iambic)
b. Feet are necessarily binary / can be unary
c. Unfooted syllables (if any) or unary feet (if any) appear on the left
/ on the right
d. Feet are quantity sensitive / quantity insensitive
e. The head foot appears on the left / on the right
f. Stress is lexically determined / predictable
g. Stress is determined on the whole word / on the last three syllables
Because each of these properties is binary, this gives us a rather vast space of
possibilites. If we want to classify an individual language, we have to navigate
this space somehow. It is important for this to have enough words, of various
lengths and with various different types of syllables.
A good way to proceed is as follows. First you try to determine whether
the language has iambs or trochees. For this you collect as many words with
an even number of syllables as you can find. If they all have a uniform stress
pattern, stress is probably not lexically determined and also not quantity-sensitive.
It becomes then fairly easy to see whether the feet are iambic or trochaic.
If the even-numbered syllables word do not show equal stress on all words,
you should first check whether stress is always on syllables which are heavy,
i.e. whether they have a long vowel or are closed by a consonant. If it looks
like this is indeed the case, the best is to first concentrate on even-numbered
words with only light syllables. This will tell give you more information about
what the ‘normal’ foot structure is. You can then assume that the feet in words
with heavy syllables will be of the same basic type (iambs or trochees).
If it is not clear that heavy syllables attract stress and if the stress seems
even randomly distributed in words with an even number of light syllables,
the language probably has lexically determined stress. If that is the case, the
analysis basically stops, although you may still want to check on long words,
whether stress is always at least on one of the last three syllables, or occurs
elsewhere.
If stress is not lexically determined, you can then proceed to words with an
odd number of syllables. First you check how many stressed syllables there
are. Suppose the word has n syllables; if the number of stresses is (n − 1)/2
(so for instance a word with 5 syllables has 2 stresses), it seems likely that
the language only has binary feet, and the remaining syllables are unfooted.
Because you already know what the type of feet are, you should now be able
to see whether unfooted syllables appear on the left or on the right.
If words with an odd number of syllables n have (n+1)/2 stressed syllables
(so a word with 5 syllables has 3 stresses), it is more likely that the language
allows unary feet. Again, given that you know what ‘normal’ binary feet are
like, you should be able to determine where this ‘extra’ unary feet is placed.
The only final parameter you have to determine, regardless of what the
foot type is, and whether or not the language is quantity-sensitive or lexically
determined, is which of the feet gets the main stress. You may safely assume
— at least in the case of the exercises to this chapter! — that this is either the
first or the last foot. Other options have sometimes been shown in languages
of the world, but they are rare and will not be taken into account here.
138 7.6. Stress typology
Some typological gaps
Hayes (1987, 1995) claims that one of the four basic types of feet (iamb vs. trochee,
quantity-sensitive vs. quantity-insensitive) which we would expect to
exist is typologically inexistent: there are no quantity-insensitive iambs. On
the other hand, most trochaic languages seem to be also quantity-sensitive.
Hayes connects this to a psycholinguistic finding (in particular Woodrow,
1909). If we expose informants to a signal ta-ta-ta-ta-ta-ta-..., and we alternate
the intensity of the ta’s, listeners will tend to group them in a trochaic fashion;
that is to say, they will tend to hear ...(táta)(táta)(táta).... On the other hand,
if we keep the intensity constant, but alternate the length of the vowels, the
listeners will tend to group the sounds as ...(tata:)(tata:)(tata:)... . The conclusion
of this is that the difference between foot types is partly determined by
the Iambic/trochaic law (Bolton, 1894):Iambic/trochaic law
(268) Iambic/trochaic law:
a. Elements contrasting in intensity naturally form groupings with
initial prominence (trochees).
b. Elements contrasting in duration naturally form groupings with
final prominence (iambs).
Trochees now should be constituents which consist of two elements with roughly
the same duration. There are two types of these, according to Hayes: we can
build feet on the basis of morae, or on the basis of syllables. In the former case,
we have a type of quantity sensitive system:
(269) Moraic trochees
Ft
dd
σ σ
µ µ
Ft
σ
dd
µ µ
In a system with moraic trochees, heavy syllables will form a foot of their
own, whereas light syllables will be grouped together. An example of this is
so-called ‘Egyptian Radio Arabic’, also called ‘Cairene Arabic’. In bisyllabic
words, stress is on the last syllable if it is (super)heavy, and otherwise it is on
the first syllable:
(270) a. Last syllable (super)heavy: sa"la:m ‘peace’ di"mašq ‘Damascus’
b. Last syllable light: "malik ‘king’ "huna ‘here’
Another possibility is to build trochees on syllables, disregarding the internal
structure. We then get a quantity-insensitive trochaic structure:
(271) Syllabic trochees
Ft
dd
σ σ
7.6. Stress typology 139
An example of this is Icelandic, where primary stress is on the first syllable of
the word, and secondary stress alternates:
(272) h"öfðing ja ‘chieftain (gen.pl)’, "akva rella ‘aquarelle’, "bíóg rafí a ‘biography’
Yet in iambs, the requirement is that the two parts of the foot are uneven in
length, and we have only one canonical foot type:
(273) Iambs
Ft
  
σ σ
dd
µ µ µ
An example of this is Tübatulabal, the language we have discussed already. It
can furthermore be observed that many languages which use iambic feet have
some rule of lengthening vowels and/or consonants to satisfy requirements
on foot structure. An example of this is provided by Menomini, a Central
Algonquian language. In this language, when a word begins with two light
vowels underlyingly, the vowel of the second syllable is lengthened; this can
be understood if we assume that these first two syllables are grouped into an
iamb:
(274) a. /ahsama:w/ → [ahsa:ma:w] ‘he is fed’
b. /netahsama:w/ → [neta:hsama:w] ‘I feed him’
It thus seems that we would have to relax our typology of feet some more to
also include a preference for uneven trochees.
More in general, it seems that the typology is more lenient for trochees
than for iambs. We could also observe that the number of attested (and wellunderstood)
trochaic systems is much larger than the number of attested iambic
systems; the latter mainly consist of native languages of (North) America. We
could, once again, wonder which conclusions we can draw from these typological
considerations. On the one hand, some might wish to argue that the
relative paucity of iambic systems is just some accident of history, and that,
given this arbitrary historical fact, it is no wonder that there is less diversity in
iambic systems: even syllabic or moraic iambs might be possible in principle,
but we simply have a much smaller opportunity of finding them actually at-
tested.
Alternatively, some have argued that there is a more principled reason
why iambic systems are so few. We could claim, for instance, that iambic feet
are not part of our inventory of possible structures. Iambic languages would
then need an alternative analysis.
140 7.7. Exercises
7.7 Exercises
1. Consider the following examples from the Pacific language Awtuw. What
kind of foot does this language have — iambs or trochees? Which foot
carries main stress?
i "ki.nik ‘sit’
ii ow.ti."ka.yæn ‘old’
iii "wa.ru.ke ‘big’
iv la.pe ‘village’
2. Consider the following examples from the Semitic language Modern
Hebrew. What kind of foot does this language have — iambs or trochees?
Which foot carries main stress?
i ˇga."dol ‘big’
ii bi."ra ‘capital city’
iii ta."am ‘tasted’
iv me. vu.ga."rim ‘adults’
3. Consider the following examples from the Austronesian language Malay.
What kind of foot does this language have — iambs or trochees? Is the
language quantity-sensitive or quantity-insensitive?
i s@n."dar ‘to snore’
ii s@."ma.di ‘concentration’
iii lak."sa.na ‘quality’
4. Consider the following examples from the Malaccan Creole of Portuguese.
What can you say about the stress system of this language?
i "ka.za ‘house’
ii sa."ba.na ‘fan’
iii ka.za."min.tu ‘wedding’
iv min."ti.ra ‘lie’
v o.r@."saN ‘oration’
vi kar."baN ‘coal’
5. Consider the following examples from the Palestian Arabic. What can
you say about the stress system of this language?
i "Sa.ˇa.ra.tun ‘a tree’
ii "ka.ta.bu ‘they wrote’
iii duk."ka:n ‘shop’
iv ba:."be:n ‘two doors’
v "ba:.rak ‘he blessed’
vi "ba:.ra.ko ‘he blessed him’
vii "ka.tab ‘he wrote’
viii ma."ka:.ti.bi ‘my offices’
6. Give OT tableaus for the derivation of stress in three of the TÜbatulabal
examples in (243) (you may choose your own words, except that you
may not choose both of the first two, since these have the same pattern).
7.7. Exercises 141
7. In this chapter it has been claimed that there are no languages which
have both iambs and trochees. Show that the OT constraints presented
here actually predict otherwise. Discuss.
8. Consider the following examples from the Indo-Iranian language Pashto.
What can you say about the stress system of this language?
i "gu.ta ‘knot’
ii gu."ta ‘pochard’
iii "vu.lam.be.d@ ‘he took a bath’
iv tS@r."gu.Ray ‘baby chick’
v sto.ma:n.ti."a: ‘fatigue’
Give a constraint ranking within OT, and draw tableaux for the first two
examples.
9. Reformulate the following constraints as special instances of the schema
(240) on page 129:
• ONSET
• NOCODA
10. Consider the following examples from Hixkaryana (Carbib); try to place
the language inside the typology of moraic vs. syllabic trochees and
iambs:
i ow.to. ho:.na ‘to the village’
ii kha. na:. nih.no ‘I taught you’
iii toh. ku. rj
e:.ho. na:.ha. ša:.ka ‘finally to Tohkurye’
11. Consider the following examples from Fijian (Austronesian); try to place
the language inside the typology of moraic vs. syllabic trochees and
iambs:
i a."to.mi ‘atom’
ii ndai. re."ki.ta ‘bazaar’
iii ndi.ko."ne.si ‘deaconess’
iv mbe:."le.ti ‘belt’
v ta. rau."se.se ‘trousers’
vi mba."sa: ‘bazaar’
12. Consider the following examples from Cahuila (Uto-Aztecan); try to
place the language inside the typology of moraic vs. syllabic trochees
and iambs:
i "ta.ka. li.ˇcem ‘one-eyed ones’
ii "táx.mu. ‘song’
iii "qa:n. ki.ˇcem ‘palo verde (pl)’
13. Try to give an analysis of Cahuila not in terms of mora’s, but in terms
of the typology at the beginning of section 7.6. Where does the analysis
fail? Can you solve this by using ranked constraints?
14. StressTyp2 is a typological database collecting information about stress
patterns in hundreds of languages. The database uses a way to encode
stress which is slightly different from the one used in this book, but you
should be able to understand it. Try to find example languages for the
following patterns:
142 7.8. Sources and further reading
• A language with iambs and main stress on the last foot.
• A language with quantity-sensitive trochees.
• A language with lexicalized stress within a three-syllable window
at the end of the word.
7.8 Sources and further reading
Section 7.2. (Hayes, 1995) is a classical text on metrical stress theory; the Pintupi,
MalakMalak, Creek and Weri data are also from that text.
(Topintzi, 2006)
Chapter 8
Prosodic structure
8.1 The phonological tree
We have seen in previous chapters that phonological words can be represented
by a tree structure: segments are organized into syllables (with some
internal structure), syllables into feet and feet into words. We can draw this as
follows:
(275)
ω
¨¨¨
rrr
F F




σ σ σ σ
This tree structure is usually called prosodic structure in phonological theory. prosodic structure
It is also usually believed that prosodic structure does not end at the level of
the phonological word, and that trees reach higher than this. There is variation
in the literature as to which higher-order levels are actually present, but
it is usually assumed that these involve at least the phonological phrase (φ), cor- phonological phrase
responding roughly to major syntactic constituents such as (large) NPs or the
main predicate VP, the Intonational Phrase (IP), corresponding roughly to sen- Intonational Phrase
tences, and the Utterance (Utt), corresponding, well, to a whole utterance of a Utterance
single speaker.
We thus get a tree structure such as the following:
(276)
Utt


IP IP


φ φ


ω ω


F F

σ σ
143
144 8.2. The Phonological Word
The hypothesis that linguistic utterances are organized in this way gets support
from various different types of evidence, as we will see in this chapter:
these structures play a role in our understanding both of phonological and of
morphological phenomena which seem to refer to them. For instance, we will
see that certain types of phonological processes only work within certain prosodic
categories: they cannot cross the boundaries of such categories. On the
other hand, there are also processes which function only across the boundaries
of phonological structures.
