MARCUS KRACHT & JENNIFER SPENADER

AN OPEN
INTRODUCTION TO
LINGUISTICS

USED BY PERMISSION
EDITED AND EXTENDED BY
JENNIFER SPENADER & HESSEL HAAGSMA
Copyright © 2022 Marcus Kracht & Jennifer Spenader

published by university of groningen

tufte-latex.googlecode.com

Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in com-
pliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/
LICENSE-2.0. Unless required by applicable law or agreed to in writing, software distributed under the
License is distributed on an “as is” basis, without warranties or conditions of any kind,
either express or implied. See the License for the specific language governing permissions and limitations
under the License.

This printing, February 2022

 Contents

1 Introduction 5

2 Phonetics 11

3 Phonology I: Features and Phonemes 27

4 Phonology II: Realization Rules and Representations 47

5 Phonology III: Syllable Structure & Stress 59

6 Phonology IV: Rules, Constraints

and Optimality Theory 69

7 Morphology I: Basic Facts 81

8 Morphology II: Morphemes

and Morphs 93

9 Syntax I: Categories, Constituents, Trees and Context Free Grammars 103

10 Syntax II: Argument Structure 115

11 Syntax III: Local Dependencies and Constraints: Selection, Agreement and

Case Marking 133
 4

12 Syntax IV: Movement and Non-Local Dependencies 143

13 Syntax V: How to Draw Syntactic Trees 155

14 Syntax VI: Binding 165

15 Semantics I: Fundamentals of Meaning and its Representation 177

16 Semantics II: Quantification: Beyond Reference 193

17 Semantics III: Basic Elements of Interpretation 205

18 Semantics IV: Scope 215

19 Semantics V: Cross-Categorial Parallelism 221

20 Pragmatics 229

21 Language Acquisition 243

22 Language Families and History of Languages 265

Bibliography 275

Index 275
1
Introduction

Language is the scaffolding of thought.

An introduction

When we exchange information with others, we usually do it via lan-

guage. Of course, other modes of communication like posture, facial
gestures, eye rolling or eye-brow raising are other ways in which we
communicate, but language is the means by which we communicate
complex thoughts.
For this reason it’s no wonder that the most natural way to com-
municate with computers would be natural language. Unfortunately,
most communication with machines is quite limited. Like a novice
speaker of a second-language, the computer carefully controls the in-
teraction in order to make sure it understands the responses, by asking
clear “yes” and “no” questions. The user responds via a mouse, touch
screen or keyboard. None of these modes of response could be consid-
ered “natural”.
But this is only how carefully designed computer programs that are
meant to be used by the general public communicate with their users.
More advanced or specialty systems can simply say there is a syntax
error, or give an error number only. Less sophisticated systems also
communicate with their users this way. For example, if my washing
machine has an error it actually just tells me an error number that I
then have to look up in a manual.
Windows asking a question that can
In the first case, with experts using specialized systems, we can be answered with “Okay” or “Cancel”.
expect the users of those systems to invest time into learning how
the computer communicates information or errors. But a drawback is
that it makes the system less transparent, and for example harder to
interact with in a stressful situation.
In the second case, interacting via a code and looking that up in a
book is a very awkward, time-consuming method to access the infor-
mation that the machine has. We would prefer to be able to simply
be told what’s wrong, or to be able to ask a clear question.
With the advent of digital assistants on our phones and smart speak-
ers in our homes, average people can now access information in spoken
natural language via different programs using a phone assistant or
smart speaker. For example, an Alexa device can be used to do sim-
6 an open introduction to linguistics

ple tasks like setting an alarm, asking what the weather is like, ask-
ing Alexa to play the more recent news broadcast or to play a song.
What’s interesting about Alexa, google assistant or Siri, is that all
the tasks and programs that these smart speakers can access are also
available on our computer’s or phones, and we frequently and easily
access them. But it’s much more convenient to simply use your voice
to ask in natural language for the same information or task. It’s not
only easier, because you don’t have to find your phone, but it also feels
more natural, because speech is the more natural, and more common
way in which we ask humans to do things for us.
For an Alexa device to speak and understand our commands re-
quired years of computational linguistic research, which built and still
builds on our understanding of human language and how language is
used as a means of communication. While many natural language pro-
cessing tasks (NLP) like speech recognition and natural language un-
Alexa and google home, smart speak-
derstanding (NLU) and natural language generation (NLG) are now
ers.
frequently dealt with using machine learning methods that rely on
large data sets, often these data sets require training and developmen-
tal data that first required linguistic processing or annotation. This
processing relies on an understanding of fundamental linguistic con-
cepts, based on decades of research. One of the goals of this book and
the course that accompanies it, is to give you a solid, if brief, intro-
duction to these linguistic concepts and the methods used in linguistic
analysis. Often linguistic analysis deals with recognizing patterns, and
uncovering hidden representations. Many linguistic analysis tasks are
really like puzzles. Along the way you will be surprised to discover you
already unconsciously know very complex things about language and
language structure. In the following sections I will give a brief sum-
mary of the major topics covered in this book, noting which topics are
most relevant and important to artificial intelligence. Keep in mind
that artificial intelligence research has two sides: on the one hand we
are trying to understand human intelligence, in order to be able to
model or recreate it. To this end, we study humans, and human devel-
opment and thought processes, and we use computational tools to do
this. On the other hand, we want to create intelligent systems that can
interact with humans in a natural, fluent way. In both these aspects
there is a need for understanding the structure and communicative
functions of language.

Phonetics

Phonetics is the study of speech sounds. We can distinguish between

acoustic phonetics, that looks at the physical properties of sounds
when transmitted, articulatory phonetics, which focuses on the
production method of the sounds, and auditory phonetics, which
focuses on our perception of speech sounds. Acoustic phonetics is rel-
evant to speech development and tasks such as speech recognition and
speech synthesis. Acoustic and auditory phonetics can explain mis-
hearings and is also relevant when developing smart auditory systems,
 introduction 7

such as smart hearing protection. Articulatory phonetics interacts

with other language processes. The way in which sounds are produced
influences the production of nearby sounds, which in turn can lead to
changes in their realization and perception. Auditory phonetics is also
concerned with how speech sounds are learned and recognized, which
also gives us insights into human learning mechanisms.

Phonology

Within a given language, the meaningful speech sounds of that lan-

guage, and the patterns within which they occur are studied in the
field of phonology. Already at this level we will be able to see how
language is made up of overt and covert representations and processes
that act upon them. Within phonology we can also introduce the first
examples of linguistic analysis and reasoning by looking at how we can
decide what the underlying ‘base’ form of a sound is in a language. The linguistics llama. There are
many linguistic memes created with
the llama, including this one which
Morphology is (admittedly only slightly) amusing
because of the confusing between the
words “pears” and “pairs”.
The study of morphology deals with the structure and formation of
words. What is a word? How do we make new words? What kind
of word classes are there? How do we manipulate the forms of words
so that they are appropriate in different contexts? Why do languages
requires these oftentimes redundant manipulations?
Computational morphological analysis is also an important step in
many natural language processing applications, and may become more
important in the future. For example, in neural machine translation,
which uses neural network machine learning to learn how to translate
between a source to a target language, an important step is segmenting
the words into meaningful parts that the system can learn to "match"
between languages. This can be done using naive methods, that do not
appeal to linguistic knowledge, or it can be done with actually morpho-
logical parsing. This can even been done by a computational linguistic
with a language they themselves do not speak, by using techniques to The linguistic llama again: In syn-
find patterns, often relying on knowledge about the phonology of the tax some intermediate structures be-
tween phrases and words are “bar”
languages cross-linguistically. structures’, which can be written as
the category plus an apostrophe. In
this meme, A’ is pronounced “A bar”.
Syntax

Words can be combined to form sentences, which then can be used

to communicate statements about situations, or to ask questions or
give commands. But the way in which words are combined is also
structured. The dog bit the man and The man bit the dog mean two
completely different things, despite containing the same words. Lan-
guages differ greatly in their syntactic forms. Inu ga otoko o kanda
in Japanese puts the verb at the end of the sentences, rather than in
the middle, and identifies the subject and object by the use of par-
ticles after their arguments. This contrasts with English which has
to rely on position to distinguish subjects from objects. Languages
8 an open introduction to linguistics

like Serbian and Finnish rely more on morphology. Computational

syntactic analysis is done by syntactic parsers. Syntactic parsing is
less aggressive and more pragmatic in their analysis choices than the
theoretical counterpart. Building syntactic parsers is a major task,
but to achieve true natural language understanding, it is a necessary
one. The goal of NLU is to understand a speaker or writer’s intention,
and this requires first understanding their meaning, and it is generally
agreed that syntactic structure feeds meaning. Working in the other
direction, from meaning to natural language generation (NLG), also
requires generating correct syntactic structures. While recent machine
learning methods have shown amazing results in this area, there is
still much difficult work ahead that requires old-fashioned syntactic
analysis.

Semantics

Semantics is the study of meaning, how it is constructed and what

additional inferences follow from a given statement. Language in iso-
lation is extremely ambiguous, and while we don’t often notice the
ambiguity at the more basic levels, such as morphology, we do notice
it at the higher level of sentences outside a clear context. How can we
represent meaning? How can we interpret sentences? These are the
topics of semantics.
Lexical semantics, historically considered the poor cousin of struc-
tural semantics by formal linguists, has found a major role in current
natural language processing applications, in particular because of the
success of techniques that rely on extracting the distributional patterns
of word usage.

Pragmatics

It simply isn’t sufficient to know the meaning of a sentence to under-

stand how it is used in communication. How contextual information,
linguistic and non-linguistic, influences our language understanding, is
the topic of study in Pragmatics. For example: You got two lollipops!
could be what a parent would say to scold a child whining about not
having gotten any candy, but could also be used by a friend to point
out a happy result after playing a bingo game with candy prizes. Our
ability to reason about how other humans communicate with us also
influences how we understand sentences said to us. If I say “Some
of the students failed the exam” most people would interpret that to
mean that not all of the students failed. But if you’ve taken an intro-
ductory course in logic (and if you think about it a bit too much) you
will agree that actually, the sentence does not say that, and it is in fact
compatible with the potential horrible state of affairs when all of the
students failed the exam. (So far, never happened, thank goodness!).
 introduction 9

Language Acquisition
How do we learn to speak our native language? Is this different from
how we learn other skills or knowledge, or how we would learn to
speak a new, second language, as an adult? The answer to both these
questions seems to be yes. Language is different from other cognitive
skills, and much research has shown that we possess preprogrammed
language structures as part of our genetic endowment, allowing us to
quickly learn our native language when we are exposed to it as children.
Language acquisition research helps us better understand human
intelligence and human learning. For this reason, understanding the
basics of how language is acquired is important. Further, we can often
better understand the final, adult state of human language knowledge
if we examine how that state was reached. Like other psycholinguis-
tic research, theories of child language acquisition have benefited from
more advanced statistical modelling and computational models of lan-
guage learning. We’ll touch briefly on that topic.

Applications of linguistic knowledge: Language and speech tech-

nology
An introduction to linguistics is a key element in any artificial intel-
ligence program in order to understand and improve existing speech
and language technological applications. In the current course we will
only be able to spend a short amount of time on these topics, but get-
ting an initial glimpse will help make the usefulness of the course more
concrete. For this reason we’ll look at machine translation and topic
identification as two applications where linguistic analysis is useful.

Machine Translation
Machine translation is the task of translating from as source language
to a target language automatically with a computer program. The
history of machine translation research is actually an interesting topic
in itself, because initial efforts to use computers to analyze text were
all focused on the task of translation. It still remains one of the major
research areas in computational linguistics today. It’s also particularly
interesting because it involves all levels of human language. For speech-
to-speech translation, both speech recognition and speech generation
is necessary. Depending on the structure of the language and the
translation technique used, pre-processing steps such as segmenting
the input into morphemes, disambiguating homographs, dealing with
issues such as dropped subjects, etc. are all tasks that require linguistic
analysis.

Topic identification
Automatic topic identification requires analyzing texts to determine
what the dominant topic is automatically. This is useful for all kinds
of applications, including twitter misinformation detection. We’ll talk
a bit about this in the last lecture.
10 an open introduction to linguistics

Final notes about this book

Over a decade ago, I was dissatisfied with the text book we were using
for this course. Alternatives were similar: large American textbooks
of 600 or more pages, where we only used a small subset. Most intro-
ductions also failed to discuss computational linguistic applications of
the included topics, and instead focussed too much on topics needed
for linguistic students. I was then quite happy when I found lecture
notes written by Markus Kracht for an introductory course on Lin-
guistics for linguistic students. It was more of a book, than lecture
notes, and the targeted audience was still linguistic students, but I
saw the possibility of using it as a base, and revising it to make a text-
book that is truly appropriate and most of all, inspiring for Artificial
Intelligence students. Writing a textbooks takes more time than you
may think (or even than I thought) and even though each year I make
some improvements, the books is still not finished. I’ve written some
of the chapters (e.g. Phonetics, the first half of Phonology, the first
two Semantics chapters and the Pragmatics and Language Acquisition
chapters are mine) and have revised the other chapters. We’ve also
added study exercises and margin notes to all chapters. Some chap-
ters from the original lecture notes from Kracht remain, though with
substantial additions and revisions throughout. This means the style
varies throughout the book, with some clearly mine, and others clearly
written by Markus.
Of course, there are some disadvantages of choosing a reader over
a polished textbook is that it has not been written with the help of
professional editors. There are errors. There are sections that could
be more clear. There are omissions. This is the cases even despite the
small army of TAs and colleagues that has helped me edit and improve
it over the years. The advantages are that this reader is designed to
support learning exactly the topics that I think are relevant for AI
students It presents the topics similar to how the lectures present the
topics. And it is free.
It is also important to note that it is a reader, not a textbook. It is
not a stand-alone product, but should be understood to be only one
part of the General Linguistics course, along with the lectures, the
tutorials, and the homework. Together I believe the materials give a
solid introduction to linguistics for AI students.
Some tips. New words and technical terms used for the first time are
usually bold, generally with a definition. Be aware that other linguists
might define things differently. If you find a different definition in
another book, it does not mean that the other books (or this book)
are necessarily wrong. There’s a lot of terminological confusion/in-

fighting in linguistics, so just be aware. The symbol in the margin
signals some material that is difficult, and optional. //

Jennifer Spenader, Groningen, Feb 6, 2022

2
Phonetics

Phonetics is the study of sounds used by humans in language.

We can distinguish between three different types of phonetics: (1)
Articulatory Phonetics is concerned with the way in which we
use our tongue, mouth and vocal track to create human sounds.
(2) Acoustic Phonetics is concerned with the physical proper-
ties of the sounds humans produce in communication. (3) Au-
ditory Phonetics is concerned with the way humans perceive
language sounds. Of course, all three of these sub areas are in-
timately connected: physical properties of sounds are perceived
categorically, and errors in perception can be explained by articu-
lation similarities that in turn can be traced back to the acoustic
signal. In this chapter we review the nature of speech sounds,
how they are produced, and how they are perceived. We will also
discuss the issue of how we record speech sounds, introducing
the International Phonetic Alphabet. This all forms an impor-
tant basis for understanding phonology, the study of meaningful
sound systems in a given language.

Spoken language is transmitted via speech sounds. Speech sounds are

those sounds that we can make with our lips, teeth, tongue, and the
different positions in the inside of our mouth. As humans, we can make
other sounds as well: sneezing or coughing also create noise. But these
are not controlled, and therefore are not meaningful in any language.
In fact, even newborn infants recognize the difference between inten-
tional speech sounds and these involuntary sounds.
Phonetics is the study of human speech sounds, both their produc-
tion and comprehension.
This chapter is about phonetics. Phonetics is concerned with the
mechanics and form of human speech as a potential carrier of meaning.
But the study of phonetics does not involve meaning or communica-
tion. What makes the sounds we can produce meaningful, or able to
communicate meaning, is studied in phonology, the topic of the next
few chapters.
However, the sound patterns in a given language, and how cate-
gorically perceived sounds are realized in practice are all related to
how sounds are produced by the vocal tract and how closely related
different features are acoustically. This means that without an under- The Vocal Tract.
12 an open introduction to linguistics

standing of phonetics, the common patterns found in the phonology of

the world’s languages will seem arbitrary, when in fact they are quite
predictable.

The vocal tract

Most speech sounds are pulmonic egressives: that means the sound
is created by using a stream of air coming from the lungs.
We inhale air, and then we release the air from the lungs in short
puffs. We are unconsciously able to control this exhaling using the
muscles of our rib cage, maintaining even air pressure. The air from
the lungs travels first through the trachea into the larynx which houses
the vocal cords. The larynx is also called the voice box. In men,
the position of the larynx is visible in the neck. It’s therefore colloqui-
ally named the Adam’s apple. The vocal cords are two small flaps
made out of tissue that can be completely open, completely closed, or
The vocal cords when open. Air
partially closed. The gap between the vocal cords also has a name: flows freely throught the glottis.
the glottis. When partially closed the vocal cords create vibrations
which become sound.
The larynx is then considered the first point of phonation, the be-
ginning of speech sounds. Air then travels through the trachea to the
pharynx. The pharynx is a tube-like chamber that connects to the
mouth and nasal passages from the trachea, if you think of the flow
of air from the lungs. But if you follow the pharynx from the mouth
cavity down towards the lungs, it actually splits into two passages: the
larynx and trachea (which goes to the lungs) and the esophagus (which
goes to the digestive tract). In order to prevent food from entering the
esophagus by mistake, there is a flap of tissue called the epiglottis
that covers the esophagus. The pharynx also has resonating proper- The vocal cords are closed, only a
small stream of air flows through the
ties, and it’s shape can be manipulated slightly by the position of the
opening.
back of the tongue, allowing it to also participate in phonation.
The velum determines whether or not air enters the nasal cavity.
If closed, air will only flow through the mouth. Once a flow of air
is present, it can be manipulated to produce different sounds. Most
of this manipulation happens in the oral cavity (the inside of your
mouth), where the tongue can be in different positions. The oral cavity
can also be opened to different degrees and the shape of the lips can
be varied. (Not so) fun fact:
Above we described sound production using an air stream expelled It is actually an unfortunate design
of the human body that the phar-
from the lungs. Sounds can also be produced with air entering the ynx leads to both the lungs and the
mouth going to the lungs but it’s less natural. These are called pul- stomach, because this means eating
and breathing are done via the same
monic ingressives. Pulmonic ingressives are generally only used par- passage way. Because the epiglottis,
alinguistically, that is they are used to convey things like attention, which is supposed to protect the air-
emotion or agreement in a conversation. For example, both in Swedish way from food entering, sometimes
fails in this function, food can get
and Dutch, pulmonic ingressives have been used to signal that one is stuck in the trachea, which leads to
listening (or agreeing) in a conversation, similar to the English use of airflow being cut off i.e. choking.
“uh huh”.
Almost everyone is familiar with the distinction between consonants
and vowels: the articulatory difference has to do with the degree of ob-
 phonetics 13

struction present. For vowels, there is no obstruction, and the different

vowels are produced by manipulating the shape of the vocal (and nasal)
cavity. Consonants have some sort of obstruction, sometimes even a
complete one. As the description of consonants is simpler than that of
vowels, we will begin there.

Consonants

Consonants are speech sounds created by (partially) blocking the flow

of air from the lungs. They are usually described by three features:
(1) the manner in which they are produced, (2) the position of their
obstruction, and (3) their voicing, i.e. whether or not vocal chord
vibration is involved.
Let’s begin by describing a class of sounds that all share the same
manner of productions: plosives. These are sounds with a complete Active and passive articulators
in an X-ray
obstruction of the expelling airflow, including the raising of the velum From: The Phonetics of Consonants
to prevent air from flowing into the nasal cavity. After pressure builds, (Fuchs and Birkholz (2019), Fig. 4a)
the obstruction is removed, and there is a a sudden release of air (hence
the name). In English, ‘p’, ’b’, ’t’, ’d’, ’k’ and ’g’ are all plosives. For
’p’, the lips are initially closed, pressure can build up in the oral cavity,
until at a certain point the lips open and the air is released.
The position of the obstruction creates different consonants. The
sounds ’p’ as in pie and ’b’ as in bell create and obstruction using both
lips, and are therefore called bilabials. But the place of obstruction
can also be at the teeth (dental), (not a sound in English), just behind
the teeth (alveolar), e.g. ‘t’ as in tie or ‘d’ as in doll, or even further
back, such as at the velum (velars, e.g. ‘k’ as in kite or ‘g’ as in girl).
So we can talk of sets of sounds as being bilabial plosives (e.g. ‘p’ and
‘b’) or velar stops (‘k’ and ‘g’).
The third defining feature of stops is their voicing. Some sounds
are produced with the active vibration of the vocal cords. These are
voiced sounds. If the glottis is open, and the vocal cords do not A voiceless alveolar plosive. Note
vibrate, we get voiceless or unvoiced sounds. For example, ‘p’, ‘t’ the representation of vocal cords as
not vibrating in the diagram.
and ’k’ are voiceless while ‘b’, ‘d’ and ’g’ are voiced. You can easily feel
the voicing of different sounds by putting your hand on your larynx (the
middle of the front of your neck, where you can feel a hard cartilage
tube under your skin) and producing a series of the same sounds, e.g.
‘p p p p p’ vs. ‘b b b b b ’. Alternatively, you can put both fingers
in your ears and produce a series of the sounds. In both cases voiced
sounds will produce a much louder, buzzing like noise and vibration.
Affricates are consonant sounds that initially cut off all air flow,
followed by a less explosive, more gradual release of air. The initial
sound in the English words jazz and chess are both affricates.
But many consonants don’t require complete obstructions. Frica-
tives allow airflow but there is friction, and the opening is very narrow. A voiced velar plosive A Sagit-
tal section of the position of the vo-
Examples of English fricatives include the sounds associated with ‘s’ cal tract for the pronunciation of the
as in soup and ‘z’ as in zebra, ‘f’ as in fun and ‘v’ as in vase. Fricatives sound associated with ‘g’ (goat) in En-
glish. Note that the vocal cords are
are created by obstructing the airflow partially. Like stops, fricatives vibrating, the velum is raised to pre-
can also be identified by the place where the constriction occurs and vent airflow to the nasal cavity, and
the point of obstruction is the velar
(the back of the roof of the mouth).
14 an open introduction to linguistics

the voicing. Thus ’f’ is a voiceless labiodental fricative, while ‘v’ is a

voiced labiodental fricative.
Plosives, fricatives and affricates together make up the class of ob-
struents, so called because they all involve a large degree of obstruc-
tion in their production.
Approximants allow air flow through a narrow constriction but
without obstruction or friction. In English sounds associated with
the letters ‘r’ (read), ’l’ (leaf ), ’y’ (young) and ‘w’ (wow!) are all
approximants.
The sounds described above are all oral consonants, but the nasal
cavity can also be involved in consonant production. Nasals are
sounds where the airflow through the mouth is blocked temporarily, A voiced dental fricative
but air is allowed to flow through the nasal cavity by having the velum Note that in a voiced dental fricative
lowered. In English, the sounds associated with ‘n’ (nose), ‘m’ (mouse) air flows through the sides and just
under the teeth, but this small gap is
and the final sound ‘g’ sound in words like long and sing are all nasals not easy to see in the drawing (look
consonants. carefully), so the diagram looks very
similar to a voiced dental stop.
A final group to be aware of is the set of stops. This term refers
to both plosives and nasals like ‘n’ and ‘m’ with restricted airflow.
In addition to these sounds, languages of the world also use trills,
taps and flaps. Trills can only be produced with good airflow past
soft tissue articulators that can vibrate in the air flow, such as the lips
or the tongue. Like plosives, there is a closure created, but different
from plosives, the closure is less tight, less firm. As pressure builds
up behind the point of closure, air is briefly released, but then as the
pressure drops, the closure re-seals. This process occurs rapidly in suc-
cession, creating the trill sound. In speech sounds, only 2-3 vibrations
are usually used, but speakers of languages with trill consonants can A voiceless labio-palatal frica-
also create the trills much longer if desired. tive. This fricative does not occur in
many languages. It has been attested
Taps and flaps are created with rapid movement of the tongue, cre- in Kham, a Sino-Tibetan language,
ating a very short obstruction that is quickly released. No air pressure and Iaai, a Polynesian language. Note
builds up behind the obstruction, which means there is also no burst that the obstruction is not complete
(fricative), and no oscillation of the
of air at release. Whether or not there is a difference between taps and vocal cords.
flaps is still a matter of debate, though originally there were proposals
that flaps are produced more with the back of the tongue than taps.
Many linguists only use the term flap nowadays, but you should be
aware that tap is another name for a similar sound. .
Above we have concentrated on the pulmonic consonant sounds,
but there are also a number of non-pulmonic consonants. While it’s
not the case that the lungs are not at all involved in non-pulmonic
consonants, lung expression is not central to their production.
Click consonants are a well-known type of non-pulmonic conso-
nant. These consonants are created by creating two closures in the
mouth, one to the front and one to the back. Clicks are thus ingres- A bilabial nasal
sives. The tongue is lowered, and then the back closure is released, Note that the velum is lowered so
that air flows through both the oral
creating a click sound. These consonants are the loudest consonants
and nasal cavities.
in human languages. The ‘tsk tsk’ expression of disapproval in English
is actually a dental click. The ‘clip clop’ sound made to imitate the
sounds of a horse trotting in English is actually an alveolar click. Thus
clicks are used in English, but not as speech sounds. In contrast, sev-
 phonetics 15

Term Articulators involved

bilabial both lips, both active and passive
labiodental active lower lip with passive upper teeth
dental active tongue tip/blade touches passive upper teeth
alveolar active tongue tip/blade touches passive front part of
alveolar ridge
postalveolar active tongue blade touches behind alveolar ridge
retroflex active tongue tip raised or curled touches passive
postalveolar
palatal tongue blade/body touches hard palate behind entire
alveolar ridge
velar active body of tongue raises to passive soft palate
(sometimes to back of soft palate)
uvular active body of tongue touches passive (or active)
uvula
pharyngeal active body/root of tongue touches passive pharynx
glottal both vocal cords involved, both active and passive

eral languages spoken in Africa, such as in the Bantu languages Xhosa

and Zulu, do use click sounds as part of their consonant inventory. Location of languages with click
consonants (from Wals):
Sounds can also be made by varying the pressure inside the oral
cavity by manipulating the glottis. For example, glottic egressives
are made by closing the air flow from the mouth, raising the velum
to prevent air from flowing into the nasal cavity, and then raising the
glottis to create higher pressure in the oral cavity. This can then be
released, creating sounds called ejectives. The Kartvelian language
Georgian uses glottic egressives. Glottic ingressives are also possi-
ble, with instead of pressure building up by raising the glottis, pressure
is reduced by lowering the glottis, creating implosives. The Chadic
language Hausa has implosives.
Before we move on to vowels, we should first step back and be more
precise in how we record the different sounds we have described so
far. Clearly there is a mismatch between how things are written in
English, and how they are pronounced. This mismatch isn’t restricted
to English, (though it might be greater in English than in many other
languages) and is one of the reasons why we have to be taught how
to spell. We could try using the same letters of the alphabet we use
in English to represent the sounds of English, but there will be a
problem with this approach: sometimes the same letter is used for
several different sounds. For example, the English word thoroughgoing
uses the letter ‘g’ to represent three different pronunciations. At the
same time, the same sound may be coded with different letters. For
example both the ‘s’ in ‘sent’ and the ‘c’ in ‘cent’ are pronounced the
same, and in fact both these words are pronounced the same.
Even if we could modify the alphabet we use in English to encode
the sounds of the world’s languages, it wouldn’t be sufficient, since
there are many more sounds than the ones used in English. This
16 an open introduction to linguistics

problem was recognized early on, and lead to the development of IPA,
the International Phonetic Alphabet. IPA was created in 1888 by the
International Phonetics Association. It’s based on the Latin alphabet,
but has been designed to make it possible to record in writing a good
approximation to all speech sounds in languages of the world. At the
end of this chapter you can find the full IPA chart from (2005) (the
most recent version), but first we’ll zoom in on the different sections
separately. Let’s first examine the part that relates to pulmonary
consonants: How to produce click consonants:
(Diagrams: JKS)

The columns show the major places of articulation, or the places

where a construction occurs, while the rows show manner of artic- Step 1: Make two obstructions, one
ulation. Visualize a saggital section on top of the IPA chart and in the front, one in the back.
you can see how the left columns show the symbols for sounds pro-
duced with obstructions at the front of the mouth, and the right side
shows symbols for sounds produced with obstructions at the back of
the throat. Remember, pulmonic consonants are described by three
features: place, manner and voicing. In order to add the third voic-
ing dimension, two symbols are presented within cells where a voicing
contrast is possible, with the left symbol representing the voiceless con-
sonant, and the right symbol is the voiced consonant. Some cells are
empty. These are sounds that are in practice not possible to articulate.
For example, a bilabial approximate is not a possible sound, since ap-
proximates by definition require construction in the middle of the oral
Step 2: Lower the tongue to increase
cavity. Note also that some distinctions are not used, often because size of the oral cavity
their differences are acoustically too similar to be distinctive. So for
example, among coronal plosives, there are no basic symbol distinc-
tions between dental, alveolar and palatal alveolar sounds. However,
fricatives do make clear distinctions between these different places of
articulation, so IPA provides different symbols for the different posi-
tions in the fricatives. This doesn’t mean that coronal plosives don’t
differ in how front or back the tongue is. In some languages or di-
alects, coronal plosives may be produced at the teeth (dental) the
alveolar ridge (alveolar) or further back (palato-alveolar). When talk-
ing about the sounds in a language, the more specific description is
used. Thus English (most speakers) and Dutch technically have alve-
Step 3: Release the front closure to
olar plosives, Japanese and French have laminal denti-alveolar create a sound.
coronal stops (laminal = the tongue is then touching just behind the
teeth (denti-alveolar)). Even within these sounds, languages or dialects
can be a bit more dental or a bit more palatal. These differences are
 phonetics 17

not as profound, but IPA still makes it possible to code them with
diacritic markings. See in the table below how more dental or apical
sounds can be coded.
A given meaningful sound in a language is called a phoneme. We’ll
talk in much more detail about phonemes in the next chapter but
briefly: phonemes are the sounds that are meaningful in language.
Alveolar voiceless stops are meaningful in English. It could be that in
rapid speech, some English alveolar voiceless stops will be produced
as if they were laminal denti-alveolar stops. But English hearers will
classify both sounds as the same meaningful sound or phoneme. And
in fact, the context in which a speech sound is produced will influence
how close the production comes to its ‘ideal target’. For example,
consider the word teetotaler (someone who does not drink alcohol).
This would include three voiceless alveolar plosives ([t]). But if you
pay attention as you say it, you will notice that whereas the first two
stops are pronounced at the front of the alveolar ridge, the third is
pronounced more palatal, more to the back. The reason is the context.
The first vowel, teetotaler is a front vowel ([i]), but a diphthong that
moves to a back vowel ([@U] and then to schwa). These diacritics also
come in handy for giving a more detailed (or narrow) transcription
of sounds in context.

Of course, it’s also important to have a representation of non-

pulmonic consonants. IPA also provides these.

Co-articulated consonants are produced with two places of ar-

ticulation at the same time. There are a number of quite common
co-articulated consonants and these are the ones IPA provides symbols
for. For example, many African languages use the voiceless labial-velar
>
stop, such as in the Igbo word for call: [kpÓ]. For other co-articulated
sounds a tie can be drawn over the two consonants.
18 an open introduction to linguistics

Consonants are described by their central features. The sound [p]

is a voiceless bilabial stop, named with voicing-place-manner.
[Z] is a voiced velar fricative. If there are additional attributes,
they are placed just before the manner descriptor, so [ph ] is a voiceless
bilabial aspirated stop. For a given language we can talk about the
languages consonant inventory, and this can be shown visually within
an IPA table. See the table of American English Consonants below: How to produced a glottalic
sound:
In order to produce an implosive b,
do as follows:

• Close your lips together so as to

pronounce a [b].
• Move your glottis downward as if
you were yawning. You should
be able to feel it move with your
fingers; if you have a noticeable
Adam’s apple, you should also be
able to see it move in a mirror.
• While ‘yawning’, open your lips
and say [ba].
• Try doing this quickly so that the
air flows into your mouth while you
pronounce the [b].
• There should be a deep hollow
American English Consonant Inventory sound, and the [a] should follow
smoothly. (From Wikipedia, Glot-
talic Consonant)
Vowels
Vowels are somewhat trickier to describe, because they are all produced
in a similar manner, with free airflow, and vary in the places where the
oral cavity is narrowed. Compared with consonants, they are also more
sonorous. Sonority is often described as being related to loudness:
sounds with more sonority, like vowels, will have a greater amplitude
(be louder) than for example, stops or fricatives. Sonority also relates
to resonance, the degree to which a sounds is able to move through the
air. More sonorous sounds have more resonance, e.g. you can ‘sing’ on
them, and they will dominate over other sounds in the environment.
There are four features that can be used to identify vowels: (1)
Place (2) Height (3) Rounding and (4) Tenseness. Let’s zoom in on
the first two features that make up the horizontal and vertical axes of
the IPA vowel diagram, shown below:

Sagittal section with Vowel chart

superimposed
 phonetics 19

Place and Height The form of the vocal tract affects the sound
that air flowing from the glottis through the mouth makes. We modify
this form mostly by modifying the placement of the tongue, limiting or
permitting airflow, narrower openings in different positions. The places
where airflow can be partial obstructed can be at the front, back or
even a central position in the mouth. Further, we can have the main
body of the tongue either closer to the roof of the mouth, creating a
American English Vowels
more closed cavity, or we can have it more open, by having it flatter
American English has fewer vowels
in the mouth. Sometimes, the same feature is described from the than British RP English.
perspective of the tongue, and some researchers instead talk of high
vowels, synonymous with closed vowels, and low vowels, which are
synonymous with open vowels. Tongue height is a horizontal feature,
and we can think of how closed or open the inside of the mouth is as
a vertical feature. The IPA charts for vowels intuitively captures this.
You can imagine a sagittal section with the mouth opening to the left
superimposed on the chart, to understand how the chart relates to the
actual moth (see margin picture).
So for English, the sound in beet and feet is the most closed front Japanese vowels Japan has a vowel
vowel (or high front vowel), while the sound in the words bat and cat system with only five vowels. Accord-
ing to WALS, in the 564 languages
is the most open front vowel (or low front vowel). There are several surveyed, 287 languages have a vowel
vowels in between (see the margin note) so you can actually say a inventory between 5-6 vowels.
sequence of words with only front vowels, and notice how with each
vowel the mouth opens a bit more, e.g. Say beet-bit-bait-bet-bat. You
can also notice the difference between front and back vowels by saying
the English words in the following sequence: Say bet-but-bought.

Term Vowel features

close compared with other vowels, tongue is high, closest to the
roof of mouth, creates smaller cavity
open compared with other vowels, tongue is low, far from the
roof of the mouth, creating a large open space
close-/open-mid intermediate positions (also: mid, upper-mid, lower-mid)
front compared with other vowels, tongue is overall forward
central intermediate position
back compared with other vowels, tongue is overall back (near
pharynx)
rounded lips protrude forward

Rounding Vowel rounding is produced by simply rounding the

lips (making an ‘o’ shape, or a flatter extendend horizontal open-
ing) thereby extending the vocal tract and changing the quality of
the vowel. This creates a contrast with the unrounded form of the
vowel, which is produced with relaxed lips. In the IPA vowel chart,
rounded vowels appear to the right of their unrounded counterparts.
Roundedness isn’t always distinctive in a language. There is also a nat-
Swedish vowels. Note the presence
ural tendency to produce front vowels with unrounded lips, while back of several rounded vowels (e.g. y,Y,ø.
vowels are more naturally produced with some rounding. English, for Also, note the triangular colons (e.g.
i:,y:,0:). These indicate that these
vowels are longer in duration than
other vowels.
20 an open introduction to linguistics

example, does not use rounding to distinguish different vowels, except

in a handful of dialects (e.g. Geordie, General South African English).
But in many languages two vowels with the same height and place
are distinguished by rounding. Swedish, for example, is a language
with many rounded and unrounded vowel pairs.
Tenseness is just as it sounds: how tense, or tight, are the muscles
of the tongue in the production of the vowel sound? Tense vowels are
produced with tighter muscles than their non-tense equivalents, which
are called lax vowels. Germanic languages such as RP English, Dutch
Standard German have distinctions based on tenseness. For example,
the Dutch vowel [I] is a closed, front unrounded lax vowel (as in blick,
a glance) which contrasts with the Dutch vowel [i], a closed, front
unrounded tense vowel (as in biet, beet). However, tenseness is not
the only difference: there is a slight difference in place and height: the
lax [I] is more central and more open. In languages with distinctions
between tense and lax vowels, tense vowels are also usually longer
than lax vowels. But there are again exceptions. In Scots English and
Yiddish, which also have distinctions based on tenseness, there is no
correlation between tense and length.
The Japanese vowel inventory is relatively small cross-linguistically.
Note again how the five vowels are almost equally distant from each
other. Swedish on the other hand shows a large vowel inventory, with a
large number of contrasts in roundedness, which in the chart is shown
with two vowels next to the same dot: the right symbol is the symbol
of a rounded vowel, and the left symbol is the unrounded version.
One particular vowel is worth discussing in more detail: the schwa.
This is a mid-central vowel ([@]), that’s produced with the tongue at
its most relaxed position. Some languages (e.g. Russian) have a schwa
sound, but in many languages the schwa is the vowel that all other
vowels are reduced to in lazy speech.

Vowel inventories

Each language has its own vowel inventory, and these can differ enor-
mously. Among European Languages, Danish is the front-runner in Daniel Jones Inventor of the cardinal
vowels.
monophthongs with 32 vowels, followed by Norwegian with 19. French,
German and Swedish have 17 and Dutch, Finnish and Icelandic have
16. There are many languages that have a small vowel inventory. In
particular, languages of the indigenous peoples of north American, and
the indigenous languages of Australia have small vowel inventories of
less than 5 vowels. In fact, the Aymara, Cherokee and Haida have only
three vowels.
Even among vowels that IPA would classify as the same vowel, the
actual vowel in a given language may differ in exactly how it is pro-
duced. We can talk about a language having a high front vowel, but
that high front vowel might be more or less high in one language com-
pared to another. These differences are signalled by the way in which
the vowel inventory for different languages is positioned in an IPA style
vowel diagram. Consider the charts for American English, Japanese Cardinal vowels 17 of Daniel Jones’
cardinal vowels, without the mid vow-
els.
 phonetics 21

and Swedish in the margin notes.

First, we can see that there are quite a few vowels, but no rounded
ones. The different vowels are fairly separate from each other. In
general, vowels will tend to be at an equal distance from each other.
This is because of the need to be understood. The production of all
sounds is influenced by the surrounding sounds, but also by features
of the speaker (fatigue, a cold). This means the actual production is
going to be a range of openness, how front or back the vowel is and
how tense it is. But the hearer still needs to recognize the intended
target, so the central ‘target’ position needs to be separated for each
vowel to prevent overlapping. Now consider the vowel inventories of
Japanese and Swedish. See the diagrams in the margin.
The cardinal vowels
Before (portable) sound recording devices existed, language researchers
still needed to record the sounds of different languages accurately. IPA
helps well with consonants and vowels to a certain degree, but to make
much more accurate transcriptions all linguists had to agree on what
sound the different points on the diagram stand for. Daniel Jones was
a phonetician who developed a method to do this, by developing the
Dutch diphthongs.
cardinal vowels. The cardinal vowels are the vowels that are pro-
duced at the extreme points of the vowel chart, as they are produced
by Dr. Jones himself. So the cardinal [i] is the most front closed
vowel that Daniel Jones could produced. After being trained to pro-
duce these cardinal vowels in the vowel chart correctly, a researcher
could go investigate a new language, and annotate the features of the
vowels of that language in a way that would allow other linguists to
reproduce the same vowels. In general, only researchers trained by
Dr. Jones himself, or trained by someone trained by Dr. Jones, could
be said to actually know the cardinal vowels. It was actually quite a Main English vowels and diph-
clever system to insure standardization of vowel transcription before thongs (arrows) in the IPA chart

recording devices were available.

Diphthongs are vowels where the tongue moves from one posi-
tion to another during the pronunciation, in essence changing from
one vowel to another. They contrast with the monophthongs we’ve
been talking about until now. Dutch has a large inventory of diph-
thongs, and this is even reflected in its orthography, where many “dou-
ble vowel“ exist. English also has many diphthongs. For example, the
words low [ao], light, [2I] and lair [Eô] all have diphthongs as their
“ “
nucleus.
Vowels are described with Height-place-lip/tenseness-(nasality).
So Y is a high front rounded vowel, while [6] is a low back vowel.
Amplitude (loudness) is shown in shading: black signals high ampli-
tude, while very light gray shows very low amplitude. The spectogram
above shows the acoustic differences between the English words “ball”
and “bar”. The formats are highlighted with a yellow line. Note how
different frequency componants are clearly visible as darker lines, that
then changes frequencies over time. These are also called formants
and you can clearly see differences between the different words. Spectogram of “ball”
and “bar”.From Hagiwara
(https://home.cc.umanitoba.
ca/~robh/howto.html)
22 an open introduction to linguistics

Acoustic features of consonants and vowels

Speech is sound, so it’s physically actually a wave form. We can vi-
sually see speech with a spectogram which shows three dimensions of
speech. (See margin note) The vertical axis shows frequency in Herz.
The horizontal axis shows time.
One acoustic feature of consonants that we will refer to in later
chapters is Voice Onset Time, or VOT. Voice onset time is the time it
take from the moment of release of stop until voicing (i.e. vibration of
the vocal cords) begins.
Voiceless stops have a VOT near zero: almost immediately after
the release, the voicing of the following vowel begins. Voiced stops
have a VOT of before zero. With voiceless stops, the vibration of the
vocal cords may begin before the release or exactly with the release.
Aspirated stops have positive voice onset times. This means that after
the release of the stop, there is a period of time without vocal cord
vibration. Voice onset times of different
Below you can examine different VOT in an oscillogram and a spec- consonant types.

togram for a voiceless alveolar fricative and a voiceless alveolar aspi-

rated fricative. Note the period where there is no voicing directly after
the release of the fricative. Note that this example shows that VOT is
relevant even for fricatives, not just stops.

Another important acoustic feature of speech is how the formants

signal different vowels. Formants are concentrations of acoustic en-
ergy at certain frequencies, and they help distinguish between different
sounds. Formants are named by their position from the lowest formant,
which is the F0 or fundamental frequency of a speech sound. The next
concentration in Herz is the first formant (F1) and then the next is
the second (F2) formant and so on. In a spectogram they appear as
darker lines at different frequencies. See the margin example.
F1 and F2 are particularly relevant for vowel production and per-
ception, and in fact based on F1 and F2 vowels can often be identified.
In the chart below you can see the formants of Japanese vowels pro-
duced by four different speakers, with F1 on the x-axis and F2 on the
y-axis. The symbols show the mean value and the oval shapes around
 phonetics 23

the symbols show the variation. This shows clearly how production of
vowels aims at a target, but that that target is not always met. We
see that for the top three vowels the individual speakers do not overlap
with themselves in their productions, but that there is some overlap in
the lower two vowels. Still, even if vowel productions are off, hearers
can use the context to determine what vowel was intended.

Summary Phonetics is the study of the sounds of human speech.

We describe the sounds by their manner of articulation. There’s a
standardize system for writing human speech sounds called IPA, the
International Phonetic Alphabet. Consonant sounds differ from vowels
mainly in the degree of obstruction used in their production. Language
differ greatly in their sound inventories.

Extra Material
It is useful to have a look at the active sound chart at http://web.
uvic.ca/ling/resources/ipa/charts/IPAlab/IPAlab.htm. You can
go there and click at symbols to hear what the corresponding sound is.
Another very useful source is the Wikipedia entry for the IPA (http://
en.wikipedia.org/wiki/International_Phonetic_Alphabet) and the
entries for individual sounds, e.g. http://en.wikipedia.org/wiki/
Voiced_Velar_Fricative.

Study Questions
1. In your own words, describe the difference between phonetics and
phonology. Do you think there is a clear divide between the two?

2. Why is there a special alphabet for phonetics? Why can’t we just

use a regular alphabet?

3. We make a distinction between ‘broad’ and ‘narrow’ phonetic tran-

scription. When do you use the first and when do you use the
latter?
24 an open introduction to linguistics

4. Which three properties or dimensions are used to classify conso-

nants? And what are the equivalent four dimensions for vowels?

5. Write the IPA symbol for the following speech sounds: You can find the complete
IPA chart here: http://www.
a. voiceless velar fricative internationalphoneticassociation.
org/sites/default/files/IPA2005_
b. glottal stop 3000px.png .

c. voiced post-alveolar fricative

d. bilabial voiced implosive
e. voiceless dental fricative
f. rounded open back vowel
g. unrounded open front vowel

6. Identify the phoneme associated with the underlined letters in the

following words. Give both the name and symbol of the phoneme.

a. phone
b. theatre
c. ceiling
Persian (Farsi) Vowel Chart.
d. very
e. complete.

7. Go to the active sound chart mentioned above and listen to the

vowels. Can you hear the differences between the vowels or do some
of them sound identical to you? If they sound the same, why do
you think this is?

8. Consider the vowel chart of Persian (Farsi) and Japanese, shown in

the margin. Compare the vowel inventory of Japanese and Persian.
First, name each vowel in each language. Second, constrast the two Japanese Vowel Chart
vowel inventories by describing (1) where and how the vowels are
similar (2) where and how the vowels are different.

9. Automatic speech recognition is one of the major tasks in language

technology, and although performance on the task has improved a
lot in recent years (take Apple’s Siri, for example), it is still far from
a ‘solved task’ (just try Youtube’s automatically generated subti-
tles). With the knowledge you know have on the relation between
sounds, phones and phonemes, can you give one of the reasons for
why this is such a difficult task?
 phonetics 25

Figure 2.1: IPA Consonant Chart

26 an open introduction to linguistics

Figure 2.2: IPA Vowel Chart

3
Phonology I: Features and Phonemes

This chapter introduces phonemes and allophones. It also ex-

plains how sound regularities can be captured by using attribute
value structures to describe sound features. We also introduce
natural classes. A natural class is a set of phones that can be
described by shared articulatory features and can be analyzed
using attribute value structures.

Meaningful Distinctions
The last chapter introduced phonetics. Phonetics is the study of the
speech sounds that can be produced and perceived by humans. This
chapter introduces phonology, the study of meaningful sound systems.

What does it mean for a speech sound to be meaningful? First,

we should note that sounds are meaningful within a given language,
not on their own. So let’s begin by simply asking what sounds are
meaningful in English.
Is a bilabial voiced plosive, [b] meaningful in English? Something
is only meaningful if it communicates something, and a bilabial voiced
plosive doesn’t communicate anything on its own. But if we instead
examine several words with the [b] sound we might be able to detect
a meaningful distinction. For example the words bat ([bæt] and ball,
[bOl] are words that start with the [b] sound. We know these words
in English refer to things in the world, but how do we know what the
contribution of the [b] sound is to this meaning? Well, we can see this xkcd comic from Randall Munroe
if we contrast the words bat with mat, for example. These words refer (Linguist)

to different objects in the world, so clearly the use of the [b] rather than
the [m] is meaningful. There are many pairs of words that contrast
[b] with other sounds, e.g. ball-call, book-look, bed-head, etc. These
contrasts are all evidence that [b] is meaningful in English. Pairs of
words (with distinct meanings) which differ only in the pronunciation
of one sound have a special name in linguistics. We call these minimal
pairs. We can use minimal pairs to identify the meaningful sounds
in a given language. These language specific meaningful sounds are
termed phonemes of that language.
Systematically examining minimal pairs to determine which speech
sounds are meaningful in a given language helps us determine the
28 an open introduction to linguistics

phonemic inventory of a given language: that is, which sounds in

the language are meaningful. For each candidate phoneme, we look
for words that differ from other words with different phones. For En-
glish we’ve already started identifying the phonemic inventory. The
minimal pairs ball-call-hall tell us that [b] is distinct from [k], that [b]
is distinct from [h] and that [h] and [k] are distinct phonemes.
Minimal pairs do not have to be words that differ in the pronuncia-
tion of their initial sounds. If two words differ only in a medial sound
or a final sound, it is also evidence that the contrasting sounds are
phonemes in the language. Thus in English cat [kæt] and cut [kAt] is
evidence that [æ]-[A] are both different phonemes in English. Similarly
the two words cat [kæt] and cad [kæd] show that English has a con-
trast between a voiceless alveolar plosive and a voiced alveolar plosive
(e.g. [t]-[d]).
Now I have actually simplified things slightly. First, what do we
actually mean by “initial sound”? Sounds are waves, and we can see
clearly from a spectrogram that even if we can pronounce call [kOl] and
ball [bOl] and perceive a difference in the initial sounds, sound develops
in time in a way that makes the actual border between the initial
plosive and the vowel not completely clear. Second, it’s actually not
the case that the words e.g. call [kOl] and ball [bOl] only differ in their
initial sound. If you recall from the previous chapter on phonetics, the
pronunciation of a given sounds is influenced by the context in which
it is produced (coarticulation). So the low back unrounded [O] vowel
pronounced after a velar plosive will be slightly more back than when
it is pronounced after a bilabial stop, where it will be slightly fronted.

In fact, there is quite a bit of variation from speaker to speaker.