Notice that the tree structures depicted in (276) mimic those which are
used within syntactic theory, but they are simpler in a number of ways: there
is a fewer number of different labels — especially given that the only constituents
that roughly correspond to syntax are the word, the phonological phrase
and the Intonational Phrase —, and secondly, they are not recursive in the
same way as syntactic structures. In syntax, we can have for instance, a sentence
containing another sentence (’[John admits that [he likes milk ]]’). Such
recursion according to most scholars is not found in phonological representa-recursion
tions: a phonological phrase always directly dominates phonological words,
not other phonological phrases.
8.2 The Phonological Word
In the previous chapter, we have already seen one type of evidence for the
existence of a phonological word: the existence of primary stress. In some
languages, there is one syllable which is clearly more prominent than all other
syllables within some domain. That domain does not correspond to the whole
string of syllables spoken by the speaker (that would be the Utterance), or a
longer stretch of it (that might be for instance the phonological phrase), but to
something roughly corresponding to what we might also call a word by other
criteria (for instance morphosyntactic ones).
But there is evidence for the existence of such a constituent also from other
points of view. A well-known example of this is so-called /s/ voicing, which/s/ voicing
we find in nothern Italian dialects. Consider for instance the following cases:
(277) a. isola ‘island’ /isola/ [izola]
b. case ‘houses’ /kas+e/ [kaze]
c. amo Sandra ‘I love Sandra’ /amo+sandra/ [amosandra]
d. asoziale ‘asocial’ /a+sotsiale/ [asotsiale]
e. toccasana ‘cure-all’ /tok:a+sana/ [tok:asana]
Underlying /s/ changes into [z] when it occurs intervocalically, as you can
see in (277a). From data like that in (277b), we know that this also sometimes
happens when one of the vowels belongs to a different morpheme — in this
case a plural ending.
The domain on which the process is defined is thus bigger than the morpheme.
However, it does not apply just anywhere as the following examples
demonstrate. In (277c), the first vowel belongs to a different word than the s.
This means that it is somehow ‘too far away’ for it to be visible. The /s/ is
between two vowels, but it does not voice to a [z].
8.2. The Phonological Word 145
The domain is thus bigger than a morpheme and smaller than a (syntactic)
phrase. This leads us to suspect that it applies within a word. However, the
forms in (277d) (a prefixed word) and (277e) (a compound) show that this
‘word’ does not conform exactly to our understanding of a morphosyntactic
word. Both of these are definitely simple words from this perspective, but s
voicing applies in neither of them.
This is where the notion of a phonological word comes in, a constituent
that is similar to that of a morphosyntactic word, but not always exactly congruent
with it. In this case, we have to posit that the separate parts of a compound
each form a separate ‘phonological word’, as do prefixes. Otherwise,
the boundaries of morphosyntactic words correspond with those of phonological
words. The phonological structures for the forms in (277) are thus as
follows, where brackets indicate phonological word boundaries:
(278) a. [(izola)]
b. [(kaze)]
c. [(amo)(sandra)]
d. [(a)(sotsiale)]
e. [(tok:a)(sana)]
The phonological analysis can now refer to these phonological constituents.
For instance, we may imagine that the constraint responsible for s voicing has
the following shape:
(279) *(...VSV...): Avoid voiceless coronal fricatives between two vowels
within the same phonological word.
It is understood that markedness constraints always are defined on a certain
domain, maybe not universally but on a language-specific basis. It is also
understood that such domains are always phonological, i.e. that phonological
markedness constraints do not refer to morphosyntactic words or phrases.
Of course the way in which a phonological word is constructed itself can
be the result of some constraints. These constraints are called alignment constraints:
they make sure that the edges of morphosyntactic words correspond alignment constraints
to the edges of phonological ones. For instance in our case, we have constraints
such as this one:
(280) a. ALIGN(X0
, L, ω, L): The left edge (L) of a morphosyntactic word
(X0
) should correspond with the right edge (L) of a phonological
word (ω).
b. ALIGN(X0
, R, ω, R): The right edge (R) of a morphosyntactic word
(X0
) should correspond with the right edge (R) of a phonological
word (ω).
c. ALIGN(ω, L, X0
, L): The left edge of a phonological word should
correspond with the left edge of a morphosyntactic word (ω).
d. ALIGN(ω, R, X0
, R): The right edge of a phonological word should
correspond with the right edge of a morphosyntactic word (ω).
146 8.2. The Phonological Word
e. ALIGN(Prefix, R, ω, R): The right edge (R) of a morphosyntactic
word (X0
) should correspond with the right edge (R) of a phonological
word (ω).
f. ALIGN(Suffix, L, ω, L): The left edge (L) of a morphosyntactic word
(X0
) should correspond with the right edge (L) of a phonological
word (ω).
g. ...
By ranking some of these constraints, we can get the right results in (278). I
leave this as an exercise to the reader (exercise 6 on page 145).
In the case of northern Italian dialects, the phonological words are always
smaller than morphosyntactic words or they have the same size. It has been
argued that we can also have phonological words that are (slightly) bigger
than morphosyntactic words. Examples of this we similarly find on the Italian
peninsula, but more to the south, for instance in Lucanian dialect.
The relevant elements in this dialect are so-called clitics, small pronoun-clitics
like elements that behave morphosyntactically as if they are (more or less)
independent words, but that phonologically seem to get integrated with the
verbal stem.
The crucial evidence comes from stress. Lucanian has trochaic feet and
assigns stress by default on the antepenultimate syllable. The same is actually
true for Standard Italian:
(281) ["vinn@] ‘sell!’ (IMP) (Lucanian)
(282) ["p rta] ‘carry!’ (IMP) (Standard Italian)
However, if we add (clitic) pronouns to these forms, the stress in Lucanian
shifts onto the clitics, whereas in Italian it stays on the verb:
(283) [vinn@- "mi-l@] ‘sell it to me!’ (Lucanian)
(284) ["p rta - me - lo] ‘carry it for me!’ (Standard Italian)
The difference is that clitics get integrated into the same phonological word as
the verb, and then stress gets assigned to the antepenultimate position within
that word. In Standard Italian, on the other hand, the phonological word
keeps the size of just the verb, and the clitics stay outside (they may form a
phonological word in their own right).
One can of course wonder what the explanatory value of the phonological
word is precisely. If we set apart prefixes and clitics, we can describe Northern
Italian s Voicing by referring to the remaining material. But saying that this
remainder actually is a constituent called the phonological word and using
this s voicing as evidence for its existence is circular. We might just as well say
that s voicing in Italian applies in a morphosyntactic word, but not across the
boundary with clitics.
This is indeed a criticism that has sometimes been raised against phonological
constituency across the word. A typical answer to this is to show that
different kinds of evidence converge on this particular solution. If there are
many phonological processes which have a similar restriction, we can see that
as an indication that we are onto something.
8.2. The Phonological Word 147
It is well-known, for instance, that Italian is not alone in keeping prefixes
outside of the phonological word. In this respect, prefixes seem to be typologically
different from suffixes; we know quite a number of languages in which
the former are more separate from the stem than suffixes than we know languages
in which this is the other way around, if the former can be convincingly
shown to exist at all.
Another example of the relative independence of prefixes we find in German.
In this language, we consonants can form the onset of a following vowel
across a stem-suffix boundary, but not across the boundary between a prefix
and the stem.
For instance, from the noun Ehr ‘honour’ [e:r] one can derive a verb by
adding the negative verbalizing prefix ent. The resulting word is entehr ‘dishonour’
[Ente:r]. The syllabification of this word is [Ent.Per]: the [t] does not
get syllabified with the following [e]; since every German syllable needs to
have an onset, this vowel gets a default glottal stop instead.
If we inflect this verb further, for instance with the infinitival ending -en,
we get entehren ‘to dishonour’ [Ente:r@n], which is syllabified as [Ent.Pe.r@n].
Notice that in this case, the [r] does skip the boundary between stem and
affix to syllabify with the initial schwa of the infinitival ending. There thus is
indeed a prefix-suffix asymmetry, and it is of the sort we are looking for: the
prefix is more independent from the stem than the suffix. This difference is not
due to a difference in morphological structure: if anything, the suffix is morphologically
‘further away’ from the stem than the prefix. Morphologically,
we would first derive entehr from Ehr, and add en after this.
As far as we know, there are no languages which work in the opposite
way, so that suffixes behave as more independent than prefixes from a phonological
point of view even if the arguable are morphosyntactically closer. In
itself this is not a complete argument in favour of the prosodic word, if only
because prosodic theory does not predict the asymmetry – the opposite could
be described as well. But using the notion of a phonological word — similar
to but not necessarily always congruent with the morphosyntactic word — at
least definitely helps us give a good description of these facts.
Vowel harmony
An important kind of evidence we need to discuss is vowel harmony: in quite a vowel harmony
lot of languages, we can divide the set of all vowels into at least two subsets,
for instance front vowels and back vowels. All vowels in a word are then
taken from the same subset. We can describe this in autosegmental terms as
saying that the feature [front] spreads within the word, as described in section
3.2.
The question, obviously is how to describe this ‘word’, and it is often assumed
that the relevant notion of the word (at least in many languages) is
indeed prosodic rather than morphological or syntactic.
One concrete example comes from Hungarian. In this language, the suffix
meaning ‘with’ takes on the form -val when the preceding stem ends in a back
vowel, but the form -vel when it ends in a front vowel:
(285) a. Front vowel: egérrel ‘with mouse’, Ágnessel ‘with Ágnes’
b. Back vowel: fogóval ‘with pincers’
148 8.2. The Phonological Word
(There is also some kind of assimilation of the first consonant of -val going
on, which we ignore.) Notice that stems themselves are not necessarily harmonic:
the name Ágnes contains a back vowel followed by a front vowel.
Furthermore, in cases such as this, there is actually variation with respect to
the choice of the vowel in the suffix. It can also adapt to the first vowel rather
than to the second (the second vowel in Ágnes can be ‘transparent’):
(286) *egérral ‘with mouse’, Ágnessal ‘with Ágnes’
However, things work differently for compounds. If we suffix a compound,
the affix can harmonize only to the second part, never to the first part:
(287) *madárlessal ‘with birdwatching’, madárlessel ‘with birdwatching’ (from
madar ‘bird’ and less ‘peek’)
The morphosyntactic structure of this form is something like [[[madár][les]]sel],
but the phonology rather behaves as if it is [madár][[les]sel]. The required
phonological structure thus is not precisely congruent with the morphological
structure: there are two separate phonological words, but there is one, complex,
morphosyntactic word. This is a classical argument for prosodic word-
hood.
A final piece of evidence for phonological words come from the phenomenon
of stuttering, which most people do sometimes; and some people do so often
that it becomes problematic for them. We typically assume that in English
functions words (such as determiners the, a or the preposition to) are grouped
in one phonological word together with adjacent lexical words (such as nouns
and verbs).
It turns out that people for some reason stutter more on function words
than on lexical words, and furthermore that they tend to stutter most if the
function word is the first within a phonological word group, as the
(288)
8.3. The Phonological Phrase 149
The data are given for different groups of patients; as you can see, for this
particular phenomenon it does not really matter whether they are (’young’,
’middle’ or ’old’) children, teenagers or adults: they all show a stronger ’standardized
stuttering rate’ on a function word when it is on the first position than
when it is in the 2nd, 3d or 4th position of the phonological word.
We suspect that the reason for this is planning speech. One does not stumble
on a function word because that word is itself problematic, but because one is
then planning the next (lexical) word. The graph above confirms this idea, but
it also shows that apparently the phonological word is a unit of planning: typically,
function words at the end of a phonological word may still be followed
by a lexical word. But since that lexical word is not in the same phonological
word, this apparently does not matter.
8.3 The Phonological Phrase
If we move one level higher up in the prosodic hierarchy, we arrive at the
level of the phonological phrase. Like in the case of the phonological word,
this constituent has a clear syntactic counterpart: the syntactic phrase (XP,
in many theories of syntax). Like in the case of the word, the phonological
phrase and the syntactic phrase do not always exactly coincide; otherwise
there would of course be no reason to distinguish the two.