Even within the same speaker, the same target phoneme is pronounced
slightly differently each time. Each time, the tongue may be a fraction
more to the front or back, there may be more or less air, there may
be more or less pressure, etc. Below is a graph showing data from 19
speakers from Ohio (from the Buckeye corpus, figure from Yao (2009)).
For each speaker, the graph shows their mean voice onset time for word
initial voiceless stops. This is the point at which vocal chords begin
to vibrate after the release of the voiceless stop. As you can see, no
two speakers are alike in their mean. Further, there is a large amount
of variation even within each speaker (see the bars). Yet the variation
is not so large that the intended sound is not understandable. This
variation in VOT is not enough to be perceived as a different sound
by speakers of English.
 phonology i: features and phonemes 29

Even though the production of any speech sound made by a speaker

is an approximate aim at a certain target, these slight differences are
not perceived as different by hearers. So for native English speakers,
[O] pronounced more front or more back is within the range of normal
variation. The important thing is that speakers of English consider
these two speech streams, ball [bOl] and call [kOl] to be two different
words that only differ in the initial consonant: they perceive them as
minimal pairs.
The general rule for identifying a minimal pair (which in turn iden-
tifies phonemic contrasts in a given language) is that you have to find
two words that are (1) the same length in the number of sounds, (2)
differ only in one sound and (3) have two distinct meanings. This
means that bat and pat can identify a phonemic contrast in English
between [b] and [p] but bat and patty cannot. Neither would bat and
patch be possible minimal pairs. The [[A:nt]=[ænt] divide
Areas in red use the [ænt] pronuncia-
Identifying phonemic inventories can sometimes be tricky. Consider tion, while areas in blue (in New Eng-
for example the case of aunt. In some dialects of English, namely land) use [A:nt].
British RP and New England English, the sister of a parent, an aunt
is pronounced as [A:nt], while in General American it is pronounced
[ænt]. These are two different pronunciations of the same word, with
the same meaning. This means that [A:nt]-[ænt] is not a minimal
pair that can be used to show that [A]-[æ] is a phonemic contrast in
English. However, in English RP, the words ant-aunt are pronounced
differently, in fact as [A:nt]-[ænt]. So this minimal pair can identify
the [A]-[æ] contrast as phonemic in RP. In General American (GA)
both words are pronounced the same, so they are not minimal pairs.
However, the [A]-[æ] is phonemic in GA: the words bot-bat [bAt]-[bæt]
are minimal pairs.
So even though the [O] in call and ball is slightly different, due to the
initial consonant, we can use the same IPA vowel symbol. In general,
if you perceive two words as only differing on one sound, then you
can be fairly certain that the different sounds are showing different
phonemes. This difference is then a phonemic contrast.
But how is that the speakers of a language perceive these contrasts
even though the contrasts are actually from a continuum? Phonemic
contrasts are possible because our brain translates continuous speech
30 an open introduction to linguistics

sounds into categories. The range of sounds that make up what we

identify as [O] is actually infinite, but there are boundaries at which
the sound waves will start to be perceived as something else. But
within the boundaries (which are also rough, and differ by language
and to a lesser degree by the speaker) our brain classifies all the sounds
as the same category, the sound [O]. This is categorical perception.
By classifying an actually infinite set of sounds within a given range
as the same category, that sound becomes a sort of symbol that can
have meaning. And in English, [O] does have meaning. We know this
because the minimal pair ball-bill ([bOl] and [bIl]) is evidence that [O]
is a phoneme (and so is [I]).
The boundaries between one category and another are somewhat
arbitrary, and can differ by language. Consider the following data from
a study looking at the perceptions of Spanish and English speakers
(Zhao, 2018). The study looked at what syllable speakers perceived
in a synthetically created syllable where the VOT was systematically
varied. Recall that voiceless plosives have a longer lag till voice onset
(a longer VOT) than voiced plosives (which often have no VOT). The
Spanish speakers’ data is in red (the higher line), and the English
speakers are shown in the lower line. At negative VOTs, both Spanish
and English speakers identified the syllable as [ba]. Note that Spanish
speakers begin to identify the stimuli as [pa] at about 10ms VOT,
but most English speakers will still classify these stimuli as [ba]. Only
when presented with stimuli that have a VOT of 15 msec do more than
half of the English speakers start to identify the sound as [ba]. There
are several online tests that illustrate categorical perception that you
can find at https://splab.net/apd/u100/.

.
Categorical perception allows us to perceive phonemes as distinct
categories, despite the continuous nature of sound. English has about
40 distinct phonemes, depending on the dialect. Each phoneme con-
trasts with many of the other 39 phonemes, but not necessarily all.
Consider the following data from Burushaski, a language spoken in
Srinagar, India. In this language there are phonemic voicing contrasts
in plosives word initially, but there are not clear minimal pairs for all
cases. For example, the word [tam] means swimming, and the word
[dan] means stone. These do not qualify as minimal pairs, but because
 phonology i: features and phonemes 31

there are other similar examples of dental plosives in similar environ-

ments (after a vowel followed by a nasal consonant), it can be argued
that the distinction between voiced and voiceless dental plosives is
phonemic in Burashaski (Example from Munshi, 2018). There are no
true minimal pairs in these cases, but we can say that the two sounds
are contrastive in analogous environments (C.A.E). Note that
it is important that the environments are similar in terms of sharing
features. So in the case of Burashaski, the analogous environments
share the same vowel and nasal final consonants. Similar sounds (as
determined by sound features), are necessary in order to make an ar-
gument for a phonemic contrast that is based on two sounds being
contrastive in analogous environments.
There are cases where two related sounds are perceived as being
part of the same category. These sounds differ quite a bit, much more
than just a question of being more front, or more back, etc., but where
the contrast is not phonemic, and even more interestingly, not per-
ceived by native speakers without it being pointed out to them. For
example, in English, the bilabial voiceless plosive [p] actually occurs
in two predictable forms: without aspiration ([p]) or with it ([ph ]).
In fact, when it occurs word initially it is aspirated, but in all other
positions it is unaspirated. See the table below:

Word Broad Narrow Word Broad Narrow Table 3.1: Different allophones of /p/
in English
pat /pat/ [ph a:t] peace /pis/ [ph i:s]
prince /prIns/ [prIns] play /pleI/ [pleI]
spin /spIn/ [spIn] pin /pIn/ [ph In]
happy /hæpi/ [hæpi] tap /tæp/ [tæp]
stamp /stæmp/ [stæmp] pump /p@mp/ [ph @mp]

The above table shows the words in standard English spelling, fol-
lowed by two transcriptions in IPA, broad transcription and narrow
transcription. Narrow transcription gives all the details of the pro-
nunciation. Broad transcription is a phonemic transcription: in other
words, the broad transcription just shows the phonemes in English.
An English speaker reading the broad transcription will pronounce
the words like the narrow transcription, but they will (without train-
ing) not be aware of the features of their pronunciation that are shown
in the narrow transcription. Most English speakers are not aware that
they aspirate word initial [p]’s. That’s because the contrast between [p]
and [ph ] is not phonemic. There are no minimal pairs in English that
contrast them. However, they are clearly related in articulation. Both
are bilabial voiceless plosives, and, as the example shows, the context
predicts where each sound occurs. None of the contexts overlap, so this
means that the two sounds are in complementary distribution.
Sounds that are in complementary distribution, are allophones.
Allophones belong to a given phoneme. The sounds [p] and [ph ] are
both allophones of the phoneme /p/. Note that we talk about both
sounds being allophones of /p/ (a phoneme, so phonemic transcrip-
tion) and not allophones of /ph /. That’s because if we look at the
32 an open introduction to linguistics

different contexts in which both sounds occur, [p] occurs in a wider

set of contexts than its aspirated cousin. In fact we can write a simple
rule that will predict perfectly where each sound should appear: Word
initially before a vowel, use [ph ]. Elsewhere, use [p]. The phone that
has the wider range of uses, (and which is also acoustically simpler
as well, because it doesn’t have the added aspiration) is sometimes
termed the elsewhere allophone and they are often designated as
the symbol of the phoneme (which is actually an abstract construct.)
Now again, we have simplified things. The [p] in stamp ([stæmp]) is
going to be slightly different than the [p] in [hæpi], because a vowel to
a plosive is going to differ from a nasal consonant going into a plosive.
These are minor differences in pronunciation that are predictable, and
common across all languages, and in the degree to which they occur
in e.g. English, is also not a phonemic distinction in other languages.
It simply has to do with the pronunciation of sounds in context, and
the variation can be subsumed under unavoidable features of coartic-
ulation.
Most speakers are completely unaware of the allophonic variations
in their language. This is because for the native speakers, this variation
is not important: they use the right forms and hear them as the same
category. Thus the words dye and dry seem to have the same initial
voiced alveolar stop in English. However, English has retraction (=A
sound is pronounced more to the back) before some alveolar sounds,
e.g. [d],[t],[n] [ì]. Thus in actuality dry is pronounced as [draI]. But
¯
English speakers are unaware of this variation.
This is a natural point to discuss transcription practices in more
detail. We’ve already introduced the ideas of broad and narrow tran-
scriptions. This has to do with the level of detail that is transcribed. If
only the most salient features of speech sounds are transcribed, with-
out smaller variations, we usually write the sounds within slashes (/·/).
Thus dye and dry would be written as /daI/ and /draI/ in broad
transcription. But if we want to note the retraction feature, we
need to include that in our transcription, in which case we are doing a
narrow transcription, and the pronunciation is written within two
brackets ([·]) e.g. [daI] vs. [draI]. The categories of broad and narrow
¯
transcription almost line up with the contrast between phonemic and
phonetic transcription. In a phonemic transcription, only variations
that are phonemic contrasts are transcribed. A native speaker reading
a phonemic transcription will of course pronounce all the typical vari-
ations, but these are not in the transcription. Phonemic transcriptions
are a kind of broad transcription, and as such are also written within
slanted lines. A phonetic transcription is recording the features of the
sounds, not their status in a particular languages. This is a kind of
narrow transcription and is written in square brackets. Thus when
we talk of sounds, divorced of their status in a particular language, we
can use phonetic (narrow) transcription. Allophonic variation will only
be identifiable in narrow transcription. Note also that for non-native
speakers, a phonetic or narrow transcription would be most helpful for
learning to pronounce sounds like a native.
 phonology i: features and phonemes 33

Back to aspiration. One of the reasons we know that the difference

between aspirated and unaspirated [p] is a potentially significant dis-
tinction (and not an arbitrary one), is from cross-linguistic data. There
are languages that make a phonemic distinction between [p] and [ph ].
Thai, for example, has a phonemic contrast in initial aspiration. In
Thai the words [paa] means forest and [ph aa] means to split. These
are thus minimal pairs. These words are evidence that [p] and [ph ] are
different phonemes in Thai.
Let’s return to English. In English we saw that the voiceless bilabial
stop [p] is aspirated when it occurs word initially, but unaspirated in
all other word positions. The phoneme /p/ thus has two allophones,
[p] and [ph ]. Actually, this isn’t just the case for the voiceless bilabial
stops. In fact, all voiceless stops in English follow the same pattern:
The phonemes /t/ and /k/ also both have an aspirated and an unaspi-
rated allophone, and these occur in the same environments as [p] and
[ph ]. This is actually the norm for allophonic variation: contextual
variation found in one sound is often not limited to that sound alone,
but usually is found in a set of sounds with similar features.

Narrow English Narrow English Table 3.2: Phonemic contrasts in as-

piration in English
[ph eIp@r] paper [spIn] spin
[th Ent] tent [steItl] state
[kh eIk] cake [klIk] click
As the table illustrates, in English all voiceless stops are aspirated
word initially before a vowel, but are unaspirated everywhere else. So
the allophonic variation we identified for /p/ above can actually be
characterized by a more general rule than the one we created above:
Word initially before a vowel, aspirate Voiceless stops. Elsewhere, they
are unaspirated. Instead of describing a specific phone, we describe a
class of sounds, the voiceless stops. And in most languages, phone-
mic and allophonic variations are to a certain degree systematic: a
distinction or conditioned variation occurs across a class of sounds.
We can see this in Thai. In Thai, aspiration is not just a phonemic
contrast for voiceless bilabial plosives. This variation is also present
for dental and velar voiceless plosives. This means that /p/,/ph /, /t/,
/th /,/k/ and /kh / are all phonemes in Thai. See the table below:

Unaspirated English Aspirated English Table 3.3: Phonemic contrasts in as-

piration in Thai
[paa] forest [ph aa] to split
[tam] to pound [th am] to do
[kat] to bite [kh at] to interrupt
Now let’s consider an example from Maasai. Consider the following
data in the table below, paying particular attention to the phones [p],
[b] and [B]. Are these distinct phonemes or is there allophonic variation
here?
Maasai or Masai is an Eastern
The way to approach this question is to examine the context of use Nilotic language spoken in South-
ern Kenya and Northern Tanzania by
of each of the three sounds. the Maasai people, numbering about
(1) Where in the words do the sounds occur? Initially, word internally, 800,000, http://en.wikipedia.org/
wiki/Maasai_language
34 an open introduction to linguistics

Table 3.4: Maasai Allophones

[p] [b] [B]
[padan] skilled in shooting [imbala] paper [ilpABit] hairs
[pus] light colored [embaruo] smoke [keBer] heavan
[Ailap] to hate [imbok] you clean ceremoniously [enaiBoSa] Naivasha lake

or word finally?
(2) Do the sounds occur before or after vowels? Which vowels? Do
they have features in common? (front, back, low, high etc.).
(3) Do the sounds differ in distribution of nasals or other consonants
next to them? If you make a list of the contexts in which each of
the three sounds occurs, you should usually be able to identify a clear
pattern.

In Maasai, [p] occurs word initially, word finally, and always before
or after a vowel. In contrast, [b] occurs only after a nasal. Finally,
in the example given, [B] only occurs between two vowels. All three
sounds are thus in complementary distribution. This means also that
they are all allophones of the same phoneme. Speakers of Maasai
perceive all three sounds to be the same sound. In the case of Maasai,
it’s a bit harder to determine how to represent the allophonic set: what
symbol should we use for the phoneme? There isn’t really evidence to
prefer one over the other in terms of the number of contexts they
occur in, so we can perhaps choose [p] to represent the set because
word initial sounds seem more primary than word internal sounds, but This joke rests on the fact that in
this is fairly arbitrary. Then we can say that /p/ is a phoneme in General American, the central vowel
Maasai that is made up of three allophones: [p],[b] and [B]. [2] is often reduced to a schwa, e.g.
[@]. However, not always. So in fast
Now we’ve covered phonemic contrasts and allophonic variation. speech, photograph will be pronounced
There is only one more important type of variation possible when as [foUt@græf], but in single syllable
words it cannot be reduced, so we do
comparing two sounds within a language: free variation. It’s possible say cut [c2t] and luck [l2k].
that two sounds can be used interchangeably. Speakers are aware of the
differences, but the distinction is not meaningful (there are no minimal
pairs. This often is found between two dialects of the same language,
but also within an individual speaker. The aunt pronunciation in
English is an example of this. While each of the two pronunciations is
dominant in different dialects in English, some speakers may freely use
both. Note that technically, sounds in free variation are also allophones
of the same phoneme. They are just not in complementary distribution
with each other.
There’s a decision tree you can use to identify whether two sounds
you are analyzing in a language are different phonemes, allophones of
the same phoneme, or if they are in free variation. See the chart below:
 phonology i: features and phonemes 35

The importance of features

Weirdly, by definition, any set of sounds can constitute a phoneme.
However, it turns out that phonemes are composed of classes of sounds
that have certain properties in common. These are defined by fea-
tures. Features are phonetic, and supposedly not influenced by cross-
language variation. What exactly is a feature? The actual features
found in the literature take a (more or less) articulatory standpoint.
Take any sound realizing English /b/. It is produced by closing the
lips, thereby obstructing the air flow (‘bilabial’) and then releasing it,
and at the same time letting the vocal cords vibrate (‘voiced’). If the
vocal cords do not vibrate we get the sound corresponding to /p/.
We can analyse the sound as a motor program that is executed
on demand. Its execution is not totally fixed, so variation is possible
Second, the motor program directs various parts of the vocal tracts,
some of which are independent from each other. We may see this as
a music score which has various parts for different ‘instruments’. The
score for the voicing feature is one of them. The value ‘+’ tells us that
the cords have to vibrate during the production of the corresponding
sound, while ‘−’ tells us that they should not. We have to be a bit
cautious, though. It is known, for example, that /b/ is not pronounced
with immediate voicing. Rather, the voicing is delayed by a certain
onset time. This onset time varies from language to language. Hence,
the actual realization of a feature is different across languages, a fact
that is rather awkward for the idea that phonemes are defined by
recourse to phonetic features. The latter should namely be language
independent. The problem just mentioned can of course be resolved
by using more finely distinguished features. But the question remains:
just how much detail do the phonetic features need to give? The answer
is roughly that while phonetically we are dealing with a continuous
scale (onset time measured in milliseconds), at the phonemic level we
are just looking at a binary contrast.
36 an open introduction to linguistics

We shall use the following notation. There is a set of so-called

attributes and a set of so-called values. A pair [ att : val] consisting
of an attribute and a value is called a feature. We treat +voiced as
a notational alternative of [ voiced : +]. An attribute is associated
with a value range. For phonology, we may assume the following set
of attributes:

place, manner, voiced, consonantal, aspirated, aperture, . . .

(3.1)

and we may assume the following set of values:

bilabial, labiodental, plosive, approximant, high, mid, +, −, . . . (3.2)

The range of values for place is obviously different from that of man-
ner, since ‘dental’ is a value of the former and not of the latter. A set
of features is called an attribute value structure (AVS). You have
already seen AVSs in the first chapter. The notation is as follows. The
attributes and values are arranged vertically, the rows just having the
attribute paired with its value, separated by a colon:
 
place : dental
 manner : fricative  (3.3)
 
voice : +
Notice that the following are also legitimate AVSs:
 
place : dental  
 place : dental  place : dental
 place : uvular  (3.4)
   
 
 manner : fricative 
voice : +
voice : +
Note that the fact that you can still
The first is identical to (3.3) in the sense that it specifies the same distinguish between [p] and [b] while
whispering shows you that the distinc-
object (the features are read conjunctively). The second however does tion is not exclusively a voicing con-
not specify any sound, since the values given to the same feature are in- trast! There are additional differences,
one of which is that the lip tension is
compatible. (Features must have one and only one value.) We say that higher in [p].
the second AVS is inconsistent. Notice that AVSs are not sounds,
they are just representations thereof, and they may specify the sounds
only partly. I add here that some combinations may be formally con-
sistent and yet cannot be instantiated. Here is an example:
" #
consonantal : −
(3.5)
voice :−
This is because vowels in English are voiced. There are a few languages,
for example Mokilese, which have voiceless vowels. To understand
how this is possible think about whispering. Whispering is speaking
without the vocal chords vibrating. In effect, whispering is the same
as systematically devoicing every sound.
Not everything goes for AVSs, however. The following AVS, for
example, is illegitimate because it assigns a value to place that is
outside of its value range.
" #
place : fricative
(3.6)
voice : +
 phonology i: features and phonemes 37

There are a number of arguments for why features exist. First and
foremost the features encode a certain linguistic reality; the features
that we have spoken about so far have phonetic content. They speak
about articulatory properties. It so happens that many rules can be
motivated from the fact that the vocal tract has certain properties.
For example, in German all syllable-final plosives are voiceless (see the
discussion on Page 53). This is the case even when there is reason
to believe that the consonant in question has been obtained from a
voiced consonant. Thus, one proposes a rule of devoicing for German.
However, it would be unexpected if this rule would turn [g] into [t].
We would rather expect the rule to turn [g] into [k], [b] into [p] and
[d] into [t]. The questions that arise are as follows:
À Why is it that we expect matters to be this way?

Á How can we account for the change?

The first question is answered as follows: the underlying rule is not
a rule that operates with a lookup table, showing us what consonant
is changed into what other consonant. Rather, it is encoded as a rule
that says: simply remove the voicing. For this to make sense we need
to be able to independently control voicing. This is clearly the case.
However, it is one thing to observe that this is technically possible and
another to show that this is actually the rule that speakers use. One
way to check that this is actually the rule is to make Germans speak
a different language. The new language will have new sounds, but we
shall observe Germans still devoice them at the end of the syllable.
Moreover, they will not randomly choose a devoiced consonant but You can hear Germans devoicing con-
will simply pick the appropriate voiceless counterpart. sonants in English, for example. The
prediction is this that—if they can at
Ideally, we wish to write the rule of devoicing in the following way. all produce these sounds—at the end
" # " # of a word [ð] will come out as [T].
consonantal : + consonantal : +
→ / # (3.7)
voice :+ voice :−
It will turn out that this can indeed be done (the next chapter provides
the details of this). This says that a consonant becomes devoiced at the
end of a word. The part before the arrow specifies the situation before
the rule applies; the part to the right and before the slash show us how
it looks after the application of the rule. The part after the slash shows
in what context the rule may be applied. The underscore shows where
the left part of the rule must be situated (and where the right part will
be substituted in its place). Here, it says: it must occur right before
#, which signals the end of a word. The way this rule operates needs
to be explained. The German word /grob/ is pronounced [gGo:p] (the
colon indicates a long vowel; G is a voiced velar fricative, the fricative
equivalent of g). The letter /b/ however indicates an underlying [b].
Thus we expect this to be an instance of devoicing. So let’s look at
[b]:
 
consonantal : +
 voice : + 
(3.8)
 
 
 place : bilabial 
manner : stop
38 an open introduction to linguistics

As the sound occurs immediately before #, the rule applies. When it

applies, it matches the left hand side against the AVSs and replaces
that part with the right hand side of the rule; whatever is not matched
remains the same.
   
consonantal : + consonantal : +
 voice : +   voice : − 
→
   
 
 place : bilabial   place : bilabial 
manner : stop manner : stop
(3.9)

Thus, the resulting sound is indeed [p]. You may experiment with
other AVS to see that the rule really operates as expected. Notice
that the rule contains [consonantal : +] on its left but does not
change it. However, you cannot simply eliminate it. The resulting rule
would be different:
h i h i
voice : + → voice : − / # (3.10)

This rule would apply to vowels and produce voiceless vowels. Since
German does not have such vowels, the rule would clash with the con-
straints of German phonology. More importantly, it would devoice
every word final vowel and thus—wrongly—predict that German has
no word final voiced vowels (counterexample: /Oma/ [oma] ‘grand-
mother’).
Suppose that you have to say this without features. It is not enough
to say that the voiced consonants are transformed into the voiceless
ones; we need to know which voiceless consonant will replace which
voiced consonant. The tie between [p] and [b], between [k] and [g] and
so on needs to be established. Because features have an independent
motivation the correspondence is specified uniformly for all sounds. As
will be noted throughout this course, some rules are not really specific
to one language but a whole group of them (final devoicing is a case
in point). This seems to be contradictory, because the rules are stated
using phonemes, and phonemes are language dependent, as we have
seen. However, the fact that language has a contrast between voiced
and voiceless is independent of the exact specification of what counts,
say, as a voiced bilabial stop as opposed to a voiceless bilabial stop.
Important is that the contrast exists and is one of voicing.
For example, Hungarian, Turkish and Finnish all have a rule called
vowel harmony. Simplifying somewhat, all rules agree that words can-
not both contain a back vowel and a front vowel. On the other hand,
the front close-mid rounded vowel of Finnish (written /ö/) is pro-
nounced with more lip rounding than the Hungarian one, also writ-
ten /ö/. Nevertheless, both languages systematically oppose /ö/ with
/o/, which differs in the position of the tongue body (close-mid back
rounded vowel). The situation is complicated through the fact that
Hungarian long and short vowels do not only contrast in length but
also in a feature that is called tension. Finnish /ö/ is tense even
when short, while in Hungarian it is lax (which means less rounded
and less close). However, even if short and long vowels behave in this
 phonology i: features and phonemes 39

way, and even if back and front vowels are different across these lan-
guages, there is good reason to believe that the contrast is between
‘front’ and ‘back’, no matter what else is involved. Thus, among the
many parameters that define the actual sounds, languages decide to
systematically encode only a limited set even though one still needs
to fill in details as for the exact nature of the sounds. Precisely this
is the task of realization rules. These are the rules that make the
transition from phonemes to sounds. They will be discussed in the
next chapter.

Natural Classes
Suppose we fix the set of attributes and values for a language. On the
basis of this classification we can define the following:

Definition 1 (Natural Class; Provisional) A natural class of sounds

is a set of sounds that can be specified by a single AVS.

This is still not as clear as I would like this to be. First, we need
to say something about the classification system used above. Let P
be our set of phonemes. Recall that this set is in a way abstract.
It is not possible to compare phonemes across languages, except by
looking at their possible realisations (which are then sounds). We
then define (using our (pre-)theoretical insights) some features and
potential values for them. Next we specify which sounds have which
value for which attribute. That is to say, for each attribute A and
value v there is a set of phonemes written [A : v ] (which is therefore a
subset of P ). Its members are the phonemes that are said to have the
value v for the attribute A. In addition, we require that the following
holds:

À For each phoneme p, and each attribute A there is a value v such

that p ∈ [A : v ], that is, p has A-value v.

Á If v 6= v 0 then [A : v ] ∩ [A : v 0 ] = ∅. In other words: the value of

an attribute for a given sound is unique.

 For every two different phonemes p, p0 there is a feature A and

values v, v 0 such that v 6= v 0 and p ∈ [A : v ] and p0 ∈ [A : v 0 ].

If these postulates are met we speak of a classification system for

P . The last condition is especially important. It says that the classi-
fication system must be exhaustive. If two phonemes are different we
ought to find something that sets it apart from the other phonemes.
This means, among other things, that for each p the singleton {p} will
be a natural class.
First, notice that we require each sound to have a value for each
attribute. This is a convenient requirement because it eliminates some
fuzziness in the presentation. You will notice, for example, that vow-
els are classed along totally different lines as consonants. So, is it
appropriate to say, for example, that vowels should have some value to
manner? Suppose we do not really want that. Then a way around
40 an open introduction to linguistics

this is to add a specific value to the attribute, call it ?, and then de-
clare that all vowels have this value. This ‘value’ is not a value in the 
intended sense. But to openly declare that vowels have the ‘non-value’
helps us be clear about our assumptions. To show the first, an intersection of no
subsets of P is defined to be identical
Definition 2 (Natural Class) Let S be a classification system for to P , so that is why P is natural. To
show the second, let H be the intersec-
P . A subset U of P is natural in S if and only if it is an intersection
tion of all sets [A : v ] that contain p.
of sets of the form [A : v ] for some attribute and some legitimate value. I claim that H = {p}. For let p0 6= p.
Then there are A, v, and v 0 such that
I shall draw a few conclusions from this. v 6= v 0 p ∈ [A : v ] and p0 ∈ [A : v 0 ].
However, p0 6∈ [A : v 0 ], since the sets
1. The set P is natural. are disjoint. So, p0 6∈ H. Finally,
for the third, let there be at least two
2. For every p ∈ P , {p} is natural. phonemes, p and p0 . Then there are
A, v and v 0 such that p ∈ [A : v ],
3. If P has at least two members, ∅ is natural. p0 ∈ [A : v 0 ] and v 6= v 0 . Then
[A : v ] ∩ [A : v 0 ] = ∅ is natural.

The Classification System of English Consonants

Table 3.5 gives a list of the phonemes of American English. The slanted
brackets denote phonemes, not sounds, but the sounds are neverthe-
less given in IPA. On the whole, the classification of phonemes looks

Consonants Table 3.5: Phonemes of English

bila- labio- den- alve- palato- pala- velar glot-
bial dental tal olar alveolar tal tal
stops vl /p/ /t/ /tS/ /k/
vd /b/ /d/ /dZ/ /g/
frica- vl /f/ /T/ /s/ /S/ /h/
tives vd /v/ /ð/ /z/ /Z/
nasals /m/ /n/ /N/
appro- lat /l/
ximants cnt /w/ /ô/ /j/
vl = voiceless, vd = voiced, lat = lateral, cnt = central

Vowels and Diphthongs

front central back diphthongs
unrounded unrounded unrounded rounded
upper high /i/ /u/ /aı/, /aU/,
lower high /ı/ /U/ /oı/
upper mid /e/ /@/ /o/ syllabic
lower mid /E/ /2/ consonant
low /æ/ /A/ /Ä/

very similar to that of sounds. But there are mismatches. There is a

series of sounds, called affricates, which are written as a combination
of a stop followed by a fricative: English has two such phonemes, [tS]
and [dZ]. Similarly, diphthongs, which are written like sequences of
vowels or of vowel and glide, are considered just one phoneme. Notice
also that the broad transcription is also hiding some diphthongs, like
[e] as in /able/. This is a sequence of the vowel [e] and the glide [j].
 phonology i: features and phonemes 41

The reason is that the vowel [e] is obligatorily followed by [j] and there-
fore mentioning of [j] is needless. However, unless you know English
well you need to be told this fact. The sequence [aı] is different in
that [a] is not necessarily followed by [ı], hence writing the sequence is
unavoidable. I shall indicate now how Table 3.5 establishes a classifi-
cation system and how it is written down in attribute value notation.
To make matters simpler, we concentrate on the consonants. There
are then three attributes: place, manner, and voice. We assume
that the features have the following values:

place :bilabial, labiodental, dental, alveolar, palatoalveolar, palatal,

velar, glottal
manner :stop, fricative, nasal, lateral approximant, central approximant
voice :+, −
(3.11)

The sounds with a given place features are listed in the columns, and
can be read off the table. However, I shall give them here for conve-
nience:

[ place : bilabial] = {/p/, /b/, /m/, w/}

[ place : labiodental] = {/f/, /v/}
[ place : dental] = {/T/, /ð/}
[ place : alveolar] = {/t/, /d/, /s/, /z/, /n/, /l/, /ô/}
(3.12)
[ place : palatoalveolar] = {/tS/, /dZ/, /S/, /Z/}
[ place : palatal] = {/j/}
[ place : velar] = {/k/, /g/, /N /}
[ place : glottal] = {/h/}

The manner feature is encoded in the row labels.

[ manner : stop] = {/p/, /b/, /t/, /d/, /tS/, /dZ/, /k/, /g/}
[ manner : fricative] = {/f/, /v/, /T/, /ð/, /s/, /z/, /S/, /Z/}
[ manner : nasal] = {/m/, /n/, N /}
[ manner : l approx] = {/l/}
[ manner : c approx] = {/w/, /j/, /ô/}
(3.13)

[ voice : +] ={/b/, /d/, /g/, /dZ/, /v/, /ð/, /z/, /Z/, /m/, /n/,
/N /, /w/, /l/, /ô/, /j/}
[ voice : −] ={/p/, /t/, /k/, /tS/, /f/, /T/, /s/, /S/}
(3.14)

So, one may check, for example, that each sound is uniquely character-
ized by the values of the attributes; /p/ has value bilabial for place,
42 an open introduction to linguistics

stop for manner and − for voice. So we have

 
place :bilabial
 manner: stop  = {/p/} (3.15)
 
voice : −

If we drop any of the three specifications we generally get a larger class,

but this is not always so. For example, English has only one palatal
phoneme, /j/. Hence we have
" #
h i place : palatal
place:palatal = = {/j/} (3.16)
manner:c approximant

Note that although a similar system for vowels can be given, I do

not include it here. The reason for this is that it turns out that the
classification of vowels proceeds along different features. We have, for
example, the feature rounded, but do not classify the consonants
according to feature. If we are strict about the execution of the classi-
fication we should then also say which of the consonants are rounded
and which ones are not.
Notice also that, although the system of classification is motivated
by phonetics, it is not completely consistent in that regard. For exam-
ple, the phoneme /ô/ is classified as voiced. However, at closer look it
turns out that the phoneme contains both the voiced and the voice-
less variant, written [ô]. In broad transcription (which is essentially The pronunciation of /bridge/ in-
˚ volves the voiced [ô], the pronunciation
phonemic) one writes [ô] regardless. The policy on notation is not al-
of /trust/ the voiceless [ô].
ways consistently adhered to; the symbolism encourages confusing [ô] ˚
and [ô], though if one reads the IPA manual it states that [ô] signifies
˚
only the voiced and not the voiceless approximant. So, technically, the
left part of that cell should contain the symbol [ô].
˚

Binarism
The preceding section must have caused you to think that for a given
set of phonemes there must be several possible classification systems.
Indeed, not only are there several conceivable classification systems,
phonologists are divided on the issue of which one to actually use.
There is a never-ending debate on the idea of binarism of features.
Binarism is the idea that features have just two values: + and −.
In this case, an alternative notation is used; instead of [att : +] one
writes [+att] (for example, [+voiced]) and instead of [att : -] one
writes [−att] (for example [−voiced]).
Although any feature system can be reconstructed using binary val-
ued features, the two systems are not equivalent, since they define
different natural classes.
Consider, by way of example, the sounds [p], [t] and [k]. They
are distinct only in the place of articulation (bilabial vs. alveolar vs.
velar). The only natural classes are: the empty one, the singletons
or the one containing all three. If we assume a division into binary
features, either [p] and [t] or [p] and [k] or [t] and [k] must form a
natural class in addition. This is so since binary features can only cut
 phonology i: features and phonemes 43

a set into two parts. If your set has three members, you can single
out a given member by two cuts and only sometimes by one. So you
need two binary features to distinguish the three from each other.
But which ones do we take? In the present case we have a choice of
[+labial], [+dental] or [+velar]. The first cuts {[p], [t], [k]} into {[p]}
and {[t], [k]}; the second cuts it into {[t]} and {[p], [k]} and the third
into {[k]}, and {[t], [p]}. Any combination of two of these features
allows for the singleton sets as natural classes. If you have only two
features then there is a two element subset that is not a natural class
The choice between the various feature bases is not easy and hotly
disputed. It depends on the way the rules of the language can be sim-
plified which classification is used. But if that is so, the idea becomes
problematic as a foundational tool. It is perhaps better not to enforce
binarism.

Study Questions
1. It is very useful to get to grips with making minimal pairs. Take
some of the sounds of English (see the table of English phonemes in
the chapter) and make some minimal pairs (that are not already in
the book) to show that these sounds belong to distinct phonemes?

2. If two speakers of the same language consistently realize a phoneme

differently (e.g. ’hard g’ ([x] or [G]) and ’soft g’ ([ç] or [J]) in Dutch),
for example due to dialectal differences, is this an example of free
variation or of conditioning by context?

3. Phonemes are sets of phones. Each of these phones is understood

as an expression of that phoneme, i.e. phonemes can be realized
in different ways. Is there a similar variation in the realization of
phones? Or are these always realized identically?

4. Why, in phonetic rules, do you sometimes find two identical attri-

bute-value pairs on both sides of the rule? (e.g. rule 3.7 on page 37)

5. There is an overview of natural classes of consonants on page 41.

You can also get this information directly from the IPA chart, how?

6. ‘Translate’ the following broad phonetic transcriptions to a conven-

tional English spelling:

a. [f@nEIks]
b. [SUô]
c. [ôaItIN]
d. [Igzæm]
e. [noU]
f. [@bæUt]

7. Give broad phonetic transcriptions of the following English words:

a. skunk
44 an open introduction to linguistics

b. heath
c. nudging
d. uncommon
e. English

8. In the following sets of four sounds, three sounds belong to the same
natural class, while the remaining phoneme, the ’odd one out’, does
not. Give the phoneme that is the odd one out and the attribute-
value pairs which describe the difference. There might be more than
one solution.

a. [d T x Z]
b. [ð Z S s]
c. [Y W u ø]
d. [I 3 2 æ]

9. Korean allophones The Korean letter ㄹ is pronounced as either

[l] or [R] depending on the environment. In other words in Korean
these are allophones of the same phoneme. Consider the following
Korean words: (in the translation provided the subject for the con-
jugated verbs is always ’he’, but this can be anything ranging from
I, you, he, we, they, up to the cat. In Korean the subject can be,
and is most of the time, omitted and the intended subject is derived
from the context.)

Korean form English gloss Korean form English gloss

[alda] to know [aRajo] he knows
[ulda] to cry [uRojo] he cries
[ulmj2n] if he cries [kWlsæ]1 well,...
¨
[mal] words [maRHada] to speak (do words)
[il] work, one, sun [iRHada] to work
[nj2n] year [illj2n] one year
[joil] weekday [iRjoil] sunday
1. [s] is a ’tense’ s - similar to s but with a stiff tongue and more air
¨

a. List the environments in which [l] and [R] occur.

b. What appears to be the the basic form and which is the derived
form? explain why.
c. Write a set of rules to explain the variation between [l] and [R].
For the sake of simplicity, let’s assume both [j] and [H] are ap-
proximants (and there are no other approximants in Korean,
except for the lateral approximant [l]).

10. Korean allophones - part 2 Have another look at the Korean

words in the table above.

a. What other variation to you spot? (Hint: it also involves [l])

b. Apply your rules to this word and see what happens.
 phonology i: features and phonemes 45

c. Write a new rule to explain this variation.

d. Which rule needs to be applied first?
e. What happens when you apply the rules in a different order?

11. A problem in any discussion of phone(me)s is that it always in-

volves a somewhat arbitrary mapping from a continuous speech
sound to a discrete symbol system (e.g. narrow or broad IPA tran-
scription). This could lead to a case where the same sound, heard The International Phonetic Al-
by two phoneticians, is annotated as a different phone(me). Can phabet is not the only phonetic
alphabet in existence; you can
you think of a way to make this process less ambiguous? find some of its precursors at
https://en.wikipedia.org/wiki/
12. In English, the orthography (the way a word is written) and phone- Category:Phonetic_alphabets.
mic transcription of a word are often far apart (e.g. the ‘gh’ in
‘enough’). Can you think of what might have caused this? And
what would the advantages and disadvantages be of changing the
orthography to make it more consistent and/or similar to the phone-
mic transcriptions?
4
Phonology II: Realization Rules and Representations

The central concepts in this chapter are those of a natural class

and of a rule. We learn how rules work, and how they can be
used to structure linguistic theory.

Determining Natural Classes

Let us start with a simple example to show what is meant by a natural
class. Sanskrit had the following obstruents and nasals

p ph b bh m
t th d dh n
ú úh ã ãh ï (4.1)
c ch , é éh ñ
k kh g gh N

To describe these sounds we use the following features and values: By the way, if you read the sounds as
they appear here, this is exactly the
cons(onantal) :+ way they are ordered in Sanskrit. The
Sanskrit alphabet is much more logi-
manner :stop, fric(ative) cally arranged than the Latin alpha-
bet!
place :bilab(ial), dent(al), retro(flex), velar, palat(al)
(4.2)
asp(irated) :+, −
nas(al) :+, −
voice :+, −

We shall omit the specification ‘consonantal’ for brevity. Also, we shall

omit ‘manner’ and equate it with ‘nasal’ (= [ nas : +]). Here is how
the phonemes from the first row are to be represented:

p ph b
     
place:bilab place:bilab place:bilab

 asp :− 


 asp :+ 


 asp :− 

     
 nas :−   nas :−   nas :− 
voice :− voice :− voice :+
(4.3)
bh m
   
place:bilab place:bilab

 asp :+ 


 asp :− 

nas :+
   
 nas :−   
voice :+ voice :+
48 an open introduction to linguistics

Let us establish the natural classes. First, each feature (a single

attribute-value pair) defines a natural class:

[ place : bilab] {p, ph , b, bh , m}

[ place : dental] {t, th , d, dh , n}
[ place : retroflex] {ú, úh , ã, ãh , ï}
[ place : palatal] {c, ch , é, éh , ñ}
[ place : velar] {k, kh , g, gh , N}
[ asp : +] {ph , bh , th , dh , úh , ãh , ch , éh , kh , gh }
[ asp : −] {p, b, m, t, d, n, ú, ã, ï, c, é, ñ, k, g, N}
[ nas : +] {m, n, ï, ñ, N}
[ nas : −] {p, ph , b, bh , t, th , d, dh , ú, úh , ã, ãh , c, ch , é, éh ,
k, kh , g, gh }
[ voice : +] {b, bh , m, d, dh , n, ã, ãh , ï, é, éh , ñ, g, gh , N}
[ voice : −] {p, ph , t, th , ú, úh , c, ch , k, kh }
(4.4)

All other classes are intersections of the ones above. For example, the
class of phonemes that are both retroflex and voiced can be formed by
looking up the class of retroflex phonemes, the class of voiced phonemes
and then taking the intersection:

{ú, úh , ã, ãh , ï} ∩ {b, bh , m, d, dh , n, ã, ãh , ï, g, gh , N} = {ã, ãh , ï} (4.5)

Basically, there are at most 6 × 3 × 3 × 3 = 162 (!) different natural

classes. How did I get that number? For each attribute you can either
give a value, or leave the value undecided. That gives 6 choices for
place, 3 for nasality, 3 for voice, and 3 for aspiration. In fact, nasality
does not go together with aspiration or with being voiceless, so some
combinations do not exist. All the phonemes constitute a natural
class of their own. This is so since the system is set up this way: each
phoneme has a unique characteristic set of features. Obviously, things
have to be this way, since the representation has to be able to represent
each phoneme by itself. Now, 162 might strike you as a large number.
However, as there are 25 phonemes there are 225 = 33, 554, 432 dif-
ferent sets of phonemes (if you cannot be bothered about the maths
here, just believe me)! So a randomly selected set of phonemes has a
chance of about 0.00005, or 0.005 percent of being natural!
How can we decide whether a given set of phonemes constitutes a
natural class? First method: try all possibilities. This might be a little
slow, but you will soon find some shortcuts. Second method: find a
description that fits all and only the sounds in your set. It has to be
of the form ‘has this feature, this feature and this feature’—so no dis-
junction, no negation. You take two sounds and look at the attributes
on which they differ. Obviously, these ones you cannot use for the
description. After you have established the set of attributes on which
all agree, determine the set that is described by this combination. If
it is your set, that set is natural, otherwise, it is not. Take the set
 phonology ii: realization rules and representations 49

{m, ph , ã}, for example:

m ph ã
     
place:bilab place:bilab place:retro

 asp :− 


 asp :+ 


 asp :− 
 (4.6)
nas :+
     
   nas :−   nas :− 
voice :+ voice :− voice :+
The first is nasal, but the others are not. So the description can-
not involve nasality. The second is voiceless, the others are voiced.
Therefore, the description cannot involve voicing. The same goes for
aspiration and place. It means that the smallest natural class that
contains this set is—the entire set of sounds. Yes, the entire set of sounds is a nat-
This example was in some sense easy: there was no feature that the ural class. Why? Well, no condition
is also a condition. Technically, it cor-
phonemes shared. To see another example, look at the set {[p], [ph ], responds to the empty AVS, which is
[b]}. Agreeing features are blue, disagreeing features are red (I have denoted by [ ]. Nothing is in there,
so any phoneme fits that description.
marked the agreeing features additionally with +):

p ph b
     
+ place:bilab place:bilab place:bilab

 asp :− 


 asp :+ 


 asp :− 

     
 + nas :−   nas :−   nas :− 
voice :− voice :− voice :+
(4.7)

It seems that we have found a natural class. However, when we extract

the two agreeing features we get the class of bilabial stops, which is
{[p], [ph ], [b], [bh ]}. This class contains one more phoneme. So the
original class is not natural.
Now, why are natural classes important and how do we use them?
Let us look at a phenomenon of Sanskrit (and not only Sanskrit) called
Sandhi. Sanskrit words may end in any of the following phonemes: p,
m, t, n, ú, k, and N. This word-final consonant changes depending on
the initial phoneme of the following word. We can use the following Sometimes the initial phoneme also
look-up table: changes. An example is /tat Ja
ri:ram/, which becomes /tac ch a
ri:ram/.
word ends in:
k ú t p N n m
.
p, ph k ú t p N n m
.
b, bh g ã d b N n m
.
t, th k ú t p N n m
.
d, dh g ã d b N n m
.
ú, úh k ú ú p N .ù
m m
. (4.8)
ã, ãh g ã ã b N ï m
.
c, ch k ú c p N .J
m m
.
é, éh g ã é b N ñ m
.
k, kh k ú t p N n m
.
g, gh g ã d b N n m
.
n, m N ï n m N n m
.
[ù] is a voiceless retroflex fricative, [J] is a voiceless palatal fricative.
There is one symbol that needs explanation. The symbol m . denotes
50 an open introduction to linguistics

a nasalization of the preceding vowel (thus it is not a phoneme in the

strict sense—see below on a similar issue concerning vowel change in
English). Despite its nonsegmental character I take it here at face
value and pretend it is a nasal.
We can capture the effect of Sandhi also in terms of rules. A rule
is a statement of the following form:

X →Y / C D
(4.9)
Input → Output Context

For the understanding of rules is important to stress that they repre-

sent a step in a sequence of actions. In the rule given above the action
is to replace the input (X) by the output (Y) in the given context. If
the context is arbitrary, nothing is written. The simplest kind of rule,
no context given, is exemplified by this rule:

a→b (4.10)

This rule replaces a by b wherever it occurs. Thus, suppose the input

is

The visitors to Alhambra are from abroad. (4.11)

then the output is

The visitors to Alhbmbrb bre from bbrobd. (4.12)

Notice that A, being a different character is not affected by the rule.

Also, b is not replaced by a, since the rule operates only in one direc-
tion, from left to right.
If we want to restrict the action of a rule to occurrences of letters
at certain places only, we can use a context condition. It has the form
C D. This says the following: if the specified occurrence is between
C (on its left) and D (on its right) then it may be replaced, otherwise
it is unaffected. Notice that this is just a different way of writing the
following rule:

CXD → CYD (4.13)

Let’s look at an example. The rules of spelling require that one uses
capital letters after a period (that’s simplifying matters a bit since the
period must end a sentence). Since the period is followed by a blank—
written ␣, in fact, maybe there are several blanks, but let’s ignore that
too—the context is ‘.␣ ’. D is omitted since we place no condition
on what is on the right of the input. So we can formulate this for the
letter a as

a → A/.␣ (4.14)

This rule says that a is changed to A if it is preceded by a blank which

in turn is preceded by a period. Alternatively we could use

.␣a → .␣A (4.15)

 phonology ii: realization rules and representations 51

Let’s return to Sandhi. As in the previous chapter, a word boundary

is denoted by #. This is not a printed character, and may in fact come
out in different ways. Also, since we are mostly dealing with spoken
language, there is no real meaning in counting blanks, so we leave the
precise nature of blank unspecified. Suppose we want to write rules
that capture Sandhi. Each cell of the look-up table then represents one
individual rule. For example, if the previous word ends in /k/ and the
following word begins with /b/, then rather than the sequence /k#b/
we will see the sequence /g#b/. Thus we find that Sandhi is among
many others the rule

/k#b/ → /g#b/ (4.16)

We can reformulate this into

k → g/ #b (4.17)

To be precise, Sandhi also erases the word boundary, so we should

actually write the rule as follows.

k# → g/ b (4.18)

However, now that we understand where I have simplified matters, we

can move on to the essential question, namely, how to best represent
the Sandhi using abstract rules.
If you do the calculations you will find that this table has 154 cases
(and I haven’t even given you the whole table). In 60 cases an actual
change occurs. It is true that if there is no change, no rule needs to be
written, unless you consider the fact that in all these cases the word
boundary is erased. However, in any case this is unsatisfactory. What
we want is to represent the regularities directly in our rules.
There is a way to do this. Notice for example the behaviour of
/k/, /ú/ and /p/. If the consonant of the following word is voiced,
they become voiced, too. If the consonant is voiceless, they remain
voiceless. This can be encoded into a single rule. Observe that the
last consonant of the preceding word is a stop; and the first consonant
of the following word is a stop, too. Using our representations, we can
capture the content of all of these rules as follows.
" # " # " #
voice:− voice:+ voice:+
→ / # (4.19)
nas :− nas :− nas :−

As explained in the previous chapter, this is to be read as follows:

given a phoneme, there are three cases. Case 1: the phoneme does
not match the left hand side (it is either voiced or a nasal); there is
no change. Case 2: the phoneme matches the left hand side but is
not in the context required by the rule (does not precede a voiceless
stop); there is no change. Case 3: the phoneme matches the left hand
side and is in the required context. In this case, all features that are Notice that every consonant which is
not a nasal is automatically a stop in
not mentioned in the rule will be left unchanged. This is the way we
this set. This is not true in Sanskrit,
achieve generality. I will return to the issue of [t] shortly. Note that but we are working with a reduced set
of sounds here.
52 an open introduction to linguistics

the rule above can also be written as follows:

" # " #
h i voice:+ voice:+
nas:− → / # (4.20)
nas :− nas :−

The omission of the voicing specification means that the rule applies
to any feature value. Notice that on the right hand side we find the
pair [ voice : +]. This means that whatever voice feature the original
sound had, it is replaced by [ voice : +].
Next, if the consonant of the following word is a nasal, the preceding
consonant becomes a nasal. The choice of the nasal is completely
determined by the place of articulation of the original stop, which is
not changed. So, predictably, /p/ is changed to /m/, /t/ to /n/, and
so on.
 
h i voice:+ h i
nas:− →  asp :−  / # nas:+ (4.21)
 
nas :+

The reason we specified voicing and aspiration in the result is that we

want the rule to apply to all obstruents. But if they are voiceless and
we change only nasality, we get either a voiceless or an aspirated nasal.
Neither exists in Sanskrit.
We are left with the case where the preceding word ends in a nasal.
The easy cases are /N/ and /m . /. They never change, and so no rule
needs to be written. This leaves us with two cases: /t/ and /n/. Basi-
cally, they adapt to the place of articulation of the following consonant
provided it is palatal or retroflex. (These are the next door neighbours,
phonetically speaking.) However, if the following consonant is voice-
less, the nasal changes to a sibilant, and the nasalisation is thrown
onto the preceding vowel.
Let’s do this one by one: /t/ becomes voiced when the following
sound is voiced; we already have a rule that takes care of it. We need
to make sure that it is applied, too. The change towards a retroflex There needs to be a system of schedul-
articulation can be captured by the following rule: ing rule applications, a theme to which
we shall return in Chapter6. If we are
" # " # " # exact and state that the rules remove
nas :− nas :− nas :− the word boundary, however, the rules
→ / # cannot be applied sequentially, and we
place:dent place:retro place:retro need to formulate a single rule doing
(4.22) everything in one step.

A similar rule is written for the palatals. It is a matter of taste (and

ingenuity in devising new notation) whether one can further reduce
these two rules to, say,
" # " # " #
nas : − nas :− nas :−
→ / #
place:dent place: α place: α
(α ∈ {retro, palat}) (4.23)

Let us finally turn to /n/. Here we have to distinguish two cases:

whether the initial consonant is voiced or unvoiced. In the voiced case
 phonology ii: realization rules and representations 53

/n/ assimilates in place:

" # " # " #
nas : + nas :+ nas :−
→ / #
place:alv place: α place: α
(α ∈ {retro, palat}) (4.24)

If the next consonant is voiceless, we get the sequence /m . / plus a

fricative whose place matches that of the following consonant. This is
the only occasion where we are putting in the manner feature, since
we have to describe the resulting phoneme.
 
nas : −  
" # nas :−
nas : +  voiced : + 
→m / #  voice :− 
   
. 
place:alv  manner:fric 
place: α
place : α
(α ∈ {retro, palat}) (4.25)

Thus, we use the rules (4.19), (4.21), (4.23), (4.24) and (4.25). Given
that the latter three abbreviate two rules each this leaves us with a
total of 8 rules as opposed to 60.

Neutralization of Contrast
There are also cases where the phonological rules actually obliterate
a phonological contrast. We’ll discuss an example of this called final
devoicing. In Russian and German, stops become devoiced at the end
of a syllable. It is such rules that cannot be formulated in the same
way as above, namely as rules of specialization. This is so since they
involve two sounds that are not allophones, for example [k] and [g] in
German. We shall illustrate the German phenomenon, which is rather
widespread. The contrast between voiced and voiceless is phonemic:

Kasse ["kas@] (cashier) : Gasse ["gas@] (narrow street)

Daten ["da:t@n] (data) : Taten ["ta:t@n] (deeds)
Peter ["pe:tEa ] (Peter) : Beter ["be:tEa ] (praying person)
(4.26)

Now look at the following words: Rad (wheel) and Rat (advice). They
are pronounced alike: ["Ka:t]. This is because at the end of the syllable
(and thus at the end of the word), voiced stops become unvoiced:
" #
h i +stop
+stop → / # (4.27)
−voiced

So how do we know that the sound that underlies Rad is [d] and not
[t]? It is because when we form the genitive the [d] does not disappear:

(des) Rades ["Ka:d@s] (4.28)

The genitive of Rat on the other hand is pronounced with [t]:

(des) Rates ["Ka:t@s] (4.29)

54 an open introduction to linguistics

This is because the genitive adds an [s] (plus an often optional epen-
thetic schwa) and this schwa suddenly makes the [t] and [d] non-final, In phonology, epenthesis means the
so that the rule of devoicing does not apply. addition of one or more sounds to
a word, especially to the interior of
a word. http://en.wikipedia.org/
Phonology: Deep and Surface wiki/Epenthesis

The fact that rules change representations has led linguists to posit
two distinct sublevels. One is the level of deep phonological repre-
sentation and the other is that of surface phonological represen-
tation. The deep representation is more abstract and more regular.
For German, it contains the original value for voicing, regardless of
the environment the consonant is in. The surface representation, on
the other hand, contains the phonological description of the sounds
that actually appear; so it will contain only voiced stops at the end
of a syllable. The two representations are linked by rules that have
the power of changing one representation into the other. In terms of
IPA–symbols, we may picture the change as follows:
["Ka:d]
↓ (Final Devoicing) (4.30)
["Ka:t]
However, what we should rather be thinking of is this:
   
−vowel +vowel −vowel
 +approximant   +open   +stop 
#
   
 +velar   +front   +labiodental 
  

+voiced +long +voiced

↓ (4.31)
   
−vowel +vowel −vowel
 +approximant  +open  +stop 
#
   
+velar +front +labiodental
  
   
+voiced +long −voiced
There is another option: we can
deny that the voicing contrast at the
We may in fact view all rules proposed above as rules that go from
end of a syllable is phonological—a
deep to surface phonological mapping. Some of the rules just add fea- voiced stop like [b] is just realized
tures while some others change features. The view that emerges is (i.e. put into sound) in two differ-
ent ways, depending on the environ-
that deep phonological structure contains the minimum specification ment. This means that the burden is
necessary to be able to figure out how the object sounds, while pre- on the phonology–to–phonetics map-
serving the highest degree of abstraction and regularity. For example, ping. However, the evidence seems to
be that the syllable-final [b] is pro-
strings are formed at deep phonological level by concatenation, while nounced just like [p], and therefore
on the surface this might not be so. We have seen that effect with final simply is transformed into [p].
devoicing. While on the deep level the plural ending (or a suffix like
chen) is simply added, the fact that a stop might find itself at the end
of the word (or syllable) may make it change to something else. The
picture is thus the following: the word /Rad/ is stored as a sequence
of three phonemes, and no word boundary exists because we might
decide to add something. However, when we form the singular nomi-
native, suddenly a word boundary gets added, and this is the moment
the rule of final devoicing can take effect.
 phonology ii: realization rules and representations 55

This setup is not without problems. Look at the English word

/mouse/. Its plural is /mice/, so the plural is formed by changing the
root vowel. How is this change accounted for? Is there are a phono-
logical rule that says that the root vowel is changed? (We consider
the diphthong for simplicity to be a sequence of two vowels of which
only the second is relevant here.) The answer to the second question
is negative. Not because such a rule could not be written, but be-
cause it would be incredibly specific: it would say that the sequence
[maUss#] (with the second ‘s’ coming from the regular plural!) is to be
changed into [maıs#]. Moreover, we expect that phonological rules can
be grounded in the articulatory and perceptive quality of the sounds.
There is nothing that suggests why the proposed change is motivated
in terms of difficulty of pronunciation. We could equally well expect
the form [maUs@z], which is phonologically well–formed. It just is not
a valid plural of the word [maUs], though English speakers are free to
change that. English used to possess more irregu-
So, if it is not phonology that causes the change, something else lar plurals. The plural of [bUk] was
once [be:k], the plural of ["tunge] was
must. One approach is to simply list the singular [maUs] and the ["tungan], and many more. These
plural [maıs], and no root form. Then [maıs] is not analysable into a have been superseded by regular for-
mations.
root plus a plural ending, it is just is a simple form. Another solution
is to say that the root has two forms; in the singular we find [maUs], in
the plural [maıs] and that the plural actually has no exponent at all,
like the singular. This last approach has the advantage that it does
not handle the change in phonology; for we have just argued that it is
not phonological in nature.
There is—finally—a third approach. Here the quality of the second
half of the diphthong is indeterminate between U and ı. It is specified
simply as a ‘lower high’ vowel. If you consult Table 3.5 you see that
there are exactly two vowels that fit this description: U and ı. So, it is
a natural class. Hence we write the representation as follows.

m a U/ı s
 
  +vowel  
+stop  " # +fric
+low  +vowel (4.32)
 +nasal   +alveol 
    
+back  +lower high


+bilab −voiced
−rounded

This representation leaves enough freedom to fill in either front or back

to get the plural or the singular form. Notice however that it does not
constitute a phoneme. In early phonological theory one called these
objects archiphonemes.
We shall briefly comment on this solution. First, whether or not
one wants to use two stems or employ an underspecified sound is a
matter of generality. The latter solution is only useful when it covers a
number of different cases; in English we have at least /this/:/these/,
/woman/:/women/, /foot/:/feet/. So there is a basis for arguing In German this root vowel change is
that we find an archiphoneme here. The second step is to identify much more widespread, and we shall
return to that phenomenon.
the representation of the plural. Obviously, the plural is not a sound
in itself, rather it is something that makes a sound into a different
sound. As a solution, we can propose a notation ← [+ front] which
56 an open introduction to linguistics

says the following: go leftward (= backward) and attach yourself to

the first possible sound. So, the plural of /mouse/ becomes represented
as follows:

m a U/ı s
 
  +vowel  
+stop  " # +fric
+low  +vowel
 +nasal   +alveol ← [+ front]
    
+back  +lower high


+bilab −voiced
−rounded
(4.33)

By convention, this is

m a ı s
 
  +vowel   
+stop  +vowel +fric
+low 
 (4.34)
+
 nasal   +lower high  +alveol 
   
 +back 
+bilab +front −voiced
−rounded

The singular then has the representation ← [+ back]. You may have wondered about the
A downside of this approach, however, is that it leaves us with a fact that the so-called root of a noun
like /cat/ was identical to its singular
more complex picture of phonological representation. It does not just form. This is not actually the case; the
contain items that define sounds but also funny symbols that act on root does not have a word boundary,
while the plural does (as nothing can
sounds somewhere else in the representation. attach itself after the plural suffix).

Notes
You may have wondered about the use of different slashes. The slashes
are indicators showing us at which level of abstraction we are work-
ing. The rule is that /·/ is used for sequences of phonemes while [·]
is used for sequences of actual sounds. Thus, one rule operates on
the phonemic level while another operates on the phonetic level. Fur-
thermore, writers make a distinction between the surface phonological
level and the deep phonological level. (There could also be a deep
phonetic and surface phonetic level, though that has to my knowledge
not been proposed.) In my own experience textbooks do not consis-
tently distinguish between the levels, and for the most part I think
that writers themselves do not mentally maintain a map of the levels
and their relationships, which is a popular source of mistakes. Since I
have not gone into much detail about the distinction between the two,
obviously there is no reason to ask you to be consistent in using the
slashes. It is however important to point out what people intend to
convey when they use them.
In principle there are two levels: phonetic and phonemic. At the
phonemic level we write /p/, and at the phonetic level we write [p].
Despite the transparent symbolism there is no connection between the
two instances of p. Phonemics does not know how things sound, it
only sees some 40 or so distinct categories of sounds. Writing /p/ is
therefore just a mnemonic aid; we could have used ♠ instead of /p/
(and get rid of the slashes, since it is now clear that we are not on the
phonetic level). Phonemes may be organised using features, but the
 phonology ii: realization rules and representations 57

phonetic content of these features is unknown to phonemics. To con-

nect the (abstract) phoneme /p/ with some sound we have realisation
rules: /p/ → [p] is a rule that says that whatever the phoneme /p/ is,
it comes out as [p]. The trouble is that [p] is ambiguous in English.
We have a different realisation of /p/ in /spit/ than in /pit/. There
are two ways to account for this. One is to write the realisation rules
so as to account for this different behaviour (V stands for vowel), for
example by writing

/p/ → [p]/s V (4.35)

Another is to translate /p/ into the ‘broad’ vowel [p] and leave the
specification to phonetics. Both views have their advantages, though
I prefer the first over the second. I have made no commitment here,
though to either of the views, so the use of /·/ versus [·] follows no
principle and you should not attach too much significance to it.
I also use the slashes for written language. Here they function in
a similar way: they quote a sequence of characters, abstracting from
their surface realisation. Again, full consistency is hard to maintain
(and probably not all that necessary). A side effect of the slashes is
that strings are easier to identify in running text.
With regards to our sketch in Chapter 1 I wish to maintain (for the
purpose of this course at least) that phonology and phonetics are just
one level; and that phonology simply organises the sounds via features
and contains the abstract representations, while phonetics contains the
actual sounds. This prevents us from having to deal with too many
levels.

Study Questions
1. In your own words, describe what a natural class is. Is an intersec-
tion of two natural classes also a natural class? What is the largest
natural class for a given set of sounds? And the smallest?

2. If we devise a set of rules to describe a phonological phenomenon,

like we did for Sandhi in Sanskrit, does the order in which we apply
the rules matter?

3. We discussed two sublevels of phonological representation, deep

and surface. What is the advantage of using a deep representation,
which, in turn, needs a system of realization rules to produce the
actual surface phonemics? Why not just use the surface represen-
tation exclusively?

4. In the subsection on deep and surface representation, three ap-

proaches are proposed to account for the problem of plurals that in-
clude root vowel change (mouse-mice). What are the three options
proposed and why does the author argue in favour of the third one?

5. For each of the following AVSs that define a natural class, give the
set of phonemes from English (see Table ?? on Page ??) that belong
to that natural class.
58 an open introduction to linguistics

a.
 
consonant:+
:bilabial 
 
 place
voice :−

b.
 
consonant:−

 height :open 

 
 rounded :− 
voiced :+

c.
" #
consonant:+
place :dental

6. Considering the following data,

a. Identify the allophones of the phoneme /n/. Make sure you do not confuse the com-
bination of I and n, In, and m!
b. Write a set of rules that describes their occurrence in this data
c. Can you give an explanation for why this variation and distri-
bution is found?

Word IPA Word IPA

note [noUt] handy [hæn”di]
inhuman [In"hju:m@n] pan [pæn]
indeed [In
”"di:d] skunk [sk@Nk]
in theory [In
”"TIri] singing [sININ]
uncommon [2N"kA:M@n] inept [Inept]
in part [Im"pA:rt] unbelievable [2mb@li:v@bl]
"

7. In English, the lateral phoneme /l/ has two realizations: a clear,

alveolar allophone [l], and a dark, velarised allophone [&l ]. Consider
the following English words given with a narrow IPA transcription
so you can see the allophonic variation of this lateral phoneme.

a. In what environments do [l] and [&l ] occur? Hint: If you don’t see the answer im-
mediately, carefully write out what
b. Write a phonological rule that captures this variation. sound occurs before and after each us-
age to try to find a pattern.

Word IPA Word IPA

lost [l6st] sold [soU&l d]
lamp [læmp] feel [fi:&l ]
license [laIs@ns] crawl [kr6&l ]
last [læst] melt [me&l t]
5
Phonology III: Syllable Structure & Stress

Words consist of syllables. The structure of syllables is deter-

mined partly by universal and partly by language-specific prin-
ciples. In particular we shall discuss the role of the sonoricity
hierarchy in organising the syllabic structure, and the principle
of maximal onset.

Utterances are not mere strings of sounds. They are structured into
units larger than sounds. A central unit is the syllable. Words consist
of one or several syllables. Syllables in English typically start with one
or more consonants, followed by a vowel or a diphthong and then one
or more consonants again. The first set of consonants is the onset,
the group of vowels the nucleus and the second group of consonants
the coda. The combination of nucleus and coda is called rhyme. So,
syllables have the following structure:

[onset [nucleus coda]] (5.1)

For example, /strength/ is a word consisting of a single syllable:

[stô [E NT]]
Onset Nucleus Coda
(5.2)
Rhyme
Syllable

Thus, the onset consists of three consonants: [s], [t] and [ô], the nucleus
consists just of [E], and the coda has [N] and [T]. We shall begin with
some fundamental principles. The first concerns the structure of the
syllable.

Syllable Structure I (Universal)

Every syllable has a non-empty nucleus. Both coda and onset
may however be empty.

A syllable which has an empty coda is called open. Examples of

open syllables are /a/ [e] (onset is empty), /see/ [si] (onset is non-
empty). A syllable that is not open is closed. Examples are /in/ [ın]
(onset empty) and /sit/ [sıt] (onset non-empty). The second principle
identifies the nuclei for English.

Vowels are Nuclear (English)

A nucleus can only consist of a vowel or a diphthong.
60 an open introduction to linguistics

This principle is not fully without problems and that is the reason that
we shall look below at a somewhat more general principle. The main
problem is the unclear status of some sounds, for example [Ä]. They
are in between a vowel and consonant, and indeed sometimes end up
in onset position and sometimes in nuclear position, for example in
/bird/ [bÄd].
The division into syllables is clearly felt by any speaker, although
there can be hesitation as to how to exactly divide a word into sylla-
bles. Consider the word /atmosphere/. Is the /s/ part of the second
syllable or part of the third? The answer is not straightforward. In
particular the stridents (that is, the sounds [s], [S]) enjoy a special sta-
tus. Some claim that they are extrasyllabic (not part of any syllable
at all), some maintain that they are ambisyllabic (belonging to both
syllables). We shall not go into that here.
A validation for the existence of rhymes can be given by looking at
verses (which also explains the terminology): words that rhyme do not
need to end in the same syllable, they only need to end in the same
rhyme: /fun/ – /run/ – /spun/ – /shun/. Similarly, the idea of the
coda is sustained by the fact that it is the domain of a rule that affects
many languages: For example, in English and Hungarian, within the There is an interesting problem
caused, among other things, by nasals.
coda the obstruents must either all be voiced or unvoiced; in German
Nasals are voiced by default. Now try
and Russian, all obstruents in the coda must be voiceless. The notion to find out what will happens when
of onset was used in Germanic verse in the Middle Ages, which used a you pronounce words with a sequence
nasal+voiceless stop in the coda, such
rhyming technique where the onsets of the rhyming words had to be as /hump/, /stunt/ and /Frank/.
the same.
It is worthwhile to remain with the notion of the domain of a rule.
Many phonological constraints are seen as conditions that concern two Rhyming words with the same on-
adjacent sounds. When these sounds come into contact, they undergo set is also called alliteration. It al-
lows to rhyme two words of the same
change to a smaller or greater extent, for some sound combinations are stem; German had a lot of Umlaut and
easier to pronounce than others, such as Sandhi, which we have dis- ablaut, that is to say, it had a lot of
cussed at length in Chapter 4. Another example is the Latin word /in/ root vowel change, making it impos-
sible to use the same word to rhyme
‘in’, a verbal prefix, which changes in writing (and in pronunciation) with itself (say /run/ – /ran/).
to /im/ when it precedes a labial (e.g. /impedire/). Somewhat more
radical is the change from [ml] to [mpl] to avoid the awkward combi-
nation [ml] (the word /templum/, for example, derives from /temlom/,
with /tem/ being the root, meaning ‘to cut’).
There is an influential theory in phonology, autosegmental phonol-
ogy, which assumes that phonological features are organized on differ-
ent scores (tiers) and can spread to adjacent segments independently
from each other. Think for example of the feature [±voiced]. The
condition on the coda in English is expressed by saying that the fea-
ture [±voiced] spreads along the coda. Clearly, we cannot allow the
feature to spread indiscriminately, otherwise the total utterance would
be affected. Rather, the spreading is blocked by certain constituent
boundaries; these can be the coda, onset, nucleus, rhyme, syllable,
foot, or word. To turn that on its head: the fact that features are
blocked indicates that we are facing a constituent boundary. So, the
existence of voicing harmony indicates that English has a coda.
The nucleus is the element that bears the stress. We have said that
 phonology iii: syllable structure & stress 61

in English it is a vowel, but this applies only to careful speech. In

everyday language use this is not always true. Consider the standard In German this is certainly so; the
verb /retten/ is pronounced ["KEth n].
pronunciation of /beaten/: ["bi:th n] (with a syllabic [n]). To my ears, "
" The [n] must therefore occupy the nu-
the division is into two syllables: [bi:] and [th n]. cleus of the second syllable. There are
" more languages like this. Slavic lan-
Phonologists have posited a number of conditions on syllable struc-
guages are full of consonant clusters
ture, partly based on the so-called sonoricity hierarchy, which is and syllables that do not contain a
shown in Table 5.1 vowel. Consider the island /Krk/, for
example.
dark vowels > mid vowels > high vowels Table 5.1: The Sonoricity Hierarchy.
vd. = voiced, vl. = voiceless.
[a], [o] [æ], [œ] [i], [y]

> r-sounds > nasals; laterals > vd. fricatives

[r]; [ô] [m], [n]; [l] [z], [Z]

> vd. plosives > vl. fricatives > vl. plosives

[b], [d] [s], [S] [p], [t]

The syllable is then organized as follows:

Syllable Structure II
Within a syllable the sonoricity follows a strict pattern: it in-
creases, then decreases again and is highest in the nucleus.

This means that the nucleus must contain at least one sound which is
at least as sonorous as any of the others in the syllable. It is called
the sonoricity peak. Thus, in the onset, consonants must be orga-
nized so that the sonority rises, while in the coda it is the reverse.
These conditions say nothing about the nucleus itself. In fact, some
diphthongs are increasing ([ı@] as in the British English pronunciation
of /here/), while others are decreasing ([aı], [oı]). This explains why
the phonotactic conditions for the onset are the opposite of those for
the coda. You can end a syllable in [ôt], but you cannot begin it that
way. You can start a syllable by [tô], but you cannot end it that way
Let me briefly note why diphthongs are not considered problematic
in English: it is maintained that the second part of the diphthong is
actually a glide (not a vowel), and as such is part of the coda. Thus,
/right/ would have the following structure:

r a j t
(5.3)
O N C C

A moment’s reflection now shows why the position of stridents is

problematic: the sequence [ts] is the only legitimate onset; according
to the sonoricity hierarchy, [st] is ruled out. Unfortunately, both are
attested in English, with [ts] only occurring in non-native words (e.g.
/tse-tse/, /tsunami/). There are even phonologists who believe that
[s] is part of its own little structure here (‘extrasyllabic’). In fact,
there are languages that prohibit this sequence; Spanish is a case in
point. Spanish avoids words that start with [st]. Moreover, Spanish
speakers like to add [e] in front of words that do anyway, to avoid
pronunciation difficulties. Rather than say [stôenZ] (/strange/) they
62 an open introduction to linguistics

will say [estôenZ]. The effect of this manoeuvre is that [s] is now part
of the coda of an added syllable, and the onset is reduced to just [tô], French is similar, but French speak-
ers are somehow less prone to add
or, in their pronunciation, most likely [tr]. Notice that this means
the vowel. (French has gone through
that Spanish speakers apply the phonetic rules of Spanish to English, the following sequence: from [st] to
because if they applied the English rules they would still end up with [est] to [et]. Compare the word
/étoile/ ‘star’, which derives from
the onset [stô] (see the Maximise Onset principle below). Latin /stella/ ‘star’.

Representing the Syllabification

The division of words into syllables is called syllabification. In writ-

ten language the syllable boundaries are not marked, so words are not
explicitly syllabified. We only see where words begin and end. The
question is: does this also hold for the representations that we need
to assume, for example, in the mental lexicon of a speaker? There is
evidence that almost no information about the syllable boundaries is
stored in the mental lexicon. One reason is that the status of sounds
at the boundary changes as soon as new material comes in. Let us
look at the plural /cats/ of /cat/. The plural marker is /s/, and
it is added at the end. Let’s suppose the division into syllables was
already given in the lexicon. Then we would have something like this:
/†cat†/, where † marks the syllable boundary. Then the plural will
be /†cat†s†/, with the plural ‘s’ forming a syllable of its own. This is
not the desired result, although the sonoricity hierarchy would predict
exactly that.
Let’s look further. We have seen that in German consonants in the
coda are devoiced. The word /Rad/ is pronounced ["Ka:t] as if written
/Rat/. Suppose we have added the syllable boundary: /†Rad†/. Now
add the genitive /es/ (for which we would also have to assume a syl-
labification, for example /†es†/, or /es†/). Then we get /†Rat†es†/,
which by the rules of German would have to be pronounced ["Ka:t P@s],
with an inserted glottal stop, because German does not have syllables
beginning with a vowel (whereas French does) and prevents that by
inserting the glottal stop. Phonemically, there is no glottal stop, how-
ever! In fact the pronunciation is ["Ka:d@s]. There are two indicators
as to why the sound corresponding to /d/ is now at the beginning of
the second syllable: (1) there is no glottal stop following it, (2) it is
pronounced with voicing, that is, [d] rather than [t].
Right away, we notice a consequence of this: syllables are not the
same as morphemes (see Chapter 7 for a definition). Neither are mor-
phemes necessarily the same as syllables or sequences thereof, nor do
syllable boundaries constitute morpheme boundaries. Morphemes can
be as small as a single phoneme (like the English plural), a phone-
mic feature (like the plural of /mouse/), they can just be a stress
shift (nominalisation of the verb /protest/ [pro"tEst] into /protest/
["protEst]) or they can even be phonemically zero (in English, you can
turn a noun into a verb (/to google/, /to initial/, /to cable/,
/to TA/). The representation does not change at all, neither in writ-
ing nor in speaking. You just verb it...
Of course, it may be suggested that the syllabification is explicitly
 phonology iii: syllable structure & stress 63

given and changed as more things come in. But this viewpoint compli-
cates matters unnecessarily. As syllabification is largely predictable,
there is good reason to believe most of it is not stored. It is enough
to just insert syllable boundary markers in the lexicon where they are
absolutely necessary and leave the insertion of the other boundary
markers to be determined later. Another reason is that rate of speech
influences pronunciation, which in turn influences syllable structure.

Language Particulars
Languages differ in what types of syllables they allow. Thus, not only
do they use different sounds, they also restrict the possible combina-
tions of these sounds in particular ways. Finnish allows (with few
exceptions) only one consonant at the onset of a syllable. Moreover,
Finnish words preferably end in a vowel, a nasal or ‘s’. Japanese sylla-
bles are preferably CV (consonant plus vowel). This has effects when
these languages adopt new words. Finns, for example, call the East
German car /Trabant/ simply [rabant:i] (with a vowel at the end and
a long [t]!). The onset [tr] is simplified to just the [r]. The same can
be seen in other loanwords: /koulu/ [kOulu] ‘school’ has lost the ‘s’,
/rahti/ [rahti] ‘freight’ has lost the ‘f’, and so on. Notice that it is
always the last consonant that remains.
English too has constraints on the structure of syllables. Some
are more strict than others. We notice right away that English does
allow the onset to contain several consonants, similarly with the coda.
However, some sequences are banned. Onsets may not contain a nasal
except in the first place (exception: [sm] and [sn]), although there are
some loanwords that break the rule, e.g. /mnemonic/ and /tmesis/.
The sequence sonorant-obstruent is also banned ([mp], [ôk], [ôp], [lt]
and so on; this is a direct consequence of the sonoricity hierarchy.
Stridents are found only in the first place; exceptions are [ts] and [ks],
found however only in non-native words. The cluster [ps] is reduced to
[s]. It also makes a difference whether the syllable constitutes a word
or is peripheral to the word. Typically, syllables have to be simpler
inside the word than at the word boundary.
Syllabification is based on expectations concerning syllable struc-
ture that derive from conditions on the well-formedness of syllables.
However, these still leave some room for variation. ["pleito] (/Plato/)
can be syllabified ["plei.to] or ["pleit.o]. In both cases, the syllables we
get are well formed. However, if a choice exists, then preference is
given to creating open syllables and/or syllables with onset.
All these principles are variations of the same theme: if there are
consonants, languages prefer them to be in the onset rather than the
coda, also known as Maximise Onset.
Maximise Onset
Put as many consonants into the onset as possible.
The way this operates is as follows: first we need to know what the
nucleus of a word is. In English this is easy: vowels and only vowels are
nuclear. Thus, we can identify the sequences consisting of coda+onset
64 an open introduction to linguistics

but we do not yet know how to divide them into a coda and a subse-
quent onset. Since the first syllable of a word can only begin with a
(null) onset, we say that a sequence of consonants x is a legitimate
onset of English if there is a word such that the largest stretch of
consonants it begins with is identical to x.

Legitimate Onsets
A sequence of sounds is a legitimate onset if and only if it is the
onset of the first syllable of an English word.

For example, [spô] is a legitimate onset of English, since the word

/sprout/ begins with it. Notice that /sprout/ does not show that [sp]
is a legitimate onset, since the sequence of consonants that it begins
with is [spô]. To show that it is legitimate we need to find another
word, e.g. /spit/. In conjunction with the fact that vowels and only
vowels are nuclear, we can always detect which sounds belong to either
a coda or an onset without knowing to which of the two it belongs.
However, if a sound occurs before the first nucleus, it must definitely
be part of an onset. In principle, there could be onsets that never show
up at the beginning of a word, but the above principle of legitimate
onsets in conjunction with Maximise Onsets states that this can never
happen. And this is how we can find our syllable boundaries. Given
a sequence of consonants that constitute a sequence of coda+onset,
we split it at the earliest point so that the second part is a legitimate
onset.
From now on we denote the syllable boundary by a dot, which is
inserted into the ordinary spelling of the word. So, we shall write
/in.to/ to signal that the syllables are /in/ and /to/. Let us now
look at the effect of Maximise Onset. For example, take the words
/restless/ and /restricted/. We do not find /re.stless/ nor
/res.tless/. The reason is that there is no English word that begins
with /stl/ or /tl/. There are plenty of words that begin with [l],
for example /lawn/. Hence, the only possibility is /rest.less/. No
other choice is possible. Now look at /restricted/. There are onsets
of the form [stô] (/strive/). There are no onsets of the form [kt],
so the principle of maximal onsets mandates that the syllabification
is /re.stric.ted/. Without Maximise Onset, /res.tric.ted/ and
/rest.ric.ted/ would also be possible since both coda and onset
are legitimate. Indeed, Maximise Onsets works towards making the
preceding syllable open, and making the next syllable have an onset.
In the ideal case all consonants are onset consonants, so syllables are
open. Occasionally this strategy breaks down. For example, /suburb/
is syllabified /sub.urb/. This reflects the composition of the word
(from Latin /sub/ ‘under’ and /urbs/ ‘city’, so it means something like
the lower city; cf. English /downtown/ which has almost the opposite
meaning!). If it truly forms an exception we might expect that there
is a representation in the mental lexicon which contains a syllable
boundary: /s2b†Äb/.
 phonology iii: syllable structure & stress 65

Stress
Syllables are not the largest phonological unit. They are themselves or-
ganised into larger units. A group of two, sometimes three syllables is
called a foot. A foot contains one syllable that is more prominent than
the others in the same foot. Feet are grouped into higher units, where
again one is more prominent than the others, and so on. Prominence is
marked by stress. There are various ways to give stress to a syllable.
Ancient Greek is said to have marked stress by pitch (the stressed syl-
lable was about 3/2 of the frequency of an unstressed syllable). Other
languages (like German) use loudness, while other languages use a
combination of the two (Swedish). Within a given word, there is one
syllable that is the most prominent. In IPA, it is marked by a pre-
ceding ["]. We say that it carries primary stress. Languages differ
with respect to the placement of primary stress. Finnish and Hun-
garian place the stress on the first syllable, French on the last. Latin
put the stress on the second-to-last (penultimate) syllable if it was
long (that is to say, had a long vowel or was closed); otherwise, it was
the syllable that preceded it (the antepenultimate). Thus, we had
pe.re.gri.nus (‘foreign’) with stress on the penultimate (gri) since
the vowel was long, but in.fe.ri.or with stress on the antepenulti-
mate /fe/ since the /i/ in the penultimate ri was short. Sanskrit
is said to have had free stress, that is to say, stress was free to fall
anywhere in the word.
Typically, within a foot the syllables like to follow each other in a
specific pattern. If the foot has two syllables, it consists either of an
unstressed syllable followed by a stressed syllable (iambic metre), or
vice versa (trochaic metre). Sometimes a foot carries three syllables
(a stressed syllable followed by two unstressed ones, a dactylus). So,
if the word has more than three syllables, it will have more than one
foot, and therefore there will be a syllable that is more prominent than
its neighbours while not carrying main stress. You can try this with
the word /antepenultimate/. You will find that the first syllable is
more prominent than the second but less so than the fourth. We say
that it carries secondary stress: [­antEp@n"2ltımEt].
The so-called metrical stress theory tries to account for stress as
follows. The syllables are each represented by a cross (×). This is a
Layer 0 stress. Then, in a sequence of layers, syllables get assigned
more crosses. The more crosses, the heavier the syllable. The number
of crosses is believed to correspond to the absolute weight of a sylla-
ble. So, a word that has a syllable of weight 3 (three crosses) is less
prominent than one with a syllable of weight 4. Let’s take the word
/assimilation/, [@­sım@"leıSn].

Layer 0 × × × × ×
(5.4)
@ sı m@ leı Sn
We have five syllables. Some syllables get extra crosses. The syllable
[sı] carries primary stress in /assimilate/. Primary stress is always
marked in the lexicon, and this mark tells us that the syllable must get
a cross. Further, heavy syllables get an additional cross. A syllable
66 an open introduction to linguistics

counts as heavy in English if it has a coda or a diphthong or a long

vowel. So, [leı] gets an extra cross. [Sn] is not heavy since the [n] is
nuclear. So this is now the situation at Layer 1:

Layer 1 × ×
Layer 0 × × × × × (5.5)
@ sı m@ leı Sn

Next, the nominalisation introduces main stress on the fourth syllable.

So this syllable gets main stress and is therefore assigned another cross.
The result is this:

Layer 2 ×
Layer 1 × ×
(5.6)
Layer 0 × × × × ×
@ sı m@ leı Sn

If larger units are considered, there are more cycles. The word /maintain/
for example has this representation by itself:

Layer 2 ×
Layer 1 × ×
(5.7)
Layer 0 × ×
meın teın

To get this representation, all we have to know is where the primary

stress falls. Both syllables are heavy and therefore get an extra cross
at Layer 1. Then the main syllable gets a cross at Layer 2. Now, if
the two are put together, a decision must be made which of the two
words is more prominent. It is the second, and this is therefore what
we get:

Layer 3 ×
Layer 2 × ×
Layer 1 × × × × (5.8)
Layer 0 × × × × × × ×
meın teın @ sı m@ leı Sn

Notice that stress is governed by a number of heterogeneous factors.

The first is the weight of the syllable; this is the deciding factor for
Layer 1 stress. Then there is the position of the main stress (which
in English must to a large extent be learned—equivalently, it must
explicitly be given in the representation, unlike syllable structure).
Third, it depends on the way in which the word is embedded into
larger units (so syntactic criteria play a role here).
In addition, morphological formation rules can change the location
of the main stress. For example, the suffix (a)tion attracts stress
([­kOm"baın] and [­kOmbı"neıSn]) so does the suffix ee (as in /employee/),
but /ment/ does not (["gavÄn] and ["gavÄnment]). The suffix /al/ does
move the accent but does not attract it (["ænEkdot] versus [­ænEk"dotal]).
Finally, we mention a problem concerning the representations that
keeps coming up. It is said that certain syllables cannot receive stress
because they contain a vowel that cannot be stressed (for example,
 phonology iii: syllable structure & stress 67

schwa: [@]). On the other hand, we can also say that a vowel is that
way because it is unstressed. Take, for example, the pair ["ôi@laız] and
[­ôi@lı"zeıSn]. When the stress shifts, the realisation of /i/ changes. So,
is it rather the stress that changes and makes the vowel change quality
or does the vowel change and make the stress shift? Often, these
questions have no satisfactory answer. In this particular example it The tendency to simplify consonant
clusters in Finnish has diminished
seems that the stress shift is first, and it induces the vowel change. somewhat in recent times. Consider
It is known that unstressed vowels undergo reduction in time. For recent loanwords such as: /trolli/
example, the reason why French stress is always on the last syllable [trol:i], ‘internet troll’ and /spammata/
[spAm:AtA], ‘to spam’, as compared
is because it inherited the stress pattern from Latin, but the syllables to older loans like /ranska/ [rAnskA],
following the stressed syllable eventually eroded away. Here the stress ‘French’, from Swedish ‘franska’. This
is likely due to an increased knowl-
was given and it drove the development! edge of foreign languages, especially
English. Similarly, Finnish dialects
spoken in close proximity to Swedish-
Study Questions speaking areas of Finland have gener-
ally been more tolerant of word-initial
1. Describe how the following terms relate to each other: coda, foot, consonant clusters than others.
syllable, stress, rhyme, nucleus, onset, open, closed.
Hint: try pronouncing the sounds
from Table 5.1 in hierarchical order
2. What are the sonoricity hierarchy and sonoricity peak used for? To
what property of articulation is the notion of sonoricity related?

3. What are the arguments for the idea that syllable information is
not stored in the mental lexicon? Are there any exceptions?

4. What principle explains the fact that we syllabify the word /pro-
prietary/ as /pro.pri.e.tar.y/, and not, for example as /prop.-
ri.e.tar.y/?

5. What factors influence the stress (or heaviness) that is assigned to

a syllable? How do these factors interact to produce the final stress
pattern?

6. For each of the following English words, give the syllable structure
of the word and mark the primary (and, if applicable, secondary)
stress.

a. restricted
b. disappeared
c. stressful
d. encyclopaedia
e. protester

7. For each of the phonetic transcriptions of English words below, give

the syllable structure and identify the onset, nucleus and coda of
each syllable.

a. [Imp6sIbl]
b. [fr2streItINli]
c. [trEZ@]
68 an open introduction to linguistics

8. We have seen that languages differ in the way they fit loanwords
into their phonology. Finnish, for example tends to adapt loanwords
so that they do not have word-initial consonant clusters and end in
a vowel (e.g. /koulu/), while English generally leaves them intact,
even if they do not fit in with general phonological patterns (e.g.
/tsunami/). What are the advantages and disadvantages of these
approaches? Which do you think is better?
6
Phonology IV: Rules, Constraints
and Optimality Theory

In this chapter, we will look at phonological rules and how to

apply them. Also, arguments for distinguishing between multiple
possible analyses will be discussed and a constraint-, not a rule-
based framework called Optimality Theory is presented.

Rules and their Application

It is important to be clear about a few problems that arise when dealing
with (phonological) rules. If you have some string, say /teatime/ and
a rule, say t → d, how should you go about applying it? Once, twice,
as often as you can? And if you can apply it several times, where do
you start?
Suppose you can apply the rule only once, then you get either
/deatime/ or /teadime/, depending on where you decide to apply
it. If you apply it as often as you can, then you can apply the rule a
second time. In this case the result is the same, no matter where you
start: /deadime/. This is not always so. Consider the rule a → x/ a;
with input /aaa/. This gives you two options: you can apply it to
the first occurrence of /a/ or to the second. The third is not eligible
because of the context restriction. If you apply the rule to the first
occurrence you get /xaa/. If you apply it to the second you get /axa/.
With /xaa/ you can apply it a second time, resulting in /xxa/; but
after this, the rule cannot be applied a third time.
Often a proper formulation of the rule itself will be enough to ensure
only the correct results will be derived, regardless of the application.
Sometimes, however, an explicit scheduling must be given, such as:
apply the rule going from left to right as often as possible. In this
chapter we shall not go into the details of this. Instead, we shall turn
to another problem, namely the interaction between two rules. Here
is a very simple example: suppose we have two rules, R1 : a → b and
R2 : b → c. We can schedule these in four different ways:

À R1 or R2 can be applied once;

Á R1 or R2 can be applied any number of times;

 R1 must be applied before R2 ;

70 an open introduction to linguistics

à R2 must be applied before R1 .

All these choices give different results. We shall illustrate this with an
interesting problem, the basic form of the past tense marker in English.

A Problem Concerning the Underlying Form

Regular verbs form the past tense by adding one of the following three
suffixes: [t], [d] or [@d] (do not confuse these with the spelling vari-
ations of the past tense marker). The choice between these forms is
determined by the form of the root.

[t] [d] [@d]

licked [lıkt] bugged [b2gd] mended [mEnd@d]
squished [skwıSt] leaned [lind] parted [pAôt@d] (6.1)
kept [kEpt] buzzed [b2zd] feasted [fist@d]
laughed [læft] played [pleıd] batted [bæt@d]

The question is: what is the source of these differences? Sure, it is pos-
sible to say that the past has three different forms and that depending
on the verb a different form must be chosen. This, however, misses
one important point, namely that the choice of the form is determined
solely by the phonological form of the verb and can be motivated by
phonological constraints of English. Therefore, the variation should
be explained in the phonological stratum, not anywhere else The facts
can be summarized as follows.

À [d] is found if the last sound of the verb is voiced but unequal to
[d].

Á [t] is found if the last sound of the verb is voiceless but unequal to
[t].

 [@d] is found if the verb ends in [d] or [t].

We mention here that it is required that the verb is regular. Irregular

verbs like /run/ and /catch/ are of course not covered by this rule. We
may think of the latter as entered in the mental lexicon as unanalysed
forms. Thus, rather than seeing /caught/ as a sequence of two forms,
namely /catch/ plus some past tense marker, we think of the form
entered as a whole, though otherwise functioning in the same way. It
is, if you will, an idiom. This idea is similar to what we saw in earlier
discussions of irregular plural formations.
It seems that the variation can be accounted for solely by applying
some general principles. First, notice that in a coda, with the exception
of the sonorants ([l], [m], [n]), all consonants agree in voicing.

Voice Agreement Principle

Adjacent obstruent sequences must either be both [ voice : +]
An obstruent is a consonant that is
or both [ voice : −] at the end of a word.
either a stop, an affricate, or a frica-
tive.
This principle already goes halfway towards explaining the choice of The principle is less general than pos-
the suffix form. It tells us why we see [d] after voiced consonants, but sible: the constraint is valid not only
it does not tell us why it is that we get [lind] rather than [lint], because at the end of a word but in any coda.
 phonology iv: rules, constraints and optimality theory 71

both of them are legitimate according to this principle. Furthermore,

this does not explain why we find the inserted schwa. The latter can be
explained as follows: suppose there was no schwa. Then the present
and the past forms would sound alike: /mendd/ would be [mEnd].
Languages generally try to avoid double consonants (although they
never completely manage). Another popular strategy is haplology, English employs the strategy to in-
the dropping of one of the consonants. sert schwa also in the plural. We
find [b2s@z] /busses/ (or /buses/),
It is still possible to use the Voice Agreement Principle if we assume not [b2s:] (/buss/, with a long /s/).
that there is just a single form of the past tense marker, instead of
three, and that variants arise only as the result of a repair. This repair
is carried out by applying some additional rules. Various analyses are
possible, based on which a system of repair rules can be devised.
Analysis 1. We assume that the underlying form is [d]. There is
a rule that devoices [d] right after a voiceless obstruent. There is a
second rule which inserts a schwa right before [d]. For the purpose of
the definition of the rules, two consonants are called similar if they
differ not at all or only in the voicing feature (for example, [t] is similar
to both [t] and [d], but to nothing else). This can be formalized as:
" # " #
+voice −voice
→ /[−voice] # (6.2a)
+obstruent +obstruent

∅ → [@] /C C0
(6.2b)
(C and C0 similar)

The symbol ∅ denotes the empty string. The first rule should actually
be written as two rules in our feature system:
" # " #
voice : + voice : −
→ /[−voice] #
manner:stop manner:stop
" # " # (6.3)
voice : + voice : −
→ /[−voice] #
manner:fricative manner:fricative

The second rule effectively says that it is legal to insert a schwa any-
where between similar consonants. Since we have two rules, there is a
choice as to which one shall be applied first. We shall initially schedule
(6.2b) before (6.2a). This means that the rule (6.2b) is applied to the
original form F0 , giving us an output form F1 , and then we apply rule
(6.2a) to get F2 . Each rule applies only once, so the final output is F2 .
This yields the following result:

root /b2gd/ /lıkd/ /mEndd/ /staôtd/

(6.2b) /b2gd/ /lıkd/ /mEnd@d/ /staôt@d/ (6.4)
(6.2a) /b2gd/ /lıkt/ /mEnd@d/ /staôt@d/

Notice that, when we say that a rule does not apply, this does not
mean that no output is generated. Instead, it means that the form is
left unchanged by the rule. Textbooks sometimes indicate this by a
hyphen, as if to say that there is no output. But there is an output, it
just happens to be identical to the input. Notice, too, that sometimes
a rule does apply, but does not actually change anything.
72 an open introduction to linguistics

Now, suppose we had instead scheduled (6.2a) before (6.2b). Then

this would be the outcome:
root /b2gd/ /lıkd/ /mEndd/ /staôtd/
(6.2a) /b2gd/ /lıkt/ /mEndd/ /staôtt/ (6.5)
(6.2b) /b2gd/ /lıkt/ /mEnd@d/ /staôt@t/
If the last consonant is /t/, the rule (6.2a) would first devoice the past
tense marker, and we get the suffix /@t/, contrary to fact. Thus, the
order in which the rules apply is relevant here. There is, however,
nothing intrinsic in the system of the rules that tells us in which order
they have to apply. This has to be stipulated explicitly.
Analysis 2. In this case, the underlying form is assumed to be [t].
In place of (6.2a), there now is rule (6.6) that voices [t] right after a
voiced obstruent or a vowel. There is a second rule which inserts a
schwa right before [d] or [t], which is identical to (6.2b)
" # " #
−voice +voice
→ /[+voice] # (6.6)
+obstruent +obstruent

root /b2gt/ /lıkt/ /mEndt/ /staôtt/

(6.2b) /b2gt/ /lıkt/ /mEnd@t/ /staôt@t/ (6.7)
(6.6) /b2gd/ /lıkt/ /mEnd@d/ /staôt@d/
If we, instead, schedule (6.6) before (6.2b), this will be the outcome:

root /b2gt/ /lıkt/ /mEndt/ /staôtt/

(6.6) /b2gd/ /lıkt/ /mEndd/ /staôtt/ (6.8)
(6.2b) /b2gd/ /lıkt/ /mEnd@d/ /staôt@t/
Once again, we see that schwa insertion must take place first to yield
the correct outcome.
Analysis 3. Now, we assume the underlying form is [@d]. There is
a rule (6.9a) that devoices [d] right after a voiceless obstruent. There
is a second rule (6.9b) which deletes schwa in between dissimilar con-
sonants.

" # " #
+voice −voice
→ /[−voice] # (6.9a)
+obstruent +obstruent

[@] → ∅ /C C0 (6.9b)
0
(C and C dissimilar)
Initially, we apply rule (6.9b) before rule (6.9a):

root /b2g@d/ /lık@d/ /mEnd@d/ /staôt@d/

(6.9b) /b2gd/ /lıkd/ /mEnd@d/ /staôt@d/ (6.10)
(6.9a) /b2gd/ /lıkt/ /mEnd@d/ /staôt@d/
If we instead schedule (6.9a) before (6.9b), this would be the outcome:

root /b2g@d/ /lık@d/ /mEnd@d/ /staôt@d/

(6.9a) /b2g@d/ /lık@d/ /mEnd@d/ /staôt@d/ (6.11)
(6.9b) /b2gd/ /lıkt/ /mEnd@d/ /staôt@t/
 phonology iv: rules, constraints and optimality theory 73

We thus conclude that schwa deletion must precede voice assimilation.

In principle, there are many more analyses. We can assume the
underlying form to be anything we like (say, even [ð] or [@Z]). However,
one clearly feels that such a proposal would be much inferior to any of
the above. But why?
The principal difference between proposals is solely the extent to
which the rules that transform them can be motivated language inter-
nally, as well as externally. This is also the criterion that will make
us choose one analysis over the others. Let’s take a careful look at
the different analyses. First, let us go back to the Voice Agreement
Principle. The only thing it says is that adjacent obstruents must
agree in voicing. It does not claim that obstruents must agree with
the preceding vowel, since we have forms like [kæt].
Since rule (6.6) repairs some of the forms without need, Analysis 2
can be said to incorporate the wrong version of the Voice Agreement
Principle. Analysis 1, on the other hand, repairs the forms if and
only if they do not conform to the Voice Agreement Principle. Now
consider Analysis 3: it does not conflict with the Voice Agreement
Principle. However, it proposes to eliminate schwa in certain forms
such as ["lıked]. Since there are words in English such as /wicked/
that actually have such a sequence, Analysis 3 seems to repair forms
that are actually well-formed. Therefore, the best analysis is Analysis
1, which proposes that the underlying form is [d].
Let us summarize: an analysis is preferred over another if it proposes
laws of change that are widely attested (schwa insertion is one of them,
and final devoicing is another). Additionally, an analysis is dispreferred
if its rules change representations that are actually well-formed. Rules
of the kind discussed here are seen as repair strategies that explain
why a form sometimes does not appear as we would expect it to. This
constitutes, by the way, a mapping from deep phonological form to
surface phonological form. There is an additional piece of evidence
that makes us go for Analysis 1, namely the following principle:
Not-Too-Similar Principle
No English word contains a sequence of subsequent similar ob-
struents.
Rule (6.2a) devoices an obstruent in coda if the preceding consonant
is voiceless, i.e. when the Voice Agreement Principle is violated. After
the rule has been applied, the offending part is gone. Rule (6.2b)
applies if there are two subsequent similar consonants, i.e. when the
Not-Too-Similar Principle is violated. After application of the rule,
again, the offending part is gone.

Which Repair for Which Problem?

The idea that we are pursuing is that deep-to-surface mappings in-
stitutionalize a repair of impossible situations. Thus, every rule is
motivated by the fact that the input structure violates a constraint
of the language and the output structure removes that offending part.
Unfortunately, this is not the complete story. Look at the condition
74 an open introduction to linguistics

that onsets may not contain more than one consonant. This constraint
exists in many languages, for example in Japanese and Finnish. But
the two languages apply different repair strategies. While Japanese
likes to insert vowels, Finnish likes to cut the onset to the last vowel.
Both repair strategies are also attested in other languages. However,
we could imagine a language that simplifies an onset cluster to its first
element: with this strategy /trabant/ would become not /rabantti/
but /tabantti/ in Finnish, and Germanic /hrengas/ would become
not /rengas/, but /hengas/ instead, and so on. This latter strategy
is however, not attested anywhere. So, we find that among the infi-
nite possibilities to avoid forbidden cluster combinations, only some
get used at all, while others are completely disfavoured.
To a certain degree, we can think of reasons why some strategies
exist and why others are not used at all. Look at schwa insertion as
opposed to schwa deletion, for example. While the insertion of a schwa
is inevitably going to improve the structure (because all languages
agree that CV is a syllable), the deletion of a schwa has the risk of
producing clusters that are illegitimate. Thus, we would expect that
there is a bias towards the repair by schwa insertion as opposed to
schwa deletion. Note, however, that all these arguments have to be
taken with care. If a rule causes stress changes, for example, there
might be good reason to reduce or eliminate a vowel.

Optimality Theory
There are two main approaches to describe the regularities of language:

+ The generative approach proposes representations and rules.

The rules shall generate all and only the legitimate representations
of the language.

+ The descriptive or model-theoretic approach proposes only rep-

resentations and conditions that a representation has to satisfy in
order to belong to a given language.

Note that generative linguists do not always propose that rules are
real, that is, in the head of a speaker, nor that derivations take place
in real time. They would say that the rules are simply a way to system-
atize the data. If that is so, however, it is not clear why we should not
adopt a purely descriptive account, describing all and only the legal
representations rather than pretending that they have been derived
in a particular way. The discussion is complicated by the fact that
many arguments drawn in favour of either analysis come from data
on child development and language learning. To interpret the data
coming from learning, we need to have a theory of the internal knowl-
edge of language (‘language faculty’). This theory is related to the
approach to language, since this internal knowledge may either consist
of representations and conditions on the representations, or in fact rep-
resentations plus a set of rules. The discussion concerning the problem
of whether we should have rules or not is probably a never-ending one.
 phonology iv: rules, constraints and optimality theory 75

Optimality Theory (OT) adds another idea to this discussion.

OT tries to do away with rules (though we shall see that this is an
illusion). Also, rather than saying exactly which representations are
legitimate, it simply proposes a list of desiderata for an optimal result.
Even if a result is not optimal, it might still be accepted if it is the best
possible among its kin. Thus, to begin, OT must posit two levels: un-
derlying representation (UR) and surface representation (SR).
We start with the UR [bætd] (/batted/). How can we then find the
correct SR? OT assumes that we generate all possible competitors and
rank them according to how many constraints from a given set they
violate and how often they do so.
Here is how it can work for the current example. We shall start out
with an SR that deviates as little as possible from the UR. To that
effect, we assign each segment a slot in a grid:

b æ t d
(6.12)
• • • •

In this grid, we may subsequently insert or delete material and we may

change the representations of the segments, but we have to keep track
of the segments throughout the derivation. We can put restraints on
the changes that can be made to the segments. Here we shall require,
for example, that they do not change order with respect to each other.
Here is an example where a segment changes: However, segements can actually
change order sometimes; this is called
b æ t d metathesis.

• • • •
↓ ↓ ↓ ↓ (6.13)
• • • •
b æ t t

Here is an example where one segment is dropped:

b æ t d
• • • •
↓ ↓ ↓ (6.14)
• • •
b æ t

And here is an example where one segment is added:

b æ t d
• • • •
↓ ↓ ↓ ↓ (6.15)
• • • • •
b æ t @ d

We formalize this as pairs hI, Oi of representations. Consider the con-

straints (called principles) on such pairs:

Recover the Morpheme

At least one segment of any non-zero morpheme must be pre-
served in the SR.
76 an open introduction to linguistics

This principle is carefully phrased. If the underlying morpheme is

empty, there need not be anything in the surface representation. But
if it is not empty, then we must see one of its segments in the SR.

Recover Obstruency
If a segment of the UR is an obstruent, it must be an obstruent
in the SR.

This principle says that the underlying segment that is hosting an

obstruent must first of all survive; it cannot be deleted. Second, the
phoneme that the segment hosts must be an obstruent too.

Recover Voicing
If a segment is voiced in the UR, it must be voiced in the SR.

Syllable Equivalence
The UR must contain the same number of syllables as the SR.