A classical argument for non-matching prosodic and syntactic structure
comes from English, where arguments have been provided for phonological
phrasing of the following type:
(289) (φthis is the cat )(φ that chased the rat )(φ that stole the cheese)
The φ here is the Greek letter Phi, and is the conventional way of abbreviation
phonological phrase. Phonologically, this sentence consists of three phrases:
the positions where I put the dashes are the positions where a speaker could
pause, and furthermore there is typically some emphasis on the last word of
every phrase.
Yet these three constituents do not necessarily correspond to syntactic constituents.
The syntactic structure would be something like this:
(290) [ this is the cat [ that chased the rat [ that stole the cheese ]]]
Only the last phrase, that stole the cheese, therefore corresponds precisely to a
syntactic constituent, but for instance that chased the rat does not in any way.
This does not mean, on the other hand, that prosodic and syntactic constituency
are completely independent from each other. In particular, every left
syntactic boundary corresponds to a boundary also in the phonology. However,
within the phonology it corresponds both to a left and to a right boundary.
The reason for this is that phonological constituents do not have recursion recursion
(embedding) in the way in which syntactic constituents do. In the syntactic
constituent one sentence (‘that stole the cheese’) forms an integral part of another
sentence (‘that chased the rat that stole the cheese’).
Phonological phrases do not contain other phonological phrases in the
same way. Instead, phonological constituents are usually thought to be restricted
by the so called Strict Layer Hypothesis:
150 8.3. The Phonological Phrase
(291) Strict Layer Hypothesis: Every foot is dominated by a phonological
word, every word is dominated by a phonological phrase, every phonological
phrase is dominated by an intonational phrase (etc.)
One cannot ‘skip layers’: feet are never dominated by a phonological phrase
directly, and one can also not go back: a word is not dominated by a foot,
and not even by a phrase. The phonological structure is organized into layers,
which look more or less like autosegmental structure.
Phonological constituents are thus derived from syntactic constituents: typically,
given a syntactic structure one can construct the phonological structure
reflecting it, but not the other way around: from the fact that (289) has three
phrases we cannot conclude which one is contained in which other one in the
syntax.
In several Bantu languages, vowel lengthening is an indication for phonological
phrases: the penultimate syllable in every phonological phrase lengthens.
The following example is from Chichewa, a language which has received a lot
of attention in the phrasing literature:
(292) a. mleéndo ‘visitor’
b. mlendó uuyu ‘this visitor’
The antepenultimate vowel of mlendó is long at the end of a phrase, but not
when some word follows it in the same phrase. (Another process that is sensitive
to phrase boundaries is tone retraction, moving a tone from a final to a
penultimate syllable, but we will ignore this here.)
The verb phrase is (usually) phrased together in Chichewa, as is the sub-
ject:
(293) a. (mwaána)(anaményá nyuúmba)
child SM-hit house
‘The child hit the house’
b. (mwaána)(anaményá nyumbá ya bwiino)) ‘
’
child SM-hit house of good
‘The child hit the good house’
As you can see, the word nymba has a different tonal distribution and different
syllable length whether it occurs at the complete end of the sentence (or
phrase) or whether some other word follows it.
Another kind of phenomenon that is often taken as evidence in favour
of phonological phrases is that some process is restricted to happen within a
word. For instance, in Bengali a word-final r assimilates (completely) to the
first consonant of the next word, but only if this word occurs in exactly the
same phonological phrase:
(294) a. (øL
morH
)(tSaL
dorH
)(taL
rakeH
)(dieL
tSeH
)
‘Amor gave the scarf to Tara’
b. (øL
motS: aL
dot:a rakeH
)(dieL
tSeH
)
‘Amor gave the scarf to Tara’ (fast speech)
8.4. The Intonational Phrase 151
One important thing to notice here is that the prosodic structure is not just
dependent on the syntactic structure, as the two examples in (294) are exactly
the same in this respect. The only difference is speech rate, the number of syl- speech rate
lables which are uttered per minute. Prosodic structure can be determined by
this: the faster one speaks, the stronger the tendency to put a lot of phonological
material in one constituent. This is true most often for prosodic structure
above the level of the word.
That invoking prosodic phrasing is not just some complicated way of expressing
that r assimilation affects fast speech more than slow speech is shown
by the fact that we can detect phrases also in some other way in Bengali: every
phrase starts with a low tone and ends in a high tone. The assimilation and
the tone pattern thus converge on exactly the same constituency.
The tones themselves are also worth noticing. Where do they come from?
So far, tones had a lexical origin, they were features of a certain word, sometimes
floating and at other times lexically prelinked to a certain syllable. But
this cannot be the case here, as any phonological phrase has the same tonal
structure, regardless of the words which are used in it.
This means that the tones belong to the prosodic constituents rather than to
the individual words: they are boundary tones, linked inherently to the edges of boundary tones
phonological phrases. Boundary tones are one type of intonation tones, tones intonation tones
which do not have the function of expressing lexical contrast but rather of
making up the tonal melody, helping the listener to parse the stream of sounds
in some initial kind of syntactic structure. This may in general be a function
of prosodic structure: to guide the listener in figuring out what the syntactic
structure is of the sentence he is trying to hear.
8.4 The Intonational Phrase
The highest levels of phonological structure is the Intonational Phrase. (In
theory, there is one level that is even higher, that of the Utterance, which com- Utterance
bines every thing a speaker says within one conversational turn. In practice,
very little phonological work has been done on this constituent, and we will
ignore it here.)
As the name suggests, the Intonational Phrase is typically the domain of Intonational Phrase
intonational phenomena, i.e. those that have to do with sentence melody. The
size of the Intonational Phrase is typically that of the full sentence (a main
clause with all dependent clauses).
One function of intonation in many languages is to denote sentence types,
such as the difference between declarative sentences and questions. Take for
instance the following two sentences from Turkish:
(295) a. Biz dün sinema-da fılm seyr-ed-er-ken Ayla Ali-yi gör-mus
we yesterday cinema-LOC film watch-AUX-AOR-ADV Ayla AliACC
see-PERF
‘Ayla saw Ali yesterday while we were watching a film at the
cinema.’
b. Biz dün sinema-da fılm seyr-ed-er-ken Ayla kim-yi gör-mus
we yesterday cinema-LOC film watch-AUX-AOR-ADV Ayla wbhoACC
see-PERF
152 8.4. The Intonational Phrase
‘Who did Ayla see yesterday while we were watching a film at
the cinema?’
The two sentences have exactly the same word order and almost exactly the
same words. Yet one is a question and the other one is not. The way in which
this difference is expressed is by intonation: like in many languages, questions
end in a relatively high pitch, as you can see in the following pictures:
(296) a.
ayladecl.jpg
8.4. The Intonational Phrase 153
b.
aylaquest.jpg
This is a so-called wh-question containing a question (’wh’) word, in this case wh-question
who. Yes/no-questions show the same high tone:
(297) a. Aynur’un Almanya-dan dön-düˇg-ün-ü bil-iyor mu-ydu?
Aynur-GEN Germany-ABL return-COMP-3SG-POSS-ACC knowIMPF
Q.PART-P.COP
Did s/he know that Aynur had returned from Germany?
154 8.4. The Intonational Phrase
b.
You can also see here that the tone is actually not necessarily attached to the
very last syllable, but to the last syllable carrying emphasis or stress. The
scholarly literature has converged on assuming that this high pitch is an autosegmental
tone which gets linked to the edge of an intonational phrase. This
is usually notated as H% — the % marks that this is a so-called boundary tone.boundary tone
Thus the wh-question above could be represented as follows:
(298) ((IP Biz dün sinema-da fılm seyr-ed-er-ken Ayla kim-yi gör-mus H% )IP
The boundary tone should be seen as floating, and attaching to the closest
vowel, in this case that of the suffix -mus. Obviously, English does not have
any lexical tones which distinguish words from each other the way e.g. thelexical tones
Bantu languages in section 3.1 (p. 41) do. But like Turkish, English does have
intonational tones, such as a High boundary tone to indicate questions.
Although intonational patterns are typically the main realisation of intonational
phrases, it has sometimes been claimed that also other kinds of phonology
are sensitive to the edges of these constituents. For instance, the Tuscan
variety of Italian has a process which turns plosives such as /k/ into fricatives
such as [h]; however it does so only inside intonational phrases, not at the
edges:
(299) a. (IP Hanno [h]atturato sette [h]anguri appena nati)IP
‘They have captured seven newly born kangaroos.’
b. (IP Almeri[h]o)IP (IP [k]uando dorme solo)IP (IP [k]ade spesso dall’ama[h]a)IP
‘Almerico, when he sleeps alone, often falls out of the ham-
mock.’
In the latter example, the comma’s of the written sentence correspond to intonational
boundaries when pronounced — also when reading aloud the English
translation you will notice that there are tonal things going on at the edges.
But the precise (left) edges of those constituents also seem to prevent /k/
from leniting. This kind of data can be taken as an indication that also fairly
‘low level’ segmental phonology, referring to individual features such as [continuant],
can be sensitive to these ‘higher-order’ constituents.
8.5. Morphological evidence for the prosodic hierarchy 155
8.5 Morphological evidence for the prosodic hierarchy
In the preceding sections, we have seen some ‘purely’ phonological evidence
in favour of the prosodic hierarchy. As pointed out already several times, this
evidence is of two types: some phonological process happens only within a
phonological constituent of a certain type, or it happens at the boundaries of
such a constituent. This is interesting evidence, but it has been pointed out
as well that it is mostly evidence for the boundaries: processes either require
those to be absent (the first type) or present (the second type).
We can find more direct evidence for the constituency of phonological objects
we have to turn to the interaction of phonology with word formation processses,
morphology. In previous chapters we have largely discussed phonology
as a world on its own, but there are many indications that phonology
interacts with the way in which words are formed in a number of ways.
We will discuss this interaction in more detail in the next chapter, but here
we concentrate on the specific phenomenon of Prosodic Morphology, types of
word formation which refers to elements of prosodic structure, such as morae,
syllables, feet and phonological words. (Higher-order prosodic constituents
play less of a role, because phonological and intonational phrases are typically
too big to match morphosyntactic words.) Examples of prosodic morphology
are infixation and reduplication; these processes are rather rare in English,
but in some other languages they abound. One basic claim of the theory of
prosodic porphology is (McCarthy and Prince, 1998):
(300) Morphological processes that refer to phonological structure use the
same prosodic structures as ordinary phonology: morae (µ), syllables
(σ), feet (Ft) and phonological words (ω).
Although English does not have a lot of prosodic phonology, it does have
one process: so-called expletive infixation. In some varieties, one can insert an expletive infixation
expletive such as bloody — or forms which are even more taboo — within a
word to give some special effect. However, famously, it is not possible to do
this at just any position in the word:
(301) a. fan-bloody-tastic
b. *fa-bloody-ntastic
c. *fantas-bloody-tic
Speakers have quite clear intuitions about what is and what is not possible.
In tests where they are asked to apply expletive infixation to new terms, they
will do so without a lot of variation:
(302) a. amalga-bloody-mated
b. *amalgam-bloody-ated
c. *amal-bloody-gamated
We can understand these judgements as a wish to keep the prosodic constituency
of the original word intact as much as possible. In (301b) and (302b), we
have inserted the expletive within a syllable (fan and ma), and the result of
156 8.5. Morphological evidence for the prosodic hierarchy
this is bad. In (301c) and (302c), we did a similar thing to a phonological foot,
which in English is always a trochee (tastic and malga respectively), which apparently
is also wrong.
We have to break open at least the phonological word in expletive infixation,
because otherwise there would be no infix; but speakers of English intuitively
seem to feel that other prosodic constituents should be respected. This in
itself is evidence that these constituents somehow are part of their knowledge
of language.
This is the kind of intuitions we find attested in prosodic morphology:
speakers show knowledge of the prosodic hierarchy in creating new words.
One process which is quite widespread and in which this happens is reduplication.