Recover Adjacency
Segments that are adjacent in the UR must be adjacent in the
SR.

These principles are too restrictive in conjunction. The idea is that

not all of them have to be satisfied, but the more the better. An
obvious idea is to do something like linear programming: for each
constraint there is a certain penalty, which is ‘awarded’ on violation.
For each violation, the corresponding penalty is added. Here, however,
the violation of a more valuable constraint cannot be made up for. No
matter how often someone else violates a lesser valued constraint, if
you violate a higher valued constraint, you lose.
To make this more precise, for a pair hI, Oi of underlying repre-
sentation (I) and a surface representation (O), we take note of which
principles are violated. Each language defines a partial linear order-
ing of the constraints (such as the ones given above). This ordering
states, given two constraints C and C 0 , whether C is more valuable
than C 0 , whether C 0 is more valuable than C, or whether they are
equally valuable.

À Suppose that π = hI, Oi and π 0 = hI, O0 i are such that for all
constraints that π violates there is a more valuable constraint that
π 0 violates. Then O is called optimal with respect to O0 .

Á Suppose that π = hI, Oi and π 0 = hI, O0 i are such that C is the

most valuable constraint that π violates; and that C is also the
most valuable constraint that π 0 violates. Then if π 0 violates C
more often than π, O is optimal with respect to O0 .

 O is optimal if it is optimal with respect to every other O0 from a

pair with the same I.

The actual output form for I is the optimal output form for I; if there
are several optimal output forms, all of them are chosen. So, given I,
if we want to know which SR corresponds to it, we must find an O
which is optimal. OT uses the following terminology: O
and O0 are called candidates and
they get ranked. Note that candidacy
is always relative to the UR.
 phonology iv: rules, constraints and optimality theory 77

Let’s apply all this to [bætd]. We rank the constraints NTS (Not-
Too-Similar), RObs (Recover Obstruency), RMorph (Recover the Mor-
pheme) and RAdj (Recover Adjacency) as follows:

NTS, RObs, RMorph > RAdj (6.16)

The first three are ranked equal, but more than the fourth. Other
principles are left out of consideration.

/bætd/ NTS RObs RMorph RAdj

[bætd] ?
[bænd] ? (6.17)
[bæt] ?
[bæt@d] ?

The forms [bætd], [bænd], [bæt] all violate constraints that are higher
ranked than RAdj. [bæt@d] violates the latter only. Hence it is optimal
among the four. (Notice that we have not counted a violation of Re-
cover Obstruency for [bæt], even though one obstruent was dropped.
This will be discussed below.) Note also that the optimal candidate
still violates a constraint.
We now turn to the form /sæpd/. Assume that the ranking is (with
VAgr = Voice Agreement, RVoice = Recover Voicing)

VAgr, RMorph, RObs > RAdj > RVoice (6.18)

This yields the following ranking among the candidates:

/sæpd/ VAgr RMorph RObs RAdj RVoice

[sæpd] ?
[sæp] ?
(6.19)
[sæmp] ?
[sæp@d] ?
[sæpt] ?

Some Conceptual Problems with OT

First, notice that OT has no rules but that its constraints are not
well-formedness conditions on representations either. They talk about
the relation between a UR and an SR. They tell us in effect that
certain repair strategies are better than others, something that well-
formedness conditions do not do. This has consequences. Consider the
form [bænd]. We could consider it to have been derived by changing
[t] to [n]:

b æ t d
• • • •
↓ ↓ ↓ ↓ (6.20)
• • • •
b æ n t
78 an open introduction to linguistics

Or we could think of it as being derived by deleting [t] and inserting

[n]:

b æ t d
• • • •
↓ ↓ ↓ (6.21)
• • • •
b æ n d

More fanciful derivations can also be conceived. Which one is the

correct one, if that can be said at all? And how do we count violations
of rules? The first derivation violates the principle that obstruents
must be recoverable, while the second does not; it does not even violate
adjacency! Of course, we could forbid that obstruents be deleted,
but this is problematic for languages that do delete obstruents (e.g.
Finnish does so to simplify onset clusters). Alternatively, we allow
obstruents to be deleted, but we now count that as a violation of the
principle of recoverability of obstruents. In that case, the second pair
also violates that principle. This illustrates the problem: what is the
punishment associated with a certain UR-SR pair? How does it get
ranked? Maybe we could take the derivation that is most favourable
for a given candidate?
This is a very subtle problem, namely: how do we actually measure
violation? The string [cat], for example, how many constraints must
we violate and how often, to get it from [staôtd]? For example, we may
drop [s], then [d] and change the place of articulation of the first sound
to [+velar]. Or may we drop [s] and [d] in one step, and then change
[t] to [k]—or we do everything in one step. How do we then count
the number of times we have violated a certain rule? Ultimately, the
idea of violating a constraint 3 times as opposed to once makes little
sense unless we assume that we apply certain rules to the process of
derivation.

Notes

The transition from underlying /d/ to /t/, /@/ or /d/ has mostly been
described as morphophonemics. This is a set of rules that control
the spelling of morphemes in terms of phonemes. In this view, there
is one morpheme, which is denoted by, say, /PAST/M and there are Recall that / · /M means: this is a unit
rules on how to realise this morpheme on the phonological level (recall on the morphological level

a similar discussion on the transition from the phonological to the

phonetic level). Under that view, the morpheme /PAST/M has three
realisations. The problem with this view is that it cannot explain the
distribution of these forms. Thus we have opted for the following view:
there is one morph and it is realised as /d/ phonologically. It produces
illicit phonological strings and there are therefore rules of combination
that change the illicit combinations into licit ones.
 phonology iv: rules, constraints and optimality theory 79

Study Questions
1. What is the advantage of using the concept of a single realisation
of a morpheme, modified by repair rules, compared to the use of
multiple realisations?

2. In the discussion of the English past tense formation, three possible

analyses and underlying forms are given, which all, given the correct
rule ordering, reproduce the English regular past tense inflection
correctly. Still, Analysis 1 is deemed to be the best one, why?

3. Describe how the following terms and concepts, all part of Op-
timality Theory, relate to each other: underlying representation,
segments, constraints, surface representation, optimality, ranking
of rules.

4. What is the main argument for dealing with the [d], [t], [@d] vari-
ation on the phonological, and not the morphological level? And
what about irregular past tense markers? Can these be explained
by phonological considerations, too?

5. Finnish distinguishes three kinds of vowels: back, front and neutral

vowels. a [A], o [o] and u [u] are back, e [E], i [i] are neutral, ä
[æ], ö [œ] and y [y] are front. The rule is as follows: within a word,
there cannot be both a front and a back vowel. (This phenomenon is
referred to as vowel harmony.) Case endings, such as the ablative
case ending, has two forms: lta and ltä. The back form is added
if there is a back vowel, the front form is added in all other cases.

a. Determine the ablative forms of satama (‘harbour’), ravintola

(‘restaurant’), hissi (‘elevator’), urheilu (‘sport’) and etelä
(‘south’).
b. Look at the pairs henkilö (‘person’) – henkilökohtainen (‘per-
sonally’), osake (‘share’) – osakeyhtiö (‘shareholder company’).
All of them are genuine Finnish words. What do they suggest for
the domain of the vowel harmony rule? How can you reconcile
this with the formulation above?
c. Neutral vowels pose a problem for a strict formulation of vowel
harmony, which would require that in a word, all vowels har-
monize, i.e. have the same quality. Moreover, there are just
two forms for the ablative suffix, with the front suffix being the
default choice; there is no neutral form. Can you suggest an ap- Hint: Use two features [±back] and
[±front], with neutral vowels being
proach using OT that gets the facts right with a strong harmony
[+back, +front]. Harmony can then
requirement? be described using an ordering of two
constraints, one for each feature.
6. One of the problems with Optimality Theory that was already
touched upon earlier is that there is always more than one way
to get from a UR to an SR. This provides problems for the ‘scoring’
of violations, since these violations are based on principles that con-
cern the changes made to get from the UR to the SR. Can you think
of a solution to this problem? Is there a way to fix or systematize
the SR derivation, so that it is consistent?
7
Morphology I: Basic Facts

Morphemes are the smallest parts of language that carry mean-

ing. Words may consist of one or several morphemes in much the
same way they consist of one or more syllables. However, the two
concepts, that of a morpheme and that of a syllable, are almost
completely unrelated in other respects.

Word Classes
Morphology is the study of the smallest meaningful units of language.
It studies the structure of words from a semantic viewpoint rather
than from the viewpoint of sound. Morphology is intimately related to
syntax, but their domains are separate. Everything that is larger than
a word falls under the scope of syntax, while morphology is concerned
with the structure within words only.
The first thing to notice is that words come in different varieties,
called word classes (also called parts-of-speech). For example,
there are verbs (/to imagine/) and there are nouns (/a car/), there
are adverbs (/slowly/) and there are adjectives (/red/).

Semantic criteria, morphological criteria or contextual (syntac-

tic) criteria?
Intuitively, one is inclined to divide words into classes according to
their meaning: verbs denote activities, nouns denote objects and con-
cepts, adverbs denote ways of performing activities and adjectives de-
note properties of things. However, language does not abide by this
method of classification. The noun /trip/, for example, denotes an
activity, even though it is a noun. Thus, the semantic criterion alone
is not accurate enough, and can sometimes be misleading.
We could try to distinguish word classes using morphological cri-
teria. We can separate nouns, verbs and adverbs/adjectives with this
type of criteria. Verbs take the endings /s/, /ed/, and /ing/, nouns
only take the ending /s/. Adjectives and adverbs, on the other hand,
do not take any such endings. They will have to be distinguished by Rye. Picture JS.
other criteria.
Thus we may propose the following criterion: a word w is a verb if
and only if we can add [z] (/s/), [d] (/ed/) and [ıN] (/ing/) and nothing
else; w is a noun if and only if we can add [s] (/s/) and nothing else.
82 an open introduction to linguistics

However, this purely morphological criterion has problems of its

own. The first is that a given word may belong to several classes;
the test using morphology alone would class anything that is both a
noun and a verb, for example /fear/, as a verb, since the noun plural
(/fears/), is identical to the third person singular form of the verb.
Changing the wording to replace ‘if and only if’ by ‘if’ does not help
either, since in that case all verbs would be classified as both verbs
and nouns. A second problem is that there can be false positives; the
word /rise/ [ôaız] cannot be taken as the plural of /rye/
[ôaı], (a kind of grain), for example. And third, not all words use
the same formation rules. There are verbs that form their past tense
not in the way discussed earlier, by adding [d], but by a root vowel
change, such as /run/, which has no form ∗ /runned/, only /ran/. Or
consider the forms of the verb /be/, which are even more problematic.
Still, we classify both these examples as verbs. That it is crucial to distinguish be-
Clearly, one cause of these problems is that the endings taken by tween semantic anomaly and syntac-
tic well-formedness was pointed out by
English nouns constitute a subset of the endings that verbs take. This Chomsky in his 1957 book. Let’s read
means morphological criteria alone is not sufficient. Another criteria the original passage:
“ (1) Colorless green ideas sleep furi-
that could be used is context of use. We take a fixed context, also
ously.
called a frame, and consider which words fit into it. A frame for
(2) *Furiously sleep ideas green color-
identifying verbs could be the following: less.
It is fair to assume that neither sen-
The governor has the bill. (7.1) tence (1) nor (2) (nor indeed any part
of these sentences) has ever occurred
If you fill the gap by a word, it is certainly a verb (more exactly a in an English discourse. Hence, in any
transitive verb, one that takes a direct object). On the other hand, if statistical model for grammaticalness,
these sentences will be ruled out on
a word can fill the gap in the next example, it is a noun: identical grounds as equally “remote”
from English. Yet (1), though nonsen-
The has vetoed the bill. (7.2) sical, is grammatical, while (2) is not
grammatical.” (Chomsky, 1957)
When we say ‘fill the gap’ we do not mean that what we get is a
meaningful sentence when we put in that word; we only mean that it
is grammatically (= syntactically) well-formed. So we could fill in the
above gap with
/cat/, but legislating felines does stretch the imagination a bit.
When we fill in /democracy/ we have to stretch it even further, and
so on. These sentences are then semantically weird, meaning their
meaning is odd, but grammatically well-formed.
Adjectives can fill the position between the determiner (/the/) and
the noun:

The governor has vetoed the bill. (7.3)

Veto cat
Finally, adverbs (/slowly/, /surprisingly/) can fill the slot just be-
Picture JS.
fore the main verb.

The governor has vetoed the bill. (7.4)

Another test for word classes, again of a more morphological nature,

is their combinability with certain affixes. Affixes are morphemes that
are not really words by themselves, but get glued onto words to change
their meaning or function. (See Chapter 8 for details.) Table 7.1 shows
 morphology i: basic facts 83

a few English affixes and lists the word classes to which they can be
attached. We see that the list of affixes is heterogeneous, and that
affixes do not always attach to all members of a class with equal ease
(/anti-house/, for example, is yet to be found in English). Still, it is
a powerful test that reveals a lot of the division of words into different
classes.

Affix Attaches to Forming Examples Table 7.1: English Affixes and Word
Classes
anti- nouns nouns anti-matter, anti-aircraft
adjectives adjectives anti-democratic
un- adjectives adjectives un-happy, un-lucky
verbs verbs un-bridle, un-lock
re- verbs verbs re-establish, re-assure
dis- verbs verbs dis-enfranchise, dis-own
adjectives adjectives dis-ingenious, dis-honest
-ment verbs nouns establish-ment, amaze-ment
-ize nouns verbs burglar-ize
adjective verbs steril-ize, Islamic-ize
-ism nouns nouns Lenin-ism, gangster-ism
adjectives nouns real-ism, American-ism
-ful nouns adjectives care-ful, soul-ful
-ly adjectives adverbs careful-ly, nice-ly
-er adjectives adjectives nic-er, angri-er

Language Technology: Part-of-speech tagging

POS-tagging is the task of identifying the correct word class of
every word in a text. It is a necessary prerequisite for almost any
natural language processing task or application. Programs that
perform POS-tagging are called taggers and there are many well-
known ones that are freely available for larger languages if you
look online.
In the current text we’ve identified 8 word classes, but generally
taggers find it more useful to make more fine-grained distinctions.
For example the Brown Corpus, the first balanced corpus of En-
glish, has some 80 POS-tags, and even a number of very frequent,
very important words are given their own tag:

• BEM (am), BER (are, art), BEZ (is)

• DT (singular determiner, this, that), DTS (plural determiner,

these, those)

• OD (ordinal number, first, second, etc.), CD, (cardinal num-

ber, one, two, etc.)

• RBR (comparative adjective, e.g. slower, hotter), RBT (su-

perlative adjective, e.g. smartest, cutest

The SUSANNE Corpus on the other hand has 355 distinct tags.
This all shows that word classes are not written in stone, and there
is much theoretical and practical debate about which ones could
84 an open introduction to linguistics

and should be recognized. At a minimum, if two proposed classes

cannot reliably be distinguished from each other they should be
merged. Experimentation has shown that while too few word
classes may mean that important information is being ignored,
too many word classes can complicate identification.
In general POS-tagging is considered a solved problem: sys-
tems generally achieve more than 97% accuracy. Further, POS-
tagging is a problem that often can be easily solved with super-
vised machine learning methods: with a large body of texts (a
corpus) of correctly tagged (= manually by humans) text, almost
any method will produce a decent result.
Word classes are language spe-
cific
Not all languages make all the
Morphological Formation distinctions discussed here. Dutch,
for example, generally does not make
Using various processes, simple words can be turned into more com- a distinction between adjectives and
adverbs, and being quick, snel and
plex ones (e.g. the famously long word ‘antidisestablishmentarian-
doing something quickly, snel is coded
ism’). Those words or word parts that cannot be decomposed into with the same lexical item. Some
smaller parts, and must therefore be drawn directly from the lexicon, proposals are even more extreme.
For example, David Gil has argued
are called roots. Roots are ‘main’ words, those that carry meaning. that there is no distinction between
Affixes are not roots, and neither are inflectional endings. An exam- nouns and verbs in Riau Indonesian,
a language spoken in the west of
ple of a root is /cat/, which is identical to the singular form /cat/.
Indonesia.
However, the latter also has a word boundary marker at the right (and
therefore should actually be written as /cat#/, but this detail is often This is a somewhat hazy definition. It
generously ignored). will become clearer through examples.
In other languages, roots are clearly distinct from every form you
get to see on paper. Latin /deus/ ‘god’, for example, has two parts:
the root /de/, and the nominative ending /us/. This can be clearly
seen if we add the other forms as well: genitive /dei/, dative /deo/,
accusative /deum/, and so on. However, we generally do not find these
roots in dictionaries. Instead, you find the words by their citation
form, which in Latin is the nominative singular. So, you find the root
in the dictionary under /deus/ not under /de/. Verbs are usually cited in their infiniti-
There are several distinct ways in which words get formed and lan- val form. This is the case in e.g. Dutch
and French, although this need not be
guages differ greatly in the extent to which they make use of them. so: Hungarian dictionaries often list
The most important ones are: them in their 3rd person singular form.
This is because the 3rd person singu-
lar reveals more about the inflection
À compounding: two words, neither an affix, become one by juxta- in Hungarian than the infinitive. This
position. Each of them is otherwise found independently. Examples saves effort!
are /goalkeeper/, /whistleblower/ (noun + verb compound) and
/hotbed/ (adjective + noun).

Á derivation: only one of the parts is a word; the other part is

only found in combination, and it often changes the word class of
the host. Examples are the affixes which we have discussed above
(/anti/, /dis/, /ment/).

 inflection: one part is an independent word, the other is not. It

does not change the category, rather it adds some specification to
the category (inflection of verbs by person, number, tense).
 morphology i: basic facts 85

Compounding
In English, a compound can be recognised by its stress pattern. For
example, the main stress in the combination adjective+noun is on the
noun if they still form two words (black board [blæk "bOôd]) while
in a compound the stress is on the adjective part (/blackboard/
["blækbOôd]). Notice that the compound now is a single word; the ad-
jective has lost its status as an adjective through compounding, which
explains the new stress pattern.
In English, multi-word constructions are favoured over compounds.
It has to be said, though, that this difference is sometimes purely
orthographical; the use (or not) of spaces between the words does
not really tell you whether you are dealing with one or two words.
For example, although one writes /rest room/, (American Eng. for
public toilet) on the phonological level the stress pattern sounds as if
it is a single, compounded word.
There are languages where the compounds can be distinguished
from multi-word constructions by other aspects than stress patterns.
German is such a language: Funny compound from
https://www.orphicpixel.com/illustrated-
Regierung-s-entwurf (7.5) by-keren-rosen/

government proposal
Schwein-e-stall (7.6)
pig sty
Räd-er-werk (7.7)
wheel work = mechanism

The compound /Regierungsentwurf/ not only contains the two words

/Regierung/ (‘government’) and /Entwurf/ (‘proposal’), it also con-
tains an ‘s’ (called ‘Fugen s’ = ‘gap s’). What makes German interest-
ing for students is that what gets inserted is not always an ‘s’ but some-
times ‘e’; /Schweinestall/ is composed from /Schwein/ and /Stall/.
/Räderwerk/ is composed from /Rad/ and /Werk/. /Schweine/ and
/Räder/ sound exactly like the plural, while /Regierungs/ is a form
of /Regierung/ that does not occur anywhere else. Note that in none
of these cases would one mistake the compound for a multi-word con-
struction.
The meaning of a compound is often not directly determinable from
the meanings of its parts. It is characteristic of compounds that they
often juxtapose some words and leave the meaning of the whole open
for interpretation (take /money laundry/ or /coin washer/, which
are generally not places where you launder money or wash coins; if you
did you would be in serious trouble). This is more true of noun+noun
compounds than of verb+noun or noun+verb compounds, though.

Derivation
English has a fair amount of derivational affixes. Table 7.1 shows some
of them. We said that derivation often changes the category of the
word, but this is not always the case. I.e. if an affix changes the word
86 an open introduction to linguistics

class, it is derivational, not inflectional, but if the word class does not
change, derivation and inflection have to be distinguished in another
way. One difference is that derivation is optional (while inflection is
not), and can be iterated (i.e. done multiple times consecutively; inflec-
tion cannot be iterated). Additionally, inflectional affixes are typically
found outside of derivational affixes. To give an example: you can
form /republican/ from /republic/. To this you can add /anti/:
/antirepublican/ and finally form the plural: /antirepublicans/.
But the plural marker /s/ is an inflectional suffix. You could have
added the ‘s’ to /republic/, but then you could not go on with the
derivation; there is no word ∗ /republicsan/. Similarly, the word
/antirepublics/ has only one analysis: first /anti/ is added and
then the plural suffix /s/ is added.
The distinction between a root word and a word formed by deriva-
tion is not always clear, either. For example, is /resist/ formed by
affixation of /re/ to a root /sist/ meaning “to sist again”? Actually,
it was in the original Latin (/re-sistere/), but in English it is not.
This is because we do not have a verb ∗ /sist/. Thus, derivation may
form words that initially are perceived as complex, but later lose their
transparent structure. This may be because they start to sound dif-
ferently because some sort of sound change has occurred, or because
the base form disappears from the modern vocabulary.
Very often language users are completely unaware of the original
origin (or etymology) of the words they use. For example, the word
/nest/ in English (e.g. as in a bird’s nest) was once was a complex
word made from two morphemes combined: ∗ /nizdo/. Here we use
an asterisk to mark that the word form is reconstructed by historical
linguists based on evidence. The word is believed to be derived from
the words ∗ /ni/ or /nether/ meaning (‘down’) and ∗ /sed/ (‘sit’). This
is what historical linguists have determined. But in modern English,
we would simply treat nest as an indivisible root, and (most) speakers
are entirely unaware that it is originally a complex form.

Inflection

To fit a word into a certain syntactic construction (or frame), it may

have to undergo some changes. In English, the verb has to get an ‘s’
suffix if the subject is third person singular. The addition of the ‘s’
does not change the category of the verb; it makes it more specific,
however. Adding inflection to a word thus narrows down the contexts
in which it can be used. Inflection is not optional; you must use the
right inflectional ending if the context requires it. In Latin, adjectives
agree in (grammatical) gender, number (plural or singular) and case
(nomative, accussative, genitive, etc.) with the noun they modify.
Not also that unlike English, Latin adjectives can follow the nouns
they modify: What’s a gloss? A gloss is a repre-
sentation of a word that is split into its
component morphemes, followed by
an aligned line containing the trans-
lation of the root word and the mean-
ing of each non-root morpheme. All
parts which have morphemic represen-
tations of their own are separated by a
dash, ‘-’ (e.g. discipul-us and student-
 morphology i: basic facts 87

second student (nominative) (7.8)

discipul-us secund-us
student-nom.sg second-masc.nom.sg
second students’ (7.9)
discipul-orum secund-orum
student-gen.pl second-masc.gen.pl
(the) girls’ (7.10)
puell-arum secund-arum
girl-gen.pl second-fem.gen.pl
(the) poets’ (7.11)
poet-arum secund-orum
poet-gen.pl second-masc.gen.pl

The last example was chosen on purpose: inflectional morphemes

marking the same meaning on different word classes do not necessar-
ily have the same form. In Latin, the nom.sig inflectional suffix that
combines with nouns is realized with the same string as the adjectival
suffix meaning nom.sig (but with the additional grammatical gender
meaning signalling masculine gender). But this form identity is not a
necessity, although it is true that the forms of adjectives are related
to to those of nouns. The word /poeta/ is an exception, because it
belongs to a class of nouns that are mostly feminine, even though it it-
self is masculine. Adjectives show the same form as this class of nouns
if agreeing with a feminine noun because of this. But since /poeta/
is masculine, the adjectives that agree with it show a different form
class, one that is identical to that of the noun class which contains
more masculine nouns. This also explains why we have not added
gender specifications to the nouns; unlike adjectives, nouns cannot be
decomposed into gender and a genderless root.
A basic skill in morphological analysis is being able to distinguish in-
flection form derivation. These two types of morphemes have very dif-
ferent functions which are relevant for further analysis of words within
a sentence or text. While the distinction is for some morphemes harder
than others, there are a number of questions about the morphological
features of a given affix that you can ask to figure out whether it is
derivational or inflectional, summarized in the chart below.

0
88 an open introduction to linguistics

0
Let’s work through a few examples to see how this works. Consider
the following two English words with suffixes.
1. govern+ment (as in the Dutch government)

2. encourage+s (as in The new coach encourages the skaters.)

Is the morpheme -ment in govern+ment a derivational or an in-
flectional morpheme? First, we should determine what is the part-of-
speech of the root morpheme, govern. This seems to be a verb: The
president governs from the capital. Then, we should determine what
the part-of-speech of the entire word is. government is a noun: you
can prefix it with a determiner, e.g. the government, it can function
as an argument (subject or object) in a sentence, e.g. The government
is corrupt, and it semantically seems to be a thing (though admittedly
abstract). So +ment seems to combine with verbs to make nouns.
Now let’s consider the chart: the very first question on the chart
has already been answered in the affirmative: +ment does lead to a
change of part-of-speech. This means it must be derivational: only
derivational morphemes can change the part-of-speech of the words
they combine with.
Let’s now consider encourages. The root morpheme encourage is
unambiguously a verb. (They encourage bloggers is fine, but “the en-
courage” is not; this root cannot be modified by a determiner to make
a noun phrase). encourages also seems to be a verb: He encourages
class participation. So now let’s consider the chart. The first answer
is of course ‘No.’: there has been no change of speech.
Now consider the second question. Can we add a derivational affix?
This question actually means can we attach a derivational morpheme
directly to the morpheme we are investigating, so after the -s. This
is a good test because inflectional morphology is always applied to
a word after all the derivations have taken place. This is because
inflections are determined by the context in which a word is used,
when the sentence is being formed. We just figured out above that
-ment can turn verbs into nouns. But we don’t seem to be able to add
it to encourages to make a word: *encouragesment. This could just
 morphology i: basic facts 89

be a fluke so we should try a few more derivational suffixes that we

know can combine with verbs just to be sure, but as you will see, they
also do not work: -able makes verbs into adjectives (drinkable), but
*encouragesable; -er makes a verb into an agent that does the action
(writer), but *encourageser. So here as well we have to conclude that
the answer is ‘No.’ The final question is: can the affix be added to
other members of the same part-of-speech? In this case what is meant
is can we add an -s to other verbs, even different verb types? Here
the answer does seem to be ‘Yes.’: read+s, eat+s, cooks. So we can
conclude that probably we are dealing with an inflection.
Now all these questions are morphological ones, related to morpho-
logical structure. We can also ask semantically what the affix seems
to do, and with a bit of investigation we might realize it marks that
the subject is the third person singular form in present-tense verbs.
It’s important to note that English has another inflectional morpheme
that has the same surface form, -s, but which combines with nouns
when they are used as plurals (alternatively we can say this inflection
marks plurality). But it is not the same morpheme as third person
singular -s.
For the examples we examined we’ve also looked only at English.
English is a bit exceptional in its morphology: it has very little in-
flectional morphology. In fact, it only has seven different inflectional
suffixes. Other languages have many more. However, there are even
languages that have less, such as Chinese which has no inflectional
morphology. English Inflections
English is a language that has very lit-
tle inflectional morphology, and usu-
ally we identify only 7 inflectional suf-
fixes.
1. noun plural /-s/
2. verb 3rd person present tense /-s/
3. verb regular past tense /-ed/
4. verb past participle /-en/ e.g. He
has eaten lunch.
5. verb present participle /-ing/ e.g.
He is eating lunch.
Analyzing word derivations with tree structures 6. adjective comparative /-er/ e.g.
bigger, softer
7. adjective superlative /-est/ e.g.
largest
We’ve seen above that the ways in which morphemes can combine is
regulated by the type of morphemes involved. This means that often
there is only one derivation that was possible to arrive at a given
morphologically complex word. We can show the deriviation process
by using tree structures. For example, the words sees and revisited.
Sees is a root form combined with an inflectional suffix. But note
that revisited is a root form that first combines with a derivational
prefix the then combine with the inflectional suffix -ed, so that the
order of derivation is revisit –> revisited. This order is then made
explicit in the tree structure. We know that this is the correct order
of derivation, and not first visted –> revisited because we know that
inflectional morphemes like the past-tense suffix -ed are added last.
90 an open introduction to linguistics

V inf.

see -s

V infl.

deriv. V -ed

re- visit
Note that re- combines with verbs to
Diagramming the morphological structure becomes more interesting make new verbs, so we know that re-
had to have combined with nationalize
when more than two morphemes are involved. Consider the derivation (there is no *renation or *renational.
of ungracefully and renationalize. : If we put it in the past-tense, then the
inflectional suffix -ed would be added
Adv last.

Adj deriv.

deriv. Adj -ly

un- N deriv.

grace -ful

deriv. V

re- Adj deriv.

N deriv. -ize

nation al
What’s interesting is that every native speaker, and also non-native
speakers, on some unconscious level know that this is the derivation of
these words. However, it requires explicit teaching (and some analysis
and practice) to be come aware of the hidden structure.
 morphology i: basic facts 91

Study Questions
1. Say you want to determine the word class (also called part-of-
speech, or POS) of a certain word, i.e. whether it is a noun, verb,
adjective or adverb (actually, there are many more (or less!) word
classes, depending on who you ask). What are the various methods
you could use? Which one do you prefer?

2. What is the main distinction between root words and affixes? How
can you check whether something is a root or an affix?

3. We identified three main ways in which more complex words can

be formed from simpler (word) parts: derivation, inflection and
compounding. How can you distinguish between compounding on
the one hand and derivation/inflection on the other? And how do
you distinguish inflection from derivation?

4. One way of identifying the part-of-speech of a word is to use a

diagnostic frame or context. For each of the following word classes,
create a diagnostic frame that can test whether a word belongs
to that word class. Make sure the frame works for all and only
instances of the given word class. For each frame, give a positive
example (shows that a word is of that class) and a negative example
(shows that a word is not of that class).

a. noun
b. preposition (e.g. /to/, /over/)
c. adjective
d. verb

5. For each of the following words, identify all morphemes which make
up the word.

a. sadness
b. unkind
c. underworld
d. kingdom
e. simplified
f. highlighters
g. irreplaceable
h. antidisestablishmentarianism

6. Categorize the following affixes (underlined) as either inflectional

or derivational.

a. rational
b. John’s
c. bigger
d. untie
92 an open introduction to linguistics

e. eaten
f. criticize
g. studied

7. The examples below show a valid, grammatical derivation and an

impossible, ungrammatical one (marked with ∗ ).

a. Divide the underlined words into morphemes.

b. Identify what type of morpheme each is (root word or affix) and
how it is added to the other parts (compounding, derivation or
inflection)
c. After doing this, what generalisation can you make about the
order of morpheme application?

I. a. left - leftist - leftists

b. left - lefts - ∗ leftsist
II. a. sleep - sleepwalk - sleepwalked
b. sleep - slept - ∗ sleptwalk
III. a. foot - football - footballer
b. foot - footer - ∗ footerball

8. In this chapter (and in the book in general), we have mainly looked

at how things are done in English (and other Indo-European lan-
guages). This sometimes allows us to make generalizations or give
definitions that work well for English, but are far from universally
true. An example of this is the way we distinguished morphology
and syntax: as one being concerned with things at the word-level,
the other with things at the sentence-level. If we look at other
languages, we see that this definition would draw the morphology-
syntax border very differently.
In Finnish, for example, a whole phrase like ‘I would have danced’
can be expressed in one word, /tanssisin/ (there are many other
languages, like the Inuit language Inuktitut, which do this to an
even greater extent). On the other side of the spectrum there are
languages like Mandarin Chinese, which use very few affixes.
Given that there is so much variation, do you think it is still sensible
to draw the morphology-syntax distinction this way? If not, can you
think of another way to determine what belongs on which stratum? In linguistics in general, it is both use-
ful and fun to apply the things you
If you also speak a language other than English, try to compare it
learn to other languages you know;
to English in this respect. Does your other language have a lot of this will deepen your understanding of
morphemes in each word? Or does it prefer to separate morphemes the material and also help you keep an
open mind with regards to the endless
as much as possible? variation of language.
8
Morphology II: Morphemes
and Morphs

Morphs and Morphemes

Before we begin, a few words on terminology. As I already said in the
introduction, we make a distinction between a morph and a mor-
pheme. In this chapter we think of morphs and morphemes as units
in the same way as phones and phonemes. As in phonology and syntax,
we shall distinguish a deep structure from a surface structure.
On the deep level, we only have morphemes, while at the surface
level, we have morphs. Morphs of the same morpheme are called allo-
morphs. What unites the morphs of a morpheme are their shared
syntactic and semantic features. An example of this is the plural
morpheme: syntactically, it turns a root noun into a plural noun; A morpheme in a language has a se-
semantically, it changes from denoting properties to denoting groups mantic and syntactic effect. Each al-
lomorph of the same morpheme will
of things. Nevertheless, it has many different surface forms in English. have the same syntactic and semantic
These are some of the ways to form the plural: effects.

À Add [z] at the end: [dOg] → [dOgz]

Á Add [s] at the end: [k2p] → [k2ps]

 Add [@z] at the end: [b2s] → [b2s@z]

à Change the root vowel from [U] to [ı]: [wUm@n] → [wım@n]

Ä Do not change anything: [Sip] → [Sip]

Å Add [@n]: [Oks] → [Oks@n]

Notice that we have used the square brackets: [·], and not the slashes:
/·/. This means that we have lost part of the abstraction that we
gained in deep phonology, which implies that we have listed more types
than necessary. We can lower the number of types, since at the deep
level, the first three cases (À – Â) become one. From the standpoint
of deep phonology the surface forms [z], [s] and [@z] are just the surface
manifestations of the deep form /z/.
Nevertheless, even if this is taken into account, there remain some
other forms that cannot be explained by surface-level variation only.
The change of root vowel, the suffix [@n] and the zero suffix belong to
this category. We therefore get the following revised list:
94 an open introduction to linguistics

À Add /z/ at the end: /dOg/ → /dOgz/; /k2p/ → /k2ps/; /b2s/ →

/b2s@z/.

Á Change the root vowel from /U/ to /ı/: /wUm@n/ → /wım@n/

 Do not change anything: /Sip/ → /Sip/

à Add [@n]: /Oks/ → /Oks@n/

Now consider the following. The plural morpheme has four different
morphs (in our example, that is) . The first is /z/, the second is root
vowel change, the third is ∅ (the empty morph) and the fourth is
/@n/. Which morph is used in a particular case depends on the noun
to which we apply the morpheme. On the deep morphology level we
only see a combination of a noun root with the plural morpheme; the
noun root-based choice of what morph to use is made at the surface
morphology level.

Kinds of Morphological Processes

The general term for grammatical morphs is affix. Generally, an affix
to a string ~x is a string ~y that puts itself somewhere in ~x. Given this
definition, it is either a prefix, a suffix or an infix. Some writers use it
in a more general sense, but we shall not do that here. Morphs are not
always affixes, however. A morph need not be a piece (= string) that
we add somewhere; it may be several pieces (transfix, circumfix), or
simply a certain kind of change in the string to which it is applied (e.g.
vowel change). The general term for all of these is morphological
change. Let’s take a look at what kinds of morphological changes
there are in the languages of the world.

Suffixes and Prefixes

A suffix is a string that is added at the end, a prefix is a string that
is added at the beginning. English mostly uses suffixes (derivational
ones like /ation/, /ize/, /ee/; inflectional ones like /s/ and /d/).
But it also has prefixes: /de/, /re/, /un/ are prefixes, for example.
It is generally agreed that—if we use an analogy with syntax here—
the affix is the head, and that it either expects the string on its right
(prefix) or on its left (suffix).
If there were only suffixes and prefixes, morphology would look like
syntax. Unfortunately, this is not the case.

Circumfixes
A circumfix consists of a part that is added to the beginning of a word
and another part that is added to the end. It is thus a combination of
prefix and suffix. The German perfect tense is marked by a circumfix.
It consists of the prefix /ge/ and a suffix which changes from stem to
stem (usually it is /en/ or /t/). The infinitive is a suffix /en/ as can
 morphology ii: morphemes and morphs 95

be derived from the table below.

Infinitive Root Perfect

sehen seh gesehen
backen back gebacken (8.1)
filmen film gefilmt
hausen haus gehaust

The two parts (the prefix and the suffix) are historically of different
origin (the prefix /ge/ did not exist originally, and is not found in
English). In present-day German, however, there is no reason to see
the two as being separate.
The superlative in Hungarian also is a circumfix:

nagy ‘great’ legnagyobb ‘greatest’

(8.2)
fehér ‘white’ legfeherebb ‘whitest’

Here, as in the German perfect, the circumfix has two identifiable

parts. The suffix is found also in the comparative:

nagy ‘great’ nagyobb ‘greater’

(8.3)
fehér ‘white’ feherebb ‘whiter’

In Romanian, we also see the comparative as part of the superlative:

we have /frumos/ (‘beautiful’) in the positive, /mai frumos/ (‘more
beautiful’) in the comparative, and /cel mai frumos/ (‘most beauti-
ful’). However, this does not prove that the superlative can be seen
as being formed from the comparative by adding a prefix. For this, it
would have to be possible to decompose the meaning of the superlative
in such a way that it is derived from the meaning of the comparative.
As far as I know, no proposals have been made that do this.

Infixes
Infixes insert themselves inside the string. Look at the following data
from Chrau.

vǒh ‘know’ vanǒh ‘wise’ (8.4)

cǎh ‘remember’ canǎh ‘left over’ (8.5)

The string /an/ is inserted after the first consonant! The string is cut
into two pieces and the nominaliser inserts itself right in between.

vǒh → v + ǒh → v + an + ǒh → vanǒh (8.6)

Chrau /’tSraU/ is a language spoken

Transfixes by some 22,000 ethnic Cho Ro people
in southern Vietnam. It belongs to
the Bahnaric branch of the Austroasi-
A transfix is an even more complex entity. We’ll look at an exam-
atic language family.
ple from Egyptian Arabic. Roots have three consonants, for example From http://www.wordtravels.com/
/ktb/ ‘to write’ and /drs/ ‘to study’. Words are formed by adding Travelguide/Countries/Vietnam/Map

some material in front (prefixation), some material after (suffixation)

96 an open introduction to linguistics

and some material in between (infixation). Moreover, all these typi-

cally happen at the same time. Let’s look at the following list:

[katab] ‘he wrote’ [daras] ‘he studied’

[Pamal] ‘he did’ [na\al] ‘he copied’
[baktib] ‘I write’ [badris] ‘I study’
[baPmil] ‘I do’ [ban\il] ‘I copy’
[iktib] ‘write!’ [idris] ‘study!’
(8.7)
[iPmil] ‘do!’ [in\il] ‘copy!’
[kaatib] ‘writer’ [daaris] ‘studier’
[Paamil] ‘doer’ [naa\il] ‘copier’
[maktuub] ‘written’ [madruus] ‘studied’
[maPmuu] ‘done’ [man\uul] ‘copied’
It requires some patience to find out how the different words have
been formed. There are variations in the patterns (just like in any
other language). One thing to be noted is that the vowels get changed
from one form of the verb to another. This explains why the vowel is
not thought of as being part of the root.

Other Kinds of Changes

Reduplication is the phenomenon where a string is copied and added.
For example, if /abc/ is a string, then its (re)duplication is /abcabc/.
Austronesian languages like to use reduplication or sometimes even
triplication for various purposes (to form the plural, to intensify a
verb, to derive nouns and so on). The following is from Indonesian:

orang ‘man’ orang-orang ‘child’

anak ‘man’ anak-anak ‘children’ (8.8)
mata ‘eye’ mata-mata ‘spy’
An example of triplication is provided in Pingelapese:

kOul ‘to sing’ kOukOul ‘singing’ kOukOukOul ‘still singing’

mejr ‘to sleep’ mejmejr ‘sleeping’ mejmejmejr ‘still sleeping’
(8.9)

Reduplication need not copy the full word. For example, in Latin
some verbs form the perfect in the following way. The first consonant
together with the next vowel is duplicated and inserted:
Pingelapese is a language spoken in
pendit ‘he hangs’ pependit ‘he has hung’ Micronesia, by approximately 3,000
tendit ‘he stretches’ tetendit ‘he has stretched’ people. It belongs to the Micronesian
(8.10) branch of the Austronesian language
currit ‘he runs’ cucurrit ‘he has run’ family. Map from CIA.
spondet ‘he promises’ spopondit ‘he has promised’

Note that the last example shows that /s/ is exempt from reduplica-
tion.

Morpheme or word?
There is no way to predict whether some piece of meaning is expressed
by a morpheme, a separate lexeme or both, and languages differ greatly
 morphology ii: morphemes and morphs 97

in this respect: some languages pack all kinds of meanings into the verb
(Inuit, Mohawk), some keep everything separate (Chinese), and most
languages do something in between. Because of this, morphology is
more important in some languages than in others.
However, even within a single language, there are varying ways to
express something. Adjectives in English, for example, have three
forms: positive (normal form), comparative (form of simple com-
parison) and superlative (form of absolute comparison). Now look
at the way they get formed:

positive comparative superlative

high higher highest
fast faster fastest
vulnerable more vulnerable most vulnerable (8.11)
common more common most common
good better best
bad worse worst

The first set of adjectives take suffixes (zero for the positive, /er/ for
the comparative and /est/ for the superlative). The second set of
adjectives take a separate word, which is added in front (/more/ and
/most/). The third set is irregular. The adjective /good/ changes
the root in the comparative and superlative before adding the suffix
(/bett/ in the comparative and /b/ in the superlative), while /bad/
does not allow an analysis into root and suffix in the comparative and
superlative. We may define a suffix for /worse/,
but it will be a one-time-only suffix,
since there is no other adjective like
Allomorphy /bad/. It therefore makes not much
sense to define a separate comparative
In the above discussions of different inflectional and derivational mor- suffix for /worse/. However, /worst/
is a debatable case.
phemes we have ignored the fact that actually sometimes the same
meaning is expressed by different forms, and sometimes the choice of
these forms is phonologically conditioned, just as allophones are. Let’s
look at an example from the Native American language Cree:

a. [tSi:ma:n] ’canoe ’ l. [nitospwa:kan] ’my pipe ’

b. [nitSi:ma:n] ’my canoe ’ m. [akimew] ’s/he counts’
c. [so:nija] ’money’ n. [nitakimen] ’I count’
d. [niso:nija] ’my money’ o. [apiw] ’s/he sits’
e. [wija:S] ’meat’ p. [nitapin] ’I sit’
f. [niwija:S] ’my meat’ q. [ispelohkew] ’s/he rests’
g. [e:mihkwa:n] ’spoon’ r. [nitispelohken] ’I rest’
h. [nite:mihkwa:n] ’my spoon’ s. [kaakimew] ’s/he will count’
i. [astotin] ’hat’ t. [nikaakimen] ’I will count’
j. [nitastotin] ’my hat’ u. [kaapiw] ’s/he will sit’
k. [ospwa:kan] ’pipe’ v. [nikaapin] ’I will sit’

In the left two columns we see examples of possessive pronouns with

different nouns. In the right columns we see different person forms for
98 an open introduction to linguistics

the present-tense of the verb. If we look at the word for canoe and
compare it to my canoe data we can already see that a prefix has been
added in the possessive form, [ni-]. We can guess that this means my.
We can then check this preliminary hypothesis by looking at the rest of
the nouns and their corresponding possessive forms. Our guess seems
to be true for examples (c.) and (d.), and (e.) and (f.) but when we
get to [e:mihkwa:n], spoon, the first-person possessive form seems to be
[nit-]. The same seems to be true for my hat. Now, just like allophones,
different allomorphs are often phonologically conditioned, so we should
look at the root form it attaches to in order to see if we can predict
when the form [ni-] appears and when [nit-] is used. It is actually quite
simple: it seems that if attaching to a vowel, [nit-] is preferred, which
makes sense from an articulation perspective: pronouncing two vowels
next to each other is hard: you either have to add a glottal stop (e.g.
[?] or you have risk turning it into a dipthong. With this hypothesis,
you can also make a very good guess as to how you would say my pipe
in Cre: [nitospwa:kan].
So now we have identified two allomorphs for the first-person pos-
sessive prefix in Cree. We can then write a rule predicting their use.
Just like with allophones, we should try to write the specific cases,
and then note that the more common form is used elsewhere. IN this
case, given that there are more consonants than vowels, it would be
logical to write a rule saying that we should use the [nit-] prefix before
a vowel, and the form [ni-] everywhere else.
Now let’s consider the verb forms. Here we see there is only one
lexical word for two (or more) English words, so we can guess that
morphological marking, rather than pronouns, are being used to signal
person. The example only shows singular forms, so we don’t have to
worry about number.
Looking at she/he counts and I count, we can see two things. First,
a similar prefix seems to be used for the first person here as in the
possessive forms: before a vowel we see [nit-] in both I count and I
sit. and I rest. We can also notice that all the first person forms end
with an [-n], whereas the third person forms end with a [-w]. Realizing
this makes it possible for us to analyze all the present tense forms
given. However, the last four forms are future tense forms, and here
we see something different. Compare s/he sits, [apiw] with s/he will
sit: [kaapiw], and what we see is the addition of a morpheme ka- before
the root. If we hypothesize that [ka- means will, then we can analyze
these forms as well, except for the prefix [ni+]. But here we can guess
that the same phonologically based allomorphic variation is probably
at work that we saw with the genitive form: Cree doesn’t like vowels
next to each other so if the root begins with a vowel the first person
morpheme is realized as [nit-, but when attaching to a consonant, [ni-
is preferred. Given this I would also guess that the double vowels in the
example are actually lengthened vowels, because they are two vowels
of the same kind (but to know for sure we would need to look this up
in a grammar of Cree.
 morphology ii: morphemes and morphs 99

Study Questions
1. Define the following terms and relate them to each other: mor-
pheme, affix, morph, allomorph, suffix, prefix, circumfix, infix, trans-
fix.

2. If morphemes relate to morphs in the same way phonemes relate to

phones, what, then, is the morphological equivalent for the relation
between phones and sounds?

3. For each of the following derivational affixes of English, give:

• The word class(es) to which they can be applied

• The word class that results after application of the affix
• Two examples of the affix in action

a. -ment
b. -ize
c. -al
d. un-
e. -ance/-ence
f. -ee
g. en-/em-

4. Words that are made up of more than two morphemes possess some
kind of internal structure, i.e. a word is not simply a linear com-
position of its morphemes. For example, the word goalkeepers has
the structure [[[goal][keeper]]s], i.e. goal and keeper are combined
first, after which the plural suffix is added. For the following En-
glish words, give the morphemes of which they consist, the type of
morpheme (root/suffix/prefix and inflectional/derivational, if appli-
cable) and the internal structure of the word (either as a tree or by
using brackets).

a. irreparable
b. replacements
c. premeditated
d. surcharges

5. Consider the following English compounds. Some of them would

typically receive only one of their logically possible interpretations,
while others could be interpreted in more than one way. For all
those with more than one plausible interpretation, give a morpho-
logical structure (either a tree structure or using brackets) for each
plausible interpretation, and explain with a paraphrase what each
means. For those which have only one sensible interpretation, give
the structure for that interpretation and also explain what the im-
plausible interpretations could have meant (if possible).

a. Dutch medical dictionary

100 an open introduction to linguistics

b. Facebook friend request

c. think tank consortium
d. voluntary student worker
This variety of the Agta language is
spoken in the Central Cagayan Val-
6. Consider the words from the Central Cagayan variety of the Agta ley on the Northern island of Luzon,
language and their English translations in the table below. in the Philippines. There are now
only about 600 speakers of this variety
of Agta, although there are perhaps
Agta English Agta English 10,000 people in the Philippines who
speak other varieties also known as
wer creek bag loincloth Agta. The Agta people now speak an
balabahuy little pig walawer little creek Austronesian language similar to other
languages spoken in the Philippines.
talobag beetle balabag little loincloth However, they are descended from the
bakbakat granny takki leg Melanesian people who were present in
the Philippines before the Austrone-
palapirak little money labang patch
sian peoples arrived. The Agta lan-
bahuy pig guage is now seriously endangered.

a. Using the information from these table, translate the following

words and phrases into Central Cagayan Agta:
I. little leg
II. money
III. little beetle
IV. little patch
V. little granny
b. What is the rule for making diminutive forms from nouns in
Agta? Identify the morpheme, and existing allomorphs.
c. Of what kind of morphological process is this an example?

7. Consider the following indefinite and definite forms of Swedish

words:

Swedish English Swedish English

by village byn the village
cykel bike cykeln the bike
grupp group gruppen the group
pilgrim pilgrim pilgrimen the pilgrim
hymn hymn hymn the hymn

a. What kind of morphological process is used to make definite

forms in Swedish?
b. Which morphs of the definite morpheme can you identify?
c. When is each of these allomorphs used? Write a set of (tentative)
rules based on phonological context (assume English pronuncia-
tion).

8. Allomorphs: Egaugnal
Consider the following verbs in present and past tense in Egaugnal
(‘language’ spelled backwards!), a fake language I created.
 morphology ii: morphemes and morphs 101

Egaugnal English Eguagnal English

kuftin run kuftin@ne ran
sekuftin rerun sekuftin@ne reran
Saçmo to arrest Saçmonade arrested
dogin to impeach dogin@ne impeached
taSmili to confide taSmilinade confided
AtinJi to sue AtinJinade sued
ginJi to faint ginjinade fainted
seJellip to reread seJellipade reread
umu to harass umunade harassed
NaN to solicit NaN@ne solicited
EduçOm to succeed EduçOmade will succeeded
atelom to highlight atelomade highlighted
kinteb to heat kintebade reheated
sekinteb to reheat sekintebade reheated

Identify all the allomorphs of Eguagnal that mark past tense.

a. These morphemes are phonologically conditioned. Write a rule

that predicts in which phonological contexts each allomorph would
be used. Make sure you make clear which allomorph is the “else-
where” allomorph.
b. What if Eguagnal borrowed the word whatsapp and made a verb
out of it. What would it’s past tense be?
c. What if Eguagna borrowed the word instagram and made a verb
out of it. What would it’s past tense be?
d. You should also be able to identify a derivational morpheme in
the list above. What is it, and what does it mean?
9
Syntax I: Categories, Constituents, Trees and Context
Free Grammars

Sentences consist of words. These words are arranged into groups

of varying size, called constituents. The structure of constituents
is a tree. In this chapter we’ll learn how to define the notion
of constituent and constituent occurrence solely in terms of sen-
tences.

Occurrences
Sentences are not only sequences of words. There is more structure
than meets the eye. Look for example at:

This villa costs a fortune. (9.1)

Visually we can see that the words are ordered. When spoken the
words will be ordered in time. This ordering is linear. It satisfies the
following word postulates:

À Irreflexivity No word precedes itself.

Á Transitivity If w precedes w0 and w0 precedes w00 then w precedes

w00 .

 Linearity For any two distinct words w and w0 , either w precedes

w0 or w0 precedes w.
Types and Tokens
There is, however, one thing about which we must be very careful. In NLP we often use the term type for
a word and the term token for each oc-
In the sentence: currence of the word in a given string.
Thus in the example there are three
the dog sees the cat eat the mouse (9.2) tokens of the word type ’the’.

we find the word /the/ three times (we do not distinguish between
lower and upper case letters here). However, the definitions above
suggest that /the/ precedes /the/ since it talks of words. Thus, we
must change that and talk of occurrences. The best way to picture
occurrences is by underlining them in the string:

the dog sees the cat eat the mouse

the dog sees the cat eat the mouse (9.3)
the dog sees the cat eat the mouse
104 an open introduction to linguistics

Occurrences of strings can either overlap, or precede each other. They

overlap when they share some occurrences of letters. Otherwise one
precedes the other. The next two occurrences in (9.4) and (9.5) over-
lap, for example, while the occurrences of /the/ above do not.

the dog sees the cat eat the mouse (9.4)

the dog sees the cat eat the mouse (9.5)

Actually we want to be very explicit when talking about occur-

rences. Let’s look at a formal definition:
Definition 3 Let ~x and ~z be strings. An occurrence of ~x in ~z is a
pair h~u, ~v i such that ~z = ~u~x~v . Given an occurrence C = h~u1 , ~v1 i of 
~x1 and an occurrence D = h~u2 , ~v2 i of ~x2 we say that C precedes D
if ~u1 ~x1 is a prefix of ~u2 ; we say that C and D overlap if C does not
precede D and D does not precede C.
Thus, the word /the/ has the following occurrences in (9.2) (with
spaces shown where important):
hε, ␣dog sees the cat eat the mousei
hthe dog sees␣, ␣cat eat the mousei (9.6)
hthe dog sees the cat eat␣, ␣mousei
If you apply the definition carefully you can see that the first occur-
rence precedes the second: ε concatenated with /the/ gives /the/,
which is a prefix of /the dog sees/. Notice the notion of overlap;
here are two occurrences, one of /the dog/ and the other of /dog
sees/. These strings, consisting of the first and second, and the sec-
ond and third occurrences of the words, must overlap. They both share
the same occurrence of the second word.
hε, ␣sees the cat eat the mousei
(9.7)
hthe␣, ␣cat eat the mousei
The word postulates simply have to be reformulated in terms of occur-
rences of words (instead of in terms of words) and then they become
correct. This has to do with the assumption that occurrences of words
(word tokens) cannot overlap. Let ~z be a given sentence, and U the
set of occurrences of words in ~z. Then the following holds.
À Irreflexivity No member of U precedes itself.

Á Transitivity Let C, C 0 and C 00 be in U . If C precedes C 0 and C 0

precedes C 00 then C precedes C 00 .

 Linearity For any two distinct occurrences C and C 0 from U , either

C precedes C 0 or C 0 precedes C.
Talk of occurrences of words is often clumsy, but it is very important
to get the distinction between a word and an occurrence of a word
straight.
A notationally simpler way is the following. We assign to each
occurrence of a word some symbol, say a number. If we use numbers,
we can even take advantage of their intrinsic order. We simply count
the occurrences from left to right.
 syntax i: categories, constituents, trees and context free grammars 105

Constituents
Consider for example /a fortune/. This is a sequence that has very
different properties than a sequence like /costs a/. We can see ev-
idence of their difference by looking at whether or not the sequence
can be replaced by a single word. We can easily replace /a fortune/
with the single word /much/ without affecting grammaticality:

This villa costs much. (9.8)

Likewise, instead of /this villa/ we can say

This costs much. (9.9)

Notice that exchanging words for sequences of words, or vice versa,

does not need to preserve the meaning; all that is required is that it
preserves grammaticality: after a replacement, grammatical English This villa costs a fortune.
Picture JS
sentences should still be grammatical English sentences. For, example
if we replace /costs much/ by /runs/ we are not preserving meaning,
just grammaticality:

This runs. (9.10)

Notice that any of the replacements can also be undone:

This villa runs. (9.11)

This costs a fortune. (9.12)

We call a sequence of words a constituent if (among other condi-

tions) it can be replaced by a single word. Further, checking if a given
sequence of words is a constituent by substituting the sequence with a
single word is called the Substitution Test.
A second condition on constituents is that they can be coordi-
nated. For example, we can replace /a fortune/ not only by /a
lot/ but also by /a fortune and a lot/. The latter construction
is called coordinated because it involves the word /and/. Checking
if a word sequence can be coordinated is called the Coordination
Test (a more precise definition will follow). Using both of these meth-
ods we can then conclude that the sentence (9.1) has the following
constituents:

{this villa costs a fortune,

this villa, costs a fortune,
(9.13)
this, villa, costs, a fortune,
a, fortune}

The visual arrangement is supposed to indicate order. However, this

way of indicating structure is not precise enough. Notice first that
a word or sequence of words can have several occurrences, and we
need to distinguish them, since some occurrences may be constituent
occurrences, while others are not.
106 an open introduction to linguistics

Abstractly, we can give each occurrence of a word a distinct number,

like this:
this villa costs a fortune
(9.14)
1 2 3 4 5
Each occurrence gets its own number. A sequence of occurrences can
now conveniently be represented as a set of numbers. The constituents
can be named as follows:

{{1, 2, 3, 4, 5}, {1, 2}, {1}, {2}, {3, 4, 5}, {3}, {4, 5}, {4}, {5}} (9.15)

Let w1 w2 w3 . . . wn be a sequence of words constituting a sentence of

English. Then a constituent of that sentence has the form wi wi+1 wi+2 . . . wj ,
for example w2 w3 , w5 w6 w7 but not w2 w4 w7 . It always involves a con-
tinuous stretch of words.
Continuity of Constituents
Constituents are continuous parts of the sentence.
Non-Crossing
Given two constituents that share a word, one must be completely
inside the other.
Words are Constituents
Every occurrence of a word forms its own constituent.
Here is a useful terminology. A constituent C is an immediate con-
stituent of another constituent D if C is properly contained in D, but
there is no constituent D0 such that C is properly contained in D0 and
D0 is properly contained in D. Our sentence (9.1) has only two imme-
diate subconstituents: /this villa/ and /costs a fortune/. The
latter has the immediate constituents /costs/ and /a fortune/. It is
not hard to see that it is enough to establish for each constituent its
immediate subconstituents. Notice also that we can extend the notion
of precedence to constituents. A constituent C precedes a constituent
D if all words of C precede all words of D. So, /this villa/ precedes
/a fortune/ because /this/ precedes both /a/ and /fortune/ and
/villa/ precedes both /a/ and /fortune/.
This coincidence opens the way to a few alternative representations.
One is by enclosing constituents in brackets:

[[[this] [villa]] [[[costs] [[a] [fortune]]]]] (9.16)

Typically, the brackets around single words are omitted, though. This
gives the slightly more legible

[[this villa] [[costs [a fortune]]]] (9.17)

Another way would be to use trees, with each constituent repre-

sented by a node, and drawing lines as given in Figure 9.1. Each node
is connected by a line to its immediate subconstituents. These lines go
down; a line going up is consequently from a constituent to the con-
stituent that it is an immediate part of. So, 8 and 9 are the immediate
constituents of 7, 6 and 7 are the immediate subconstituents of 3, and
so on. It follows that 6, 7 and 8 are subconstituents of 3, which is to
say that 3 consists of /costs/, /a/ and /fortune/. Read x < y as “x is dominated by y”
 syntax i: categories, constituents, trees and context free grammars 107

1 Figure 9.1: An Unlabelled Syntactic

• Tree

2 • • 3

4 • • 5 6 • • 7

this villa costs 8 • • 9

a fortune

Definition 4 A tree is a pair hT , <i where T is a set and < is a set

of pairs of elements from T such that (a) if x < y and x < z then
either y < z, or y = z or z < y and (b) there is an element r such
that x < r for all x 6= r (r is called the root).

The tree in Figure 9.1 consists of the nodes 1, 2, 3, . . . , 9 (we ignore

the stuff below them; this concerns word forms and is of no interest to
syntax; we add these things for orientation only). The relation < is as
follows: 2 < 1, 3 < 1, 4 < 1, 5 < 1, 6 < 1, 7 < 1, 8 < 1, 9 < 1; 4 < 2,
5 < 2; 6 < 3, 7 < 3, 8 < 3, 9 < 3; 8 < 7, 9 < 7; no other relations hold.
Notice also that the order in which elements follow each other is
also reflected in the tree, simply by placing the earlier constituents to
the left of the later ones.
Standard Constituency Tests
There are a number of well known ways to identify constituents,
and the text above introduced two: Substitution and Coordina-
tion. Here’s an overview:

The Substitution Test

The most well known constituency test is the substitution test.
If you can replace a string by a single word then you have shown
that this string can function syntactically as a single unit. For
example:

• Luke went [to Hoth]. –> Luke went there.

• [The old man] left. –> Obi left.

• Han [shot first]. –> Han did.

The Coordination Test

For the Coordination Test test you combine the string you are
testing with another similar string coordinated with and and see
if the coordinated string still behaves like the original string, e.g.

• Ed was [in the bar].

• Ed was [in the bar and in the park].

108 an open introduction to linguistics

But the coordination test does not always work, e.g.:

• Tim [bought some] fruit.

• Kim [bought more] fruit.

• *Tim bought some and bought more fruit.

In general the coordination test is actually not as easy to apply

as the other tests.
The It-cleft Test (Clefting)
A constituent should be able to be clefted, which means it can be
put into a special focus construction with It was X that. . . that
exists in English. So from the sentences Dogs shouldn’t like eating
cheese we can identify a number of constituents, e.g. :

• It’s [eating cheese] that dogs shouldn’t like.

• It’s [dogs] that shouldn’t like eating cheese.

However, this type of focusing sometimes sounds unnatural, which

might make it hard to judge if a word sequences passes or fails
the test.
Answer Ellipsis Test
A constituent should be able to function as a short answer to a
question. For example, given the sentences Luke wanted to travel
to Alderaan we can identify a number of constituents:

• A: What did Luke want to do? B: Travel to Alderaan.

• A: Where did Luke want to go? B: To Alderaan.

The Passivization Test

The Passivization test is a variation on a number of different tests
based on movement (Topicalization is another one). If you can
passivize the string it’s probably a constituent.

• The Sand People took [the robot]. –> The robot was taken
(by the Sand People).

• Ed reported [that the stormtroopers were coming]. –> That

the stormtroopers were coming was reported (by Ed).

Beware!
All these tests have to be used with caution. Not all constituents
pass all tests. In general you should always try several tests to
confirm that what you think is a constituent, actually is.

Categories

A context is a pair h~x, ~y i of strings. We denote them in a more visual

way as follows:

~x ~y (9.18)
 syntax i: categories, constituents, trees and context free grammars 109

An example of a context is

this a fortune (9.19)

Consider the following sentence:
(Notice the blanks left and right of the gap!) Given a sentence, every
Bob ate [too many apples] yes-
substring is uniquely identified by its context: the part that precedes terday.
it and the part the follows it. The missing part in (9.19) is /villa
costs/. So, every substring can be denoted by its left context and its Let’s see with three tests that this is a
constituent:
right context. Substituting a substring by another is extracting the
1. Clefting It was too many apples
context and then putting back the replacing string. Replacing /villa that Bob ate yesterday (not too
costs/ by /misses/ gives many pears).
2. Answer ellipsis What did Bob
this misses a fortune (9.20) eat yesterday? Too many apples.
3. Passivization Too many apples
Not a good sentence, but grammatical. (Does this show, by the way were eaten yesterday (by Bob).

that /villa costs/ is a constituent ...?) And the same three tests fail with:
Bob ate too many [apples yesterday].
We agree to call arbitrary strings of English words constituents
• Clefting *It was apples yesterday
just in case they occur as constituents in some sentence. The wording that Bob ate too many.
here is important: I do not say that they have to occur everywhere • Answer ellipsis *What did Bob
as constituent. It may happen that a sequence occurs in one sentence eat too many of? Apples yesterday.

as a constituent, but not in another. Here is an example. The string • Passivization *Apples yesterday
were eaten too many (by Bob).
/you called/ is a constituent in (9.21) but not in (9.22):

This is the man you called. (9.21)

This is a song for you called "Sweet Georgia Brown".
(9.22)

To know why this is to try to substitute /you called/ by /you

called and Mary liked/. So, first of all we need to talk about se-
quences that occur as constituents. If they do, we call the occurrence
a constituent occurrence.
We are now ready to make matters a bit more precise. A string is a
sequence of letters (or sounds). A string language is a set of strings.
The set of sentences of English is an example. Call this set E. We
say that a string ~x is a (grammatical) sentence of English just in Picture JS
Context dependent constituency
case ~x is a member of E. So, /We stay at home./ is a (grammatical) Here’s another example of how con-
sentence of English while /Cars a sell if./ is not. Based on this stituency is context dependent.:
set we define constituents and constituent occurrences and all that. • I saw the fox.

First we give some criteria for constituent occurrences. ~x has a • I saw the fox that wore a hat.

constituent occurrence in ~z only if there are ~u and ~v such that: The string the fox is a constituent in
the first sentence, but not in the sec-
ond. First note that in the second sen-
À ~z = ~u␣~x␣~v ; tence the string the fox cannot be re-
placed with a pronoun, e.g.
Á ~z ∈ E (that is, ~z is a grammatical sentence);
• *I saw it that wore a hat.

 either ~u␣~v ∈ E or there is a single word w ~ v ∈ E;

~ such that ~u␣w␣~ Instead it’s seems that the entire
string the fox that wore a hat is a con-
stituent. IT can be replaced with a
à ~u␣~x␣and␣~x␣~v ∈ E. pronoun, e.g. I saw it., and it seems
to allow answer ellipsis and passiviza-
The reason why in à we use ~x twice rather than using a different tion, e.g.:
conjunct is because in a formal test we want to be sure that the string • Q: What did you see? A: The fox
is of the appropriate type: only constituents of the same type can that wore a hat.
• The fox that wore a hat was seen.
• *The fox was seen that wore a hat.
110 an open introduction to linguistics

be conjoined. Of course, it is pretty unnatural to conjoin the same

constituent with itself, e.g. ?? This is the man and the man. But for
a formal definition it is better to be correct than natural.
In case À – Ã are satisfied we say that h~u, ~v i is a constituent
occurrence of ~x (in ~z). This list is incomplete. In other words, there
are more criteria that need to be satisfied before we say that ~x has a
constituent occurrence. But this list of requirements suffices for now. In (9.21), ~
u is This is the man, ~x is
Let us apply this to (9.21) and (9.22). With ~x = you called we you called and ~v is the empty context
to the right of the sentence.
have in (9.21) the occurrence

hThis␣is␣the␣man␣, ␣.i (9.23)

À and Á are obviously met. That  and à are met is witnessed by

the following.

This is the man. (9.24)

This is the man you called and you called. (9.25)

We try the same on (9.22). The following sentence is ungrammatical,

so à fails.
∗
This is a song for you called and you called (9.26)
"Sweet Georgia Brown".

Definition 5 A category (of a given language) is a set ∆ of strings

such that any constituent occurrence of a member of ∆ can be replaced
by any other member of ∆ yielding once again a constituent occurrence.

Thus, given that ~x and ~y are members of the same category, if ~u~x~v is
a grammatical sentence of English, so is ~u~y~v and vice versa.
The definition of a category is actually similar to that of a phoneme;
phonemes were defined to be substitution classes that preserve mean-
ing. In syntax we define substitution classed based on the ability to
preserve grammaticality (more or less, since we have the caveat about
‘constituent occurrences’ because we are dealing with the substitution
of strings for strings, not just of an item for another item). We give
an example. The intransitive verbs in the 3rd person singular form a
category:

{falls, runs, talks, . . . } (9.27)

By this definition, however, /talks/ and /talk/ are not members of

the same category, for in (9.28) we cannot replace /talk/ by /talks/.
The result would simply be ungrammatical.

Mary and Paul talk. (9.28)

It seems that the number of categories of any given language must

be enormous if not infinite. It is certainly true that the number of
categories is large, but it has a highly regular structure which we shall
unravel in part for English.
 syntax i: categories, constituents, trees and context free grammars 111

Now that we have defined the constituents, let us go back to our

tree in Figure 9.1. Each of the nodes in that tree is a constituent,
hence the sequence belongs to some category. Omitting some detail,
the categories are given in Figure 9.2. (We call S a sentence, D a
determiner, DP a determiner phrase, NP a noun phrase, VP a verb-
phrase, V verb.)

S
Figure 9.2: A Labelled Syntactic Tree
•

DP• • VP

D• • NP •V •DP

this villa costs D • • NP

a fortune

Context Free Grammars

Now look at Figure 9.2. If the labelling is accurate, the following
should follow: any sequence of a determiner phrase followed by a verb
phrase is a sentence. Why is this so? Look at the tree. If DP is the
category to which /this villa/ belongs, and if VP is the category to
which /costs a fortune/ belongs, then we are entitled to substitute
any DP for /this villa/, and any VP for /costs a fortune/:
  

 this villa 
  costs a fortune.
 


 a car 
  walks. 


 it 
  catches the bus.
 


 
  

tomorrow’s sunshine sings praise to the Lord.
(9.29)

None of the sentences you get are ungrammatical, so this actually

seems to work. We state the fact that a DP followed by a VP forms a
sentence in the following way:

S → DP VP (9.30)

We have seen earlier statements of the form ‘something on the left’

→ ‘something on the right’, with a slash after which some conditions
on the context were added. Conditions on the context are absent
from (9.30); this is why the rule is called context free. It can be
used no matter what the context is. However, notice one difference,
namely that the thing on the left is always a single symbol, denoting a
category, while the thing on the right is a sequence of symbols, which
may each be either a category or a word. A grammar is a set of such
rules, together with a special symbol, called start symbol (usually
112 an open introduction to linguistics

S). Consider by way of the example the following grammar. The start
symbol is S, the rules are

S → DP VP (9.31a)
DP → D NP (9.31b)
D → a | this (9.31c)
NP → villa | fortune (9.31d)
VP → V DP (9.31e)
V → costs (9.31f)

Here, the vertical stroke ‘|’ is a disjunction. It means ‘or’. For example,
the notation D → a | this is a shorthand for two rules: D → a and
D → this.
This grammar says that (9.1) is of category S. How does it do that?
It says that /this/ is a determiner (D; Rule (9.31c)), and /villa/ is
a noun phrase (NP; Rule (9.31d)). By Rule (9.31b) we know that the
two together are a DP. Similarly, it tells us that /a fortune/ is a DP,
and that /costs a fortune/ is a VP. Finally, using Rule (9.31a) we
get that the whole is an S.
Given a set of rules R, we define a derivation as follows. An R-
derivation is a sequence of strings such that each line is obtained
from the previous by doing one replacement according to the rules of
R. If the first line of the derivation is the symbol X and the last the
string ~x we say that the string ~x has category X in R. The following
is a derivation according to the set (9.31a) – (9.31f).

S (9.32)
DP VP
D NP VP
D NP V DP
D NP V D NP
this NP V D NP
this villa V D NP
this villa costs D NP
this villa costs a NP
this villa costs a fortune

Thus, in this rule system, /this villa costs a fortune/ is of cate-

gory S. Likewise, one can show that /this fortune/ is of category DP,
and so on. We can call this sequence of replacements a replacement
derivation.
A context free grammar consists of (a) an alphabet A of terminal
symbols, (b) an alphabet N of nonterminal symbols, (c) a nonterminal
S ∈ N , and (d) a set of context free rules X → ~x, where X ∈ N . Notice
that the strings DP, VP, V and so on are considered to be symbols;
but they are not symbols of English. They are therefore nonterminal
symbols. Only nonterminal symbols may occur on the left of a rule.
But nonterminal symbols may occur on the right, too.
 syntax i: categories, constituents, trees and context free grammars 113

The role of the start symbol is the following. A string of termi-

nal symbols is a sentence of the grammar if there is a derivation of
it beginning with the start symbol. The language generated by the
grammar is the set of all sentences (using only terminal symbols!) that
it generates. Ideally, a grammar for English should generate exactly
those sentences that are proper English. Those grammars are incredi-
bly hard to write.
In fact, the goal of syntax is to write a grammar (set of rules) that
generates all the grammatical sentences and doesn’t generate any un-
grammatical sentences. This is why we talk of generative grammar,
though that phrase has come to refer to work in the Chomskian tra-
dition.

Notes
Let’s return to our definition of constituent occurrence. Context free
grammars provide strings with a constituent analysis. However, not
every such analysis conforms to the definition of constituents. To make
everything fall into place we require that the grammars of English have
the following universal rule schema. For every category symbol X they
contain a rule

X → X␣and␣X (9.33)

Furthermore, for every X they must have a rule

X→w
~ (9.34)

for some word. If that is so, the criteria À – Ã will be met by all
constituents assigned by the grammar.

Study Questions
1. Show with at least two constituency tests that the bracketed words
in the following examples are constituents, if possible. If they are
not constituents, show how they fail at least two tests.

a. C3P0 kicked the ball [to the boy].

b. Everything was [destroyed by] the deathstar.
c. [The moons] shone in the empty sky.
d. Leia didn’t know [that Luke was her brother].
e. Too many [fighter pilots] were poorly trained.
f. Han [liked to] act cool.

2. Identify all the constituents in the following sentences:

a. This horse likes sugar.

b. The moons hang in the sky.

3. Create a labelled syntactic tree for the following three sentences.

Make sure you get the constituents right. Hint: Structurally the
sentences are completely parallel with the one in Figure 9.2. Make
sure you can see that.
114 an open introduction to linguistics

a. This spaceship costs a fortune.

b. The stormtrooper shot the intruder.
c. A jedi uses a lightsaber.

4. Create rules for a Context Free Grammar that can generate the
following set of sentences.

a. the dog barks

b. the cats meow
c. the bird sang
d. the bunnies jumped
e. the dog chases the cat

Your CFG should produce the following sentences as well:

a. *the bunnies meow

b. *the cats chases the dog
c. *the dog sang the cat.

But actually, you do not want your CFG to produce these sentences.
How could you prevent this? Explain your answer by showing ex-
amples of how your CFG-rules would have to be modified.
10
Syntax II: Argument Structure

This chapter introduces a general schema for all phrase structures

in every language: X-bar syntax. In X-bar syntax, every word
projects a phrase consisting of up to two arguments, and any
number of adjuncts. Finally we’ll introduce movement to give a
uniform analysis to all sentences.
Identifying word classes
Word classes, also called parts-of-
speech (POS), can be identified with
Lexical Categories three different types of criteria:
1. Semantic content
In this chapter we will delve deeply into syntax, and show the begin-
2. Morphological features
nings of a formal system for analyzing syntactic structures. We shall
3. Distributional properties
put to use our notation of attribute value structures (AVSs), but we’ll
You may have learned that a noun is
also look at tree structures to more easily see hierarchical relationships. a ‘person, place, or thing’ or that a
We start off with a few general purpose rules and then refine them as conjunction (and, or) joins two things.
we go along. First, words can be classified into a handful of categories, These are semantic criteria. Mor-
phological criteria have to do with
the so-called lexical or major categories. The ones we shall be using what inflectional and derivational af-
are: noun (N), verb (V), adjective (A), adverb (Adv), preposition fixes can be added to a given word.
For example, many nouns in English
(P), complementizer (C), determiner (D), and tense (T). Not all can be pluralized by adding -s, e.g.
classes have single words in them, but most of them do: the plural of goat is goats. There are
exceptions however, e.g. child/chil-
N car, house, storm, insight dren. By looking at what other words
a given word can combine with, we
V run, eat, hasten, crawl can also find evidence that they be-
A greedy, raw, shiny, cheerful long to a particular word class (POS-
class). For example, nouns can usually
Adv very, steadily, allegedly, down
(10.1) be prefixed with a determiner like this
P in, for, about, below or that, e.g. this house, that cat. A
C that, which, because, while verb can occur with an auxiliary verb,
such as can or might, e.g. can speak,
D a, the, this, those might come. For each part of speech it
T will, can, could, should is usually possible to identify criteria
related to these three types.
Our first attribute is cat, and it has the values just displayed (so
far, they are N, V, A, Adv, P, C, D, T). A corpus is a collection of texts used
for linguistic study. Often corpora are
carefully selected according to some
Subject and Object criteria and then preprocessed so that
they all have a consistent form to make
The next distinction we want to make is that between a word and them more useful for certain tasks.
a phrase. We have made that distinction earlier, when we called
certain words determiners and certain constituents DPs (= determiner
phrases). The distinction is intimately connected with the notion of
argument structure. The argument structure tells us for each word
116 an open introduction to linguistics

what kinds of constituents it combines with to form a phrase. Take

the verbs /run/ and /accuse/. The difference between the two is that
the first is happy to combine with just one DP, say /the sailors/ to
form a sentence, while the latter is not:
The sailors run. (10.2)
∗
The sailors accuse. (10.3)
To make the latter into a grammatical sentence, you need to supply
two more DPs, one for the one who is accused and one for what he is
accused of:

The sailors accuse the captain of treason. (10.4)

We say that /run/ takes one argument and the verb /accuse/
takes three arguments. It means that in order to get a grammatical
sentence, /run/ only needs to combine with one DP, /accuse/ needs
three. (There are differences in the degree to which /accuse/ needs
any one of its arguments; this is a matter we shall not go into.) Riau, Indonesia (from Wikipedia)
Verbs always have at least one argument in a sentence: the subject. Verb-centric view of syntax
The view of syntax we are advocating
The subject argument always occurs in the nominative case. You can for in this chapter views other argu-
recognize it by the form of pronoun that you have to use. The pronouns ments as originating in the verb. Both
the subject and any objects are pro-
/she/, /he/, /it/, /I/, /they/ are subject pronouns and they are
jected from the verb up to the phrase
in nominative case, while the object pronouns /her/, /him/, /me/, (and sentence) level. This view also
/us/ and /them/ all are in accusative case. (The form /you/ is both holds that the types of arguments that
a verb takes are coded in the mental
nominative and accusative, so we cannot use it to show that subjects lexicon.
are always in nominative case.) Whenever a pronominal argument has
to be in accusative case, then it must be an (direct) object.

They run./∗ Them run. (10.5)

They accuse the captain of treason. (10.6)
∗
Them accuse the captain of treason. (10.7)
The sailors accuse us of treason. (10.8)
∗
The sailors accuse we of treason. (10.9)
Unless adverbs are present, English puts the subject right before the
verb, and the object right after it. So, this actually is another diag-
nostic for subject and object. The third argument is /of treason/.
This is formed by using a preposition (/of/), so it is called a PP (=
preposition phrase). There are as many preposition phrases as there
are prepositions. We shall deal with them a little later.
A verb is called intransitive if it has no direct object. Otherwise
it is called transitive. Hence, /run/ is intransitive, /accuse/ is tran-
sitive. Verbs may also be both; /break/ and /eat/ can occur both
with and without a direct object. Such a verb is both transitive and
intransitive; we say that it is used transitively if there is a direct object
(as in (10.10)), and intransitively otherwise (as in (10.11)).
The children are eating the cake. (10.10)
The children are eating. (10.11)
 syntax ii: argument structure 117

The distinction between transitive and intransitive tells us whether or

not the verb needs a direct object as an argument.
We observe that intransitive verbs are distributionally equivalent to
the combination of transitive verb+direct object:

 
   ran. 
 The child
   
 lost the battle of Waterloo. 

Napoleon

 My neighbour’s dog 
 ate the cake. 


 

is beautiful.
(10.12)

This is important because it tells us two things: transitive verbs plus

(direct) objects form a constituent, and this constituent can be re-
placed by an intransitive verb. The Substitution Rule then tells us
that the verb in an intransitive verb phrase and transitive verb phrase
must be of the same constituent type. They are both VPs.

Projection

How does syntax know that a verb like ran doesn’t take an object
argument and uses a CGF rule like VP –> V without an object, but
that a verb like greet must have an object argument? Where does this
information about the structures associated with lexical items come
from? Modern syntax believes that the structures associated with
words are projected from the words themselves, with the specifics of
what type of projections each word allows coded in the mental lexicon.
Thus:

• A word in the lexicon of category X projects phrases of category

XP

• A phrase of category XP is a projection of a word from the lexicon

of category X

This means that a verb phrase like fight the empire gets its structure
from the projection of the verb fight.
VP

V DP

fight D NP

the N

empire

In the mental lexicon, the word and the structure are stored, with
empty nodes for the other arguments. We know what type the other
arguments are, because we know that fight takes an object argument:
118 an open introduction to linguistics

VP

V DP

fight D NP

In the trees we’ve drawn above, we’ve only show binary branching.
In our context-free grammar in the previous chapter we also only have
rules that would lead to binary branching because we only have rules
where a non-terminal LHS can be rewritten into two symbols on the
RHS. But we could have rules that lead to three branches, and it might
seem that this would make sense given sentence structures having a
subject, a verb, and an object, see below.
Flat S

DP V DP

Luke fought the empire

But we could also simply restrict trees to having binary branching:

Binary Branching S

S One of the reasons to prefer binary branch-

DP VP

Luke V DP

fought the empire

ing is because it better reflects constituency structures. For example,

a verb and its object is a constituent because it passes a number of
constituency tests:

• Substitution (Who fought the empire?) Luke did (=fought the

empire).

• Answer Ellipsis What did Luke do? Fought the empire.

If we choose the flat structure, we have no node in our graph that

corresponds to the verb-object constituent. We can make a similar
argument with DPs. Consider the following two structures for this red
lightsaber.
 syntax ii: argument structure 119

Flat DP

DP

D ADJ N

this red lightsaber

Binary Branching DP Note that the binary branching ex-
amples is a simplified example that
DP doesn’t show full X-bar structure,
which has yet to be introduced.

D NP

this ADJ N

red lightsaber

We should prefer the binary structure, because we know that red

lightsaber can function as a constituent, e.g. Vader likes this red
lightsaber and Kylo likes that one.

Verb-centric syntax
Verbs can restrict the types of sub-
But actually the binary branching structure we have above is not quite jects they can occur with. For ex-
right. We know that an intransitive verb does take one argument: a ample, only animate things can eat
or revise things, so *The lamp ate
subject argument. Further, a transitive verb takes two arguments, a and *The daisy revised the speech are
subject and an object. These are arguments of the verb, but in the weird. This is further evidence that
features of the subject argument are
binary branching examples above, we’ve somehow let the sentence itself controlled by the verb.
generate the subject argument. But actually, the verb, which describes
the event that the sentence communicates, also constraints the types
of subjects it can occur with so it makes sense to treat everything as
projected from the verb. In essence, that means we think of sentences
as a kind of verb phrase. And because every phrase should be a project
of a lexical category, it fits with our goal to get a uniform analysis that
we get rid of S as a start symbol (because there is no lexical category
S) and instead make VPs our start symbol.
VP

DP VP

Luke V DP

fought the empire

The projection of a transitive verb should actually show two argu-

ments. Note however, that if we assume this, a full sentence is then
120 an open introduction to linguistics

the same as a verb phrase. In a Context-Free Grammar (CFG) we

would then have a VP as our start symbol. So now the verb projects
a DP object and a VP, and the higher VP that’s the sentence and a
DP subject. The only problem we have now is that we have two kinds
of VPs: the higher VP that covers the entire sentence and a lower VP
that includes the verb and the object. We need to distinguish between
these two VPs. We cannot solve this by adding feature, because the
features of a phrase are projected from its head, and the head is the
verb, it’s hard to see how we can do this because clearly the VPs are
related, and are projected from the same verb.
It looks like we need an intermediate category, between a VP and
the Verb.
Linguists have therefore proposed the following scheme. Constituents
have two attributes: a major category (CAT) and a projection level
(PROJ). The levels are 0, 1, and 2. They are also called bar levels
(because they used to be denoted by overstrike bars). Another nota-
tion is: D0 for determiner, 0 level; D1 , D or D0 for determiner first or
intermediate projection; and D2 , D or D00 for determiner level 2, or
determiner phrase.
Words are uniformly assigned level 0, and phrases are level 2, and
the intermediate level will be level 1.

Level Contains ? Projection? Example

0 Word Minimal V (‘verb’)
1 Intermediate Bar V (‘V-bar’)
2 Phrase Maximal VP (‘VP’)

So what we used to call a determiner (D) is now a D0 , and the

representation of it and a determiner phrase (DP) is like this:
" # " #
cat :D cat :D
D0 = DP = (10.13)
proj:0 proj:2
Remember: the subject will be a DP
The need for an intermediate level will become clearer below. Now to the left of a verb in English, and
an intransitive verb has no DP object,
we can write new CFG rules. The following rules are proposed for while a transitive verb will have one
English, with attribute value structure rules where the projection level (and a ditransitive verb will have two).
We won’t really do anything with di-
is treated as an attribute with a value, and more traditional CFG transitives here.
notation to the right of the same rules.):
" # " #" #
cat :V cat :D cat :V
→ VP → DP V0
proj:2 proj:2 proj:1
(10.14)
" # " #
cat :V cat :V
→ V0 → V0 (10.15)
proj:1 proj:0
" # " #" #
cat :V cat :V cat :D
→ V0 → V DP
proj:1 proj:0 proj:2
(10.16)
Substitution: Note that these are
The start symbol is momentarily VP. Now, the rule (10.14) says substitution rules. Substitution is one
of the three major syntactic processes.
that a sentence (= VP) is something that begins with a determiner
 syntax ii: argument structure 121

phrase and then has a level 1 V, a level one verbal projection. (10.15)
is for intransitive verbs, (10.16) is for transitive verbs. So actually we
need to add an attribute to make sure that only transitive verbs get
objects. To do this we’ll add an attribute trs with values + and −. TRS = Transitivity, TRS : + = tran-
sitive while TRS : − = intransitive.
 
" # cat :V
cat :V
→  proj:0  (10.17)
 
proj:1
trs :−
 
" # cat :V " #
cat :V  cat :D
→  proj:0  (10.18)

proj:1 proj:2
trs :+
 
cat :V
 → sit | walk | talk | · · · (10.19)
 
 proj:0
trs :−
 
cat :V
 → take | see | eat | · · · (10.20)
 
 proj:0
trs :+

Notice that we did not specify transitivity for V1 . This is because we

have two V1 rules, one that requires a transitive verb and one that
requires and intransitive verb.
It is therefore unnecessary to specify in the above rules transitiv-
ity for V1 s because the RHS of the rules are only applicable to the
appropriate type of verb.
Some verbs can take sentential complements as arguments. The
category C, like /that/, projects CPs, complementizer phrases.

John believes that the earth is flat. (10.21)

John knows that two plus two is four. (10.22)

These sentences are also referred to as a kind of ‘object’, a sentential

object. They can be replaced by /this/ or /that/, showing that they
are either subjects or objects. If they follow the verb, they must be
objects. It’s also possible to have a sentential subject:

That the earth is flat is believed by John. (10.23)

That the cat was eating his food annoyed the child.
(10.24)

This means that at some point we will need to add rules to allow
certain verbs to take sentential complements which will be CPs.

Oblique Arguments and Adjuncts

Examples of oblique arguments
Verbs can have other arguments besides subjects and objects. These The oblique arguments are marked in
arguments are often marked by a preposition, are often optional, and brackets. Observe that leaving out
the oblique argument still results in a
are called oblique arguments. Consider the verb /accuse/ again. In grammatical sentence.
addition to a direct object it can also take a PP (prepositional phrase) • John replied [to the letter].
expressing subject matter. This PP must begin with the preposition • Ed left [with Jo].
/of/. There are verbs that require a PP with /on/ (for example • [To whom] did you speak?
• I lost my phone [in the train].
122 an open introduction to linguistics

/count/), others a PP with /about/ (for example /think/), and so

on. Let us add a new attribute prep whose value can be any of the
prepositions of English. Then /accuse/ will get the following syntactic
category:
 
cat :V
 proj :0 
(10.25)
 
 trs :+ 
 

prep:of

This tells us that /accuse/ wants a subject (because all verbs do),
a direct object (because it is transitive) and a PP opened by /of/
(because this is how we defined the meaning of prep:of). To get the
positioning of the phrases right, we propose to add the following rules:
    
cat :V cat :V cat :P
 proj :1  →  proj :1   proj :2  (10.26)
    
prep:of prep:of prep:of
   
cat :P cat :P
 proj :2  →  proj :1  (10.27)
   
prep:of prep:of
   
cat :P cat :P " #
 cat :D
 proj :1  →  proj :0  (10.28)
  
proj:2
prep:of prep:of
 
cat :P
 proj :0  → of (10.29)
 
prep:of

This works similarly for all prepositions. Notice that we have used the
feature [prep: of] also for the preposition itself to make sure that the
right preposition is used.
The rules say that the PP is to be found to right of the direct object,
if there is one. This is generally the case:

They found the gold in the river. (10.30)

∗
They found in the river the gold. (10.31)
The pilot flew the airplane to Alaska. (10.32)
∗
The pilot flew to Alaska the airplane. (10.33)
They found the gold in the river.
Indeed, we can observe that the above rules hold for any arbitrary
prepositional phrases, so we simply replace the preposition /of/ by a
placeholder, say α. The rules then look like this. We show the result
of this replacement for Rule (10.26) only, as rule (10.34).
    
cat :V cat :V cat :P
 proj :1  →  proj :1   proj :2  (10.34)
    
prep:α prep:α prep:α

The idea is that in this rule schema, α may be instantiated to any

appropriate value. In this case the appropriate values are the English
 syntax ii: argument structure 123

prepositions. Once we choose α, every occurrence of α must be re-

placed by the same value. For example, the following is not a correct
instance of (10.34):
    
cat :V cat :V cat :P
 proj :1  →  proj :1   proj :2 (10.35)
    

prep:about prep:on prep:about

Another important feature to notice is that the PP does not pass

directly to P0 +DP, there is an intermediate level 1 projection as shown
by rule (179). The reason is not apparent from the data given so far,
and some syntacticians dispute whether things are this way. However,
for our purposes it makes the syntax more homogeneous. It’s also
important to note that adding a PP does not decrease the level of the
result. A transitive verb plus PP has the same level as the verb itself,
namely Level 1. This is actually a defining characteristic of adjuncts.
Adjuncts do not change the level of the projection. PPs are adjuncts.
Note that arguments do: a DP that combines with a verb (V) creates
a V-bar constituent, and a V-bar that combines with a DP creates a
VP.

Similar structures for all phrases

You may have noticed that the structure of PPs looks similar to that
of transitive verb phrases, except that the subject is generally missing.
Amazingly, this similarity exists across categories. Consider NPs. An
individual noun can function as an NP, but nouns themselves can also
take optional arguments. Consider the following: Adjunction This is one of the three
major types of syntactic processes.
Adjunction is used to add information.
In Adjunction, the unit that is added
the counting of the horses (10.36) to keeps the same function (or cate-
gory) that it had. Adjuncts are usu-
some president of the republic (10.37) ally Prepositional Phrases, Adverbial
a trip to the Philippines (10.38) Phrases such as yesterday, and subor-
dinate clauses.
this talk about the recent events (10.39)

However, in English, nouns cannot take direct arguments. The verb

/count/ is transitive (/count the horses/), but the gerund form
(counting, a noun) takes an object only by combing with a PP with
/of/.
We can account for this in the same way as we did for verbs. We
allow nouns to additionally have a feature [prep : of], or [prep : in],
and so on. And if they do, they can (but need not) combine with a
PP with the corresponding preposition.
    
cat :N cat :N cat :P
 proj :1  →  proj :0   proj :2  (10.40)
    
prep:of prep:of prep:of
   
cat :N cat :N
 proj :1  →  proj :0  (10.41)
   
prep:of prep:of
124 an open introduction to linguistics

These two rules are abbreviated as follows.

     
cat :N cat :N cat :P
 proj :1  →  proj :0   proj :2  (10.42)
     
prep:of prep:of prep:of

The brackets around an item say that it is optional. For completeness,

let us note that we have the following rules.
 
cat :N
 proj :0  → counting | brother | election | chief | · · ·
 
prep:of
(10.43)
 
cat :N
 proj :0  → doubt | question | talk | rumour | · · ·
 
prep:about
(10.44)

X-bar syntax
Let’s step back and consider for a moment what our ideal syntactic
theory would be. We can probably agree that a theory with general
rules will be preferred to a theory with many exceptions. This is in part
because a simpler theory is preferred over a complex theory (Occam’s
razor). But stronger yet, linguists believe that a natural theory will
give a uniform analysis to structures based a few, general principles.
After much research, a general structure has been proposed for syn-
tactic phrases. It’s called X-bar syntax, where “X” stands for a word
class from which structure will project. X-bar syntax claims that all
phrases have the same structure. They have a core that is of a certain
word class in the language. This core, X, is called the head of the
phrase. This head can be modified by a phrasal structure to its left,
called the specifier position. Take a simple sentence for example,
like Luke fought the empire. We said above that the start symbol
for our grammar treats a sentence as a VP. In this VP then, the head
must be the verb itself, fought. The specifier is then the subject con-
stituent. Each phrase can also contain a phrasal complement. In this
example it would be the empire. Question: What is the specifier of the
What is then proposed is that all phrases have the following struc- verb in a sentence like John sleeps.?
Note that, because phrasal projec-
ture: tions (maximal projects) have their
source in the lexical information of
their head, the word class of a phrase
XP
and its head will always be the same.
So it’s not possible to have a rule
where XP –> Z’ YP. A Z word, if a
(YP) X0 head, will always project a ZP.

specifier X (YP)

head complement
 syntax ii: argument structure 125

We can then write these with CFG notation:

XP → (YP) X0 (10.45a)
0
X →X (YP) (10.45b)
In X-bar theory all phrases will
The projection of X on the right is called the head. Subjects are have the same structure (though
specifiers of verbs, direct objects complements of verbs. The YP not all positions will be used!). The
structure of a DP is as follows:
in Rule (10.45a) is called the specifier of the phrase, the one in
Rule (10.45b) is called the complement. DP
We note the following general fact about English.
ZP D
In English, specifiers are to the left, complements to the right of
the head.
Specifier D YP
All these terms are relational. An occurrence of a phrase is not a
complement per se, but the complement of a particular phrase XP in head Complement
a particular construction. Now English, like all languages, also allows
adjuncts. A phrase or a bar-level category can combine with another The structure of an NP will be as
phrase as an adjunct. So to make our rules complete we need the follows:

following three adjunct rules. NP

ZP N
XP → XP YP (10.46a)
X0 → YP X0 (10.46b)
Specifier N YP
0 0
X →X YP (10.46c)

Recall that adjuncts are always optional. Further, you can have head Complement
multiple adjuncts. The rules above allow this by permitting e.g. an
And the structure of a DP with an
X-bar to expand to another X-bar by using e.g. (10.46b):
NP complement will be as follows:
DP ]
X
ZP D
YP X
Specifier D NP
X YP
head ZP N
If we look at other languages we find that they differ from English
typically only the relative position of the specifier, complement and
Specifier N YP
adjunct, not the hierarchy. Thus some languages make use of the
following rules:
head Complement
0
XP → X YP (10.47a)
XP → YP XP (10.47b)
0
X → YP X (10.47c)

It’s also the case that the order of combining categories can change
from category to category, even within the same language. German
puts the verb at the end of the clause, so the complement is to its left.
The structure of nouns and PPs is however like that of English nouns.
Japanese and Hungarian put prepositions at the end of the PP. (That
126 an open introduction to linguistics

is why they are called more accurately postpositions. Because these

two categories (prepositions and postpositions) actually only differ by
position, but not by function, we have a more general category, ad-
positions. An adposition is either a preposition or a postposition.
Sanskrit puts no restriction on the relative position of the subject (=
specifier of verb) and object (= complement of the verb). Pro-VP Examples
These are also called VP-ellipsis:
Sentence Structure I. (Universal) All syntactic trees satisfy • John likes milk and I do too [=like
X-bar syntax. This means that Rules (10.45a) – (10.47c) hold milk]

for all structures. • Ed likes milk as much as Jo does

[=likes milk]
It is important to be clear about the sense of “holds” in this principle. • Tim speaks Ewe and so do I.
We use it with the sense of admissibility. A set of rules R of the [=speak Ewe]

form X → Y Z or X → Y is said to hold for a labelled tree (or a set

thereof) if, whenever a node in a tree has category X, then it either has
no daughter, or only one daughter with label Y and the rule X → Y
is in R, or it has two daughters with labels Y and Z (in that order)
and the rule X → Y Z is in R. This looks like the definition of the
previous chapter, but it is crucially different, because here we are not
saying how we get the trees (important when we study movement) and
we also do not require that there exists a context free grammar for the
trees (the language can be far more complex, but still the rules are
said to hold).

Pro-Forms and other non-projectors

This chapter began by presenting different words and their categories.
The words in this list were lexical, and they are able to project phrases
(nouns project NPs, prepositions project PPs, etc.) That word classes
can project phrases is true in general, but there are a number of im-
portant exceptions to this generalization.
A major type of exception are pro-forms. Pro-forms are words
that actually stand in for an unspecified word or phrase. Pronouns
make up one class. They can replace a DP. (This suggests that we
should actually call them pro-DPs, but we don’t. The name pronoun
was coined before anyone had thought of the concept of DPs). Other
examples are /one/ (pro-NP) and /do/ (pro-VP).

John wrote a good essay. Paul also wrote a good one.

(10.48)
John wrote an essay and Paul did as well. (10.49)

There are a number of additional examples. For example, the question

words: /who/, /what/, /where/, /why/ and so on, and the deictics
(/this/, /thus/, /so/ and so on) which also are phrasal. /who/ and
/what/ are DPs, /why/ and /where/ are PPs.

Category T and tensed phrases

We haven’t said anything however about Category T, which included
auxiliary verbs like will, would, can, could and so on. These verbs
 syntax ii: argument structure 127

are different from other verbs in several ways. For example, in the
third person singular form they never take the inflection ending -s.
This category can project a phrase. Let’s consider a sentence with an
auxiliary verb.

C3PO can speak Bocce. (10.50)

Given our rules, it would seem that we need to also allow a VP to

take a TP complement.
VP

DP TP

C3PO T V0

can V

speak Bocce
But then we seem to now have rules like the following:

• VP –> DP V

• VP –> DP TP

• TP –> T V

• V –> V DP

This however ruins the uniformity of the system that we had. There
seem to be three major problems. First, there is no T-bar level, and
instead we have two different kinds of VPs, VPs with modal and VPs
without. Second, we’ve now separated our VP from the V, but the
VP is supposed to be the projection of the V. How did that TP get
in between them? Third, the subject argument is now separated from
the verb in sentences with auxiliaries, but it is also supposed to have
been generated as an argument of the verb. All this suggests that this
is not the right analysis.
One solution to regain uniformity would be to simply argue that all
verb phrases have TPs, and actually that what we previous called a
VP, was actually a TP.

C3PO speaks Bocce. (10.51)

I speak Bocce. (10.52)

Let’s return to our constituency tests. If we use the Answer Ellipsis

test with the sentences above we get the following:
128 an open introduction to linguistics

C3PO speaks Bocce. What does C3P0 do? Speak Bocce.

(10.53)
C3PO speaks Bocce. What does C3P0 do? *Speaks Bocce.
(10.54)
C3PO can speak Bocce. What can C3P0 do? Speak Bocce.
(10.55)

Notice that the correct elliptical answer is not speaks, but speak. The
third person singular inflectional affix -s is not part of the constituent!
Instead, it seems to be like can. That’s exactly what we are going
to assume. That inflectional morphemes that mark person or tense
(such as the regular past tense suffix -ed) are of category T, along
with auxiliary verbs. But we have to deal with problems two and
three as well. We don’t want to separate the verb from the subject or
the VP. A solution then could be to simply put the TP on top, and let
TPs take VP complements (instead of the other way around), so then
C3PO can speak Bocce will get the following structure:

TP

T VP

can DP V

C3P0 V DP

speak Bocce

But actually, we don’t have the uniformity of X-bar structure any-

more. Let’s fix that.
TP

(Spec) T

T VP

can DP V

C3P0 V DP

speak Bocce

So now we solved problems two and three. As for problem one, sure
we can put inflectional markers in T, and for verbs that don’t take
any inflections we’ll just add a zero-inflection, and then we’ve solved
the problem of not having a uniform analysis. However you may have
noticed that we’ve introduced a new problem: the surface order
of the sentence and the order of the words in the tree is no
 syntax ii: argument structure 129

longer the same!!! This seems to be a big problem.

What we need is to get the subject to occur before the TP. To do
this we need a place to put it, which means there must be a structure
above the TP. We’ll use S again, but this is just temporary. What
we will then do is make the same type of distinction that we made in
phonology between deep structures and surface structures. Because
verbs determine their arguments, we want the subject to be generated
from the verb, and in the deep structure it should be there with the
verb. But in the surface structure we want it to be linearly before the
tense marker, T. So what we’ll do is move it. Movement is the third
major syntactic process (after substitution and adjunction).

Step 1: S

TP

(Spec) T

T VP

can DP V

CP30 V DP

speak Bocce
Step 2: S

TP

DP T

C3P0 T VP

can DP V

tCP 30 V DP

speak Bocce
We will need to restrict movement, so there are some constraints.
We want to remember where things came from, so we’ll leave some-
thing behind, which we’ll call a trace. Further, it’s actually the case
that you can only move things to positions that accept that sort of
category. So we can move the DP to Spec-TP because that’s a po-
sition for a phrase (recall from X-bar structure, a YP can go there).
This is phrasal movement. Finally, there’s a structural constraint that
forces movement to be leftward and upwards: moved elements must
C-command their traces. C-command is something we’ll go into more
detail in the next chapter.
Now the current structure, with movement, will work fine for sen-
130 an open introduction to linguistics

tences like C3PO can speak Bocce. But we still have a word order
problem with sentences like C3PO speaks Bocce in that their analysis
becomes C3PO -s speak Bocce. There are two problems with this sen-
tence. The first is that the words are not in the right order. But the
second is that we said in the chapter on morphology that inflectional
affixes cannot stand alone. So it’s a bit odd that we let -s just hang
there as a node. It actually needs to be attached to the verb. But
hey, maybe we can solve both of these problems with movement as
well! What if we let the verb move to T? V and T are both the same
level of category, Projection level 0, and this movement would satisfy
the c-command requirement. We then let the verb and the inflectional
affix merge to end up with the correct surface structure: C3PO speaks
Bocce. Actually, this type of movement is called Head-to-head move-
ment and in the following diagram you can see it’s identifeid with a
dashed red line. Try drawing the steps involved for yourself.
We still have one final problem. We have S as a start symbol, but
this ruins our nice X-bar uniformity, and it means we have a phrase
project that doesn’t come from a lexical entry. It would be much nicer
if this S was something us. A potential candidate is a CP. Recall that
CPs are sentential complements to verbs like know and discover. It
would make everything even more uniform if all sentences, whether
matrix sentences or sentential complements, belong to the same cate-
gory! Win-win! Let’s therefore assume that all sentences are CPs, and
matrix sentences have an empty C category.
CP

C TP

(that) DP T

C3P0 T VP

speak+s DP V

tC3P 0 V DP

tspeak Bocce

We now make a very important conclusion:

Sentence Structure II. (Universal) The start symbol is CP.

 syntax ii: argument structure 131

A note on PPs and adjunction

Since PPs are adjuncts, they can be repeated any number of times.
For example, we can say:

∗
The sailors accused the captain on the ship. (10.56)

We can then continue adding more PPs:

∗
The sailors accused the captain on the ship during the storm at midnight.
(10.57)

But PPs that are selected by the verb cannot be repeated:

∗
The sailors accused the captain of treason of cruelty.
(10.58)

The fact that adjuncts can be repeatedly added is a special property

called the Iterability of Adjuncts, and is actually a test to determine if
something is an adjunct or a complement. But the feature system that
we’ve developed in this section allows for arbitrary repetition for any
PP. This could be restricted to not include PPs selected by the verb but
we would then need a feature system that checks each verb and PP. We
won’t go into the details here, as they are not revealing (and not pretty
either). But this ultimately leads to a system called Generalised
Phrase Structure Grammar (GPSG). This later developed into
Head Driven Phrase Structure Grammar (HPSG).