In this process, part of a word is doubled to reveal some special meaning.reduplication
The following is an example from Ilokano. As in many languages, reduplication
expresses plurality on nouns in this language:
(303) káldíN ‘goat’ kál-káldíN ‘goats’
púsa ‘cat’ pús-pusa ‘cats’
kláse ‘class’ klas-kláse ‘classes’
jyánitor ‘janitor’ jyan-jy’anitor ‘janitors’
róPot ‘litter’ ro:-róPot ‘litter (pl.)’
trák ‘truck’ tra:-trák ‘trucks’
In these examples, the copied part (the reduplicant) has been italicized. It isreduplicant
easy to see that this reduplicant is a heavy syllable σµµ in all cases. The idea is
that the plural suffix in Ilokano takes this shape: it is an ‘empty’ heavy syllable
which has to be materially filled with segments from the base. Note, by the
way, that the mora theory is a convenient way to express this.
We do not just copy the first syllable of the stem. The first syllable of pusa
presumably is pu. In order to fill the heavy syllable template, however, we
have to add the s which is part of the second syllable. The reason why the
vowel is lengthened in ro:-róPot and tra:-trák has something to do with preferences
of syllable structure, which we will not discuss here.
We thus imagine the derivation of a reduplicated plural in Ilokano in the
following way:
(304) Input Output
σµµ ‘plural’ + klase ‘class’
σ
¢
¢
¢



dd
µ µ
k l a s
σ
¢
¢
¢



µ
k l a
σ
¢
¢
¢
µ
s e
Languages can also choose to specify light syllables (σµ) as the reduplicant,
and as a matter of fact Ilokano provides an instance of this as well, in a suffix
which meens ‘covered with’:
(305) buneN ‘buneng’ si-bu-buneN ‘carrying a buneng’
jyaket ‘jacket’ si-jya-jyaket ‘wearing a jacket’
pandiliN ‘skirt’ si-pa-pandiliN ‘wearing a skirt’
8.5. Morphological evidence for the prosodic hierarchy 157
Again, this is not just a process of copying a syllable, witness the last example:
we do not copy all material from the heavy syllable pan in the stem, but only
just enough to fill the light syllable template.
Higher-order structure can also function as a template for reduplication.
In Diyari we copy a Foot to derive various morphological effects:
(306) wil”a wil”a-wil”a ‘woman’
kanku kanku-kanku ‘boy’
kulkuNa kulku-kulkuNa ‘to jump’
tj
iplarku tj
ilpa-tj
iplarku ‘bird species’
Nankan”t”i Nanka-Nankan”t”i ‘catfish’
Again, we see that it is not just the first two syllables (or the first foot) of the
word which are copied: the final syllable of the reduplicant is always open,
even if the second syllable of the base is not.
Reduplication can also sometimes be total: we then copy the whole phonological
word. Indonesian plural formation is a case in point:
(307) wanita woman wanita-wanita women
mašakarat society mašakarat-mašakarat societies
It will come as no surprise that the phonological word is not always exactly
the same as the morphosyntactic word, and in particular that prefixes are not
always copied together with the stem. In the Bantu language Kihehe, we find
a reduplication process that has an inchoative meaning – it denotes the start of inchoative meaning
an event:
(308) kú-haáta ‘to ferment’ kú-haáta-haáta ‘to start fermenting’
kú-gohomóla ‘to cough’ kú-gohomóla-gohomóla ‘to start coughing’
Everything is copied, including the stem and the ending -a in these words; but
the prefix is not. We can thus say that Kihehe reduplicates the phonological
word; and that such words are formed in a way which is parallel to that of
Italian.
Axininca Campa gives another example of a language in which we copy
a phonological word, but with a certain restriction. In this language, we find
the following pattern:
(309) kawosi-kawosi ‘bathe’
koma-koma ‘paddle’
osampi-sampi ‘ask’
osaNkina-saNkina ‘write’
If the stem starts with a consonant, it is completely reduplicated; if it starts
with a vowel, it is reduplicated except for the initial vowel. The reason for this
presumably is syllable structure: by not copying the initial vowel, we avoid
‘unnecessary’ violations of the constraint ONSET, requiring every syllable to
start with an onset consonant.
(310) REDUPLICATEMAX (RM): Reduplicate all material from the stem.
158 8.5. Morphological evidence for the prosodic hierarchy
(311) a.
/osampi/+RED NODELETION ONSET RM
o.sam.pi.o.sam.pi **!
o.sam.pi.sam.pi * *
sam.pi.sam.pi *! *
b.
/kawosi/+RED NODELETION ONSET RM
ka.wo.si.ka.wo.si
ka.wo.si.wo.si *!*
Emergence of the Unmarked
We see a very important effect here, which distinguishes OT in a favourable
way from parametric theories. In a theory of the latter type, we would need to
say that the ONSET parameter is set ‘on’ in Axininca Campa: witness words
such as osampi, the language allows onsetless syllables. But then we cannot
explain why we all of a sudden find a restriction on them in reduplicated
forms.
In OT, the situation is different: the constraint ONSET is sufficiently lowranked
— below the relevant faithfulness constraint — to make its effect invisible
in ordinary words. But in reduplicated words, faithfulness is no longer
important (the first vowel of o is present anyway), so that now all of a sudden
we can see the universal constraint ONSET can be seen at work. This is
called an effect of the emergence of the unmarked (TETU), and it is at present theTETU
strongest argument in favour of OT over parametric theories.
TETU effects abound in reduplicative systems. For instance, Sanskrit usually
allows complex onsets, but when these onsets are reduplicated (in the
perfective forms of verbs), they are simplified. Reduplication in this case is to
a word.
(312) a. pa-prath
-a ‘spread’
b. ma-mna:-u ‘note’
c. sa-swar ‘sound’
d. da-dh
wans-a ‘scatter’
The constraint which shows up here in the reduplicated form is the one against
complex onsets:
(313) a. RED=σµ: The reduplicative suffix is a monomoraic syllable.
b.
RED + /prath
/ NODELETION *COMPLEX RED=σµ RM
pra.pra.th
a **! th
a
pa.pra.th
a * rth
a
pa.pa.th
a *! th
a
pa.th
a.pra.th
a * *!
I introduced a small piece of new notation in this tableau: in the last column, I
list the segments which violate the constraint. I could just as well have given
8.5. Morphological evidence for the prosodic hierarchy 159
one asterisk for every segment, but this notation gives slightly more insight
into what is actually going on.
We also find TETU effects at the level of segmental structure, for instance
in Akan (the facts have been slightly simplified):
(314) a. siP → si.siP ‘stand’
b. seP → si.seP ‘say’
c. buP → bu.buP ‘bend’
d. soP → su.soP ‘seize’
This pattern looks very much from what we have seen for reduction. In terms
of Element Theory we could state the following markedness and faithfulness
constraints:
(315) a. NOCOMPLEXVOWEL: Only allow primary vowels (markedness)
b. KEEPA: Don’t delete the element A
These constraints again interact to get a TETU effect:
(316)
RED + /seP/ KEEPA NOCOMPLEXVOWEL RM
se.seP **!
si.seP * |A|
si.siP *!
Infixation and shape restrictions
A second well-known example of a prosodic morphological process next to
reduplication is infixation, the positioning of an affix not at the left (prefix) infixation
or right (suffix) edge. Consider the placement of the third person singular
masculine possessive suffix -ka in Ulwa (a language from Nicaragua):
(317) bas ‘hair’ bás-ka ‘his hair’
ki: ‘stone’ kí:-ka ‘his stone’
su:lu ‘dog’ sú:-ka-lu ‘his dog’
asna ‘clothes’ as-ka-na ‘his clothes’
sana ‘deer’ sana-ka ‘his deer’
amak ‘bee’ amak-na ‘his bee’
sapa: ‘forehead’ sapa:-ka ‘his forehead’
siwanak ‘root’ siwa-ka-nak ‘his root’
ana:la:ka ‘chin’ ana:-ka-la:ka ‘his chin’
At first sight, it looks as if -ka sometimes behaves as a suffix, but sometimes
it is also inserted inside the word. On closer inspection, the generalisation is
that -ka comes after the first syllable of the word, if that syllable is heavy, and
otherwise it comes after the second syllable. An insightful way to see this, is
to say that the possessive behaves as a suffix to the first (iambic) foot of the
word.
160 8.5. Morphological evidence for the prosodic hierarchy
The hypothesis of Prosodic Morphology is that infixation is always of this
type: it never means putting an affix just in some random position inside the
stem: it is always prefixed or suffixed to a prosodic constituent. This type of
analysis is also often used to show the advantages of Optimality Theory. Look
at the following examples from Tagalog with the infix -um:
(318) um-alis ‘leave’
t-um-awag ‘call PERF. ACTOR TRIGGER
gr-um-adwet ‘graduate’
In this case, um sometimes looks like a genuine prefix, and sometimes it looks
like an infix; the generalisation is that it is prefixed if the word starts with a
vowel and infixed otherwise. Within OT, we can give an elegant description of
these facts: by infixing, we prevent an unnecessary violation of the constraint
NOCODA. We do this at the cost of violating a (new) instance of an Alignment
constraint, one forcing the left edge of the affix to be aligned to the left edge
of the word; in other words, making the affix to behave like a real prefix.
(319) a. ALIGN(-um-, L, ω, L): The left edge of -um- should correspond to
the left edge of the word (count violations in segments).
b.
/um+tawag/ NODELETION NOCODA ALIGN
um.ta.wag **!
tu.ma.wag * *
ta.wu.mag * **!*
u.ta.wa *!*
c.
/um+alis/ NODELETION NOCODA ALIGN
u.ma.lis *
a.um.lis **! *
a.lu.mis * *!*
u.ma.li *!
Notice that it follows from the principles of the theory that there is a relation
between the shape of the infix — VC — and its infixal behaviour. We predict
that there could not be a language where an affix mu could display the same
behaviour:
(320) Non-existing language:
a. mu+alis → a.mu.lis
b. mu+tawag → mu.ta.wag
The reason is that infixation in this case does not help:
(321)
/mu+alis/ NOINSERTION ONSET ALIGN
mu.a.lis *
a.mu.lis * *!
8.5. Morphological evidence for the prosodic hierarchy 161
No matter where we place the infix, there will always be a violation of the constraint
ONSET; and this hypothetical language will allow onsetless violations,
since it has hypothetical words of the shape alis.
Infixation and reduplication are sometimes combined. For instance, in
Samoan the σµ reduplicant is prefixed to the stress foot:
(322) a. fa.náu → fa.na.náu ‘be born’
b. a.lófa → a.lo.lófa ‘love’
The following paradigm (from Timugon Murut) is also of interest in this con-
nection:
(323) a. bulud → bu-bulud ‘hill/ridge’
b. dondo → do-dondoP ‘one’
c. indimo → in-di-dimo ‘five times’
d. ompod → om-po-pod ‘flatter’
This combines the two types of prosodic morphology we have seen so far:
infixation and reduplication. The affix clearly reduplicates a light syllable of
the stem; but in some cases (if the stem starts with a vowel) it also is infixed
inside the stem, so that it does not reduplicate the first syllable, but the second
one.
This example is of interest, since it seems to violate a generalisation we
just made: that there are no phonological infixes of the shape CV. The reason
is that in cases of infixed reduplication we do avoid unnecessary violations of
the constraint ONSET. If we would copy the first syllable, we would create an
‘unnecessary’ onsetless syllable, which can be avoided if we copy the second
one instead. Still, we would like the infix to be as much to the left — as much
as a prefix — as possible:
(324) a.
/ompod/+RED NOINSERTION ONSET ALIGN
om.om.pod **!
om.po.pod * om
b.
/bulud/+RED NOINSERTION ONSET ALIGN
bu.bu.lud
bu.lu.lud bu!
Hypercoristics formation
A third process of prosodic morphology next to reduplication and infixation is
nick-name (hypocoristics) formation. In many languages, shorter versions can hypocoristics
be used of personal names, for instance to express affection. In these cases, the
hypocoristics assume the shape of some well-described prosodic constituent:
162 8.5. Morphological evidence for the prosodic hierarchy
(325) Name Hypocoristic
ti tiiˇcan
šuusuke šuu-ˇcan *šuusu-ˇcan
yoosuke yoo–ˇcan *yoosu-ˇcan
taizoo tai-ˇcan *taizo-ˇcan
kinsuke kin-ˇcan *kinsu-ˇcan
midori mii-ˇcan *mi-ˇcan
mit-ˇcan
mido-ˇcan
wasburoo waa-ˇcan *wa-ˇcan
wasa-ˇcan
sabu-ˇcan
wasaburo-ˇcan *wasabu-ˇcan
The Japanese hypocoristic consists of a shortened version of the original name
plus the suffix -ˇcan. These examples show is that not just any shortening will
do; we can observe that all the correct hypocoristics consist of an even number
of morae, whereas the wrong versions all have an odd number of morae. In
terms of the typology of stress feet from the previous class, this implies that
the base to which -ˇcan is attached will consist of a number of moraic trochees.
We can observe the restriction to bases of a certain shape also outside the
domain of hypocoristic formation or prosodic morphology proper. For instance,
Dutch has two productive plural suffixes, -@n and -s. The first one is
generally chosen after stems ending in an unstressed syllable, and the second
one after a stressed syllable (underlining marks stress):
(326) a. genie ‘genius’ genieën ‘geniuses’
fabriek ‘factory’ fabrieken ‘factories’
b. familie ‘family’ families ‘families’
pieper ‘potato’ piepers ‘potatoes’
Why do we observe this distributional paradigm? Notice that because of this
effect, plural words tend to end in a (syllabic) trochee: a stressed syllable followed
by an unstressed syllable.
(327)
Ft
dd
σ σ
pi p@rs
Ft
dd
σ σ
Z@ ni @n
There is an importance difference between the Japanese hypocoristics suffix
-ˇcan and the Dutch plural suffix: the former requires its input to take a specific
trochaic shape, whereas the latter makes sure that the output has this particular
shape. Both of them are in support of the claim — which typological study
seems to have confirmed — is that when morphology requires morphemes or
words to have refer some specific shape, such shapes are always taken from
the stock of prosodic phonology. To some extent, this is of course not surprising:
phonological
8.6. Exercises 163
8.6 Exercises
1. ‘Prosodic structure is derived from syntactic structure, but not the other
way around.’ Explain.
2. ’The phonological structure is organized into layers, which look more or
less like autosegmental structure.’ (p. ??) Explain.
3. The Australian language Yidiñ has a process lengthening the antepenultimate
syllable of words with an odd number of syllables. This process
also applies to suffixed words:
(328) a. gudagagu ‘dog-PURP’
b. muãam ‘mother-ABS’
c. muãa:mga ‘mother-PURP’
However, some clusters of suffixes behave like independent units for
this process:
(329) guma:ri daga:ñu ‘to have become red’ (from /gumari/ ‘red’, /daga/
(INCHOATIVE) and /ñu/ (PAST)
a. Construct an argument in favour of the phonological word from this
example.
b. Forms which behave like (329), typically include bisyllabic suffixes.
Can you think of a reason why?
4. Korean has a process of intersonorant of obstruents (so a voiceless obstruent
becomes voiced between two vowels or sonorants). Study the
following examples, and decide what is the prosodic domain of voicing,
and how this (roughly) corresponds to morphosyntactic structure.
(330) a. /ap@ci/ [ab@ji] ‘father’
b. /kı cip/ [kıéip] ‘that house’
c. /motın kırim/ [modın gırim] ‘every picture’
d. /Suni-ıj cip/ [Sunoıj éip] ‘Suni’s house’
e. /kırim-ıl pota/ [kırim-ıl boda] ‘look at the picture’
f. /kæka canta/ [kæga canda] ‘de dog is sleeping’ (lit. ‘dog
sleeps’)
g. /horaNi-wa kojaNi/ [horaNi-wa kojaNi] ‘the tiger and the cat’
5. In Kinande, verbs can be reduplicated to mean (for instance) that the
action that is described is done little by little. In the following, you see
how the reduplication works for stems of different shapes:
(331) a. Consonant-initial huma huma-huma ‘beat’
humira huma-humira ‘beat for’
humirana huma-humirana ‘
b. Vowel-initial esa eses-esa ‘play’
oha ohoh-oha ‘pick’
164 8.6. Exercises
c. Monosyllabic swa swaswa-swa ‘grind’
ta tata-ta ‘bury’
6. Order the constraints in (280) in such a way that you can derive the
different phonological forms of (278). You may need a special constraint
for the compound. Can you think of a form for this constraint?
7. Here are a few examples of reduplication in Mokilese. What is the size
of the reduplicant?
(332) reduplicated
pOdOk pOdpOdOk ‘plant’
mwiNe mwiNmwiNe ‘eat’
kasO kaskasO ‘throw’
poki pokpoki ‘beat’
8. Look at the following examples of Spanish names and corresponding
hypocoristics.
(333) Name Hypocoristic
Alejandro Ale
Asunci´on Asun
Isabel Isa
Jeronimo Jero
Rodrigo Rodri
What is the template for this process of hypocoristic formation?
9. Example (308) shows a few instances of reduplication in Kihehe. Here is
a further example:
(334) /k´u-ita/ [kwi:ta] ‘to pour’ kwi:ta-kwi:ta ‘to pour a little’
As you can see, in this case the prefix is copied with the stem. What
could be the reason? Give a simple OT analysis; you can freely make up
the relevant constraints.
10. Try to reanalyse the examples from English and Bengali in section 8.3
without referring to phonological phrases, but only to syntactic constituents.
Does the difference give you a reason to prefer one analysis over
the other?
11. The examples below give the beginning and the end for a sign meaning
‘shop’ in Israeli Sign language and for the beginning and the end for
a sign meaning ‘shop there’ in the same language. Discuss how such
data could be used to study phonological word structure in (Israeli) sign
language.
(335) a.
8.6. Exercises 165
b.
12. In English, clusters of obstruents tend to assimilate in voicing, either
progressively (336a) or regressively (336b). However, this happens only
to obstruents within the same word, not across word boundaries (336c).
(336) a. twelve [twElv] + th [T] → twelfth twElfT
b. cat [kæt] + s [z] → cats [kæts]
c. The cat zooms [kæt] + zooms [zums] → [tz] (*[ts], *[dz])
If you find a sequence of two obstruents of which one is voiced and the
other voiceless in English, you can be sure there is a word boundary
between them. Is there any reason to think this is a phonological word
boundary?
13. Example (288) (p. 148) shows a graph of the distribution of stuttering in
function words in different positions within the first phonological word of
the utterance. The following shows a graph from the same study on the
same distrbution but now in later phonological words.
(337)
As you can see, the two patterns are not very different. What does this
mean for our evidence with respect to planning and prosodic structure?
166 8.7. Sources and further reading
8.7 Sources and further reading
Section 1.1. The classical book about prosodic phonology is In recent years,
the idea that phonological structures are not recursive has come under attack;
see
The paper on stuttering and phonological words from which the figures in
(308) and (308) are taken is
Chapter 9
Phonology and morphosyntax
9.1 The difference between word phonology and phrasal
phonology
A natural-language utterance obviously consists of more than a string of segmental
slots, organized in prosody and autosegmental structure. There are
also other types of structure: morphological, syntactic and semantic. These
are not studied by phonologists, but on the other hand the different dimensions
of linguistic structure clearly interact; and the interaction of phonology
with these other modules is the topic of study. We have already seen so in
the previous chapter: the higher levels of the prosodic hierarchy clearly have
some relation to morphology and syntax.
It is usually assumed that these are the only two levels with which phonology
interacts. In particular semantics does not seem visible for the phonology:
there are very few attested cases in which phonology responds differently
to words or phrasing having one class of meaning than to another (this
would be the case, for instance, if the OCP would only apply in sentences
with a universal quantifier). Although there are cases of, e.g., sound symbolism
and onomatopoeia in which the sounds in the word correspond in some
way to some aspects of the meaning of the word, there has been very little
phonological study on such aspects, and the literature that exists does not
seem very well-integrated into the rest of the field so far. This may obviously
be an oversight of the field, but at present there is no reason to assume that
semantics and phonology are directly connected.
Syntax and morphology in turn seem to interact with phonology in different
ways. For all intents and purposes, the relation of syntax to phonology
is unidirectional: phonological structure — in particular at the level of the
phonological phrase and higher — is determined by syntactic structure, but
not vice versa. There are no syntactic processes which refer to phonological
shape: no rules which would demand, for instance, that all words beginning
with a nasal should move to the beginning of the sentence. Syntax is in this
sense phonology-blind. phonology-blind
The relation between phonology and morphology seems to be a more intimate
one. We have already seen in the previous chapter that allomorphs can
be selected based on the phonological shape of the stem. In this way, thus
the phonology can influence the morphology. But we have also seen that the
167
168 9.2. Two layers of phonology
phonological word in many cases is a reflex of a morphological structure, and
in this way morphology influences phonology.
It is usually assumed that the consequence for the grammatical model is
that syntax precedes phonology, whereas phonology and morphology interact.
Furthermore, it seems quite clear that there these are two types of phonology,
having slightly different properties: one applying at the word level and another
one applying at the sentence level.
9.2 Two layers of phonology
Consider the following facts of Dutch phonology:
(338) /Ik hEb @t/ [Ik.hE.p@t] ‘I have it’
/hEi hAd @t/ [hEi.hA.t@t] ‘He had it’
/Ik hEb @r/ [ik.hE.p@r] ‘I have her’
The question which raises about these examples is: why are the final obstruents
of heb and had devoiced? We know from section 3.4 that Dutch has
a process of syllable-final devoicing: the feature [voice] is not allowed to appear
in the syllable coda. But the strange thing is that in cases such as this, the
obstruents in question do not appear in the coda: they are in the onset of the
next syllable, since Dutch syllables cannot begin with a schwa.
The point about these forms of course is that the obstruents are at the end
of a syllable if we syllabify these words disregarding their syntactic context.
One of the ideas of the theory of Lexical Phonology is that this idea is essentially
correct: there are (at least) two phonologies — each an Optimality
Theoretic system — which are serialised: first we apply phonology to words
(lexical phonology) and the output of this is then, after the operation of syntax,
applied to phrases (postlexical phonology):
(339) input
↓
Lexical Phonology (Gen + Eval)
↓
output of LP = input of P-LP
↓
Post-Lexical Phonology (Gen + Eval)
↓
output
In the Dutch case at hand, this works as follows. The input form is { /hEb/,
/@t/, /Ik/ } (the order is arbitrary). The output of the Lexical Phonology will
be { [hEp], [@t], [Ik] }, with still an arbitrary order. This will also be the input of
the Post-Lexical component, except that the syntax will determine word-order
at some point. The output of this component will be [Ik.hE.p@t]. The [p] will
be in the onset, but there is no reason why it should turn into a [b]; as far as
the Post-Lexical phonology is concerned, [p] is completely faithful.
Given this serialisation, there are still two possibilities: the Lexical and the
Postlexical grammar could be internally exactly the same (be composed of the
9.2. Two layers of phonology 169
same ranking of the same constraints) or they could be completely different.
The empirical facts seem to point in the second direction in many languages.
For instance, in French, the surface syllable structure seems much more complex
than what is allowed at the word level. Although there are no French
words which start with the cluster *[dvr], we do find such clusters postlexic-
ally:
(340) Henri devrait partir ‘H. would have to leave’ [˜aridvrEpartir]
It thus seems that the restrictions on syllable structure at the postlexical level
are more relaxed than at the lexical level: whatever the constraints against
complex clusters are, they are more highly ranked lexically than they are
postlexically. For example:
(341) a. Lexical ranking: COMPLEX FAITHFULNESS
b. Postlexical ranking: FAITHFULNESS COMPLEX
Another type of evidence is that there are phonological processes which only
happen in the lexicon. An example of this is so-called compound voicing or
rendaku in Japanese. In compounds, initial voiceless consonants of the second rendaku
member get voiced (under certain conditions which do not need to concern
us here):
(342) tama ‘ball’ teppoo dama ‘bullet’
sono ‘garden’ hana zono ‘flower garden’
However, rendaku does not apply within phrases: if tama or sono occur somewhere
else in the sentence, it does not voice. We can obviously describe this
process by assuming that this voicing only applies within a smaller domain
such as the phonological word and not in larger domains; however, phonologists
have compiled a list of general differences between lexical and postlexical
processes:
(343) a. Lexical processes are usually sensitive to lexical information (lexical
exceptions etc.), postlexical are not sensitive to such informa-
tion.
b. Lexical processes change one phoneme into another, postlexical
rules are typically about allophonic changes.
c. The phonetic motivation for postlexical processes is often completely
transparent, whereas this is not the case for lexical rules.
d. Lexical processes are absolute, postlexical processes are gradient:
they can apply more or less, corresponding˜e.g. to the sociolinguistic
situation.