Study Questions
1. What are the three major syntactic processes?

2. Write out the CFG rules that capture the information AVS struc-
tures 10.26-10.28.

3. Write attribute value rules that you would need to add to 10.26-
10.29 in order to parse a prepositional phrase headed by the prepo-
sition to, e.g. to the lighthouse. Hint: you will need a verb that can
take to, like run. You will need four rules.

4. Identify the specifier and the complement of the verb in the follow-
ing sentences.

a. The man saw the bird.

b. A sparrow ate seeds.
c. He annoyed me.

5. Look at the attribute-value structure rules. Can you identify which

rules involved adjuncts and which involve arguments? (hint look at
the LHS and RHS levels).
132 an open introduction to linguistics

6. The complement of a DP is an NP. Draw the tree structure for a

DP with an NP. Make sure you put in all intermediate levels.

7. In the text we gave tree structure examples of the X-bar based rules
that can form DPs, with NP complements. Now write out the same
rules using AVS structures.

8. Draw a tree structure for the DP the sparrow

9. Draw a tree structure for the sentence CP3O talks. You will need
to use both VP-to-Spec TP and Head-to-head movement.
11
Syntax III: Local Dependencies and Constraints: Se-
lection, Agreement and Case Marking

The organisation of phrases is as much a matter of placement as

a matter of morphological marking. When a head (for example a
verb) wants an argument it also determines some of its morpho-
logical features. In this chapter we learn about the important
concepts of selection, case and agreement. Selection is the
fact that heads want an argument of a particular kind; case is
the feature that identifies a phrase as a particular argument of
the head. Finally, agreement is something that both the head
and its argument(s) share.

Grammatical Features
Nouns come in two varieties: singular and plural. The feature of being
singular or plural is called number. We have already mentioned a few
ways in which the singular and plural of nouns are formed. For syntax,
the different ways of forming the plural are not relevant; only the fact
of whether a noun is singular or plural is relevant. A way to represent
the number of a noun is by adding an attribute num, with values sing
and pl (in other languages there will be more...). This means that the
syntactic representation of ‘mouse’ and ‘mice’ is

   
cat :N cat :N
mouse :  proj:0  , mice :  proj:0  (11.1)
   
num :sing num :pl

This we rephrase by using the following rules:

 
cat :N
 → mouse (11.2)
 
 proj:0
num :sing
 
cat :N
 proj:0  → mice (11.3)
 
num :pl

An alternative notation, which is used elsewhere and should now be

self-explanatory, is DP[sing] and DP[pl]. The attribute num can then
134 an open introduction to linguistics

be omitted. We shall use this type of notation without further warning.

It is to be thought of as abbreviatory only.
Noun phrases in English, like nouns, can also be both singular and
plural. Whether or not an NP is singular or plural can be seen by
looking at the head noun.

fearful warrior (singular) (11.4)

fearful warriors (plural) (11.5)
brother of the shopkeeper (singular) (11.6)
brothers of the shopkeeper (plural) (11.7)

In the first example the head noun is /warrior/, in the second it

is /warriors/. This indicates that in the first example the NP is
singular, and it is plural in the second. Thus, the number feature is
passed up, so to speak, to the phrase from the head of the phrase.
Notice that the determiner is different depending on whether the NP
is singular or plural.

this fearful warrior (11.8)

∗
these fearful warrior (11.9)
∗
this fearful warriors (11.10)
these fearful warriors (11.11)

We say that like nouns, determiners have a singular and a plural form
and that the determiner agrees with the NP complement in number.
This will be put into the rules as follows. The letter α may be instanti-
ated to any legitimate value, in this case sing or pl. Notice that within
a single rule, each occurrence of α must be replaced by the same value.
However, whether or not the letter α also occurs in another rule is of
no significance. Grammatical Gender vs. Nat-
ural Gender Languages generally
make two distinctions: features that
mark natural gender mark whether
   
cat :D cat :D or not an animate indvidual is male
 proj:2  →  proj:1  (11.12)
   
or female. Grammatical Gender is a
num :α num :α type of noun classification system that
     manifests as agreement between syn-
cat :D cat :D cat :N tactically related words in an utter-
rance. Some languages have mascu-
 proj:1  →  proj:0   proj:2  (11.13)
    
line and feminine (e.g. French, Urdu),
num :α num :α num :α some languages have common and
  neuter, (e.g. Dutch, Swedish) while
cat :D some languages instead distinguish be-
 → this | the | a | · · · (11.14) tween Animate and Inanimate, (e.g.
 
 proj:0
Basque, Hittite). Some languages, like
num :sing Swahili, have 18 noun classes.
 
cat :D
 proj:0  → these | the | ∅ | · · · (11.15)
 
num :pl

Here, ∅ is the empty string. It is needed for indefinite plurals (the

plural of /a car/ is /cars/).
The choice of the determiner controls a feature of DPs that is some-
times syntactically relevant: definiteness. DPs are said to be definite
 syntax iii: local dependencies and constraints: selection, agreement and case
marking 135

if, roughly speaking, they refer to a specific entity given by the con-
text or if they are uniquely described by the DP itself. Otherwise they
are called indefinite. The determiner of definite DPs is for example
/this/, /that/ and /the/. The determiner of indefinite DPs is for
example /a/ in the singular, /some/ in singular and plural, and ∅ in
the plural.
In other languages, nouns have more features that are syntactically
relevant. The most common ones are case and gender. Latin, for ex-
ample, has three genders, called masculine, feminine and neuter.
In English the three genders have survived in the singular 3rd person
pronouns: /he/, /she/, /it/. The Latin noun /homo/ (‘man’) is mas-
culine, /luna/ (‘moon’) is feminine, and /mare/ (‘sea’) is neuter. The
adjectives have different forms in each case (notice that the adjective
likes to follow the noun, but it does not have to in Latin):

homo ruber (red man) (11.16)

luna rubra (red moon) (11.17)
mare rubrum (red sea) (11.18)
Nouns have many different declension classes (morphology deals with
them, syntax does not), and there are rules of thumb as to what gender
nouns in a certain declension class have, but they can fail. Nouns end- Here we are looking at the form of the
ing in /a/ are generally feminine, but there are exceptions (/nauta/ word to determine what grammatical
gender it has. This is not a reliable
‘the seafarer’, /agricola/ ‘the farmer’ are masculine). Similarly, ad- predictor in all languages.
jectives have many declension paradigms, so the forms can vary. But
for each adjective there are forms to go with a masculine noun, forms
to go with a feminine noun, and forms to go with a neuter noun. We
therefore say that the adjective agrees in gender with the noun. It is
implemented by installing a new attribute gender, whose values are
‘m(asculine)’, ‘f(eminine)’ and ‘n(euter)’. English is said not to have
any forms of gender agreement. Moreover, English gender is said to
be semantic, that is, based on what a noun denotes. To know whether
you should use /he/, /she/ or /it/, you have to ask yourself whether
the referent is male, female, or neither. The semantics of gender is not The case where sex is unknown (or is
fully congruent with sex; ships and other vessels are an exception in not to be revealed by the speaker) is
another case, and it is different from
that you must use /she/ to refer to them, even though they have no the case where the sex is known or the
sex. This does not mean, though, that gender is not semantic; it means notion of sex does not apply, for exam-
ple in the case of cars. In some vari-
that its meaning is not fully determined by sex. A different case is pre- ants of English, for example, /they/ is
sented by German, where gender is demonstrably not fully semantic. used to avoid communicating the sex,
All diminutives are neuter, regardless of what they refer to. In referring as opposed to /it/ when the notion
does not apply.
to a small man, for example, you can either say /der kleine Mann/
(‘the-masc.nom.sg small man’) or /das Männchen/ (‘the-neut.sg.nom
man-dimin’).
Latin nouns also have different cases. There are five cases in Latin:
nominative (for subjects), accusative (for direct objects), dative, geni-
tive and ablative. Just as verbs select PPs with a particular preposition
in English they can also select a DP with a particular case. If the verb
is transitive, the DP is generally in the accusative, but it can be dative
(/placere/+DP[dat] ‘to please someone’), ablative (/frui/+DP[abl]
‘to enjoy something’), and genitive (/meminisse/+DP[gen] ‘to remem-
136 an open introduction to linguistics

ber someone/something’). There exist verbs that take several DPs

with various cases. For example, /inferre/ ‘to inflict’ (with perfect
/intuli/) wants both a direct object and a dative DP.

Caesar Gallis bellum intulit. (11.19)

Caesar Gauls-dat war-acc inflict.upon-perf
Caeser inflicted war on the Gauls.

This is just like English /inflict/, which wants a direct object and a
PP[on].
For us, a grammatical feature is anything that defines a syntactic
category. You may think of it as the syntactic analogue of a phonemic
feature. However, beware that elsewhere the usage is a little different.
Grammatical category and projection level are typically discarded, so
that grammatical feature refers more to the kinds of things we have
introduced in this section above: number, gender, definiteness, and
case. A fifth one is person. (This list is not exhaustive.)
The position of subject can be filled with DPs like /the mouse/,
/a car/, but also by so-called pronouns. Pronouns are distinct from
nouns in that they express little more than that they stand in for a DP.
However, they have many different forms, depending on the grammat-
ical feature. In addition they show a distinction in person. Across
languages, there is a pretty universal system of three persons: 1st, 2nd,
and 3rd. First person means: includes speaker. Second person means:
includes hearer, and third person means: includes neither speaker nor
hearer. So, /I/ is first person (it includes me and no one else); it is
also singular, because it refers to just one thing. The plural /we/ is
used when one refers to several people, including speaker. By contrast,
/you/ is used for individuals or groups including the hearer (there is no Note that /we/ might also include the
distinction between singular and plural). The third person pronouns hearer, in addition to the speaker and
other people. The distinction between
distinguish also gender in the singular (/he/, /she/, /it/), but not /we/ with and without the hearer is
in the plural (/they/). Moreover, as we explained above, pronouns not made explicit in English, but must
be derived from the context. Other
distinguish nominative from accusative. Thus, there are more mor- languages do make this distinction ex-
phological distinctions in the pronominal system in English than there plicit, using different pronouns for the
first person plural with and without
are in the ordinary DPs.
the hearer.

More on Case
Cases have many different functions. One function is to indicate the
nature of the argument. A verb has a subject, and the case of the
subject is referred to as nominative. The direct object has a case
that is referred to as accusative. Some people believe that English
has cases (because pronouns still reflect a distinction between subject
and object: /she/ : /her/, /he/ : /him/, and so on). On the other
hand, this is confined to the pronouns, and nouns show no distinction in
case whatsoever. This is why most people say that English has no case.
Strictly speaking, it means only that there is no distinction in case.
(One might say: there is one case and only one. This is useful. For
example, there is a famous principle of syntactic theory which states
that nouns need case. If there is no case, this principle fails.) Chinese
 syntax iii: local dependencies and constraints: selection, agreement and case
marking 137

is another example of a language that has no cases. These languages

make no distinction between subject and object in form; nevertheless,
one can tell the difference: the subject precedes the verb, and the
object follows it (both in English and in Chinese). Verbs can have The equivalent of case in a language
more than two arguments, and many more adjuncts. To distinguish lacking it is position.

between them some kind of marking is needed. In English this is done

by means of prepositions. For example, there often is an argument
towards which the action is directed or for which it is performed (the
‘goal’) and it is given by a PP opened by /to/ (= PP[to], for example Theta roles are the thematic roles
/talk to someone/). The goal is also called indirect object. Latin of arguments, classifying arguments
by their semantic role in the event/s-
has a case for this, the dative. There is, from a global viewpoint, not tate denoted by the verb. For exam-
much of a difference whether the goal is encoded by a case or by a PP. ple, many action verbs have an Agent
and Patient, and in a simple active
Languages can choose which way to go. sentence these will be realized as the
Another important case is the genitive. It marks possession. En- grammatical subject and grammatical
glish has basically two ways to mark possession (apart from obvious object, e.g. John hit Bill. But in the
passive, the grammatical subject be-
ones like /which belongs to/). One is the so-called Anglo-Saxon comes the Patient, and the Agent role
genitive, formed by adding an /’s/ (/my neighbour’s car/). The can remain unexpressed if desired, e.g.
Bill was hit.
other is a PP opened by /of/ (/the car of my neighbour/). The
genitive is used a lot in English. Nouns that have arguments that are
not PPs put them in the genitive:

the election of the chancellor (11.20)

the peak of the mountain (11.21)
Napoleon’s destruction of the city (11.22)

Notice that two of these nouns have been obtained from transitive
verbs. The rule in English is that the noun can take any of the ar-
guments that the verb used to take (though they are now optional).
However, the subject and the object must now appear in the geni-
tive. The PPs on the other hand are taken over as is. For example,
/destroy/ is transitive, so the noun /destruction/ can take two geni-
Map of Hungary. (From Wikipedia)
tives, one for the subject and one for the object. The verb /talk/ takes
a subject, an indirect object and subject matter, expressed by a PP
headed by /about/. The latter two are inherited as is by the noun
/talk/, while the subject is put in the genitive. Alternatively, it can
be expressed by a PP[by].

John talked to his boss about the recent layoffs.

(11.23)
John’s talk to his boss about the recent layoffs (11.24)
talk by John to his boss about the recent layoffs
(11.25)

Subject-Verb Agreement
English displays a phenomenon called subject-verb agreement. This
means that the form of the verb depends on the grammatical features
of the subject. The agreement system is very rudimentary; the only
138 an open introduction to linguistics

contrast that exists is that between singular 3rd and the rest:

She runs. (11.26)

They run. (11.27)

Notice that since the verb does agree in person with the subject it has
to make a choice for DPs that are not pronominal. It turns out that
the choice it makes is that ordinary DPs trigger 3rd agreement:

The sailor runs. (11.28)

This applies even when the DP actually refers to the speaker! So, So if Jennifer is lecturing in class and
agreement in person is (at least in English) not only a matter of what she (rather oddly) refers to herself
with a full DP, e.g. “Jennifer is lec-
is actually talked about, but it is also a syntactic phenomenon. There turing”, she still uses the third person
are rules which have to be learned. singular form of the verb to be, e.g.
*“Jennifer am lecturing”.
Other languages have more elaborate agreement systems. Let us
look at Hungarian. The verbal root is /lát/ ‘to see’.

Én látok. Mi látunk. (11.29)

I see We see
Te látsz. Ti látatok.
You(sg) see You(pl) see
Ö lát. Ök látuk.
He/she/it sees They see

Hungarian has no distinction whatsoever in gender (not even in the

pronominal system; /ö/ must be translated as ‘he’, or ‘she’, or ‘it’,
depending on what is talked about). However, it does distinguish
whether the direct object is definite or not. Look at this (translation
is actually word by word):

Én látok egy madarat. (11.30)

I see a bird
Én látom a madarat. (11.31)
I see the bird

The subject is the same in both sentences, but the object is indefinite
in the first (a bird) and definite in the second (the bird). When the
object is indefinite, the form /látok/ is used, to be glossed roughly as
‘I see’, while if the object is definite, then the form /látom/ is used, to
be glossed ‘I see it’. Hungarian additionally has a form to be used when
subject is first person singular and the direct object is 2nd person:

Én lát-lak. (11.32)
I see-sub:1sg.ob:2sg

The period is used to say that the form

/lak/ is a single morpheme expressing
two syntactic agreement facts.
Who Agrees with Whom in What?
Agreement is pervasive in some languages, and absent in others. Chi-
nese has no agreement whatsoever, English has next to none. The
 syntax iii: local dependencies and constraints: selection, agreement and case
marking 139

most common type of agreement is that of verbs with their subjects.

Some languages even have the verb agree with the direct object (Hun-
garian, Mordvin (a language spoken in Russia, related to Hungarian),
Potawatomi (an American Indian language)). Other languages have
the verb agree in addition with the indirect object (Georgian). Agree-
ment is typically in person and number, but often also in gender.
Above we have seen that definiteness can also come into the picture.
Adjectives sometimes agree with the nouns they modify (Latin, Ger-
man, Finnish), sometimes not (Hungarian). There is no general pat-
tern here. This is one of the things that one has to accept as it is.

Obligatoriness
Morphology is obligatory. A noun has a case in a language with cases,
it has a number in a language with numbers, and so on. This creates
a certain predicament about which I have spoken earlier in connection
with gender. Let me here discuss it in connection with number. The
intuitive approach is this. An English noun is either singular or plural.
If it is singular it means ‘one of a kind’ and if it is plural it means
‘several of a kind’. So, /car/ means ‘one car’, and /cars/ means
‘several cars’. But this immediately raises questions.

À Are there remaining cases to be dealt with?

Á What if the notion of number does not apply?

 What if the speaker cannot (or does not want to) communicate the
number of things involved?

Let’s deal with these questions one at a time:

À. The most obvious case that we left out is ‘zero of a kind’; but others
also come to mind, such as fractions, and negative numbers.
Á. Mass nouns (/water/, /iron/), which we shall discuss later, have
no obvious plural, since we cannot count what they refer to.
Â. The most evident situation is a question; but there are many cir-
cumstances in which we do not know how many, such as if there was
a crime and we talk about the people who did it.
In all of these cases, the language comes up with more or less explicit
recipes as to how to solve the problem. These fixes are to some degree
arbitrary. Consider the first point:

They gave me a quarter point for my answer. (11.33)

They gave me 0.25 points for my answer. (11.34)
They gave me 1 1/4 points for my answer. (11.35)
They gave me 1.25 points for my answer. (11.36)

What we see is that decimals expansions are treated differently from

fractions.
When you ask for the identity of some person, you often do not know
how many there are. In English you do not have to commit yourself,
you use /who/. However, even though the number is unknown, /who/
140 an open introduction to linguistics

itself triggers singular agreement in the verb.

Who drank my coffee? (11.37)

In Hungarian, there are two question words, a singular /ki/ ‘who-sg’

and a plural /kik/ ‘who-pl’. If you use the latter you indicate you
expect the answer to be several. If you are not using a question, the
plural sounds less committal than the singular, for example (11.38)
commits you just one, while (11.39) does not.

The murderer of the shopkeeper. (11.38)

The murderers of the shopkeeper. (11.39)

So Whoever murdered the

The important lesson to learn is that all categorisations leak. There shopkeeper is again non-specific
for the number of murderers.
simply is no overarching system to classify everything. Consequently,
languages must implement a few strategies of dealing with the un-
known cases.

Feature Percolation

When we look at the distribution of number in English, we find that

we get the following rules:

DP[ num : α] → (YP) D0 [ num : α] (11.40)

DP[ num : α] → YP DP[ num : α] (11.41)
DP[ num : α] → DP[ num : α] YP (11.42)
0
D [ num : α] → D[ num : α] YP[ num : α] (11.43)
0
D [ num : α] → D[ num : α] (11.44)
D0 [ num : α] → (YP) D0 [ num : α] (11.45)
0 0
D [ num : α] → D [ num : α] YP (11.46)

This is the same also for the projection of N and V. There is in general
no connection between the features on the YP and that of the D, with
one exception: the determiner passes the number feature on to the NP
complement. It is the same also with gender features, and with case.
Might this therefore be a general feature of the X-bar syntax?
First, recall the rules (10.17) and (10.18). What we did not establish
then was the identity of the V0 with respect to its transitivity. We use
the following diagnostic: in a coordination only constituents which
have the same features can be coordinated. If this is so then the
following sentence would be ungrammatical if we called /kicked the
ball/ as a transitive V0 .

John kicked the ball and fell. (11.47)

The solution is to class both as intransitive V0 . Thus the rules are as

 syntax iii: local dependencies and constraints: selection, agreement and case
marking 141

follows.
   
cat :V cat :V
 →  proj:0  (11.48)
   
 proj:1
trs :− trs :−
   
cat :V cat :V " #
 cat :D
 →  proj:0  (11.49)
  
 proj:1
proj:2
trs :− trs :+

The full story is to distinguish two kinds of features: selectional

features (like trs and prep) and agreement features. Selectional
features are not passed from head to phrase, while agreement features
are.
It is to be noted, though, that coordination does not follow X-bar
syntax (we have discussed the schema above). Furthermore, agreement
features are exempt from the identity requirement in coordination, i.e.
we can coordinate singular and plural DPs:

John and Mary danced. (11.50)

William and the firemen saluted the governor. (11.51)

Study Questions
1. Identify the head of the NP in the following DPS.

(a) this happy camper

(b) the sister of the suspect
(c) these generous benefactors

2. Dutch has a common (‘de’) and neuter (‘het’) grammatical gender

system: het kind (the child), de auteur. Write DP production rules
using AVS structures parallel to 10.12, 10.13 and 10.14 for Dutch.

3. Recall that selectional features are not passed from head to phrase,
while agreement features are. Show with AVSs like (10.47) and
(10.48) which features would percolate (or not) for a Dutch DP
with neuter gender, like het kind.
12
Syntax IV: Movement and Non-Local Dependencies

Even though it is possible to give a syntactic analysis of questions

by only using context-free rules, by far the most efficient analysis
is in terms of transformations. Typically, a transformation is the
movement of a constituent to some other place in the tree. This
chapter explores how this works and some of the conditions under
which this happens.

Movement
We have learned that in English the transitive verb requires its direct
object immediately to its right. This rule has a number of exceptions.
The first sentence below, (12.1), displays a construction known as top-
icalisation, the second, (12.2), is a simple question using a question
word.
Pilots Harry admires. (12.1)
Which country have you visited? (12.2)
We could of course give up the idea that the direct object is to the
right of the verb. It would be possible to argue that, in the case
of topicalisation, the sentence simply takes the Object-Subject-Verb
order. But the facts are quite complex, especially when we turn to
questions. For example, for question words, no matter what kind of
constituent the question word replaces (subject, object, indirect object
and so on), it occurs sentence initially, even if it is not the subject.
Alice has visited Madrid University in spring to learn Spanish.
(12.3)
What has Alice visited in spring to learn Spanish?
(12.4)
Who has visited Madrid in spring to learn Spanish?
(12.5)
When has Alice visited Madrid University to learn Spanish?
(12.6)
Why has Alice visited Madrid University in spring?
(12.7)
We see that the sentences involving question words (interrogative sen-
tences) differ from their declarative counterparts, (12.3), in that the
144 an open introduction to linguistics

question word is in the initial position, and the verb is in second place.
There is a way to arrive at an interrogative sentence in the following
way. First, insert the question word where it ought to belong according
to our previous rules. Next, move it to first position. Now move the
auxiliary (/has/) into second place. We can represent this as follows,
marking removed elements in red, and newly arrived ones in blue: In the example to the left, Where
has been moved from the position to
(1) Alice has travelled where in spring to learn Spanish? the right of travelled to the sentence-
initial position. Then, it has has been
moved from the position to the left of
travelled to the second position.
(2) Where Alice has visited where in spring to learn Spanish?

(3) Where has Alice has visited where in spring to learn Spanish?

A more standard notation is this:

Alice has travelled where in spring to learn Spanish?

(12.8)
Where Alice has travelled in spring to learn Spanish?
(12.9)
Where has Alice travelled in spring to learn Spanish?
(12.10)

(The underscore just helps you to see where the word came from. It
is typically neither visible nor audible.) The first notation is more
explicit in showing you which element came from where (assuming
they are all different). However, neither notation reveals the order in
which the movements have applied. It turns out, though, that this is
irrelevant.
The plus side of this proposal is that it is actually very simple.
However, we need to be clear about what exactly we are doing. Super-
ficially, it seems that what we are doing is deriving one surface sentence
of English from another surface sentence of English. (This was in fact
the way transformations were originally thought of.) However, we are
actually going to claim that the process is more substantial. We in-
stead will claim that every surface sentence of English is derived in a
two stage process. First, we generate a structure in which every head
projects its phrase, satisfying its requirements (such as verbs project-
ing a VP consisting of a subject and, if transitive, an object, and the
necessary PPs). After that is taken care of, we apply a few transfor-
mations to the structure to derive the actual surface string. Recall
that this is what we did with phonological representations. First we
generated deep representations and then we changed them according
to certain rules. Thus, we say that the context free grammar generates
deep syntactic representations, but that the rules just considered
operate on them to give a final output, the surface syntactic rep-
resentation. The surface syntactic representation is what speakers
actually say. The rules that turn deep structures into surface struc-
tures are the (syntactic) transformations.
 syntax iv: movement and non-local dependencies 145

Wh-Movement
When talking about the transformations responsible for putting ques-
tion words at the front of a sentence we usually talk about a type
of movement, called Wh-Movement. Question words are also re-
ferred to as wh-words, since they all start with /wh/ (/who/, /what/,
/where/, /why/, etc.). At first blush one would think that syntactic
transformations operate on strings; but this is not so. Suppose the
original sentence was not (12.3) but

Alice has visited which famous city in Mexico (12.11)

to wait for her visa?

Then the output we expect on this account is (12.12). But it is un-

grammatical. Instead, only (12.13) is grammatical.
∗
Which has Alice visited famous city (12.12)
in Mexico to wait for her visa?
Which famous city in Mexico has Alice (12.13)
visited to wait for her visa?

It is the entire DP containing the question word that gets moved.

There is no way to define this by referring to strings of words; instead
it is defined on the basis of the tree. To see how, let us note that the
sentence (12.3) has the following structure. (Some brackets have been
omitted to enhance legibility.)

Alice [has [[visited [which famous city in Mexico] (12.14)

[to wait for her visa]]]]?

Now, /which famous city in Mexico/ is a constituent (it passes for

example the tests À – Ã). Moreover, it is the object of the verb
/visited/. The word /which/ is a determiner, and the smallest phrase
that contains it is the one that has to move.

Wh-Movement I. (Preliminary)
Only phrases can be moved by Wh-Movement. What moves is
the smallest phrase containing a given wh-word. It moves to the
beginning of a clause (= CP).

This specification is imprecise at various points. First, what happens

if there are several wh-words? In English only one of them moves
and the others stay in place; the choice of the one to move is a bit
delicate, so we shall not deal with that question here. In some other
languages (e.g. Romanian, Bulgarian, Hungarian are examples) all of
them move. Second, what happens if the wh-word finds itself inside a
sentence that is inside another sentence? Let us take a look.

Mary thinks you ought to see what city? (deep struct.) (12.15)

Here the wh-phrase moves to end of the main sentence (and notice
that something strange happens to the verb too):

What city does Mary think you ought to see? (12.16)

146 an open introduction to linguistics

However, some verbs dislike being passed over. In that case the wh-
phrase ducks under; it goes to the left end of the lower sentence.
∗
What city does Mary wonder you have seen? (12.17)
Mary wonders what city you have seen. (12.18)

So, let us add another qualification.

Wh-Movement II. (Preliminary)

The wh-phrase moves to the beginning of the leftmost phrase
possible.

We shall soon see below that this is not a good way of putting things,
since it refers to linear order and not to hierarchical structure.

Verb Second
Wh-movement is a movement of phrases. In addition to this there
is also movement of zero level projections, or heads. (It is therefore
called head movement.) A particular example is verb movement.
Many languages display a phenomenon called Verb Second or V2.
German is one of them. Looking at individual word occurrences, the
verb is not always in second place. But if we look at constituent
structures in a tree, we will see that the verb does appear in the second
place. Here is a pair of sentences with word-to-word translation.

Hans geht in die Oper. (12.19)

Hans goes into the opera
Der Lehrer ist erfreut, weil Hans in die Oper geht.
the teacher is pleased, because Hans into the opera goes (12.20)

The main verb is /geht/ (‘goes’). In the first example it is in second

place, in the second example it is at the end of the sentence. Notice
that in the second example there is a CP which is opened by /weil/
(‘because’). It is called subordinate, because it does not display the
same kind of order as a typical clause. Now one may suspect that the
verb simply occupies a different place in subordinate clauses. However,
if we look at an auxiliary plus a main verb, matters start to become
more complex.

Hans will in die Oper gehen. (12.21)

The opera Hans will go to.
Hans wants into the opera go
Der Lehrer ist erfreut, weil Hans in die Oper (12.22)
gehen will.
the teacher is pleased, because Hans into the opera go wants

Only the auxiliary (/will/) is found in the second place in the main
clause. Based on a number of additional facts, this can be explained
as follows. The verb is at the end of the clause in deep structure.
In a subordinate clause it stays there. Otherwise it moves to second
position.
 syntax iv: movement and non-local dependencies 147

Now, what exactly is ‘second position’? It cannot be the second

word in the sentence. In the next example it is the fifth word (the dot
in the transcription shows that /im/ is translated by two words: ‘in’
and ‘the’).

Der Frosch im Teich ist kein Prinz. (12.23)

the frog in.the pond is no prince

Obviously, it is not the second word, it is the second constituent of

the sentence. Once again we find that the operation of movement is
not described in terms of strings but in terms of a hierarchical tree
structure.
English has a phenomenon similar to this. Notice that I said the
subject is in the specifier of VP. That works fine for a simple tense but
creates word order problems for complex tenses. For we would like to
posit that /will/ is a tense head (so here is one of the words that go
into T0 ). If that is so, the basic structure is this:

[[∅ [[will [John [see Alice]V0 ]VP ]T0 ]TP ]C0 ]CP (12.24) “Please help me. I am a prince!”

This gives us the following base order:

will John see Alice (12.25)

This time it is the subject, however, that moves to specifier of TP, so

that we get:

[[∅ [[John will [ see Alice]VP ]T0 ]TP ]C0 ]CP (12.26)

This is now extended to all kinds of clauses as follows. We propose

that T0 consists of tense and subject agreement. How can this be? In
the future tense we say, for example, /will see/, while in the present
tense we say /sees/. Blurring the distinction between syntax and
morphology, we segment /sees/ into /see/ and /s/, the latter the
exact equivalent of /will/. We propose therefore that /will/ as well
as /s/ are T0 heads.
Thus, by analogy, the original present tense sentence is this:

[[∅ [[s [John [see Alice]V0 ]VP ]T0 ]TP ]C0 ]CP (12.27)

Thus we discover another reason for movement: morphological in-

tegrity. The agreement morpheme /s/ cannot survive by itself, and
the verb moves up to attach to it.

[∅ [∅ [[see+s [John [ Alice]V0 ]VP ]T0 ]TP ]C0 ]CP (12.28)

We shall not concern ourselves with the exact structure of /see+s/.

Thus, we propose that across the board the V0 is void of any tense
and inflectional feature; while the future tense and negation require a
separate T0 in the overt sentence, present tense and past tense only
have one word. For example, since /saw/ is the past tense of /see/,
we analyse it as /see+PAST+3Sg/. The sequence /PAST+3Sg/ is
148 an open introduction to linguistics

hosted in T0 , and the root /see/ moves up to compose with it.

[[PAST+3Sg [John see Mary]VP ]T0 ]TP (12.29)
[[see+PAST+3Sg [John Mary]VP ]T0 ]TP (12.30)
[John [see+PAST+3Sg [ Mary]VP ]T0 ]TP (12.31)
There are various other reasons why this is a good analysis, having
to do with placement of negation and adverbs. On the other hand, it
raises questions about the compatibility of categories; for we require
that after movement the categorical structure is still in accordance
with X-bar syntax, so V must still project a VP, not a CP. I spell this
principle out in full.
Categorical Transparency
The structure obtained after movement must conform to X-bar
syntax.
Since the movement of the verb into T0 does not change any projections
(it is just movement of a constituent), we will retain a T0 node, and
so the head of this phrase must still be a T. There is a fix for this
problem, which I shall discuss below. Notice that that the T0 contains
material in it. Moreover, when the verb moves, it does not replace the
material but rather adds itself on the left of it. The resulting string
is therefore still of the same category. We shall look at the details of
this later.
The dance now continues. When a question is formed, not only does
the question word (or the phrase containing it) move to first position,
the auxiliary moves next to it. For rather than just getting (12.33) by
moving /what/, we get (12.34), which results in a second movement,
of the auxiliary /have/.
[[∅ [you will [ see what]VP ]TP ]C0 ]CP (12.32)
[what [∅ [you will [ see ]VP ]TP ]C0 ]CP (12.33)
[what [will+∅ [you [ see ]VP ]TP ]C0 ]CP (12.34)
The evidence is clear (on the basis of what we have said so far) that
the tense head /will/ must also move. If it does, we need to supply
a reason, as we did for the movement of the verb to T0 . It is believed
that the empty element in C0 actually is a question morpheme, call it
Q. The derivation therefore looks more like this:
What will you see?
[[Q [you will [ see what]VP ]TP ]C0 ]CP (12.35)
[what [Q [you will [ see ]VP ]TP ]C0 ]CP (12.36)
[what [will+Q [you [ see ]VP ]TP ]C0 ]CP (12.37)
In relative clauses, however, this movement does not happen. Calling
the head of a relative clause RC, this is the structure:
...the man [[RC [you will [ see who]VP ]TP ]C0 ]CP (12.38)
...the man [who [RC [you will [ see ]VP ]TP ]C0 ]CP (12.39)
We can say that the difference between Q and RC is that although
both are empty, only RC can be on its own, while Q wants to attach
to a host, and so triggers movement.
 syntax iv: movement and non-local dependencies 149

How Movement Works

We have already seen that movement must respect X-bar syntax.
Whatever the end result, it must conform again to X-bar syntax. We
have to begin with an important definition. Recall that a tree is a
pair hT , <i, where T is a set, the set of nodes and < the relation ‘is
(properly) dominated by’. (I use ‘is dominated by’ synonymously with
‘is properly dominated by’. This relation is never reflexive: no node
dominates itself.) x and y are said to be comparable if x = y or
x < y or x > y.

Definition 6 Let hT , <i be a tree and x, y ∈ T be nodes. x c-

commands y if and only if x and y are not comparable and the mother
of x dominates y.

This can be phrased differently as follows. x c-commands y if x is

sister to z and y is below z (or identical to it). Thus, c-command is
with any sister and its heirs. Recall the tree from Chapter 9, repeated
here as Figure 12.1. Here is a complete list of which nodes c-command
which other nodes:

c-commands
1 −
2 3, 6, 7, 8, 9
3 2, 4, 5
4 5
(12.40)
5 4
6 7, 8, 9
7 6
8 9
9 8

1 Figure 12.1: A Syntactic Tree

•

2 • • 3

4 • • 5 6 • • 7

this villa costs 8 • • 9

a fortune

The relation of c-command is inherited by the strings that corre-

spond to the nodes. For example, /this villa/ c-commands /costs
a fortune/ and its subconstituents.
Now, movement is such that the constituent that is being moved is
moved to a place that (i) is empty, and (ii) c-commands its original
150 an open introduction to linguistics

CP
•
Figure 12.2: Before Movement (Step
0)
• C0

C0 • • TP

T0
Q •

T0 • • VP

will DP • • V0

Simon • V0 • DP

like
which cat

CP
•
Figure 12.3: After Movement (Step 1)

DP • • C0

• C0 • TP
which cat
T0
Q •

T0 • • VP

will DP • • V0

Simon • V0 • DP

like ∅
 syntax iv: movement and non-local dependencies 151

place. To make this proposal work we assume that instead of lexical

items we can also find ∅ below a terminal node.
Trees 12.2 and 12.3 describe a single movement step, for the move-
ment that is otherwise denoted using strings as follows:

Will Simon like which cat? (12.41)

Which cat will Simon like ? (12.42)

s We have also added the labellings. The constituent consisting of

/which cat/ (which contains the node dominating these words plus
everything below it) is moved into a c-commanding position. It can
go there only if the node is absent. Thus, movement creates its own
constituent, but it must do so in agreement with X-bar syntax. In
place of the earlier constituent a node dominating ∅ is left behind.
(We shall say a little more about this element in the next chapter.)
For the whole story to go through it is assumed that the X-bar
grammar produces a number of choices for positions that can be used
for elements to move into. (You do not move into nodes that are
already there.) One such node is typically the specifier of CP. The
positions that an element moves into must also match in label (cate-
gory). For example, the C0 -position (which we have not shown above)
is also often empty, but a phrase cannot go there, otherwise the labels
do not match. It is believed that the verb moves there in the rele- A phrase has projection level 2.
vant examples from (12.3). You can check that X-bar syntax places
no restriction on the category of the specifier except that it must be a
phrase. Thus, wh-phrases can go there. The solution has an advantage
worth mentioning. CP has only one specifier, and therefore only one
of the wh-phrases can go there. The others have to stay in place.
Now, why do we also need the condition of c-command? To an-
swer this, one should realise that without restrictions the moving con-
stituent may be put into any position that is free. For example, take
this deep structure

which salesperson wonders should he promote which

(12.43)
product

The surface structure is

Which salesperson wonders which product he should

(12.44)
promote?

There are two specifiers of CP:

[CP [VP which salesperson wonders [CP [TP [T0 should [VP he
(12.45)
promote which product]]]]]]

Many possibilities are open now including a movement of the phrase

152 an open introduction to linguistics

/which salesperson/ downwards to the right of /wonders/.

∗
Wonders which salesperson he should promote (12.46)
which product?

As we have said, if the CP is filled by one wh-phrase, the other stays

in place, so this should be grammatical. But it is not. And the reason
for that is that movement has to be upwards, and into c-commanding
position.
Other possibilities (mentioned above that still need to be excluded)
is that the lower wh-phrase moves to the higher specifier (and /do/-
insertion):
∗
which product does which salesperson wonder he should
(12.47)
promote

To exclude this, we require first that every wh-phrase can only move to
the next CP. If it is unavailable, the wh-phrase stays in place. Second,
we shall say that /wonders/ blocks the movement of the wh-phrase.

(For Addicts:) Head Movement

I shall briefly discuss the problem of the movement of the verb to T0 ,
and that of the auxiliary to C0 . We have said that the moving head
combines with the already existing head in the structure. But how
can we make this compatible with X-bar syntax? First of all, by the
facts we need to assume that /sees/, which is the same as /see+s/,
actually is again a T0 . On the other hand, /see/ is a V0 , and /s/ is
a T0 . Thus we seems to have this structure:

[seeV0 sT0 ]T0 (12.48)

This is exactly what we shall propose. We shall supplement our X-bar

with one more rule, which covers the above case. Namely, we shall
allow adjunction also to zero level projections; however, adjunction is
only on the left, and only zero level projections can adjoin. The result
is again a zero level projection (since this is adjunction):

X0 → Y 0 X0 (12.49)

Additionally, we require that when a head is moved it must move to the

next c-commanding head and adjoin there. Adjunction is a movement
that splits up a single node of category X into two nodes with category
X, the upper one having the moved constituent and the lower X node
as its daughters. It is clear that the rule (12.49) makes adjunction to
heads an option. Let us perform this step with Figure 12.3.
Notice that we should now consider that all heads in the initial
structure are filled by either some lexical material or by a phonetically
empty head (as we have effectively required). This then gives the
structure in Figure 12.4. In order for everything to fall into place we
now need to adjust our notion of c-command. We shall say that a
 syntax iv: movement and non-local dependencies 153

CP
•
Figure 12.4: After Head Movement
DP •
(Step 2)
• C0

which cat
• C0 • TP

T0
T0 • C0 • •

will Q • T0 • VP

∅ DP • • V0

Simon • V0 • DP

like ∅

within a structure [Y 0 X 0 ]X 0 , all heads have the same c-command

domain, which is the one of the entire constituent.
The derivation is complete after we have moved the subject to the
specifier of TP.
CP
•
Figure 12.5: After Movement of Sub-
DP •
ject (Step 3)
• C0

which cat
• C0 • TP

T0
T •
0
C •
0 DP • •

will Q Simon T0 • • VP

∅ DP • • V0

∅ • V0 • DP

like ∅

Notes
It should be clear that a proper formulation of a seemingly simple
idea such as moving a wh-phrase to the front of a sentence needs a
lot of attention to detail. And this is not only because the facts are
unpredictable. It is also because what looks like a simple proposal can
154 an open introduction to linguistics

become quite complex once we start looking harder. This applies for
example to programming, where something that starts out as a simple
program can become quite long because we need to make sure it works
properly in all cases.

Study Questions
1. What is the Deep Syntactic Structure according to the account
given in this chapter before all movement has taken place:

a. When will Hans help Gretel?

b. Who will Hans help?
c. Which witch did Hans trick?
d. Who did Hans help?

2. For the Deep Structures you gave above, what is the string after
wh-movement has applied (e.g. at Step 1)?
13
Syntax V: How to Draw Syntactic Trees

Tree structures are an explicit representation of the implicit hi-

erarchical structure of sentences. Drawing trees is the most
common method for displaying the internal make-up of phrases,
clauses and sentences. but drawing trees requires practice. In
this chapter we go over again how to draw syntactic trees for a
subset of English sentences according to a simplified version of
Government and Binding theory. It also details when and how
you have to use intermediate X-bar projections, providing addi-
tional examples and explanations.

If you can identify word categories (parts-of-speech) and constituents,

then you should also be able to diagram sentences, given that you know
how the different structures fit together.
Let’s start by drawing trees of some simple constituents. Recall the
constituency test that we talked about in class called The Answer-
Ellipsis Test: if a sequence of words can stand alone as an answer
to a question, then it is a constituent. Consider the following small
dialogue about the fellowship of the ring:

Speaker A: What is the fellowship trying to do to the ring?

(13.1)
Speaker B: Destroy it. (13.2)

This dialogue identifies destroy it as a constituent. Destroy it is a verb

phrase, a VP for short. The head of a VP is a verb, destroy and in
this example, it is the DP complement. We can draw a tree structure
illustrating the relationship between the head and the complement:
VP

V DP

destroy it

Introducing the Tense Phrase (TP)

Some books use the term IP (inflec-
What would happen if Speaker A had asked a slightly different ques- tional phrase), instead of TP. We will
tion? E.g.: use TP throughout.
156 an open introduction to linguistics

Speaker A: As for the ring, what does the fellowship intend?

(13.3)
Speaker B: To destroy it. (13.4)

The fact that the sequence of words to destroy it can be used as

an answer to a question shows that it is also a constituent, though
you should check this claim with other constituency tests. What con-
stituent is it? To is an infinitive marker. One of the functions of the
infinitive marker to is to identify the phrase as non-finite. In Govern-
ment and Binding theory, to is therefore considered a tense marker,
and the phrase that it heads is a Tense Phrase (TP). It has the
same distribution as other lexical items that mark tense information.
We can add a branch to our existing structure to make the following
new tree: Observe that infinitival to and
TP auxiliaries like should and can occur
in the same position, e.g.
(a) It’s important [that Frodo should
T VP have the ring.]
(b) It’s important [for Frodo to have
the ring.]
to V DP
Observe also that both allow ellipsis
of their constituent VP, e.g. (a) and
destroy it (b).
To sum up, this phrase is a TP, headed by the infinitive particle to. (a) Frodo doesn’t want to leave the
The VP destroy it is the complement of to. shire but he should.
(b) Frodo doesn’t want to leave the
What we have now seen is that we can successively build up struc- shire but he has to.
tures by combining pairs of categories to form even larger phrases.
In generative grammar, Merge is the term used to describe this suc-
cessive combination of items to create complex structures. If we now
merge to destroy it with the verb trying we can make another com-
bination. The dialogue below confirms with a constituency test that
trying to destroy it is also a constituent.

Speaker A: What is the fellowship doing with the ring?

(13.5)
Speaker B: Trying to destroy it. (13.6)

The head of this phrase must be the verb trying, so the entire phrase is
thus a verb phrase. This verb then must take the TP phrase to destroy
it as its complement.
VP

V TP

trying T VP

to V DP

destroy it
 syntax v: how to draw syntactic trees 157

Problems with diagramming full sentences: introducing X-

bar

We have already made quite a long phrase. But we have not yet
analyzed a complete sentence. Consider the expanded dialogue here:

Speaker A: What is the fellowship of the ring trying to do about the ring?
(13.7)
Speaker B: They are trying to destroy it. (13.8)

If we assume that clauses and sentences are formed the same way
phrases are, we can also assume that the same type of merge operation
can be applied to form the full clause. So we should expect that we
can first merge the auxiliary verb are with the VP trying to destroy it,
and then merge They with the result.
Auxiliary verbs like are belong to the category T, just like the in-
finitive particle to. This would seem to suggest that trying to destroy
it is a projection of the auxiliary verb are and the entire sequence is a
TP, an auxiliary phrase.
But there is a problem. Are trying to destroy it isn’t actually a
constituent. Check for yourself that it doesn’t pass any constituent
tests. This means it cannot be a TP. Speaker A: What is the fellowship try-
If Are trying to destroy it is not a constituent, then it must be some ing to do? Speaker B: *Are trying to
destroy it.
kind of incomplete phrase. In what way does it seem to be incomplete?
It seems that auxiliaries require a subject, and are trying to destroy it
does not have a subject. Consider other auxiliary verbs like can and
would. These are also not constituents. Other examples of T-bar structures:
Following previous syntactic analysis, we will assume that a lexical am leaving on a jet plane.
has gone.
item from category T, like the auxiliary expression are, can merge can congratulate the president
with a VP to form an incomplete auxiliary expression. We will call
this incomplete expression ‘T-bar’, also written as T, but also as T’,
for easier typesetting. When the T’ structure merges with a subject
DP we can form a complete auxiliary phrase. This means we will get
the following structure:
TP

DP T

They T VP

are V TP

trying T VP

to V DP

destroy it
Again, we will describe the auxiliary are as the head of a phrase, and
158 an open introduction to linguistics

the VP trying to destroy it as the complement. We now describe

the pronominal determiner They as the specifier of the auxiliary verb
are. The head features of infinitival to in-
Now you may be wondering, what is it that tells us that are is the dicate that it is an infinitive particle;
its specifier features allow it to be used
kind of auxiliary that will require a specifier? First, realize that a without any overt specifier subject and
tree diagram only diagrams information about categorical constituent its complement features indicate that
it selects a complement headed by a
structure. It is only a partial representation of the entire grammatical verb in its infinitive form.
structure. We know that, in addition to belonging to a finite set of
word categories (parts-of-speech), words also have other features. We
saw this in our discussion of s-selection and c-selection.
In relation to syntax, syntacticians often talk of head features, which
describe intrinsic grammatical properties of words, complement fea-
tures, which describe what complements are possible, and specifier
features which define the kinds of specifiers a word can take. For the C is a functional category. The com-
auxiliary are we know that it has the head feature present, in that it plementizers that and if are [+FIN],
e.g they can only be used to intro-
has a present tense form, that it selects a complement that has a V duce a finite clause. For is an in-
head with the participial inflection -ing, and that it requires a specifier herently infinitival complementizer [-
FIN], and can introduce a clause with
that has nominative case and plural features.
the infinitive to but not a finite clause.
NOTE: Don’t confuse complements,
which are sentences (usually subordi-
Dealing with complementizer phrases (CPs) nate clauses), with the position ’com-
plement’ in a tree structure, where
What about words like that, for and if ? We can treat these a comple- ’complement’ is the counterpart of
specifier, and refers to the branch to
mentizers, that create Complementizer Phrases (CPs). We can expand the right of the X-bar.
on the above structure to create a CP in the following way:

Speaker A: What are you saying?! (13.9)

Speaker B: That they are trying to destroy it! (13.10)

CP

C TP

That DP T

they T VP

are V TP

trying T VP

to V DP

destroy it
Now we see that we have a sentence with a CP at the top. Actually,
all sentences are CPs, not just subordinate clauses. However, the C
position is empty in declarative main clauses in English. In English
yes-no questions, the CP head position is filled with the tensed head,
moved there via head-movement. For example:
 syntax v: how to draw syntactic trees 159

CP

C TP

Will DP T

you T VP

twill V DP

marry the farmer

In this case what we see is that the T from the TP can move to the
specifier-CP position to make a question. This makes sense, given that
questions and answers are similar to each other, e.g. a possible answer
to the above question is “I will marry the farmer”.

Notes on X-bar

A major point at the end of chapter 3 in the textbook was that all
phrases have a uniform structure. We called this X-bar structure, and
it looks like this:
XP Recall that there are three types
of DPs: Names, which are dia-
grammed as nouns with an empty de-
Specifier X terminer, Pronouns, which are deter-
miners without an empty noun phrase
complement, and full NPs, which are
X YP generally made up of a determiner plus
a common noun, and where we might
Head Complement get adjective adjuncts.
All phrases follow this X-bar structure, except for the Determiner
Phrase, DP, which is a special exception. The X-bar expansion of
a DP does not have a D-bar. Instead, in the place where you would
expect to find the D, there is an NP. This NP is special too, as it al-
ways has D as its specifier. This D is the head of the whole Determiner
Phrase. A complete, empty DP thus looks like this:
DP

Specifier D

D NP

head Specifier N

N Complement
An example of a phrase that follows the regular X-bar structure is the
PP below:
160 an open introduction to linguistics

PP

Specifier P

P Complement
However, you should know that this structure is not limited to one
X’. Sometimes, you need multiple X-bar levels to be able to correctly
diagram your sentence. This allows for one or more extra arguments
in between the specifier and the head, so-called adjuncts. In the
following tree, ZP is in the adjunct position:
XP

Specifier X

ZP X

adjunct X YP

Head Complement
You may have noticed that in the tree derivations in the previous
sections, we only use this extensive X-bar structure in one position:
for the TP headed by are. This is in fact the only place in the above
derivation where using the entire X-bar expansion is actually necessary.
We could have added the expansions everywhere, and if we had, we
would have gotten trees like the following:
 syntax v: how to draw syntactic trees 161

CP

TP

DP T

D T VP

D NP are V

They ... V TP

trying T

T VP

to V

V DP

destroy D

D NP

it ...
As you can see, there are numerous empty specifiers and complements.
These positions could contain lexical items, but for the sentences we
will be diagramming, they most often will not. What kind of items
could go in there then? Well consider this, for example:
162 an open introduction to linguistics

CP

TP

DP T

D T VP

D NP are V

They ... V TP

trying AP T

hard T VP

to AP V

quickly V DP

destroy QP D

all D NP

the N

rings
When the spec position is empty, which is frequently the case for many
DPs and TPs, the X-bar level is sometimes left, letting the head to
appear as a sister node to the complement, both directly under the
XP:
XP

X YP

Head Complement

Sometimes, when the phrase has no complement either, you can use
the even shorter form with the triangle notation, as we have done in
previous trees, mostly for DPs. For example:
 syntax v: how to draw syntactic trees 163

CP

TP

DP T

They T VP

are running
If an exam or homework question says that you should diagram using
(full) X-bar structures, then you should not use the shorthand form.
Otherwise, the shorthand form is fine (except of course, when you need
the X-bar structure to correctly diagram the sentence).
One final note: We have focused on learning to diagram tree struc-
tures for a small number of phrase and sentence types, without much
variation. But of course, X-bar structure must be applicable to all
possible structure in language. One question that has come up is how
to deal with multiple modifiers of the verb. Consider the sentence :

(1) Han read a book in the café on Tatooine.