It turns out that Rendaku satisfies most of these conditions. There are for instance
exceptions to it, it deals with phonemes, and it is not entirely clear from
a phonetic point of view why an initial consonant would sometimes have to
be voiced. Since processes often converge along these lines, it seems reasonable
to assume that natural language really has two different components: one
for objects of the size of words or smaller, and one for objects of larger size.
170 9.2. Two layers of phonology
By way of further illustration, two processes in English phonology are
sometimes used: the lexical process of trisyllabic shortening, and the postlexical
process of flapping.
Trisyllabic shortening is the name for a process which shortens a vowelTrisyllabic shortening
when it is followed by two other vowels:
(344) a. div/aj/n + ity → div[ˇı]nity
b. gr/e:/de + ual → gr[ˇE]dual
c. cl/ij/r + ify → cl[ˇæ]rify
Note that the quality of the vowel also changes, but that is something which
is usually abstracted away from. We can describe this by the following con-
straint:
(345) TROCHAIC MAXIMUM: A long vowel cannot be followed by two short
vowels.
According to all three criteria in (343), trisyllabic lengthening (TSL), as a re-trisyllabic lengthening (TSL)
sponse to TROCHAIC MAXIMUM, is a lexical process because it satisfies the
criteria for such a processes:
a. It is subject to lexical factors. There are many words to which is does not
apply, such as nightingale and ivory: in each of those words we have a long
vowel followed by two short vowels. In particular, TSL can only be seen at
work with certain Latinate suffixes, such as -ity, -ual and -ify. Other suffixes
do not trigger the process at all:
(346) br[e:]very, m[aj]tily, p[aj]rating
And even if a suffix will normally trigger TSL, it might not do so in certain
individual cases: ob[i:]sity does not turn into *ob[ˇE]sity. The process thus is
not (fully) productive.
b. Any change for which TSL is responsible will turn one phoneme of the
language into another phoneme. All the sounds we have mentioned so far
{aj, i:, e:, ˇa, ˇae, ˇe,...} represent (distinct) phonemes of the language.
It is said that lexical processes are structure preserving: they never cre-structure preserving
ate sounds which do not already exist underlyingly and in this way they
preserve the structure of the sound inventory of the language, in a way in
which postlexical processes don’t always do: they may add new sounds.
c. It is not very easy to describe the process of TSL in a ‘natural’ way. As a
matter of course, the formulation in (345) is hardly satisfactory; we would
like to reduce it to what we already know about e.g. trochaic feet. But has
proven to be not very easy to find a good formulation of the process in
question in these terms. Why would it be bad to have a long vowel and
then two short syllables? Those conditions seem to be rather specific. Of
course, one does not want to have structures that are longer than binary,
but why would such structures be particularly bad if they contain a long
vowel? We can also observe that a long vowel followed by a short vowel
is an unattested trochee (see section 7.6). But then the question is, why is
such a structure bad when it is followed by another short syllable?
9.2. Two layers of phonology 171
It is also not easy to find evidence for TSL in other languages of the
world, which is another indication that (345) is a rather suspicious constraint
from a typological perspective.
d. The process is absolute: you do not shorten the vowel just a little bit, or
more or less, or more when you talk to your boss then when you talk to
your mother. The vowel of divinity is always shortened in exactly the same
way.
TSL contrasts in all these respects from another process of (American) English
phonology — so-called flapping, whereby an intervocalic coronal stop /t/ or
/d/ turns into a sonorant flap, which we will write as [D], between a stressed
vowel and a following unstressed vowel:
(347) a. a/t/om → a[D]om (cf. a/t/ómic)
b. mee/t/ + ing → mee[D]ing
c. what/t/ + is wrong → wha[D] is wrong
Since we assume that English has trochaic feet, the relevant context seems
to be that the coronal stop is foot internal. We can formulate the following
constraint that could be responsible for flapping:
(348) FLAP: (Coronal) sonorants are preferred over (coronal) stops in the
middle of a foot.
Let us check how this constraint, and therefore this process fares with respect
to the three criteria in (343):
a. The process is not subject to any lexical factors, it does not care what word
we are talking about: all relevant underlying coronal stops in all words
will flap if they occur in the right context. There are no lexical exceptions.
As the examples in (347) show, the process furthermore occurs both
within morphemes as well as across morpheme and even word boundaries.
Whenever we have a stressed vowel on the left, an unstressed vowel
on the right and a coronal stop in the middle, that stop can flap.
b. The output [D] is not an independently occurring phoneme of English;
there are no minimal pairs distinguishing /t/ (or /d/) from a hypothetical
/D/. If we set up an inventory of underlying consonants of English,
we will thus not include a /D/. In that sense flapping is thus not structure
preserving. We actually only find this sound as a result of flapping,
and untrained speakers are typically unaware of the fact that they say a
different consonant in atom than in atomic.
c. Also, FLAP has a natural phonetic explanation: onsets of stressless syllables
tend to ‘assimilate’ in sonority to the surrounding vowels. This type of
process — lenition — is also something which is well attested in the world’s lenition
languages.
d. The process is gradient in two ways. First, one can flap more or less: the
sound /D/ is not completely stable, and is sometimes more like the original
/t/ or /d/, and at other times more sonorant. Furthermore, speakers
may vary what kind of flap they produce under different circumstances,
for instances depending on the formality of the situation.
172 9.2. Two layers of phonology
It thus looks as if the Lexical Phonology is a different type of grammar from
the Post-Lexical Phonology; while the former is much closer to the lexicon, i.e.
the way in words are stored — which explains the sensitivity to exceptions
and to morphological structure, and the structure preservation — the latter is
much closer to the phonetics — hence the naturalness, and the gradience. Especially
for the extreme cases of both, some scholars have argued that they are
‘not real phonology’, and the boundaries are definitely difficult to determine:
• For lexical phenomena, some say that they are really lexical, hence listed
in the lexicon, and not the result of some grammar. A speaker of English
has clˇarity in her head, and she does not produce this on line from clear
and -ity. A less favourable resukt of this line of thinking is that many
regularities cannot be accounted for.
• For postlexical phenomena, some say that they are really phonetic: flapping
is just an automatic consequence of the way the human speech organs
work. A potential problem with this is that it cannot account for
the fact that languages may differ (also) with respect to post-lexical phenomena,
and this is not something we want to account for in ‘universal’
phonetics.
Obviously, if we divide up the whole domain of phonology in this way, there
is nothing for a module of phonology proper to do: everything is either just
stored in the lexicon, or it is the consequence of phonetic processes. Somebody
who wants to defend such a position, and there are some respectable people
who do, has the burden to show for a whole literature of phonology how
things can be carved up again. Notice that the whole work still needs to be
done, so even if you do not call it phonology, it is still is worthwile studying
it, albeit under a different label.
It is a consequence of the organisation of the grammar in (339) that lexical
processes should precede the post-lexical processes. The English phenomena
we have just considered do not provide us with a lot of evidence about this,
simply because the two processes do not interact. They happen independently
from each other, independent of which comes first.
But in cases where there is evidence that there should be ordering of theordering
two processes, it basically always goes in the right direction and the lexical
process comes first. For instance, two things happen in the following Polish
examples:
(349) bup bob-u ‘bean’
xut xod-u ‘pace’
kot kot-a ‘cat’
vus voz-u ‘cart’
In the first place, Polish is subject to syllable-final devoicing, a process we
have already seen operative in many other languages. This process is thus
‘natural’ at leat from the typological point of view. It is also exceptionless in
Polish and it replaces phonemes by other phonemes and is therefore structurepreserving.
It thus is a paradigm case of a postlexical process (although it does
not seem to be gradient).
In the second place, there is a process raising underlying /o/ to [u] when
it occurs before a word-final voiced consonant, explaining the alternation in
9.3. Lexical Levels 173
e.g. bup from underlying /bob/. A form such as kot illustrates that raising
never happens if the form underlyingly has a voiceless obstruent.
Raising is a classic case of a lexical process. It has lexical exceptions: for
instance the words mob [mop] and snob [snop] do not undergo it. It talks about
two phonemes, [u] and [o]. Finally, there seems no phonetic motivation for
this type of raising in this particular context, and the process is also not gradient
in any way.
We have a clear indication that these processes must be ordered in the
sense that they do not happen at the same time, but one after the other. In
particular, when raising applies, it is as if devoicing is not there yet, since
raising distinguishes between /xod/ and /kot/. If we would have only one
grammar, we cannot formulate the right constraint for raising to apply. Such
a constraint would need to say ‘do not have a mid vowel before a voiced
consonant’, but if devoicing applies at the same time the underlyingly voiced
consonant will be devoiced and the constraint will not apply.
This only works well if we have to grammars, and raising applies first and
then devoicing:
(350) a. /xod/ /kot/
raising xud n.a.
devoicing xut n.a.
[xut] [kot]
b. /xod/ /kot/
devoicing xot n.a.
raising d.n.a. n.a.
*[xot] [kot]
The comparison of Dutch and Polish teaches us, by the way, that a process
(final devoicing) which is lexical in one language may be postlexical in another
one.
The difference between lexical and postlexical phonology has also been observed
in languages that do not belong to the Indo-European language family.
9.3 Lexical Levels
Usually, it is assumed that phonological processes start their life in natural
language as postlexical processes, only to turn lexical in the course of time.
9.4 The life cycle of phonological processes
Usually, it is assumed that phonological processes start their life in natural
language as postlexical processes, only to turn lexical in the course of time.
9.5 Cyclicity
We can thus see that phonology is not monolithic. There are at least two phonological
modules. In the architecture of the grammar, it would be logical to
say that one precedes the syntactic component, whereas the other follows it.
174 9.5. Cyclicity
The former thus interacts with morphology, and the latter with the output of
syntax.
We will now concentrate on the lexical phonology. Until now, we have
concentrated in this course on words which consist of one or two morphemes.
But what happens if there are more than two morphemes in the input, viz. a
stem plus more than one affix? In principle, there are two possibilities: we
add all the affixes at the same time, or we first add one affix, then we apply
phonology, and then we add another affix. It is one of the basic tenets of
so-called cyclic phonology — usually incorporated into Lexical Phonology —
that the latter is the case. The model is called cyclic, because we go in a circle:
we add a suffix, then we apply phonology, then we add another suffix, etc. In
this class we will assume that the phonology is the same constraint ranking
every time.
A classical example is the difference between the English words condensation
and compensation. (The relevance of the example is not uncontested, but
we will use it here for illustrative purposes.) These words have virtually the
same segmental makeup, but there stress pattern is different, and so is schwa
reduction as a consequence of this: cond`Ensátion versus còmp@nsátion. According
to this analysis, the reason for this is that the former noun is related to the
verb cond´Ense, whereas the latter is related to còmp@nsáte. These relations are
reflected in the stress structure of the verbs.
The cyclic account of this would run along the following lines. We first
derive the stress structure of the verbs, then we attach the suffixes and — on
the next cycle — we get the stress structure of the nouns.
(351) underlying compensate condense
↓ ↓ stress assignment
còmp@nsáte cond´Ense
↓ ↓ nominalisation
còmp@nsáte+ion cond´Ense+ation
↓ ↓ stress assignment
surface còmp@nsátion cond`Ensátion
The working of the cycle is one of the key components of traditional generative
phonology. Within generative syntax (minimalism) it still survives in the
form of derivational phases; it may be considered as one of Chomsky’s most
important contributions to our insight into linguistics.