This sentence has both a transitive verb with an object argument,

plus two modifiers of the reading event: it happened in a café, and
on Tatooine. We can interpret the cafe as being modified or we can
interpret this as two different modifiers of events. We’ll just look at
the latter. Basically for any bar-level we can add more bar-levels,
as long as it ends with a head. So the beginning of the diagram is
shown below. After seeing this example you should be able to apply
it to other structures, though there are seldom questions with such
complex structures in the homework or on the exam.
VP

DP V

Han V PP

V PP on Tatooine

V DP in the cafe

read a book
164 an open introduction to linguistics

Study Questions
1. What words could appear in the Spec-VP position between are
and trying in the example sentence They are trying hard to quickly
destroy the rings.

2. The sentences Han read a book in the cafe on Tatooine has (at
least) two possible syntactic structures. Diagram the VP so that on
Tatooine modifies the cafe rather than the reading event.
14
Syntax VI: Binding

Pronouns do not only refer to a particular person or thing, they

often refer to a particular constituent. This provides coherence to
a text. Binding theory is about the interpretation of pronouns as
well as the way in which they can be linked to other parts of the
sentence. Key concepts are Principles A, B and C in addition to
c-command and binding.

Pronouns
In this chapter we look at a phenomenon that is most intimately con-
nected with c-command, namely binding. Binding is, as we shall see,
as much of a semantic phenomenon as a syntactic one, but we shall
ignore its semantic aspects as much as we can. What is important here
is that there are several different kinds of DPs: ordinary DPs, names
and pronouns. Pronouns can be either reflexive (like /myself/,
/yourself/ etc.) or not (/he/, /she/, etc.). In addition, there are
also reciprocals (/each other/). In generative grammar, the term
anaphor is used to refer to both reciprocals and reflexives but this is
not the usage elsewhere, where an anaphor is an expression that refers
to something that another expression already refers to (and therefore
also non-reflexive pronouns generally count as anaphors).
In English, the reflexive pronouns exist only in the accusative. There
is no ∗ /hisself/, for example. There a number of other important The pronoun /her/ can both be gen-
pronouns, such as demonstratives (/this/, /that/) and relative itive and accusative, so /herself/
doesn’t clearly illustrate this fact. On
pronouns (/who/, /what/, /which/), which are used to open a rela- the other hand, it is /ourselves/ and
tive clause. not ∗ /usselves/. English has not
been completely consistent in arrang-
I could not find the CD [which you told me about]. (14.1) ing its case system.

There is a new film [in which no one ever talks]. (14.2)

The enclosed constituents are called relative clauses. They are used
to modify nouns, for example. In that respect they are like adjectives,
but they follow the noun in English rather than preceding it, e.g.:
∗
I could not find the which you told me about CDs. (14.3)
∗
There is a new in which no one ever talks film. (14.4)

The relative clause is opened by a relative pronoun, /which/, to-

gether with the preposition that takes the pronoun as its comple-
166 an open introduction to linguistics

ment. One should strictly distinguish the relative pronoun from the
complementizer /that/ as found in sentential complements (/believe
that/). The complementizer occupies a different position (C0 , Spec-
CP). Some relative clauses can be opened by /that/.

I could not find the CD [that you told me about]. (14.5)

Complementizers: where did
This use of /that/ seems to be a hybrid between the complementizer they come from? It is common
and a relative pronoun (recall that /that/ is also a demonstrative). for the complementizers of a language
to develop historically from other syn-
(And syntacticians are divided on the question whether to call it a rela- tactic categories (a process known as
tive pronoun or a complementizer in this usage.) The relative pronoun, grammaticalization). Across the lan-
guages of the world, it is especially
for example, can move together with its preposition (/the film in common for pronouns or determin-
which I played piano/), but you cannot say /the film in that I ers to be used as complementizers
played piano/; instead, you have to say /the film that I played (e.g., English that). Another frequent
source of complementizers is the class
piano in/. Typically, a relative clause either has a relative pronoun or of interrogative words. It is espe-
a complementizer; this has become an iron rule of generative grammar cially common for a form that other-
wise means what to be borrowed as a
("Doubly filled COMP filter"). But facts are different. Some dialects complementizer, but other interroga-
of German give us evidence that the relative pronoun and the com- tive words are often used as well; e.g.,
plementizer really can co-occur. In Swabian German, /wo/ is used colloquial English I read in the pa-
per how it’s going to be cold today,
as a complementizer for relative clauses, accompanied by a relative with unstressed how roughly equiv-
pronoun. alent to that). English for in sen-
tences like I would prefer for there
to be a table in the corner shows a
die Leit die wo beim Daimler schaffet (14.6) preposition that has arguably devel-
the people who that at.the Daimler work oped into a complementizer. The se-
quence for there in this sentence is not
the people who work at the Daimler Benz factory. a prepositional phrase under this anal-
ysis. (from Wikipedia, Complemen-
tizer, 3/28/2015)
As with questions we like to think of the relative clauses as being
derived from a structure in which the relative pronoun is in the place
where the verb expects it.

the CD [∅ you told me about which] (14.7)

a new film [∅ no one ever talks in which] (14.8)

The position of specifier of the CP is vacant, and the relative pro-

noun wants to move there. (The position of C is also empty, but we
like to think that there is a silent C that sits there. Anyway, phrases
can never go there.) Sometimes the relative pronoun goes alone (being
a DP, hence a phrase, it can do that), sometimes it drags the P along.
The latter is known as Pied-Piping, from the fairy tale of the piper
The Pied Piper
who promised the city to get rid of the rats...
Different languages have different systems. Latin does not distin-
guish reflexive and irreflexive pronouns in the 1st and 2nd person,
simply using the object forms. A separate reflexive is also unnecessary
for the first and second person, e.g.I like me cannot have an interpre-
tation different than I like myself because I, me and myself must all
be interpreted as the current speaker, while in the third person there
is a clear meaning difference, e.g. He likes him can means something
different than He likes himself (though binding principles will also rule
this ambiguity out, as we will soon see). Latin does however have two
 syntax vi: binding 167

3rd person pronouns, /is/ (‘he’), and /se/ (‘himself’). The reflex-
ive exists in all cases but the nominative. It is the same both in the
singular and the plural, see below:

nom −
gen sui
dat sibi (14.9)
acc se
abl se
Latin also has another set of pronouns called possessive. They are used
to denote possession. They are just like adjectives. For example,

equus suus (14.10)

horse-nom his-nom
equum suum (14.11)
horse-acc his-acc

Binding
These different kinds of expressions each have their distinct behaviour.
Look at the following three sentences.

John votes for John in the election. (14.12)

John votes for himself in the election. (14.13)
John votes for him in the election. (14.14)

There is an understanding that the two occurrences of the word /John/

in (14.12) point to two different people. If they do not, we have to use
(14.13) instead. Moreover, if we use (14.13) there is no hesitation as to
the fact that there is an individual named /John/ which casts a vote
for the same individual (also named /John/). If we use (14.14), finally,
we simply cannot mean the same person, /John/, by the word /him/.
John cannot be self-voting in (14.14), he votes for someone else.
There are several ways one may try to understand the distribution
of full DPs, pronouns and reflexives. First of all, however, let us notice
that a reflexive pronoun expects the thing it refers to be the same as
something else in the sentence. The expression that denotes this is
called the antecedent of the pronoun. In (14.13), for example, the
antecedent of /himself/ is /John/. To express that some constituent
refers to the same object we give them little numbers, called indices,
like this.
∗
John1 votes for John1 in the election. (14.15)
John1 votes for himself1 in the election. (14.16)
∗
John1 votes for him1 in the election. (14.17)

We have already assigned grammaticality judgements. Needless to say,

any other number (say, 112, 7, 56 or 34) would have done equally well,
so that as far as syntax is concerned (14.15) is the same as
∗
John34 votes for John34 in the election. (14.18)
168 an open introduction to linguistics

However, in (14.12) the two occurrences of /John/ are not meant to In the books you often find letters i,
point to the same individual. If they are supposed to point to different j and k in place of concrete numbers,
but this not a good idea since it sug-
individuals, we write different numbers: gests that plain numbers are not ab-
stract enough. But in fact they are.
John1 votes for John2 in the election. (14.19)

The numbers are devices to tell us whether some constituent names

the same individual as some other constituent, or whether it names a
different one.
Non-reflexive pronouns seem to encourage a difference between sub-
ject and object in (14.14), and similarly with names (14.12). However,
things are tricky. In (14.20) the genitive pronoun /his/ could be taken
to refer not only to someone different from John, but also to John him-
self. Just an aside: the reflexives cannot be
used in genitive, they only have ac-
cusative forms. This may be the rea-
His lack of knowledge worried John. (14.20)
son why we do not find them here, but
the theory we are going to outline here
So, it is not really the case that when we have a pronoun that it must tries a different line of argumentation.
be used to talk about a different person than the others in the sentence
(= that it must have a different index).
The conditions that regulate the distribution of these expressions
are at follows. First, we define the notion of binding.

Definition 7 A constituent X binds another constituent Y if X c-

commands Y and X and Y have the same index.

Binding is an interesting mixture between semantic conditions (car-

rying the same index, hence talking about the same individual) and
purely structural ones (c-command). A note of warning is in order.
Constituents are strings, but we talk here as if they are nodes in a
tree. This confusion is harmless. What we mean to say is this: sup-
pose that x and y are nodes and the corresponding constituents are X
and Y . Then if x c-commands y and has the same index, then x binds
y, and X binds Y . So, X binds Y if there are nodes x and y such that
x binds y and X is the constituent of x and Y the constituent of y.
Now we are ready to say what the conditions on indexing are.

Principle A.
A reflexive pronoun must be bound by a constituent in the same
CP (or DP).
Principle B.
A non-reflexive pronoun must not be bound by a constituent
inside the same CP.
Principle C.
A name must not be bound.

For example, (14.12) can be used successfully to talk about two in-
dividuals named John. So, the rules alone cannot tell us whether a
sentence is grammatical, but the rules do tell us sometimes what the
possible meanings are.
To understand the role of c-command, we need to look at the struc-
ture of some sentences. The subject c-commands the object of the same
 syntax vi: binding 169

verb, since smallest constituent containing the subject is the VP, which
also contains the object. However, the object does not c-command the
subject, since the smallest constituent containing it is only the V0 . So
the following is illegitimate no matter what indices we assign:

∗
Himself voted for John. (14.21)

The subject also c-commands all other arguments. This accounts

for the correctness of the following.

The queen was never quite sure about herself. (14.22)

The students were all talking to themselves. (14.23)

Notice that the principles not only say that these sentences are fine,
they also tell us about the assignment of indices. They claim that
(14.24) is fine but (14.25) is not.

The queen1 was never quite sure about herself1 . (14.24)

∗
The queen1 was never quite sure about herself2 . (14.25)

This is because the reflexive (/herself/) must be bound inside the

sentence; this means there must be some antecedent c-commanding it The queen being pensive.
having the same index. In (14.24) it is /the queen/, but in (14.25)
there is no such constituent. Thus, the overall generalization seems to
be good. Notice that Principle A says that the binder must be in the
same clause. Thus the following contrast is explained.

[John thinks [that Alice likes herself.]CP ]CP (14.26)

∗
[John thinks [that Alice likes himself.]CP ]CP (14.27)

In both (14.26) and (14.27), the reflexive pronoun must be bound

inside the lower CP, by Principle A. However, the only binder can be
/Alice/, and not /John/, and thus (14.27) is ungrammatical.
Now we look at pronouns. The situation where a non-reflexive pro-
noun should not be used in the same sentence is when actually a re-
flexive would be appropriate according to Principle A. For example,
the following sentences are ruled out if meant to be talking about the
same individual(s) in the same sentence (that is, if /her/ refers to /the
queen/ and /them/ refers to /the students/).

∗
The queen1 was never quite sure about her1 . (14.28)
∗
The students1 were all talking to them1 . (14.29)

By the same token, if we use numbers we can also write:

∗
The queen1 was never quite sure about herself2 . (14.30)
∗
The students1 were all talking to themselves2 . (14.31)

Here, different numbers mean that the expressions are meant to refer
to different individuals or groups. Principle B talks about binding
170 an open introduction to linguistics

inside the same CP. It does not exclude that a pronoun is bound in a
higher CP. This is borne out.

[John1 thinks [that Alice3 likes him1 .]CP ]CP (14.32)

Notice that the contrast between a reflexive and a non-reflexive pro-

noun only matters when the antecedent is c-commanding the pronoun.
We repeat (14.20) below:

His1/2 lack of knowledge worried John1 . (14.33)

Here, /John/ is the antecedent. We can take the sentence to mean

that John is worried about his own lack of knowledge, or that he is
worried about someone else’s lack of knowledge. In none of the cases
would a reflexive pronoun be appropriate.
Let us now change the position of the two:

∗
John1 ’s lack of knowledge worried himself1 . (14.34)

Here, /John/ does not c-command the pronoun. But the reflexive must
be bound by something that c-commands it inside the clause. This
can only be /John’s lack of knowledge/. But if we think that, we
would have to say /itself/ rather than /himself/. Next, why is
(14.35) fine?

John1 ’s lack of knowledge worried him1 . (14.35)

Even though it has the same index as /John/, it is not c-commanded,

hence not bound by it. Having the same index nevertheless means
that they refer to the same thing (John), but for syntax binding takes
place only if c-command holds in addition. Hence, the next sentence
is also fine for syntax:

John1 ’s lack of knowledge worried John1 . (14.36)

Admittedly, we would prefer (14.37) over (14.36), but it is agreed that

this is not a syntactic issue.

His1 lack of knowledge worried John1 . (14.37)

Movement Again
We have argued earlier that an element can only move into a position
that c-commands the earlier position. It seems that binding theory
could be used in this situation. However, in order to use binding
theory, we need to have an element that is being bound. Recall from
previous discussions that when an element moves it leaves behind an
empty element. But why is that so? To see this, look at movement
out of a subordinate clause.

Who do you think answered which question? (14.38)

∗
Who do you think which question answered? (14.39)
 syntax vi: binding 171

In these sentences, it is the subject of the subordinate clause which

moves to the specifier of CP of the main clause. If that was all that
happened, we would expect that the specifier of the subordinate CP
would be empty and so the object /which question/ would have to
move into it. But this is not the case. Thus we conclude that the
specifier of the lower CP is not empty. It is filled by something that
we cannot hear. That thing is the element that /who/ left behind,
when it moved to the higher specifier. So the full proposal is this:
movement of a wh-phrase must be first to the specifier of CP of the
same clause. After that it is to the specifier of CP of the next higher
clause; and so on. Each time it moves, it leaves behind an empty
element, called a trace. The constituent in the new position and the
trace will be co-indexed. Finally, we declare that traces are some sort
of pronouns, and hence must be bound.
Condition on Traces.
All traces must be bound.
If we assume this much it follows that movement must inevitably be
into a position that c-commands the original position. It does some-
thing else, too. It ensures that whatever thing we choose to interpret
the moved element by, it will be used to interpret the trace.

Pilots1 Harry admires t1 . (14.40)

The relevant structure of (14.38) is now as follows.

[Who1 do you think [t1 [answered2 [t1 t2 which (14.41)

question]TP ]C0 ]CP ]CP

The t1 in specifier of CP blocks the movement of /which question/

there. However, though movement there is impossible, it is also not
required, so the sentence is grammatical.
One may wonder why this is a good theory. It postulates empty
elements (traces), so how can we be sure that they exist? We do
not see them, we do not hear them, so we might as well assume that
they do not exist. Furthermore, (14.40) is not the sentence we shall
see in print or hear, so this actually adds structure that is seemingly
superfluous.
Opinions on this diverge. What is however agreed is that empty
elements are very useful. We have used them occasionally to save our
neck. When nouns get transformed into verbs, no change is involved.
Other languages are not so liberal, so it is not the nature of nouns
and verbs that allows this. It is, we assume, an empty element which
English has (and other languages do not) which can be attached to
nouns to give a verb. At another occasion we have smuggled in empty
complementizers. You will no doubt find more occasions on which we
have made use of empty elements.

Binding And Agreement

Pronouns distinguish not only case and number but also gender in
English. Crucially, when a pronoun becomes bound it must agree in
172 an open introduction to linguistics

number and gender with its binder (not in case though).

John voted for himself. (14.42)

∗
John voted for herself. (14.43)
∗
John voted for themselves. (14.44)
The committee members voted for themselves. (14.45)
Mary voted for herself. (14.46)

The fact that agreement holds between binder and pronoun means
that certain indexations are not licit. Here is an example.
∗
Her1 lack of knowledge worried John1 . (14.47)
∗
Their disinterest in syntax1 bothered John1 . (14.48)

Also, we have said that a pronoun cannot be bound inside the same
CP. But it can be bound outside of it:

John1 told his boss2 that he1/2 looked good. (14.49)

In the previous example, both indices are licit, since binding by /John/
or by /boss/ is compatible with agreement. However, in the following
sentence one of the options is gone.

John1 told Mary2 that she∗ 1/2 looked good. (14.50)

For if the pronoun is bound by /John/, it must agree with it in gender,

which it does not. So it can only be bound by /Mary/.

The Difference Between Morphology and Syntax

Some languages will use separate words to mark syntactic structure,
while other languages will modify the morphology of the head word.
This means that the difference between PPs and case marked DPs can
be seen as a difference between morpheme and lexeme (or morphology
and syntax). Finnish, for example, has a case to express movement
into a location, and a case for movement to a location:

Jussi menee talo-on. (14.51)

Jussi goes house-into = Jussi goes into the house.
Jussi menee talo-lle. (14.52)
Jussi goes house-to = Jussi goes to the house.

When it comes to other concepts—like ‘under’ and ‘over’—Finnish

runs out of cases and starts to use postpositions:

Jussi menee talo-n alle. (14.53)

Jussi goes house-gen under = ‘Jussi goes under the house.’
Jussi menee talo-n yli. (14.54)
Jussi goes house-gen over = ‘Jussi goes over the house.’

Hungarian does have a case for being on top of, but in general the
situation is quite analogous to Finnish.
 syntax vi: binding 173

Thus, the cut-off point between morphology and syntax is arbitrary.

However, the two may be shown to behave differently in a given lan-
guage so that the choice between morphological and syntactical means
of expression has further consequences. We have seen, for example,
that the comparative morpheme /er/ is a suffix—so it is added after
the word. However, the comparative lexeme /more/ wants the adjec-
tive to its right. The latter is attributable to the general structure of
English phrases. The complement is always to the right. Morphemes
are exempt from this rule. They can be on the other side, and gen-
erally this is what happens. For example, verb+noun compounds in
English are formed by placing the verb after the noun: /goalkeeper/,
/eggwarmer/, /lifesaver/ and so on. If these were two words, we
should have /keeper goal/, /warmer egg/, and /saver life/.
The reason why we do not get that is quite interesting and related
to syntax. English used to be a language where the verb follows the
object (as is the case in German). It then changed into a language
where the verb is to the left of the object. This change affected only
the syntax, but did not spread toe morphology. Contrastingly, French
forms compound the other way around (/casse-noix/ lit. cracker-nut
= ‘nutcracker’, /garde-voie/ lit. guard-way = ‘gatekeeper’). This
is because when French started to form compounds, verbs already
preceded their objects. You are however warned that we do
not really know that what is nowa-
days written as /u/ is the same sound
The Latin Perfect—A Cabinet of Horrors as what is now written /v/; this is
an inference, partly based on the fact
This section serves to illustrate that the same meaning can be signalled that the Romans used the same letter.
Probably pronunciation was different,
by very different means; and sometimes by more than one at the same but the difference was, for all we know,
time. To start, there is a large group of verbs that form the perfect not phonemic.
stem by adding /v/. (3rd person does not specify gender; we use ‘he’
only as an abbreviation of the longer ‘he/she/it’.) The ending of the
3rd singular perfect is /it/. The ending in the present is /at/, /et/
or /it/, depending on the verb.

amat ‘he loves’ amavit ‘he has loved’

delet ‘he destroys’ delevit ‘he has destroyed’ (14.55)
audit ‘he hears’ audivit ‘he has heard’

We also find a group of verbs that form the perfect by adding /u/.

vetit ‘he forbids’ vetuit ‘he has forbidden’

(14.56)
habet ‘he has’ habuit ‘he has had’

The difference is due to modern spelling. The letters /u/ and /v/ were
originally not distinct and denoted the same sound, namely [u], which
became a bilabial approximant in between vowels. The difference is
thus accounted for by phonemic rules.
Other verbs add an /s/. The combination /gs/ is written /x/.

regit ‘he reigns’ rexit ‘he has reigned’

augit ‘he fosters’ auxit ‘he has fostered’ (14.57)
carpit ‘he plucks’ carpsit ‘he has plucked’
174 an open introduction to linguistics

There are verbs where the perfect is signalled by lengthening the vowel
(lengthening was not written in Roman times, we add it here for illus-
tration):

iuvit ‘he helps’ iūvit ‘he has helped’

(14.58)
lavat ‘he washes’ lāvit ‘he has washed’

Sometimes this lengthening is induced by a loss of a nasal:

rumpit ‘he breaks’ rūpit ‘he has broken’

(14.59)
fundit ‘he pours’ fūdit ‘he has poured’

Finally, there are verbs which signal the perfect by reduplication:

currit ‘he runs’ cuccurrit ‘he has run’

(14.60)
tendit ‘he stretches’ tetendit ‘he has stretched’

Now, things are made difficult by the fact that certain verbs use a
combination of these. The verb /frangere/ (‘to break’) is an example
of this: it uses loss of nasal, accompanied by vowel lengthening, plus
ablaut (vowel change):

frangit ‘he breaks’ frēgit ‘he has broken’ (14.61)

The lengthening of the vowel could be attributed to the loss of the nasal
(in which case it is called compensatory lengthening). However,
the next verb shows that this need not always be the case.
The verb /tangere/ uses a combination of reduplication, loss of
nasal, and ablaut, without lengthening the vowel):

tangit ‘he touches’ tetigit ‘he has touched’ (14.62)

The root is /tang/. The nasal is dropped, yielding /tag/. Redupli-

cation yields /tetag/ (actually, it should give ∗ /tatag/.) Finally, the
vowel changes to give /tetig/. The change of vowel is quite common
in the reduplicating verbs since the reduplication moves the stress away
from the root.

What Syntax Gets To See

We have seen above that the comparative of adjectives is sometimes
formed by adding a morpheme and sometimes by using a separate
word. It is unfortunate, however, that there is such an asymmetry.
For from a syntactic perspective, there is not much of a difference
between the two. It is for this reason—and others—that people have
sought to expand the scope of syntax downwards into morphology.
Linguists once used to think of the form /thinks/ as a verb form
with some features, and thus as a syntactically indecomposable whole.
Nowadays, however, one sees /thinks/ as syntactically complex. This
brings us back to Chapter 12, Page 152. There we said that the reason
for head movement was that the verb form /thinks/ is syntactically
complex, but that its parts are distributed in the sentence in the wrong
way. Given that the string /s/ wants to be added to the end of a verb,
 syntax vi: binding 175

and that the verbal root cannot be pronounced as such, the solution is
to move the verb upwards, and form a complex head [C [V think][C s]].

The analysis of words as complex heads has proved to be a very

fruitful idea. With verbs actually hosting quite a number of other
morphemes (agreement (in person, in number, in gender), voice (active
or passive), aspect (completed or ongoing activity), definiteness, to
name a few), the scope of syntax expanded massively at the expense
of morphology.

Study Questions
1. Add indices to the referential expressions in the following sentences
to show binding. Clearly identify any potential ambiguities.

a. John washed himself.

b. John’s brother washed him.
c. Jayne thinks that Alice likes her.
d. Mary hid Jayne’s book behind her.
15
Semantics I: Fundamentals of Meaning and its Rep-
resentation

Semantics is the study of meaning. Lexical semantics is the study

of word meaning, and how words relate to each other. Formal
semantics focuses on structural meaning. Formal semantics is
intimately connected with logic and reasoning.

What is meaning?
What is meaning? What does it mean to understand a word, or a
statement that someone makes? How do we know that we understand
the same things when we hear the word blue? When I remark that
the sky is very blue today, and you agree, how do we then know that
we both have the same understanding of the word “blue”? These are
fundamental questions about what meaning is.
But in order to analyze meaning and deal with it in our applications,
we have to also consider how we can represent it.
These are all questions dealt with in semantics, the study of mean-
ing. Humans have attempted to understand what meaning is for cen-
turies. The type of semantic theory that is most relevant for compu-
tational applications is called model theoretic truth-conditional seman-
tics. This is also where most formal semantic research is carried out.
In this chapter, we’ll introduce the fundamental ideas of this theory
and how it represents meaning.
First let’s consider how we understand a simple word. Take a con-
crete noun like apple. How do we understand that someone is referring
to an apple when they say the word “apple”, and not, for example, a
pear or a chair? One (rejected) theory of linguistic meaning is that
when you hear a word, it conjures up a picture of what that word can
refer to in your mind. Now for the word apple, this doesn’t seem to be
such a crazy idea. But what about verbs? For the word eat, we might
also be able to envision in our mind what action is involved, and if our
speaker says to us, I ate an apple, we might see that action occurring
in our minds and thus be able to understand it.
This theory of meaning suggests that when we listen to someone,
we see a stream of pictures in our mind, like watching a film that is
constantly playing. For speech production, this theory claims we rely
178 an open introduction to linguistics

on visual images to link our thoughts to the words we want to say. But
the hypothesis that meaning is related to visual pictures is problematic
when we examine it further. First, how do we know that we are all
using the words to mean the same thing? My experience with apples
may be different than yours, so does that mean that the picture of
an apple I have in my mind will be different than yours? And if that
is the case, will it affect our ability to use language to communicate?
Maybe for apples it won’t matter so much, but what if the meaning
of the word ‘red’ to me is what you call the color ‘orange’? At what
point do divergences in our mental imagery start to impede smooth
communication? Second, how did we learn the meaning of words? We
are not born knowing the words in our language or the concepts that
they apply to, and in fact, depending on the language we are exposed
to as children, we learn different words (and perhaps different concepts,
but more on that later). We know that we learn the meaning of most
concrete words through our interaction with the world: mothers point
to or give their children apples and say apple or ringo or pommes
and children then interact with the world and the object and learn
it. So even if we did rely on mental imagery to understand sentences,
we would still need the world around us to mediate our acquisition of
the words and their associated images. Note though, that this solves Use vs. Mention
part of the problem of determining whether or not different individuals When you use a word with its mean-
ing, we don’t have to mark it in any
use words to mean the same things: because we all live in the same special way. But when we talk about
physical world, and are introduced to words via our interaction with the word itself, we say that we are
mentioning it, and then we need to
that world, we can be reassured that despite different experiences, the signal this. That’s why in the text
meanings we ascribe to many words will be similar among speakers of when I talk about the word “Jan”, I
put quotes around it, or sometimes we
the same language.
use italics to signal mention.
But a visual imagery theory of meaning ignores the fact that we can
learn the meaning of new words by the context in which they occur,
without any visual imagery necessary. If I say to you “I have a pet
wug”, without any imagery you can still understand something about
wugs (that they can be pets, that they are a kind of animal) without
any visual imagery. Clearly the relationships between known words,
and constraints they signal about the meaning of unknown words, are
also part of meaning that doesn’t seem to be handled well with visual
imagery alone.
A visual imagery theory of meaning also ignores the fact that peo-
ple have very strong intuitions about whether or not a given sentence
correctly describes a situation in the world or not. Surely this re-
lationship between the world and language must be relevant for our
understanding of language.
A final problem with a theory of meaning based on mental images
is defining the meaning of abstract words like confidence, nothing, the,
above and loud. What about negation, in sentences like There is no
such thing as a pink elephant or I never eat breakfast. For these and
other reasons (for example, blind people also learn to speak their native
language, so visual imagery is not a necessity for language) this simple
idea was rejected years ago. Hopefully I’ve convinced you that the
idea that visual imagery moderates linguistic meaning becomes more
 semantics i: fundamentals of meaning and its representation 179

and more cumbersome the more we explore it. I present it only as a

comparative example to better understand an alternative theory: a
referential theory of meaning, also called the denotational theory
of meaning, where the meaning of words is their reference.
What is reference? Reference is the ability of a sign (=word) to be
used to pick out something in the world. You can think of reference as
a kind of linguistic pointing, though instead of using your index finger
you can simply use words. The easiest examples are unique “pointers”
such as proper names. So the name “Barack Obama” refers to (picks
out) one unique individual in the world, the former U.S. president
named “Barack Obama”. The name “Jan”, one of the most common
names in Dutch, can be used to pick out an individual known as “Jan”.
Because we use language in often clearly defined contexts, it’s fine that
many of the proper names we use are not actually unique in our lan-
guage, because they are often unique in the context in which they
are being used. So if I start talking about “Jan”, our mutual friend,
you will naturally assume I am using the name “Jan” to refer to our
mutual friend, and not some Jan that I know but you do not. This
example also illustrates that even though many words are ambiguous
without a context, within a context, where participants in the conver-
sation share information (share common ground), words are not (or
seldom) ambiguous.
Now what is the reference of common nouns like apple and chair?
When I say the word “apple”, the word can be used to refer to any
member of the set of apples in the world. If I say “Look! An apple”,
I am using the phrase “an apple” to refer to an individual member of
the set of apples that you can see. Again, context is important here. If
I say The apples are all rotten! you can usually assume I do not mean
all the apples in the world, but some relevant set of apples, such as the
ones in the fruit bowl we are currently looking at. Of course, if there
was some world-wide apple blight, which causes all the apples in the
world to rot, then I could use the same statement to mean that all the
apples in the world are rotten. But then the context would hopefully
make this clear. These examples are of concrete common nouns. Of
course, things get more complicated when we move to more abstract
meanings, so let’s begin with the concrete. Set Theory Overview:
• Things in a set are the members
of the set
Interpretation within a model • The empty set, ∅, has no members
– This is the set of unicorns,
The context in which different linguistic expressions are evaluated leads the set of human-elephant hy-
to different interpretations, which is in part defined by the shared brids, the set of elves enrolled
in ATW, etc.
knowledge of the conversants. At this point it seems that to truly
have a formal way to interpret words in a context, we have to have • The number of members in a set is
called its cardinality.
someway to model the context, or in other words, to make a model of
– E = John, Paul, George, Ringo,
the world within which our interpretations will be made. A common |E| = 4
formal way to model the context of interpretation is to use set theory. – S = 1,2,3,4,5,6, |S| = 6
So we say that the word apple picks out the set of things that are
• An element is either a member of
apples, and the adjective red picks out the set of things that are red. a set, or not
We can even model simple sentences with intransitive verbs using set – George ∈ E
– 4∈
/E
180 an open introduction to linguistics

theory.

(1) Jan walked.

(2) Han shoots
(3) Jan talks and Peter walks or talks.

In our model, example (1) will be true when the individual named
John (so an individual member) is a member of the set of those who
walked. We often write it in a form of predicate logic, e.g. walked(john).
For example (2), we’ll simply say shoots(han). For (3), we can use con-
nectors from logic to create talks(j) ∧ walks(p) ∨ talks(p) (ignoring for
the moment that we need to specify the scope of each operator).
What I have now been doing is describing, in set-theoretic terms,
how the model of the world has to look for the examples statements
to be true. So actually, in addition to the idea of reference, and the
model organized with set theory, I’m also relying on a theory of truth.
At this point we really should introduce some terminology and be-
come more consistent with how we talk about different things. Sen-
tences, or more precisely, statements, are linguistic strings. They
communicate something (in a context) that we will call a proposi-
tion. We need to make a distinction between the sounds or words
strung together that makes a statement, and the proposition it com-
municates. We can now say something about what makes a good
translation from one language to another: good translations commu-
nicate the same propositions. So the statement in (4) in English and
the statement in Dutch in (5) express the same proposition and will
thus be true in the same models and false in the same models.

(4) Darth Vader is scary.

(5) Darth Vader is eng.

The same statement can communicate more than one proposition, be-
cause a statement can be ambiguous.

(6) Three ewoks captured one imperial soldier.

How many imperial soldiers were captured? There are two possible
interpretations, or propositions, which can be paraphrased as the fol-
lowing:

(7) Three ewoks individually captured one imperial soldier each, so

all together three soldiers were captured.
(8) Three ewoks working as a group captured one imperial soldier
together.

Now I explained the possible propositional meaning of a statement in

English (Example (6)), using English paraphrases in (7) and (8)). One
of the problems with doing semantics is that we need to talk about
the meaning of words and statements in an language, such as English,
using English. We say that English is our , the object of our study, but
 semantics i: fundamentals of meaning and its representation 181

we then use English as a to talk about the object language, English.

If we use the same language as both our object and metalanguage,
we create problems. We need to be able to distinguish between the
meaning of a word or sentence, and the concept or proposition it ex-
presses. We also need this to be clear, so that we don’t get mixed up.
A second problem has to do with paraphrasing itself. We could, for ex-
ample, paraphrase our English sentences in German, and use German
as our metalanguage. But the problem with this method is that some-
times it’s hard to get a good paraphrase. What should the paraphrase
of (4) be for example? That Darth Vader causes a negative feeling
called fear in people? Is this the right way to describe something that
is scary?
An easier way that bypasses the need to define words is to simply
mark somehow that we mean the meaning, not the word (sign), and
ignore the tricky business of specifying the meaning unless it is our
focus of attention. Two ways to do this are with apostrophes or with
double brackets, e.g.

(9) Darth_Vader’ is’ scary’.

(10) [[Darth_Vader is scary]]

We can then discuss whether or not the proposition expressed by the

statement in (4) is a true or false proposition (the (9) with respect
to a given model. So with this distinction, (4) is a sentences, that
technically does not have a truth value, while (9) or (10) are proposi-
tions that (4) can express, and they can be true or false w.r.t. a given
model. Sometimes we ‘skip the middleman’ and in a somewhat sloppy
way talk about sentences being true or false, but technically that is
incorrect. This will also be problematic in cases of ambiguity (e.g.
(6)). Propositions thus have truth values. In this short introduction
to semantics we will only deal with simple statements (declarative
sentences), but we should mention that besides statements, language
also has questions and commands. Questions ask for information,
and thus cannot be said to be true or false. Commands tell people
to do something, and also cannot be said to be true or false. How
semantics deals with these locutions is beyond an introduction. Note
also that only a complete statement can be true or false, contribut-
ing a proposition, so something like a noun phrase in (11) cannot be
evaluated in a model.

(11) a red apple

So now that we’ve dealt with terminology, we need to have a principled

way to determine if a given proposition is true or false in our model.
Let’s look at at some intransitive sentences more closely:

(12) John walks.

(13) Mary laughed.

Syntactically, these sentences are intransitives with only one argument,

182 an open introduction to linguistics

the subject DP. Each proposition expressed by the statements is true

if and only if the individual that is referred to by the subject DP is
a member of the set denoted by the verb. This seems to always be
the case with intransitives. Thus we can write a simple rule for the
interpretation of intransitive sentences.

Interpretation rule for intransitive sentences

with a proper name as subject
DP V’intransitive –> VP’(DP’)
V’(DP’) is true iff DP’ ∈ V’

We can also write rules for a simple copula sentence with an adjec-
tive phrase ascribing a feature to an individual. For example:

Interpretation rule for copula + adjective

with a proper name as subject
DP is AP –> AP’(DP’) An unidentified flying object near a
AP’(DP’) is true iff DP’ ∈ AP’ windmill in the Netherlands. Source:
secret Dutch government archive.

Note that the set-theoretic relationships between the words in a

statement are made clear by the syntax of the sentence: we need to
know what word is the verb to identify what set the subject needs to be
a member of, and we need to know what the subject is as well. This is
the connection between syntax and semantics: the syntactic structure
of a sentence is essential to its interpretation. Unfortunately, this is
about as far as we can get without adding to our technical apparatus.
In order to deal with for example transitive sentences, we will need to
start using more sophisticated functional application, which requires
using lambda calculus, which is beyond the scope of what we can do
in a short introduction, But the two rules above should give an idea
of how a comprehensive semantic theory can be developed.
Let’s take stock. What does our semantic theory have now?

• A theory of reference (how we link linguistic symbols to sets in a

model)

• A model in which propositions can be interpreted (determined to

be true or false) relying on set theory

• Truth values which identify whether or not propositions are consis-

tent with the model

Basically what we now have is a model-theoretic referential theory of meaning.

But this is still not really going to be sufficient.

Truth alone is insufficient: truth-conditional semantics

Now we have a way to interpret statements. We can determine (us-
ing an appropriate set-theoretic model) whether they are true or false
 semantics i: fundamentals of meaning and its representation 183

within the model. But this actually is going to be problematic. For

example, if I say a statement such as the following:

(14) The Dutch government has proof that aliens have visited earth.

Do you understand what the sentence above means? Hopefully the

answer is ‘yes’, and I don’t mean: do you understand the implications
of this information, but simply do you understand what the words in
the sentence communicate?
But do you know if the statement is true or false? Here the answer
must undoubtedly be ‘no’. I really have no inside track on secret Dutch
government intelligence. But the point is this: you can understand
what this sentence means without knowing whether or not it is true
or false. And that’s actually often the case. Of course, we usually
cooperatively assume that what people tell us is true, or at the very
least, that they believe it is true. But actually whether it is true or
not, doesn’t matter in order to understand it. In fact, what matters
is if you can imagine what the world would look like if the sentences
was true (or false).
There are many things about the world that we believe or even do
not believe, yet we are still able to satisfactorily communicate about
them. With a truth-conditional theory of meaning, we are not required
to be omniscient.

Compositionality

We’ve talked about the meanings of words, and how they can (at least
for a lot of simple words) be interpreted in set theory, but we haven’t
paid much attention to how this happens. Consider the following two
statements.

(15) Fido bit Jan.

(16) Jan bit Fido.

These two statements, even though they are made up of the same
three lexical items, actually express very different propositions. That’s
because the subject position, which without further information can
be interpreted as the position to the left of the verb in English, is the
agent that carries out the action. The verb to bite is transitive, so it
subcategorizes for an object argument, and that appears to the right
of the verb in a simple statement. These two examples thus differ
syntactically as to which name is used as a subject and which as an
object. These syntactic differences are meaningful because these two
examples also differ in who is the biter and who is bitten.
This illustrates a further point: that the meaning of a proposition
is not simply derived by adding up the meanings of the component
words, but the way in which they are combined is also relevant to
determining the meaning. The semantics of natural language adheres
to what is called the Principle of Compositionality, first proposed
by the German philosopher, Gottlieb Frege.

Gottlieb Frege, father of semantics.

Frege also invented predicate logic.
184 an open introduction to linguistics

The Principle of Compositionality

The meaning of a statement is a function of the meaning of it’s
parts and how they are put together.

So this means that meaning is not composed from a kind of unstruc-

tured bag of words, but with a structure that identifies important re-
lationships, such as modification, dominance, and/or argument struc-
ture. In modern semantic theories, syntactic structure feeds semantics
with relevant structural information that guides the semantic compo-
sition that can then be interpreted. Let’s see another example of this,
but this time with a complex sentence made up of several conjoined
propositions:

(17) Greedo shot and Greedo missed or Han shot.

In the famous Cantina scene from Star Wars (Episode IV: A New
Hope), Han Solo is sitting across the table from Greedo, and there is
some ambiguity about what happened. Did Greedo shoot and miss, Han Solo and Greedo in the original
making it possible for Han then to shoot? Or did Han simply shoot movie.

first? (17) could be used to describe the two possible situations, but
without a comma or some intonation to disambiguate the syntactic
structure associated with this reading, the sentences is also ambiguous.
The structural ambiguity arises from the difference between conjunc-
tions and disjunctions, and whether you take the conjunction and to
be the main conjunct, or the disjunctive or to be the main conjunct,
even when the truth value of the component propositions is kept the
same. Greedo shot’ and Greedo missed’ is true (represented by ‘1’ in
the table), while Han shot(first)’ is false (represented by ‘0’):

(Greedo shot or Greedo missed) and (Han shot)

(1 or 1) and (0)
(1) and (0)
(0)

(Greedo shot) or (Greedo missed and Han shot)

(1) or (1 and 0)
(1) or (0)
(1)

I will refrain from revealing my own interpretation of the Cantina

events. What’s important about this example is that it shows that
knowing the truth value of the component parts of propositions in a
sentence isn’t sufficient for determining the truth value of the com-
plex proposition. How they are combined matters. This is actually
a principal of semantics called the Principle of Compositionality.
Constituent parts of propositions by themselves be complex having
component parts, but we will also be able to interpret them. This
gives our semantics the kind of recursiveness that we also saw in our
syntax. Compositionality thus gives language two very important fea-
tures. First, it’s part of what makes it possible for us to understand in-
 semantics i: fundamentals of meaning and its representation 185

finitely many linguistic expressions, because we can calculate from the

expression to its meaning. Second, it make language learnable. There
is not a finite list of constructions and their associated meanings that
have to be learned, but general rules of construction combined with the
meaning of the building blocks that together create a rich (actually in-
finite in theory, but constrained by processing limitations), generative
grammar that can be grasped by learning only a finite set of rules.

Extensions and intensions

So far we have been working with what’s called and existential theory
of meaning. An extensional theory of meaning says that the meaning
of a word is the set of things it can be used to refer to in the real world.
Names refer to individuals, while nouns, verbs and adjectives refer to
sets. This seems like a plausible idea at first, but with a little more Han Solo and Greedo in a model
thought we easily see that it won’t actually scale-up. (Lego)
Frege was also one of the first people to write about the problems
with an extensional theory of meaning when applied to proper names,
and to suggest a possible solution, which we use to this day. Frege dis-
cussed the problem with an example about Phosphorus and Hesperus
From the northern hemisphere, there is a very bright star that can
be seen in the east, where the sun rises. The ancient Greeks named
the star ‘Phosphorus’ (also known as the morning star.). Similarly, a
very bright star can be seen in the west, where the sun is setting, and
the ancient Greeks named it ‘Hesperus’ (alternatively known as the
evening star). The ancient Greeks, because they believed incorrectly
that these were two different heavenly bodies, gave them two different
names. But around 350 BC they made a discovery: Hesperus and
Phosphorus were both the same object: the planet Venus!
Now how do the objects in the world and the words we can use to
refer to them in this story work if we have an existential theory of
meaning? Let’s consider two statements.

(18) Hesperus is Hesperus.

(19) Hesperus is Phosphorus.

The first sentence (1) probably seems quite silly. It’s clearly a tautol- The morning star.
ogy, because everything is clearly themselves. But it’s also true: the
object in the world picked out by the name ‘Hesperus’ is the same
object picked out by the same name (and in fact this seems to be quite
a basic prerequisite we have to agree on to even have an existential
theory of meaning!). But the second statement, within an existential
theory of meaning, is also a tautology. It’s saying that the object that
is referred to with ‘Hesperus’ is the object referred to by ‘Phosphorus’.
Since both names refer to the planet Venus, it’s again basically saying
that Venus is Venus.
But, this is problematic. Clearly we don’t want our semantic theory
to treat scientific discoveries as tautologies. The Greeks before 350
B.C., who didn’t know that their morning and evening stars were both
Venus, had a different relationship to the same object. They considered

An Ancient Greek Astronomer

186 an open introduction to linguistics

the object by one name to have different features than the same object
when called by another name (because they didn’t know it was the
same object). Frege’s conclusion was then that actually, there must be
more to the meaning of a word than it’s ability to identify objects in
the world. How we relate to the object, what features we ascribe to
it, how we learned about it, how it appears in a certain context, are
an additional part of it’s meaning. In order to deal with this, Frege
added another layer of meaning to each word: it’s intension.
Under an intensional theory of word meaning, words have an ex-
tension, which identifies what they pick out in the world. But words
also have an intension, which describes its features and how it is un-
derstood. This new aspect of meaning lets us analyze Hesperus and
Phosphorus. For the ancient Greeks, these two words have quite dif-
ferent intensions. One appears in the morning, one in the evening, etc.
We know they have the same extension, but the ancient Greeks did
not. When they discovered that both referred to Venus, this specified
the extension.
The introduction of intensions also solves another problem related
to the set-theoretic interpretation of words. Under and extensional
theory, the meaning of the words zombie, elf and ewok are all members
of the empty set, because they are things that do not exist in the
world. There is therefore no difference between them in extensional
meaning. However, an ewok and a zombie have very different features,
and clearly need to be distinguished!
So let’s again take stock. What are the features of our semantic
theory?

• A theory of extensional reference (how we link linguistic symbols

to sets and individuals in a model that serves as a proxy for the real
world)

• A theory of intensions, which keeps track of how we understand or

what we know about words

• Compositionality: the recognition that the way in which we com-

bine words can result in different meanings

• A model of the world in which propositions can be interpreted (de-

termined to be true or false) relying on set theory

• Truth conditions which tell us what our model has to look like for
a proposition to be true or to be false

Another very important feature of our theory, and the analysis we

then give to grammatical sentences, is that we should be able to under-
stand what follows from a proposition or multiple propositions, that
is, how the world (or our model) must be structured if the propositions
are true.

Entailments
Entailment can be defined formally as follows: A entails B, if and only
if in every situation in which A is true, B is also true. Consider the
 semantics i: fundamentals of meaning and its representation 187

following sentences.

(20) Han shot a gun and Greedo shot a gun.

(21) |= Han shot a gun.

This is clearly the case. If (20) is true, given the nature of conjunction,
(21) must also be true.
Now consider the following:

(22) Han and Leia are dating.

(23) |= Han is dating.
(24) |= Leia is dating.

If (22) is true, does it mean that (23) and (24) is true? That does seem
to be the case. In any world where (22) is true, that the individuals
are also dating has to be true. We can then say that (22) entails (using
the symbol for entailment: |= ) examples (23) and (24). Note however,
that entailment is not symmetric (though it sometimes can be): if Han
is dating, that does not mean that Han and Leia are dating. And note
as well, that most people reading (22) will assume that the speaker
means that Han is dating Leia and vice versa. However, the sentence
doesn’t actually say that, and the sentence is also consistent with a
situation where e.g. Han is dating Betty and Leia is dating Barney.

(25) Han Solo is tall and handsome.

(26) |= Han Solo is handsome.

If example (25) is true, then example (26) must also be true. If you
accept (25) as true then the rules of language, the rules of word mean-
ings and how they combine when a conjunction is used, what this then
entails, force you to agree that (26) must also be true. The above
examples showed entailment relationships that arise in part because of
the the conjunction and. Let’s now have a look at some more entail-
ment relationships with adjectives.

(27) Han is a Corellian pilot.

(28) |= Han is a pilot.

Han Solo was born on the planet Corellia. The sentence (27) entails
the simpler, less informative statement in (28). This is the same for
many adjectives. Confirm for yourself that the following statements
entail the simpler statement without the modifying adjective:

(29) Andre is a Dutch singer. |= Andrea is a singer.

(30) I bought a red dress. |= I bought a dress.
(31) That’s a cool skateboard. |= That’s a skatebord.

These adjectives actually belong to a class of adjectives called inter-

sective adjectives. That’s because anything that is claimed to be
a member of the set of the noun modified by the adjectives (e.g. the
188 an open introduction to linguistics

set of pilots in (27)) must also be a member of the set described by

the adjective (e.g. the set of Corellians in (27)), which means that
the set evoked by the noun and the set evoked by the adjectives must
intersect. Not all adjectives are intersective, however. Consider the
following two examples:

(32) Dr. Janssen is a suspected criminal.

(33) Dr. Janssen is a criminal.

If (32) is true, even though we again have an adjective modifying a

noun in the object argument, we cannot now conclude that Dr. Janssen
is a criminal. That’s because adjectives like suspected, alleged and po-
tential are non-intersective adjectives. They do not license entailments
that the modified individual can be described by the noun alone.

(34) Dr. Janssen is a fake dentist.

(35) 6|= Dr. Janssen is a dentist.
(36) I have a toy gun in my desk drawer at work.
(37) 6|= I have a gun in my desk drawer at work.

I certainly would not want Dr Janssen to treat my teeth, because

he is not a dentist. You also do not have to fear coming to talk to
me in my office, because a toy gun is not a gun. The words fake
and toy are anti-intersective adjectives, Anti-intersective adjectives
actually indicated that the object that they are modifying cannot be
described by the noun that is being used to modify it. So a fake dentist
is not a dentist and a toy gun is not really a gun. Types of propositions:
These examples of different types of adjectives illustrate an impor- Tautology: A proposition that’s al-
ways true by virtue of its form.
tant point. Earlier we saw that words of similar word classes that be- Contradiction: A proposition that’s
have similarly syntactically, often are also similar in how they semanti- always false by virtue of its form.
Contingency: A proposition that is
cally interact with the meanings of the other words they modify or are true in some models and false in oth-
modified by. But there are also finer distinctions, in addition to part- ers.
of-speech. Intersective adjectives differ greatly from non-intersective
and anti-intersective adjectives in terms of their entailment relation-
ships. This means that while syntactic information, such as whether or
not an adjective modifies a certain noun, will give you a large amount
of information about meaning and entailment, it doesn’t tell you every-
thing. Syntax supports semantic interpretations but it isn’t sufficient
to explain them all.

Knowledge of language versus knowledge of the world

(38) There are some kittens in the basket.
(39) There are some young cats in the basket.
(40) A man walked into a bar.
(41) A male individual walked into a bar.