A more complicated — and realistic — example comes from Palestinian
Arabic (Brame, 1974; Kager, 1999b; Kiparsky, 2000). This language has a process
deleting unstressed high front vowels /i/ in all syllables of the word
except for the last one:
(352) a. /fihim/ ‘to understand’ (verb stem)
b. /fihim/+ → [f´ıhim] ‘he understood’
c. /fihim/+/na/ → [fh´ımna] ‘we understood’
d. /fihim/+/u/ → [f´ıhmu] ‘they understood’
Now consider the following form:
9.5. Cyclicity 175
(353) /fihim/+ + /na/ → [fih´ımna], *[fh´ımna] ‘he understood us’
/i/ deletion does not take place here, even though its phonological conditioning
is met. The reason for this is that structures with object clitics are built on
the basis of structures which occur as independent words, and we have the
following principle governing derivations:
(354) Natural Bracketing Hypothesis. A substring ψ of a string φ is a domain
of cyclic rule application in phonology only if it shows up elsewhere as
an independent word sequence which enters compositionally into the
determination of the meaning of φ.
Since fíhim ‘he understood’ occurs independently in Palestinian Arabic
(e.g. in the sentence fíhim il-walád ‘he understood the boy’), and since it enters
compositionally into the determination of [fih´ımna] — we have to calculate
what the former means in order to understand what the latter means —, we
also phonologically determine the structure of ‘he understood’ in order to get
the structure of the whole form. On the other hand, fíhim ‘he understood’ itself
derives from the verbal stem fihim, but this is not a separately pronounceable
word, hence it cannot influence the structure of the stem in any way.
Note that the example looks very much like the English case discussed
above, except that in English we were dealing with vowel reduction rather
than vowel deletion. There is another difference as well: for the English case
we could assume that it would be the stress which is transferred from the base
to the extended form and that there would be some secondary stress on the [E]
in condensation, blocking its reduction. Yet this is not true for the Palestinian
dialect of Arabic. There is no trace of secondary stress in the cliticised form.
This becomes apparent in parallel forms with other vowels than /i/. These
vowels are not deleted, and words like d. arábna are ambiguous between the
meaning ‘he hit’ and ‘he hit us’.
A basic assumption of the model is that cyclic phonology is always lexical
(because cyclicity involves sensitivity to morphological structure). Is the
reverse also the case? The attentive reader who studies the following Dutch
facts will notice that this cannot be the case:
(355) a. Derivation: he/b/+erd → [hEb@rt] (‘greedy person’)
b. Inflection: he/b/+en → [hEb@n] (‘have’ PLUR)
We have seen above that devoicing crucially applies lexically in Dutch. These
facts show, on the other hand, that it cannot apply cyclically, otherwise it
would apply in the stem heb. The usual way of thinking about this is that
the lexical phonology actually itself consists of two subphonologies: a cyclic
component first, which interacts with the morphology in the way we have
seen for English and Palestinian Arabic, and a ‘word’ phonology after this,
which applies to whole words before they are then inserted into the syntax.
All in all, the standard model of Lexical Phonology thus looks as follows:
176 9.5. Cyclicity
(356) input
↓
Cyclic Phonology (Gen + Eval)
↓
output of CP = input of WP
↓
Word Phonology (Gen + Eval)
↓
output of WP = input of P-LP
↓
Post-Lexical Phonology (Gen + Eval)
↓
output
Exercises
1. On p. 172 it is claimed that the fact that lexical phonology is close to the
lexicon explains the sensitivity to exceptions and to morphological structure,
and the structure preservation, whereas the nearness of postlexical
phonology explains why the latter is more natural and more gradient.
Explain both points.
2. There is some overlap in predictive power between prosodic phonology
and Lexical Phonology. Discuss.
References
Kenstowicz (1994) gives a nice overview of the literature on differences between
lexical and postlexical phonology.
Bibliography
Abbas, N. (2009). Nasalization in saraiki. MA Thesis, University of Essex.
Adger, D. (2003). Core Syntax. A Minimalist Approach. Oxford: Oxford University
Press.
Alidou, O. (1997). A linguistic study of Nigerian language games. Ph. D. thesis,
Indiana University.
Anderson, J. and C. Ewen (1987). Principles of Dependency Phonology. Cambridge:
Cambridge University Press.
Anderson, S. R. (1985a). Phonology in the Twentieth Century: Theories of Rules
and Theories of Representations. Chicago: University of Chicago Press.
Anderson, S. R. (1985b). Phonology in the Twentieth Century. Theories of Rules and
Theories of Representations. Chicago and London: The University of Chicago
Press.
Backley, P. (2011). An Introduction to Element Theory. Edinburgh: Edinburgh
University Press.
Bakovic, E. (2000). Harmony, Dominance and Control. Ph. D. thesis, Rutgers.
Bat-El, O. (2011). Semitic templates. In M. van Oostendorp, C. Ewen, E. Hume,
and K. Rice (Eds.), The Blackwell Companion to Phonology, Volume V, Chapter
108, pp. 2586–2608. Oxford, UK and Cambridge, Mass.: Wiley-Blackwell.
Baudouin de Courtenay, J. (1895). Versuch einer Theorie phonetischer Alternationen:
Ein Capitel aus der Psychophonetik. Strassburg/Crakow: Trübner.
Bermúdez-Otero, R. (2005). Diachronic phonology. In P. de Lacy (Ed.), The
Cambridge Handbook of Phonology. Cambridge: Cambridge University Press.
Bermúdez-Otero, R. (forthcoming). Stratal Optimality Theory. Oxford: Oxford
University Press.
Blake, B. J. (1969). The Kalkatungu Language: A Brief Description. Canberra:
Australian Insitute for Aboriginal Studies.
Bolton, T. (1894). Rhythm. American Journal of Psychology 6, 145–238.
Bonatti, L., M. Pena, M. Nespor, and J. Mehler (2005). The role of consonants
and vowels in continuous speech processing. Psychological Science.
Botma, B. (2004). Phonological Aspects of Nasality. An Element-Based Dependency
Approach. Ph. D. thesis, University of Amsterdam.
Brame, M. (1974). The cycle in phonology: stress in palestinian, maltese and
spanish. Linguistic Inquiry 5, 39–60.
Bybee, J. (2001). Phonology and Language Use. Cambridge: Cambridge University
Press.
Campbell, A. (1959). Old English Grammar. Oxford, UK: Clarendon Press.
Carreiras, M. and C. J. Price (2008). Brain activation for consonants and vowels.
Cerebral Cortex 18(7), 1727–1735.
177
178 Bibliography
Carstairs-McCarthy, A. (1999). The origins of complex language: an inquiry into
the evolutionary beginnings of sentences, syllables and truth. Oxford: Oxford
University Press.
Charette, M. (1991). Conditions on Phonological Government, Volume 58 of Cambridge
Studies in Linguistics. Cambridge: Cambridge University Press.
Chierchia, G. (1986). Length, syllabification and the phonological cycle in
italian. Journal of Italian Linguistics 8(5-33).
Chin, S. B. (2006). Realization of complex onsets by pediatric users of cochlear
implants. Clinilcal Linguistics and Phonetics 20(7-8), 501–508.
Chomsky, N. (1968). Language and Mind. New York: Harcourt Brace Jovan-
ovich.
Chomsky, N. (1986). Knowledge of Language: Its Nature, Origin and Use. New
York: Praeger.
Chomsky, N. (1995). The Minimalist Program. Cambridge, Mass.: MIT Press.
Chomsky, N. (2004). Beyond explanatory adequacy. In A. Belletti (Ed.), Structures
and Beyond: The Cartography of Syntactic Structures, Vol. 3. Oxford: Oxford
University Press.
Chomsky, N. and M. Halle (1968). Sound Pattern of English. Harper and Row.
Clements, G. N. (1990). The role of the sonority cycle in core syllabification. In
J. Kingston and M. Beckman (Eds.), Papers in Laboratory Phonology I: Between
the Grammar and the Physics of Speech, pp. 283–333. Cambridge, UK: Cambridge
University Press.
Clements, G. N. and B. Hume (1995). The internal organization of speech
sounds. In J. Goldsmith (Ed.), The Handbook of Phonological Theory, pp. 245–
306. Cambridge: Blackwell Publishers.
Clements, N. and S. J. Keyser (1983). CV Phonology: A Generative Theory of the
Syllable. Cambridge, Mass.: MIT Press.
Clements, N. and E. Sezer (1982). Vowel and consonant disharmony in turkish.
In H. van der Hulst and N. Smith (Eds.), The Structure of Phonological
Representations, pp. 213–255. Dordrecht: Foris.
Conklin, H. (1956). Tagalog speech disguise. Language 32, 136–139.
Coulmas, F. (2003). Writing systems. An introduction to their linguistic analysis.
Cambridge, UK: Cambridge University Press.
Crosswhite, K. (2001). Vowel Reduction in Optimality Theory. New York: Rout-
ledge.
Cutler, A., J. Mehler, D. Norris, and J. Seguí(1986). The syllable’s differing
role in the segmentation of french and english. Journal of Memory and Language
25, 385–400.
Davis, S. (1990). Italian onset structure and the distribution of il and lo. Linguistics
28, 43–55.
Davis, S. and S.-H. Shin (1999). The syllable contact constraint in korean. an
optimality-theoretic analysis. Journal of East Asian Linguistics 8, 285–312.
de Lacy, P. (Ed.) (2007). The Cambridge Handbook of Phonology. Cambridge:
Cambridge University Press.
Dresher, B. E. (2009). The Contrastive Hierarchy in Phonology. Cambridge: Cambridge
University Press.
Duanmu, S. (2009). Syllable structure. The limits of variation. Oxford, UK: Oxford
University Press.
Ewen, C. and H. van der Hulst (2001). The Phonological Structure of Words. An
Introduction. Cambridge: Cambridge University Press.
Bibliography 179
Féry, C. (2003). Liaison and syllable structure in french. Manuscript, Potsdam
University.
Fikkert, P. (1994). On the acquisition of prosodic structure. Ph. D. thesis, University
of Leiden.
Freidin, R. (2012). Syntax. Basic Concepts and Applications. Cambridge: Cambridge
University Press.
Fromkin, V. (Ed.) (1973). Speech errors as linguistic evidence. Mouton.
Goldsmith, J. (1990). Autosegmental and Metrical Phonology. Oxford: Basil
Blackwell.
Goldsmith, J. (to appear). The syllable. To appear in the second edition of The
Handbook of Phonological Theory.
Gouskova, M. (2003). Falling sonority onsets, loanwords, and syllable contact.
CLS 37.
Gussenhoven, C. and H. Jacobs (1998). Understanding phonology. London:
Arnold.
Gussenhoven, C. and H. Jacobs (2005). Understanding Phonology. London:
Hodder Arnold.
Gussman, E. (2002). Phonology. Analysis and Theory. Cambridge Textbooks in
Linguistics. Cambridge: Cambridge University Press.
Hale, M. (2003). Neogrammarian sound change. In B. Joseph and R. Janda
(Eds.), The Handbook of Historical Linguistics, pp. 344–368. Oxford: Blackwell.
Hale, M. and C. Reiss (2000). Phonology as cognition. In N. Burton-Roberts,
P. Carr, and G. Docherty (Eds.), Phonological Knowledge: Conceptual and Empirical
Foundations, pp. 161–184. Oxford: Oxford University Press.
Hammond, M. (1999). The Phonology of English. Oxford: Oxford University
Press.
Harris, J. (1994). English Sound Structure. London: Blackwell.
Harris, J. (2005). Vowel reduction as information loss. In P. Carr, J. Durand,
and C. Ewen (Eds.), Headhood, Elements, Specification Headhood, Elements, Specification
and Contrastivity. Phonological Papers in Honour of John Anderson, pp.
119–132. Amsterdam: John Benjamins.
Harris, J. and G. Lindsey (1995). The elements of phonological representation.
In J. Durand and F. Katamba (Eds.), Frontiers of Phonology, pp. 34–79.
London and New York: Longman.
Haugen, J. D. (2003). Allomorphy in yaqui reduplicatio. In L. M. Barragan
and J. D. Haugen (Eds.), Studies in Uto-Aztecan, Volume 5, pp. 75–103. MIT
Working Papers on Endangered and Less Familiar Languages.
Hayes, B. (1987). A revised parametric metrical theory. Northeastern Linguistics
Society 17, 274–289.
Hayes, B. (1989). Compensatory lengthening in moraic phonology. Linguistic
Inquiry 20.
Hayes, B. (1995). Metrical stress theory. Principles and Case Studies. University
of Chicago Press.
Hayes, B. (1999). Phonetically-driven phonology: the role of optimality theory
and grounding. In M. Darnell, E. Moravscik, M. Noonan, F. Newmeyer,
and K. Wheatly (Eds.), Functionalism and Formalism in Linguistics, Volume I:
General Papers, pp. 243–285. Amsterdam: John Benjamins.
Hayes, B. (2009). Introductory phonology. Malden, Mass., and Oxford, UK:
Wiley-Blackwell.
Hoffmann, C. (1963). A grammar of the Margi language. London: Oxford University
Press.
180 Bibliography
Hogg, R. (1992). A Grammar of Old English, Volume !: Phonology. Oxford, UK
and Cambridge, Mass.: Blackwell.
Holt, D. E. (2004). Optimization of syllable contact in old spanish via the
sporadic sound change metathesis. Probus 16(1), 43–61.
Hulst, H. v. and J. v. Weijer (1995). Vowel harmony. In J. Goldsmith (Ed.), The
Handbook of phonological theory, pp. 495–534. Oxford: Basil Blackwell.
Humbert, H. (1995). Phonological Segments. Their Structure and Behaviour. Ph.
D. thesis, University of Amsterdam.
Hyman, L. (1976). Phonologization. In A. Devine and L. D. Stephens (Eds.),
Linguistic studies offered to Joseph Greenberg on the occasion of his sixtieth birthday,
pp. 407–418. Saratoga: Anma Libri.
Hyman, L. (2008). Universals in phonology. The Linguistic Review 25, 85–137.
Itô, J. (1986). Syllable theory in prosodic phonology. Ph. D. thesis, University of
Massachusetts at Amherst.
Ito, J. (1986). Syllable Theory in Prosodic Phonology. Ph. D. thesis, University
of Massachusetts, Amherst. Published 1988. Outstanding Dissertations in
Linguistics series. New York: Garland.
Itô, J. and A. Mester (2003). Japanese Morphophonemics. Markedness and Word
Structure. Cambridge, Mass.: MIT Press.
Iverson, G. K. and J. C. Salmons (2011). Final devoicing and final laryngeal
neutralisation. In M. van Oostendorp, C. J. Ewen, B. Hume, and K. Rice
(Eds.), The Blackwell Companion to Phonology, Volume III, pp. 1622–1643. Oxford,
UK: Wiley-Blackwell.
Kager, R. (1999a). Optimality Theory. Cambridge: Cambridge University Press.
Kager, R. (1999b). Surface opacity of metrical structure in optimality theory.
In B. Hermans and M. van Oostendorp (Eds.), The Derivational Redidue in
Phonological Optimality Theory. Amsterdam: John Benjamins.
Kaye, J. (1989). Phonology. A cognitive view. Cognitive Science Series. Tutorial
Essays. Hillsdale, New Jersey and Hove and London: Lawrence Erlbaum
Asociates.
Kaye, J. (1992). Do you believe in magic? The story of s+C sequences. SOAS
Working Papers in Linguistics and Phonetics 2, 293–313.
Kaye, J., J. Lowenstamm, and J.-R. Vergnaud (1985). The internal structure
of phonological elements. a theory of charm and government. Phonology
Yearbook 2, 305–328.
Kaye, J., J. Lowenstamm, and J.-R. Vergnaud (1990). Constituent structure and
government in phonology. Phonology 7, 193–232.
Kayne, R. S. (1994). The Antisymmetry of Syntax. Cambridge: The MIT Press.
Linguistic Inquiry Monograph Twenty-Five.
Kennedy, R. (2011). Reduplication in sanskrit. In M. van Oostendorp, C. J.
Ewen, E. Hume, and K. Rice (Eds.), The Blackwell Companion to Phonology,
Volume 5, Chapter 119, pp. 2855–2878. Oxford, UK: Wiley-Blackwell.
Kenstowicz, M. (1994). Phonology in Generative Grammar. Oxford, UK: Black-
well.
Kenstowicz, M. (2007). Salience and similarity in loanword adaptation. a case
study from fijian. Language Sciences 29, 316–340.
Keyser, S. J. and P. Kiparsky (1984). Syllable structure in finnish phonology. In
M. Aronoff and R. Oehrle (Eds.), Language Sound Structure. Studies in Phonology
Presented to Morris Halle by his Teacher and Students, pp. 7–31. Cambridge,
MA: MIT Press.
Bibliography 181
Kiparsky, P. (2000). Opacity and cyclicity. The Linguistic Review 17, 351–367.
Kiparsky, P. (2003). Accent, syllable structure and morphology in ancient
greek. In E. M. Athanasopoulou (Ed.), Selected papers from the 15th international
symposium on theoretical and applied linguistics. ISTAL.
Knight, R.-A. (2012). Phonetics. A Coursebook (Cambridge ed.). Cambridge
University Press.
Kornfilt, J. (1977). Turkish. New York: Routledge.
Krämer, M. (2009). The Phonology of Italian. The Phonology of the World’s
Languages. Oxford: Oxford University Press.
Lahiri, A. and H. Reetz (2010). Distinctive features: Phonological underspecification
in representation and processing. Journal of Phonetics 38, 44–59.
Lass, R. (1976). English phonology and phonological theory. Cambridge: Cambridge
University Press.
Levelt, C. C. and M. van Oostendorp (2007). Feature co-occurrence constraints
in l1 acquisition. In B. Los and M. van Koppen (Eds.), Linguistics in the Netherlands
2007, pp. 162–172. Amsterdam and Philadelphia: John Benjamins.
Liberman, I. Y., D. P. Shankweiler, F. W. Fischer, and B. Carter (1974). Explicit
syllable and phoneme segmentation in the young child. Journal of Experimental
Child Psychology 18, 201–212.
Lieber, R. (2009). Introducing Morphology. Cambridge: Cambridge University
Press.
Lynch, J. (1974). Lenakel Phonology. Ph. D. thesis, University of Hawaii. Also,
University of Hawaii Working Papers in Linguistics vol. 7.1.
MacKay, D. G. (1972). The structure of words and syllables. evidence from
errors in speech. Cognitive Psychology 3, 210–227.
McCarthy, J. and A. Prince (1993a). Prosodic morphology i. constraint interaction
and satisfaction. Available on ROA 485-1201.
McCarthy, J. and A. Prince (1998). Prosodic morphology. In A. Spencer
and A. M. Zwicky (Eds.), The Handbook of Morphology, pp. 283–305. Oxford:
Blackwell.
McCarthy, J. J. (1979). Formal problems in Semitic phonology and morphology. Ph.
D. thesis, MIT. Published by Garland Press, New York, 1985.
McCarthy, J. J. (2008). Doing Optimality Theory. Cambridge: Blackwell.
McCarthy, J. J. and A. Prince (1993b). Generalized alignment. Morphology
Yearbook, 79–153.
Meringer, R. and K. Mayer (1895). Versprechen und Verlesen. Göschense Ver-
lagsbuchhandlung.
Mielke, J. (2004). What ambivalent segments can tell us about the universality
of distinctive features. Boston. LSA 78.
Myers, S. (1987). Tone and the structure of words in Shona. Ph. D. thesis, University
of Massachusetts at Amherst, Amherst, MA.
Nespor, M. and I. Vogel (1986). Prosodic Phonology. Dordrecht: Foris.
Nijen Twilhaar, J. (1990). Generatieve fonologie en de studie van Oostnederlandse
dialecten. Amsterdam: P.J. Meertens-Instituut.
Odden, D. (1980). Associative tone in shona. Journal of Linguistic Research 1,
37–51.
Odden, D. (1986). On the role of the obligatory contour principle in phonological
theory. Language 62, 353—383.
Odden, D. (1987). Dissimilation as deletion in chuckchi. In A. Miller and
J. Power (Eds.), Proceedings from ESCOL, Volume 3, Ohio, pp. 235–246. Ohio
State University.
182 Bibliography
Odden, D. (2005). Introducing phonology. Cambridge: Cambridge University
Press.
O’Roerke, E. (2008). Intonation in quechua: Questions and analysis. In International
Congress of Phonetics Sciences. Satellite workshop on ‘Intonational Phonology:
Understudied or Fieldwork Languages.
Parker, S. (2002). Quantifying the Sonority Hierarchy. Ph. D. thesis, University
of Massachusetts at Amherst.
Parker, S. (2008). Sound level protrusions as physical correlates of sonority.
Journal of Phonetics 36(1), 55–90.
Prince, A. and P. Smolensky (1993). Optimality theory: Constraint interaction
in generative grammar. Manuscript, Rutgers University and University of
Colorado at Boulder. Available at ROA and published 2004 at Blackwell.
Radford, A., M. Atkinson, D. Britain, H. Clahsen, and A. Spencer (2009). Linguistics.
An Introdution. Cambridge: Cambridge University Press.
Rehg, K. and D. Sohl (1981). Ponapean Reference Grammar. Honolulu: University
of Hawaii Press.
Rhee, S. (2002). Empty Nuclei in Korean. Ph. D. thesis, University of Leiden.
Sakel, J. and D. L. Everett (2012). Linguistic Fieldwork. Cambridge: Cambridge
University Press.
Sapir, D. (1965). A Grammar of Diola-Fogny. Cambridge, UK: Cambridge University
Press.
Scheer, T. (2004). A Lateral Theory of Phonology. Volume I: What Is CVCV, and
Why Should It Be? Berlin: Mouton de Gruyter.
Sezer, E. (1986). An autosegmental analysis of compensatory lengthening in
turkish. In L. Wetzels and E. Sezer (Eds.), Studies in compensatory lengthening,
pp. 227–250. Dordrecht: Foris.
Sievers, E. (1881). Grundzüge der Phonetik. Leipzig: Breitkopf und Härtel.
Steriade, D. (1988). Greek accent. a case for preserving structure. Linguistic
Inquiry 19, 271–314.
Topintzi, N. (2006). Moraic Onsets. Ph. D. thesis, University College London.
Trubetzkoy, N. S. (1939). Grundzuge der Phonologie. Prague: Travaux du cercle
linguistique de Prague 7.
van Oostendorp, M. (1998). Bergen op zoom diminutives in phonological theory.
In S. Barbiers, J. Rooryck, and J. van de Weijer (Eds.), Small words in the
big picture; Squibs for Hans Bennis. Leiden: HIL.
van Oostendorp, M., C. J. Ewen, E. Hume, and K. Rice (Eds.) (2011). The
Blackwell Companion to Phonology. Oxford: Blackwell.
van Oostendorp, M. and J. van de Weijer (2005). Phonological alphabets and
the structure of the segment. In M. van Oostendorp and J. v. Weijer (Eds.),
The Internal Organization of Phonological Segments, Volume 77 of Studies in
Generative Grammar. Berlin: Mouton De Gruyter.
Vennemann, T. (1988). Preference laws for syllable structure and the explanation
of sound change. With special reference to German, Germanic, Italian and Latin.
Berlin: Mouton De Gruyter.
Watson, J. (2011). Word stress in arabic. In M. van Oostendorp, C. J.
Ewen, E. Hume, and K. Rice (Eds.), The Blackwell Companion to Phonology,
Volume V, Chapter 124, pp. 2990–3018. Oxford, UK: Wiley-Blackwell.
Wetzels, L. (1995). Mid vowel alternations in the brazilian portuguese verb.
Phonology 12, 281–304.
Williams, E. (1976). Underlying tone in margi and igbo. Linguistic Inquiry 7,
463—484.
Bibliography 183
Woodrow, H. (1909). A quantitative study of rhythm. Archives of Psychology 14,
1–66.
Zsiga, E. (2011). Local assimilation. In M. van Oostendorp, C. J. Ewen,
E. Hume, and K. Rice (Eds.), The Blackwell Companion to Phonology,
Volume II, pp. 1919–1944. Oxford: Wiley-Blackwell.