Some things are necessarily true given the relationships between

lexical meanings within a language. A speaker of English must accept
 semantics i: fundamentals of meaning and its representation 189

that kittens are young cats, and that men have the feature of being
male„ and if someone dies they are dead (zombies and vampires ex-
cluded of course). These are lexical entailments: in all possible worlds
where (38) is true, (39) must also be true. If you don’t accept these
things then you are not speaking English. These have to do with the
basic features of meaning of these words which all native speakers have
learned and accept. So in interpreting any given sentence, there is a lot
of lexical ‘baggage’ that comes with it in the form of the information
we know about how the concepts signalled by the words we use relate
to each other in the world.
Consider the following sentences:

(42) The sun is the sun.

(43) My kitten is a puppy. Lexical semantics: Giving a for-
mal analysis to words in the lexicon
Example (42) is an example of what is called a tautology. These is tricky, and most formal semantic
research has instead focused on spe-
are statements that are always true in every possible world given their cial words (like quantifiers) and on
structure and the meaning of the lexical items that they contain. Ex- structural relations related to syntax.
However for natural language process-
ample (43) is a contradiction, which is false in every possible world ing, lexical semantics has become ex-
given its structure and the meaning of the words it contains. These are tremely important, though currently
special, uninformative sentences. Now, in normal conversation we can the solutions for dealing with mean-
ing with large text collections is to
sometimes use these statements in a meaningful way, but only if we either use distributional lexical infor-
re-interpret the meaning of some of the words. Someone could utter mation (what words co-occur with
what words) or large databases such
(42) to mean The sun is going to shine no matter what you do or use as WordNet or the generative lexicon
(43) to mean My kitten is rolling around on the ground like a puppy. or other manually created lexicons.
But without any context, these statements are not meaningful. We
should examine what meaningful means here. A meaningful proposi-
tion would be one that contributes new information, information that
has the potential to be true or false, to the conversation. In doing so
it updates the information that the conversational participants share,
i.e. it updates their common ground.
Statements that are meaningful in this way are also called contingent
statements, that is whether or not they are true or false is contingent
on the way the world looks.

(44) All kings are dictators.

Example (44) is not necessarily true. In fact, the king of the Nether-
lands, William-Alexander, is not a dictator because the Netherlands
is a constitutional monarchy. You might accept the statement Histor-
ically, all kings are dictators, but it is not necessarily true, but true
because of our understanding of the world and the history of our world.

Study Questions

1. Basics of meaning
Consider the following three claims. For each claim, criticize or de-
fend the claim. (State at the beginning which you are doing!) and
then motivate your argument with real examples from natural lan-
guage. Be sure to use correct terminology and mention any relevant
190 an open introduction to linguistics

principles.

(a) Given some combination of phonemes, we can never predict the

meaning of the combination
(b) Given some combination of morphemes, we can sometimes pre-
dict the meaning of the combination.
(c) Given some combination of words in a sentence, if we know the
words and their grammatical relations, we can always predict the
meaning of the sentence.

2. Entailment
Consider the following sentences

(a) None of my students have visited the Wadden-islands.

(b) None of my students have visited Terschelling.
(c) None of my Frisian students have visited the Wadden-islands.
(d) Some of my students have visited the Wadden-islands.
(e) Some of my students have visited Terschelling.
(f) Some of my Frisian students have visited the Wadden-islands.
(g) All of my students have visited the Wadden-islands.
(h) All of my students have visited Terschelling.
(i) All of my Frisian students have visited the Wadden-islands.

(a) (a) Does (1) entail (2)?

(b) (b) Does (1) entail (3)?
(c) (c) Does (4) entail (5)?
(d) (d) Does (4) entail (6)?
(e) (e) Does (7) entail (8)?

3. Referential Theory of Meaning

(a) Explain what the Referential Theory of Meaning is.

(b) Explain why the word “hobbit” and the word “barbapapa” refer
to the same thing in a referential theory of meaning.
(c) Who was Gottlieb Frege? (Hint: “Gottlieb Frege was the father
of....”)
(d) Explain why Frege thought that meaning must be more than
just reference.
(e) What was Frege’s solution to dealing with meaning that is more
than just reference?

4. Adjective meaning
 semantics i: fundamentals of meaning and its representation 191

(a) Explain the difference between the adjective pretend in a pre-

tend celebrity and Korean in a Korean celebrity, referring to the
examples.
(b) Create an interpretation rule for intersective adjectives.

5. One of the things we have not really touched upon here is how we
assign meaning to words. Implicitly, we have assumed that words
like /John/, /think/ and /yellow/ have meaning because they re-
fer to something in the real world. This idea is called a ‘referential
theory of meaning’. It works well in most cases, but it becomes prob-
lematic when we talk about things that are not part of our world,
such as fictional characters, magical creatures or a Napoleonic vic-
tory in the battle of Waterloo. A solution to this is to distinguish
between extension, i.e. that what is referred to, and intension,
i.e. what is referred to in the mind of the speaker. Now, consider
the following true statements and answer the questions:

I. Clark Kent is Superman.

II. Superman is Superman.

a. Explain why purely extensional theory of meaning would treat I

as a tautology (just as II is).
b. How does the notion of intension help us to make I and II have
different meanings?
c. How does the notion of intension help us to distinguish between
vampires and elves?
16
Semantics II: Quantification: Beyond Reference

Every day may not be good but there is something good in every
day.
-Anonymous

Quantifiers

Quantifiers are words that select a quantity of something in the world

(or in a model), making it possible for us to talk about that quantity.
They are not referential, unlike proper names and common nouns.
Common natural language quantifiers select different amounts of indi-
viduals: all, each, some, many, less than three etc. But we also have
words that let us quantify over places, e.g. someplace, everywhere, and
times as well, e.g. sometime, every day, often, seldom. We’ll focus on
the analysis of quantification over individuals in this chapter.
When the DP is a proper name, a sentence of the form [DP VP],
is true if and only if, [[DP]] is a member of the [[VP]]. But what
if the DP is not a proper name? An intransitive sentences with a
quantified subject does not state that the subject is a member of the
set contributed by the predicate. Let’s look more closely at the class
of arguments called quantifier phrases (QPs).

(1) Every jedi meditates.

(2) Some Ewok can speak English.
(3) Most jedi wear brown.

First a little terminology. In analyzing sentences with quantifica-

tion, semanticists talk of the restrictor, which identifies what things
we are talking about, and is usually contributed by the nominal ar-
gument that is syntactically modified by the quantifier, and the nu-
clear scope, which is contributed by the predicate (the predicate is
everything minus the subject argument). In the above examples, the
linguistic material that contributes the restrictor has been underlined
once, and the predicate, which contributes the nuclear scope or scopal
set, is underlined twice.
Early in formal approaches to semantics, researchers believed that Three jedi knights, two of the three
it would be possible to extend predicate calculus and use it as a formal wearing the traditional brown robes.
representation of meaning. Two of the researchers most influential in
194 an open introduction to linguistics

trying to develop a semantics of natural language along this line were

Richard Montague and Barbara Partee. Predicate logic seemed to
provide everything desirable: it already has a model-theoretic inter-
pretation, it has a formal apparatus, it interfaces well with set-theory,
etc.
However, the interpretation of quantifiers is one area where it’s clear
the predicate logic falls short. The predicate logic quantifier for all (∀)
and the existential quantifier (∃) only take one argument. In order to
communicate the information conveyed by every and some in natural
language, additional operators have to also be used. For existential
quantification, we rely on conjunction, stating that there is an indi-
vidual that belongs to both the A set and the B set, e.g. for (2) we
say that there is some individual that is both a member of the set of
Ewoks and a member of those that can speak English. For the uni-
versal quantifier we must rely on implication, which means that ‘for
everything it holds that if it is a jedi, it meditates’.

(4) ∃x (Ewok(x) & English_speaker(x))

(5) ∀x(jedi(x) =⇒ meditates(x))

But this way of interpreting quantifiers doesn’t work well when ex-
tended to the entire range of natural language quantifiers. For exam-
ple, consider most in (6). Can we simply make a quantifier M, and
then treat it similarly to either a universal or an existential quantifier?
The answer is no, but let’s look why. Clearly we need to relate things
(the jedi and the brown-wearers) so we need to use either a conjunction
or an implication.
Let’s consider what interpretations each conjunct would lead to with
most. We’ll consider the truth of (6) in one of three possible models
presented in the Table below.

(6) Most Jedi wear brown.

Model Non-Jedi + wearing blue Jedi Jedi + wearing brown T or F?

m1 0 10 7 T
m2 20 10 7 T
m3 8 10 7 T

First, confirm for yourself that the natural language sentence in (6)
is true in each of the three models. Next, we have to have some def-
inition of the quantity meant with most. Let’s take a rather simple
definition and simply say that most means more than half. With this
definition, we can then consider what would happen if most as a quanti-
fier M would be interpreted similarly to how existential quantifiers are
interpreted in predicate logic, relying on conjunction to relate the sets.
This would mean that something like: M x(jedi(x) & wear_brown(x)).
But this means that more than half the individuals in the model are
Jedi that wear brown. In m1 this would be true, but in m2 and m3 , this
sentence would be false, because there are more individuals that are
 semantics ii: quantification: beyond reference 195

not brown-wearing Jedi than individuals that are. But if 7 of 10 Jedi

wear brown, then clearly this sentence in natural language is true in
these two models, so defining most this way would lead to a definition
of most that does not capture the natural language meaning.
If we instead take M to work like a universal quantifier, and rely on
implication to relate the sets involved, then it should mean something
like: M x(jedi(x) =⇒ wear_brown(x)). But this then means that if
more than half the individuals in the model are Jedi, then they wear
brown. But in model m2 , the majority of individuals are not Jedi.
However, this seems to be quite irrelevant to interpreting the natural
language sentence, where we really want to first zoom in on the set of
Jedi, and then check to see if the set of Jedi have a certain property
(more than half wearing brown). Actually, whenever there are more
Jedi than non-Jedi, the antecedent will be false, but the expression as a
whole will then be true, because of the truth-table for implication. So
using implication with M is not going to get us the natural language
meaning either.
So there seems to be no way to translate most into predicate logic.
There are also numerous other problems with dealing with other nat-
ural language quantifiers in predicate logic as well. For this reason, it
was a major innovation in natural language semantics in 1981 when
Barwise and Cooper (1981) proposed a new idea, grounded in set the-
ory. Their proposal was that natural language quantifiers can be seen
as signalling the relationship between two sets. The type of quantifier
tells us what that relationship is. So example (1):

• Every Jedi meditates is true if the set of jedi, [[jedi]] is a subset

of the set of those who meditate, [[meditate]]

• every A B = True iff A ⊂ B .

The shorthand Quantifier A B identifies the quantifier and the

two sets, A and B, that are in a relationship. Set A is the restric-
tor, or ‘restrictor set’ and set B is the information contributed by the
predicate.
The advantage of treating all quantifiers as a relationship between
two sets is that we can now give more plausible definitions to all the
quantifiers in natural language. Consider the definitions for no, some
and even more complex quantifiers such as at least five:

• No jedi marries. is true if the intersection between the set of

Jedi and the set of married people is empty.

• no A B = True iff (A ∩ B) = ∅ .

(7) Some Ewoks speak English.

• Some Ewoks speak English is true if the intersection between the

set of Ewoks and the set of English speakers is not empty.

• some A B = True iff (A ∩ B) 6= ∅ .

(8) At least five Jedi knit.

196 an open introduction to linguistics

• At least five Jedi knit is true if intersection between the set

of Jedi’s and the set of knitters has at least five members.

• At least A B = True iff |(A ∩ B)| ≥ 5

This analysis of quantifiers as signalling the relationship between

two sets is called generalized quantifier theory, because it gen-
eralizes to all natural language quantifiers. But another question we
can ask is if it is necessary for all natural language quantifiers? Is
this analysis overly complex for some quantifiers? Some researchers
have argued as much. Natural language quantifiers can be divided
into two groups: weak quantifiers and strong quantifiers. Even though
the strong-weak distinction is generally considered semantic, there are
some syntactic tests to determine whether a quantifier is strong or
weak. Weak quantifiers, like some, no, a few and at least four can oc-
cur in existential contexts, but strong quantifiers cannot. In English,
the syntactic environment created by There are . . . sentences is an
existential environment. See examples (9) and (10). There are similar
syntactic environments that robustly select for weak DPs in other lan-
guages as well. For example, Partee and her students (in 2001, Lecture
notes 4) have identified a similar context in Russian show in example
(11):

(9) There are some/three/more than three spaceships in orbit.

(10) *There are all/every/most/neither spaceships in orbit.
(11) U nego est’ sestra/sestry/sester . . .

Weak quantifiers don’t just differ from strong quantifiers in their

ability to occur in existential contexts. They also do not need a re-
lational analysis: For Every X does Y it must be the case that every
member of the set X is also in Y. You can also give a similar relational
analysis to weak quantifiers, e.g. Some X are Y means that some mem-
ber of the set X is also a member of Y. But you can also simply treat
the two sets as a conjunction (so a set with more specific features) for The Rocinante in orbit (The Ex-
panse).
any of the weak quantifiers: For all the sentences in (9) you can simply
consider whether the restriction of a certain quantity holds for the set
of orbiting spaceships (e.g. Are there three orbiting spaceships? Are
there more than three orbiting spaceships?). Note that this doesn’t
work with the strong quantifiers. We cannot grammatically ask ‘*Are
there all orbiting spaceships?’ or ‘*Are there neither orbiting space-
ships?’. For these quantifiers the distinction between Set A and Set B
is relevant, and the relation they code is asymmetric.
However, even though there are semantic and syntactic differences
between weak and strong quantifiers, semanticists still use generalized
quantifier theory to analyze them. There is an advantage to having
a uniform analysis for all members of a given word type. However
the distinction is still something to be aware of because it might af-
fect language processing and acquisition. For example, perhaps people
are faster at interpreting weak quantifiers, because they signal a sim-
pler relationship? Children have trouble learning the full meaning of
 semantics ii: quantification: beyond reference 197

Total Ewoks died in battle survived Table 16.1: Example situations outlin-
ing Ewok casualties
100 20 80
100 40 60
100 49 51
1,000 499 501
1,000,000 200,000 800,000

universal quantifiers and Geurts (2003) has proposed that it may be

because they incorrectly interpret universal quantifiers as if they were
weak quantifiers, rather than strong. So the distinction may still be
relevant, even if the favored formal analysis uses an analysis that treats
both types uniformally.

World knowledge and comparative quantifiers

Even with a generalized theory of quantification, there are still natural
language quantifiers whose meanings aren’t entirely captured by the
relationship between two sets. A very clear case is still the proportional
quantifier most:

(12) Most rebels speak English.

We can give a definition that simply says that if more than half of the
restrictor set is also a member of the scopal set, then the sentence is
true, and indeed, this is also a very common way to define most in
generalized quantifier theory:

• True if more than half of the rebels speak English.

• most A B = True iff (2 x |A ∩ B|) > |A|.

However, whether you are comfortable with this definition or not

seems to depend on world knowledge and the cardinality of the sets.
For example, consider the following statements:

(13) Most Ewoks survived the battle of Endor.

(14) Most Ewoks speak English

In which of the following situations are you comfortable saying that

the examples (13) or (14) is true?
Most people find it hard to accept the idea of exactly half, the
larger the numbers are. They also find that this definition of most
is less acceptable when the situation has to do with something really
terrible, such as dying in battle, where every death is bad, compared
to something good, like bilingualism. In the negative situation, it
seems that even a small proportion of deaths shouldn’t be trivialized by
focusing on the survivor rate, whereas with something positive, this is
perfectly fine. Are these nuances part of the semantic meaning of most?
Or are they more pragmatic or related to our cultural heritage? One
way to figure this out would be to ask speakers of different languages
what their intuitions are. In any case, the example makes clear that
198 an open introduction to linguistics

knowing the set-theoretic relationships that constrain the quantifiers

will not be sufficient for using them appropriately.

Conservativity
Crosslinguistically, the same inventory of quantifiers seems to be present
in most languages. Further, there are number of semantic features of
quantifiers that we also find cross-linguistically. Because quantifiers in
all languages even those with very different grammatical structures,
have these same features, there are two intriguing possibilities. First,
it could be that we are pre-programmed to have quantifiers with these
features, and this is part of our genetic linguistic endowment. Alter-
natively, it could be that these features are easier to learn, or maybe
sentences with these features easier to verify (determine whether or not
they are true in a given context). One of the most interesting features
is conservativity which we will examine in more detail. Consider the
following definition:

• conservativity A relation Q named by a determiner is conservative

if and only if, for any properties A and B, relation Q holds between
A and the things in (A ∩ B)

A conservative quantifier is easier to verify. It means that we only

have to look at members of set B that are also members of set A in
order to evaluate the truth of a quantified sentence. This means that
we do not have go check whether whatever claim the sentence makes
also holds for the members of B that are not members of A.

(15) Every Ewok danced.

The example (15) is of the form Every N VP. Every N VP is true iff
[[N]] ⊂ [[VP]]. Example (15) is thus true iff [[Ewok]] is a subset of the
set [[danced]], and false if there are members of the set of Ewoks that
are not also members of the set of those who danced. To understand
this sentence, you do not have to know anything about individuals
who danced that were not Ewoks. Every is thus conservative, because
Every Ewok danced means that same thing as Every Ewok is an Ewok
that danced. Let’s consider example (8), (At least five Jedi knit.) This
sentence is true iff the intersection between the set of individuals that
are Jedi, and the set of knitters, has a cardinality of at least five. To
determine whether or not this sentence is true in any given context,
it is not necessary to consider any knitters that are not Jedi. The
example is thus synonymous with At least five Jedi are Jedi who knit.
This means it is conservative.
Now consider the meaning of a quantifier I made up: Korgat, given
in set theory notation below.
Korgat AB is true iff ¬ (B ⊆ A)
Let’s consider the following two pictures while we discuss the mean-
ing of this quantifier.
 semantics ii: quantification: beyond reference 199

Given the definition of the meaning of the quantifier Korgat, is the

following sentence true or false with respect to Picture A and Picture
B?

(16) Korgat chicks are sledding.

First we have to check for Picture A. Korgat A B is true iff it is

not the case that B is a subset (or the same set) as A. That basically
means that there must be some B that are not A. Set A is the set of
chicks (supplied by the noun modified by korgat) and Set B is the set
of sledders. Are there sledders that are not chicks? The answer is yes,
the dog sledders do not belong to the set of chicks. So with respect to
Picture A, the sentence is true. What about with respect to picture
B? Here we see that there are no sledders that are not chicks. In fact,
the set of sledders is the subset of the set of chicks. With respect to
Picture B, example sentence (16) is false. What could we do to modify
the picture to make it true? We could, in fact, give the dog a sled.
Now is the quantifier Korgat that I made up conservative or non-
conservative? Can we determine the truth value of sentences with
korgat by simply examining the restrictor set (Set A) and members of
the scopal set that are also members of the restrictor set (the intersec-
tion of Set A and Set B?) The answer is no, because we have to check
all members of set B. This means that korgat is not conservative.
It seems to be a cross-linguistic universal that all natural language
quantifiers that are determiners have the semantic feature of con-
servativity. However, we do still sometimes want to express non-
conservative meanings, so there are non-conservative quantifiers, but
they are not syntactically determiners. The adverb only is a non-
conservative quantifier. Consider the following sentence.

(17) Only chicks are sledding.

Example (17) is true w.r.t. Picture B above: only chicks are sledding.
But to confirm this, you have to examine all the members of the set
of sledders, the scopal set. Only is therefore not conservative. As an
adverb, only is not syntactically modifying the noun chicks, either.
Only also associates with focus, a hard to define phonetic feature
in spoken language. Consider example (18). This sentence first limits
interpretation to the chicks, and then says that they are only doing one
activity, as members of the set of sledders, not other activity sets. It
therefore compares activities and singles out which ones the chicks are
doing. That’s quite different from (17), which says that among those
that are sledding, they are all chicks. Example (17) thus zooms in on
200 an open introduction to linguistics

the sledders, and then says that being a sledder only applies to chicks.
Example (17) will be wrong if you have a non-chick sledder, whereas
example (18) will be wrong if you have a chick doing something else
(skiing?).

(18) The chicks are only sledding.

You can make other differences in meaning even without moving only
around in the sentence (it’s actually an adverb, so it can occur in any
position that an adverb can occur in). Consider the following:

(19) Obi-wan only gave Luke a lightsaber. (He didn’t teach him
how to use it.)
(20) Obi-wan only gave Luke a lightsaber. (He didn’t give one to
Han or anyone else.)
(21) Obi-wan only gave Luke a lightsaber. (He didn’t give him
anything else.)

This shows an aspect of meaning that can be expressed by intonation.

In each sentence, the same information, that Obi-wan gave Luke a
lightsaber, is communicated. But additional information about how
to interpret the only is communicated via intonation.
So what we’ve discussed in the current section is that all natural
language quantifiers can be interpreted as relations between two sets,
but that’s not entirely necessary for their analysis, though it is de-
sirable for other reasons. We’ve also looked more closely at one of
the semantic features of quantifiers that seems to be universal for a
certain syntactic subclass of quantifiers: the property of conservativ-
ity. Finally, we talked about one non-conservative quantifier: only,
noting also that it has another interesting feature: it associates with
focus. Now let’s look at some of the predictable entailment properties
of natural language quantifiers.

Entailment properties of quantifiers

Recall that one of the goals of semantic theory is to discover what
follows from a given true statement. Quantifiers actually make very
rich semantic contributions, allowing us to infer many additional state-
ments from each one. Consider the following examples:

(22) No man does pilates. |= No man does power pilates.

(23) At most ten men ate. |= At most ten men ate steaks.
(24) Less than three knights drank an alcoholic beverage. |= Less Obi-wan Kenobi giving Luke Sky-
than three knights drank a beer . walker his first lightsaber in Episode
IV (Star Wars)
The quantifiers No and At most ten allow inferences from a superset
to a more specific subset, as shown in examples (22) and (23). Both
are what is called monotone decreasing in their second argument.
This means that whatever is claimed to hold in the predicate (which
contributes the second set), also holds for any subset, i.e. anything
 semantics ii: quantification: beyond reference 201

more specific. Monotonicity thus involves inferences from supersets to

subsets and vice versa.

(25) All farmers ate. 6|= All farmers ate steak.

(26) Most farmers ate. 6|= Most farmers ate steak.

The quantifiers all and most are not monotone decreasing in their
second argument. They are in fact monotone increasing, which can
be seen in the next examples:

(27) All farmers ate steak. |= All farmers ate.

(28) Most farmers ate steak. |= Most farmers ate.

This means we can make inferences from subsets to supersets for the
second argument of quantifiers all and most.
We can write formal definitions of these monotonicity properties:
mon ↓ Det is monotone decreasing in its right argument:=
for all A, B, B’: if B’ ⊆ B, and Det A B, then Det A B’.

mon ↑ Det is monotone increasing in its right argument :=

for all A, B, B’: if B ⊆ B’, and Det A B then Det A B’.
Monotonicity also applies to lexical relations.

(29) Some lawyers are thieves. |= Some lawyers are criminals.

(30) No lawyers are thieves. |= No lawyers are car thieves

Monotonicity properties can of course apply to both sets involved in

quantification, which is why we talk of them applying to the left (e.g.
Set A, set contributed by the restrictor) or the right argument (Set B
or the set contributed by the scopal set, from the predicate).

(31) All farmers ate steak. |= All Dutch farmers ate steak.
(32) Some farmers ate steak. 6|= Some Dutch farmers ate steak.
(33) Some Dutch farmers ate steak. |= Some farmers ate steak.

So we can see that in (31) we can make inferences to more specific

restrictor sets. This means that the quantifier all is monotone de-
creasing in its left argument. This is not the case for (32), which is
not monotone decreasing. Instead it is monotone decreasing, as can
be seen if we flip the example, as shown in (33)
Are some monotonic inferences easier than others? Geurts and van
der Slik (2005) studied how easy it was for people to make upward en-
tailing and downward entailing inferences. They found that inferences
from subsets to supersets was easier than from supersets to subsets,
i.e. monotone increasing is easier than monotone decreasing. They
also found that downward entailing quantifiers with cardinal numbers
like More than six or less than five were the hardest.
Not all quantifiers are are monotonic however, Consider the follow-
ing sentences, which do not license any inferences. :
202 an open introduction to linguistics

(34) Exactly two children came home late.

(35) Exactly two children came home.
(36) An even number of children came home late.
(37) An even number of children came home.

Monotonicity properties are relevant for other linguistic constructions,

and in fact, monotone decreasing contexts are able to license negative
polarity items, like any and ever, which can also occur in the scope of
negation. Consider the following examples from Ladusaw (1979):

(38) No child who has ever seen Claude has smiled at him.
(39) Every child who has ever seen Claude has smiled at him.
(40) *Several children who have ever seen Claude have smiled at
him.
(41) No child has ever smiled at Claude.
(42) *Every child has ever smiled at Claude.

A quantifier is monotone decreasing either in its left or right arguments,

and because every is monotone decreasing in it’s left argument, only
when the negative polarity item occurs in the QP is it licensed (e.g.
(39), and not when it occurs in the predicate, e.g. (42)

Study Questions
1. If we look closely at the semantics of a sentence, there is often more
than one possible way to interpret it. Nevertheless, communication
problems due to ambiguity in meaning rarely occurs in real life. Can
you think of some reasons for why this is? How does the field of
pragmatics relate to this issue?

2. One of the tasks that researchers within the field of Natural Lan-
guage Processing (NLP) have tried to do automatically is called
paraphrase detection, i.e. the automatic filtering of paraphrases
from a text collection or summary, for example. How could you
use what you know about semantics and propositions to solve this
problem?

3. Just as the sentence /Peter and Sue get married/ has multiple
possible interpretations, so can the sentence /I love Sachertor-
te/. Why, then, do we call the first sentence ambiguous, but not
the second?

4. Quantifiers such as /none/, /some/, and /all play very interesting

roles when it comes to semantics. To get an idea of this, look at the
sentences below and answer the questions.

I. None of my students have visited the Wadden Islands.

II. None of my students have visited Terschelling.
III. None of my Frisian students have visited the Wadden Islands.
 semantics ii: quantification: beyond reference 203

IV. Some of my students have visited the Wadden-islands.

V. Some of my students have visited Terschelling.
VI. Some of my Frisian students have visited the Wadden Islands.
VII. All of my students have visited the Wadden-islands.
VIII. All of my students have visited Terschelling.
IX. All of my Frisian students have visited the Wadden Islands.
Note: ‘A entails B’ means that B is a
logical consequence of A.
a. Does I entail II?
b. Does I entail III?
c. Does IV entail V?
d. Does IV entail VI?
e. Does VII entail VIII?
f. Does IX entail VI?
g. Does IX entail II?

(a) Entailment and monotonicity Consider the following sen-

tences:
i. Some students have visited Asia.
ii. Some students have visited southeast Asia.
iii. No students have visited Antarctica.
iv. No students have visited North Antarctica.
v. Less than three student have visited Antarctica.
vi. Less than three students have visited North Antarctica.
(a) What kind of entailment relationship exists between (1) and
(2) and how is this described in terms of monotonicity?

(b) What kind of entailment relationship exists between (3) and

(4) and how is this described in terms of monotonicity?

(c) What kind of entailment relationship exists between (5) and

(6) and how is this described in terms of monotonicity?

5. Quantifiers and conservativity

Consider the following quantifier I created, slever:

Slever A B = True if and only if set B has two more

members than set A.

Now consider the sentence Slever bunnies are dancing.

Given the definition, explain why this sentence does or does not
describe Picture A, and Picture B. Is Slever a conservative quan-
tifier? Explain your answer.
204 an open introduction to linguistics
17
Semantics III: Basic Elements of Interpretation

The things that exist come in different forms. There are objects,
time points, events, truth values, and so on. These properties are
fundamental. Truth values are different things than events which
in turn differ from objects. What basic elements are necessary
in order to account for meaning? This is a question semantics is
also concerned with.

On What There Is
With language we can seemingly talk about everything that exists.
The range of things that we can talk about is immense. However, this
vast expressive power creates challenges of its own. For to the human
mind, the things around us are not all of the same nature. There are
physical objects, people, places, times, properties, events and so on.
Languages reflect this categorisation of things in different ways. For
example: many languages classify things into various groups, called
genders or (noun) classes. This classification proceeds along several
lines. Many languages distinguish animates from inanimates, for ex-
ample. And within the animate category, make a further distinction
between people and animals, and within the people between the male
and the female, and so on. There is no objective reason for any particu-
lar division. Why should we care to distinguish humans from animals,
or animals from stones? Clearly, the answer is that language is shaped
to serve the needs of humans, not dogs or stones. And furthermore:
language answers to the need of a society and its culture.
The division of things into genders or classes is realised in different
ways in different languages. For some languages it is part of the mor-
phology (French, for example), while for others it is irrelevant (e.g.
Hungarian). People have attached a lot of importance to that dif-
ference; the so-called Sapir-Whorf-Thesis states that the language
we speak has a profound influence on the way we think. The reason-
ing is that, if gender distinctions are obligatory, then we need to pay
more attention to them and consequently consider this a feature worth
attending to.
Nowadays, it is thought that the influence of language on percep-
tion has been overstated. The fact that the Hopi language does not
distinguish past from future does not mean that monolingual Hopi
206 an open introduction to linguistics

speakers cannot tell the difference or are in any way less likely to
understand it. Nevertheless, there are areas where the thesis has its
merits, for example in the categorical perception of sounds. Recall
that distinctions not in your native languages require effort to hear.
Another example is spatial language: it has been claimed that there
are languages where there are only absolute spatial prepositions of the
type /north/, /south/, and that the speakers of these languages are
far better at tracking absolute directions than speakers of languages
which also have things like /left/ and /right/.
Here, we are more concerned with how the classification of things
is relevant for language in general and semantics in particular. The
elements that seem to be necessary are reflected in the basic semantic
types that most semantic theories recognize. We will discuss a few of
them in this chapter.

Number: Individuals and Groups

Look at the following contrast:

John and Mary met. (17.1)

The senators met. (17.2)
The committee met. (17.3)
∗
John met. (17.4)

Evidently, in order to meet, you need to meet with somebody. But

apparently, you can use the verb with a singular subject if it means
that several people meet, in the sense that they meet with each other.
One way to account for this difference is to say that the verb /meet/
needs a group of people as its subject. But non-people things like
committees, lines and other things can also meet, so this is not the
whole story. Clearly, it is not just a matter of grammar, since it cannot
be decided by syntactic means alone whether a sentence is well formed,
notice, for example, the distinction between (17.3) and (17.4). So, a
better solution would be to say instead that /meet/ needs a group of
things, without further qualification.
There are basically two ways to form groups: you use the plural or
you use /and/. Thus, there are two kinds of basic DPs: those that
denote individuals, and those that denote groups. It is however to be
noted that singular DPs rarely denote groups directly; rather, they de-
note some entity that is constituted (among other things) by its mem-
bers (like a committee or a parliament). They therefore can be seen
as group-denoting DPs only in a derived sense. The primary source of
group-denoting expressions is not the category of group-denoting sin-
gular DPs, but rather plural DPs and DPs coordinated with the help of
/and/. Sometimes, verbs specifically require groups as subjects while
other verbs take both, for example:

Paul is eating. (17.5)

Paul and the neighbouring cat are eating. (17.6)
 semantics iii: basic elements of interpretation 207

In the case of /eat/ we may think of it as being an activity that

is basically performed by each individual alone, not by the group as
a whole. Verbs that are understood to be this way, are said to be
distributive. This means , in the present example, that the fact that
Paul is eating and that the neighbouring cat is eating is enough to
make the sentence (17.6) true. We can also say that the following
inference is correct:

Paul is eating.
The neighbouring cat is eating. (17.7)
∴ Paul and the neighbouring cat are eating.

In general, if a verb V is distributive, then the following inferences go

through:

A Vs.
A and B V. A and B V.
B Vs. (17.8)
∴ A Vs. ∴ B Vs.
∴ A and B V.

Now, if there are groups of people, are there also groups of groups?
There are! This causes some interesting ambiguities. For example, you
can get married as a group of two people by getting married to each
other. This involves a one-time event that makes you married to each
other. A group of two people can also get married by each marrying
someone else. The sentence (17.9) can be read in these two ways. This
is due to the fact that /get married/ is a verb that can take both
groups and individuals as its subject.

John and Sue got married. (17.9)

John got married. (17.10)

Note that the above tests of distributivity go through only if we in-

terpret it in the latter sense. But if we understand it in the non-
distributive sense (John and Sue get married to each other), the infer-
ence does not go through, of course, since that information cannot be
derived from the premisses. Now, let’s form a group of groups:

John and Alison and Bill and Sue got married. (17.11)

There is a reading which says that John marries Alison, and Bill mar-
ries Sue. This reading exists in virtue of the following: /John and
Alison/ denotes a group, and so does /Bill and Sue/. The verb
applies to such groups in the sense ‘marrying each other’. Since our
default idea is that of marriage between exactly two people, if several
groups get married to each other, this must means that all groups (of
two) get married separately.
There are also verbs that encourage a group of groups reading:

The ants were marching eight by eight. (17.12)

Here, we think of a group of groups of ants, each smaller group con-

sisting of eight ants in a line. In the alternative reading, we would
have multiple groups of 8 ants, all marching separately.
208 an open introduction to linguistics

Note that the same pattern can be observed with /meet/:

The senators meet.

The congressmen meet. (17.13)
∴ The senators and the congressmen meet.
However, reader beware: the conclusion has a reading where the sub-
ject is a single group which meets in a one-time event. This is not
what can be concluded from the premisses. All that can be derived
from the premisses is that the senators met (with each other) and that
the congressmen met (with each other), which validates only the read-
ing of the conclusion where there is no meeting between senators and
congressmen.

Time
Properties can change over time. I can be tired in the evening, but the
next morning I might not be tired any more. Some properties have an
inbuilt time dependency. For example, a cat is a kitten only in certain
stages of its life. When it is old enough, by definition, it ceases to be
a kitten, even though it never ceases to be a cat.
Here, we picture time as consisting of a continuum of time points
on a line. We take t and t0 as time points, and we write t < t0 to say
that t is prior to t0 , and t > t0 to say that it is after t0 . For any two
time points t and t0 , either t = t0 , t < t0 or t > t0 . This trichotomy is
reflected in the three basic tenses of English. The present is used for
something that happens now, the past is used for something that has
happened before now, and the future is used for something that will
happen later.

John runs. (present) (17.14)

John ran. (past) (17.15)
John will run. (future) (17.16)

We make references to time in various ways. One is with words like

/now/ and /yesterday/. /now/ refers to the very moment of time
when it is uttered; /yesterday/ refers to any time point that is on the
day before today. Today on the other hand is the day that contains the
time point referred to by ‘now’. Other words require some additional
calculation.

John realized on Monday that he had (17.17)

lost his wallet the day before.

We do not know exactly when these things happened, but we do know

that John’s realisation happened in the past (because we used past
tense); and it happened on a Monday. His losing the wallet happened
just the day before that Monday, so it was on a Sunday. Suppose we
replaced /the day before/ by /yesterday/:

John realized on Monday that he had (17.18)

lost his wallet yesterday.
 semantics iii: basic elements of interpretation 209

Then John’s realisation is in the past, and it is on a Monday. His losing

his wallet is prior to his realisation (we infer that, among other things,
from the phrase /had lost/ which is in a tense called pluperfect).
And it was yesterday when he lost the wallet. So, the most straightfor-
ward interpretation is that today is Monday and yesterday John lost
his wallet, and today (Monday) he realizes that. An alternative, but
less intuitive interpretation is that today is Tuesday and he realized
yesterday (Monday) that on that same day he had lost his wallet. Actually, the phrase /on Monday/ is
That time is linear and transitive accounts for the following infer- dispreferred here. We are not likely to
say exactly what day of the week it is
ence patterns: when it is today or yesterday or tomor-
row. But that is not something that
A Ved. semantics concerns itself with. I can
A Ved before B. say: let’s go to the swimming pool on
B Ved. Thursday even when today is Thurs-
B Ved before C.
∴ A Ved before B or A Ved after B day. It is just odd to do so.
∴ A Ved before C.
or A Ved at the same as B.
(17.19)

Location
Space is as important in daily experience as time. We grow up think-
ing spatially; how to get furniture through a door, how to catch the
neighbour’s cat, how to not get hit by a car, etc. All these things
require coordination of actions and thinking in time and space. This
is the reason why language is filled with phrases that in one way or
another refer to (points in) space. The most evident expressions are
/here/, which functions in the same way as /now/, and /there/, which
is analogous to /then/. /here/ denotes the space that the speaker is
occupying at the moment of utterance. This concept is also relevant
for the meaning of words such as words /come/ and /go/. If someone
is approaching me right now, I can say:

He is coming. (17.20)

But I would not say that he is going. That would imply he is moving
away from me now. German uses verbal prefixes (/hin/ and /her/)
for a lot of verbs to indicate whether movement is directed towards
the speaker or not.
Unlike time, space is not linear. It is organized differently, and lan-
guage reflects that. Suppose I want to say where a particular building
is located on campus. Typically we phrase this by giving an orienta-
tion and a distance (also known as ‘polar coordinates’). For example,

200 m southwest of here (17.21)

gives an orientation (southwest) and a distance (200 metres). The

orientation is either given in absolute terms, or it is given relative to
the way one (or someone else) is positioned (now or at another point
in time), for example /to the right/. For example, to understand
the meaning of what I am saying when I say /Go to the right!/ you
have to know which way I am facing.
210 an open introduction to linguistics

Worlds and Situations

We have started out by saying that sentences are either true or false.
So, any given sentence such as the following has to be either true or
false:

Paul is chasing a squirrel. (17.22)

Napoleon lost the battle of Waterloo. (17.23)
Kittens are young cats. (17.24)

We can imagine with respect to (17.22) that it is true right now or that
it is false. In fact, we probably do not even know. With (17.23) it is
the same, although if we have learned a little history we will know that
it is true. Still, we might find ourselves thinking ‘what if Napoleon had
actually won the battle of Waterloo ...’. Thus, we are able to imagine
a situation that is different from the actual situation. The technical
term is world. Worlds decide the truth value of every sentence one
way or another. There are worlds in which (17.22) is true and (17.23)
is false, others in which (17.22) is false and (17.23) is true, others in
which both are false, and again others in which both are true. And
one of those worlds is the one (and only) world we live in.
Seemingly then, any sentence can be true or false in some world.
But this is not quite true. Sentences like (17.24) are different, because
they are always true. Of course, you could imagine otherwise, but
this would be tantamount to violating the rules of the language the
sentence is expressed in. If I were to say ‘suppose that kittens are not
young cats but in fact old rats ...’ what I ask you is to change the
way English is understood. Then I am not talking about a different
world, but a different language. Worlds have an independent existence
from the language that is being used. We say then that (17.24) is
necessarily true, just like /4+7=11/.
The denotation of a word like /cat/ in this world is the set of all
beings that are cats. This can change from world to world. We can
imagine a world that has absolutely no cats. (If we go back in time,
there was a time when this was actually true.) Or one that has no
mice. But we do not suppose that, just because there are different sets
of cats in different worlds, the meaning of the word changes—it does
not. That’s why you cannot suppose that kittens are old rats. We say
that the meaning of the word /cat/ is a function cat0 that for each
and every world w gives some set, cat0 (w ). We of course understand
that cat0 (w ) is the set of all cats in w. (Some people use the word
intension (sic) for that function.) Likewise, the intension of the word
/rat/ gives for each world w the set of all rats in w, and likewise for
the word /kitten/. It is a (world-independent) fact of English that:

kitten0 (w ) ⊆ cat0 (w ), kitten0 (w ) ∩ rat0 (w ) = ∅ (17.25)

There are plenty of words that are sensitive not just to their denotation
 semantics iii: basic elements of interpretation 211

but also to their meaning:

John doubts that Homer has lived. (17.26)

Robin thinks that Napoleon actually won Waterloo.
(17.27)

Nobody actually knows whether or not Homer has existed. Still we

think that the sentence ‘Homer has lived.’ has a definite answer (some
ghost could tell us...). It is either true or not. We can hold beliefs
that settle the question one way or another, regardless of whether
the sentence is factually true or not. Robin, for example, might be
informed about the meaning of all English words, and yet be a little
weak on history. So she holds that Napoleon won Waterloo. John
might believe the opposite, and Robin might believe that Homer has
lived, while John does not. Different people, different opinions. But
to disagree on the fact that kittens are cats and not rats means not
using English any more.

Events

When I sit behind the computer typing on the keyboard, that is an

activity. You can watch me do it and describe the activity in various
ways. You can say:

He is typing on the keyboard. (17.28)

He is fixing the computer. (17.29)
He is writing a new book. (17.30)

Both (17.28) and (17.29) can manifest as the same physical activity:
me sitting behind the computer and typing something in. Whether or
not I am fixing the computer by doing so, time will tell. But in theory,
I could be fixing the computer just by typing on the keyboard. The
same goes for writing a book. Thus, one and the same physical activity
can be a manifestation of various different things. In order to capture
this insight, it has been proposed that verbs denote particular objects
called events. There are events of typing, as there are events of fixing
the computer, and events of writing a book. Events are distinct from
the activities that they manifest as.
Events will figure in Chapter 19, so we shall not go into much detail
here. There are a few things worth knowing about events. First, there
are two kinds of events: states and processes. A state is an event
where nothing changes:

Lisa knows Spanish. (17.31)

Harry is 8 feet tall. (17.32)

These are examples where the truth of something is asserted at a mo-

ment of time, but there is no indication that something is changing.
212 an open introduction to linguistics

By contrast, the following sentences talk about processes:

Paul is running. (17.33)

The administrator is filling out the form. (17.34)
The artist is making a sculpture. (17.35)

In the first example, Paul is changing place or posture. In the second,

the administrator is writing on a piece of paper, the act of which
changes that very piece of paper. In the third example a lump of
material changes into a statue. Events have participants whose number
and contribution can vary greatly. A process always involves someone
or something that undergoes change. This is called the theme. In
(17.33) the theme is Paul, in (17.34) the theme is the form, in (17.35)
the theme is the sculpture. Events usually have a participant that
makes the change happen; in (17.33) the actor is again Paul, in (17.34)
it is the administrator, in (17.35) it is the artist. There need not be
an actor, just as there need not be a theme; but usually there is a
theme. Some events have what is called an experiencer. In the next
sentence, Jeremy is the experiencer of hate:

Jeremy hates football. (17.36)

Notice that experiencer predicates express states of emotion, so they

fall into the category of verbs that express states rather than processes.
Another class of participants are the beneficiaries; these are the ones
for whose benefit an action is performed, like the wife of the violinist
in the following example:

The violinist composed a sonata for his wife. (17.37)

The list of participant types is longer, but these ones are some of the
most common ones.
Processes are sometimes meant to finish in some state and some-
times they are not. If you are said to be running, no indication is made
as to when you will stop doing so. If you are said to be running a mile,
then there is an inherent point that defines the end of the activity you
engage in. Some verbs denote the latter kind of events: /arrive/,
/reach/, /pop/, /finish/. The process they denote is finished when
something specific happens.

Mary arrived in London. (17.38)

The composer finished the oratorio. (17.39)

In (17.38) the arriving of Mary happens at a more or less clearly defined

time span, say, when the train gets close to the station up to when it
comes to a halt. Similarly for (17.39), where the finishing is the last
stretch of the event of writing the oratorio. So, you can write an
oratorio for a long time, maybe years, but you can only finish it in a
considerably shorter time, at the end of your writing it.
 semantics iii: basic elements of interpretation 213

Study Questions
1. What is the difference between distributive and non-distributive
verbs? What is your natural interpretation of the following verbs,
distributive or non-distributive?

a. fight
b. sleep
c. talk
d. sit

2. Some sentences, such as /Napoleon lost the battle of Water-

loo/ are largely unambiguous, but nevertheless do not have a fixed
truth value, as it might be false in the context of an alternate history
novel, for example. Other sentences, such as /A joey is a young
kangaroo/, do have a fixed truth value. Why is this?

3. How are the terms event, process, state, theme and experiencer re-
lated?
18
Semantics IV: Scope

Different analyses of sentences give rise to different c-command

relations between constituents. These in turn determine different
interpretations of sentences. Thus one of the reasons why sen-
tences can mean different things is that they can have different
structures.

Scope ambiguities can arise when there are two quantified expres-
sions that can be in different dominance relations. Consider for exam-
ple the following sentences:

(1) Everyone in this room speaks two languages.

There are two different potential interpretations of sentence (1): Either

everyone in the room can speak two languages, possibly different ones,
or there are two specific languages, say e.g. Dutch and English, that
everyone in this room can speak. In the first interpretation, the surface
relations match the scope dominance relations. We call this surface
scope. In the second interpretation, the DP two languages takes scope
over the universally quantified subject Everyone. We call this inverse
scope. This potential ambiguity in scope relations is also exploited to
create jokes. Consider the following:

(2) In the United States a man is run over by an automobile every

seven minutes. That poor man.
(3) Bono addressing a crowd at a concert: "Every time I clap my
hands, a child dies of malnutrition. Voice in the crowd: “Stop
clapping!”

Actually, it’s not just quantified expressions, but also other semantic
operators that can enter into scope (read: dominance) relations with
other operators.
Scopal relations can also be entered into between adverbs and quan-
tified expressions, and by quantified phrases and intensional verbs (Ex-
amples from Tunstall, 1981):

(4) Ronny quickly made all the beds.

(5) Ronny completed the task of making the beds in a short amount
of time. quickly >all the beds
216 an open introduction to linguistics

(6) Ronny made each of the beds in a short amount of time, but it
took a while to complete the whole task. all the beds > quickly
(7) Sophie thinks a friend will come by tomorrow.
(8) Sophie thinks some friend or other will come by. think > a
friend
(9) There is a particular friend, say Emma, that Sophie thinks will
come by. a friend > think

Negation can enter into different scopal relations with different DPs.
Consider the following example:

(10) Jan didn’t solve one problem.

Example (10) is ambiguous between a reading where one problem takes

wide scope over the negation (there is one problem that Jan didn’t
solve) or a reading where the negation takes scope over the DP, e.g.
Jan didn’t solve any problems. But scope can be tricky: not all oper-
ators can enter into multiple scopal relations that result in ambiguity.
Consider also the following examples:

(11) The butler didn’t murder three people.

(12) The butler didn’t murder at least three people.

Example (11) is ambiguous. It can have the surface scopal reading that
the butler didn’t murder three people, (but perhaps two or one!) or
that there are three people that we can identify, that the butler didn’t
murder. However, the very similar example (12) is not ambiguous as it
does not permit an inverse scope reading, and can only mean that that
the butler did not murder three people (again, maybe two or one).
Formally we can say that if a logical operator Op1 c-commands
another operator, Op2, than the reading where Op1 takes scope over
Op2 is available. That’s the surface scope reading. But the reading
where Op2 takes scope over Op1 is not always available. It turns
out that this relates to monotonicity properties. Non-decreaseing VPs
can usually have inverse scope readings, whereas monotone decreasing
VPs tend not to allow inverse scope readings. This again illustrates
the relevance of monotonicity properties to language.
Different quantifiers therefore have different tendencies to prefer
different scopes. For example, each tends to want to take wide-scope
more than every. In fact we rank different quantifiers in terms of their
preferences for wide scope:

each >> every >> all >> most

We can then confirm this with some examples:

(13) Mike knows a song for all holidays.

(14) Mike knows a song for every holiday.
(15) Mike knows a song for each holiday.
 semantics iv: scope 217

Because each likes to take wide-scope, it seems to be quite impossible

to get the reading for (15) that there is one single song for all holidays
that Mike knows, but this reading is possible with every and all. So
the potential scopal relations are related to the type of quantifiers.

Relative clauses
Let’s consider some other cases where scope makes a difference in in-
terpretation. Look at the difference between (18.1) and (18.2):

This is the only car we have, which has (18.1)

recently been repaired.
This is the only car we have which has (18.2)
recently been repaired.

The part /which has recently been repaired/ is a clause that func-
tions like an adjective; it is called a relative clause. Unlike adjec-
tives, relative clauses follow the noun. Notice that /we have/ also is
a relative clause, though it is somewhat shortened (we could replace it
/which we have/).
Suppose you go to a car dealer and he utters (18.1). Then he is
basically saying that he has only one car. Moreover, this car has been
under repair recently. Suppose however he utters (18.2). Then he
is only saying that there is a single car that has been under repair
recently, while he may still have tons of others. It is clear from which
dealer you want to buy. To visualize this difference, we indicate the
scope of the operator /the only/:

This is the only [car we have], which has (18.3)

recently been repaired.
This is the only [car we have which has (18.4)
recently been repaired].

In the first sentence, /the only/ takes scope over /car we have/. It
has scope over /car we have which has recently been repaired/
in the second sentence. It is not necessary to formalize the semantics
of /the only/. We need only say that something is the only P , if and
only if (a) it is a P , and (b) nothing else is a P . So, if /the only/ takes
scope only over /car we have/, then the car we are talking about is
a car the dealer has (by (a)), and there is no other car that he has (by
(b)). Thus, he has only one car. If the scope is /car we have which
has recently been repaired/, then the car we are talking about
has recently been repaired (by (a)), it is one of the dealer’s cars (also
by (a)), and there is no other car like that in the dealer’s possession.
So, there may be other cars that the dealer has, but they have not
been repaired, and there may be other cars that were not the dealer’s
but that have been repaired.
The difference in structure between the two is signalled by the
comma. If the comma is added, the scope of /the only/ ends there.
218 an open introduction to linguistics

The relative clause is then said to be non-restrictive; if the comma

is not there, the relative clause is said to be restrictive.
If we relate this to X-syntax again, we see that non-restrictive rel-
ative clauses must be at least D0 -adjuncts, because they are not in the
scope of /the/. In fact, one can see that they are DP-adjuncts. Let
us see how this goes. First we notice that /the/ can be replaced by a
possessive phrase:

Simon’s favourite bedtime story (18.5)

Paul’s three recent attacks on squirrels (18.6)

The possessors (Simon, Paul) are DPs themselves, so they are put in
the specifier position of the DP, quite like the determiner /the/. Now, Notice that even though the possessive
the actual structure we are proposing is this: and the determiner cannot co-occur in
English this is not due to the fact that
they compete for the same position.
[DP [DP Paul][D0 [D ’s][NP three recent attacks on squirrels]]] In Hungarian, for example, they can
co-occur:
(18.8)
Mari-nak a cipő-je (18.7)
Now, take a DP which has a non-restrictive relative clause: Mary-dat the shoe-poss3sg
Mary’s shoe
Simon’s only bedtime story, which he listens to (18.9)
Literally translated, this means
carefully Mary’s the shoe her. The possessive
/Marinak/ (in the dative!) occurs
before the actual determiner.

This DP is identical to what we saw earlier, except that we have Notice that the ’s is analysed here as
/Simon’s only/ . In 18.9, /Simon’s only/ takes /bedtime story/ the D-head; there are other possibili-
ties, for example taking it to be a case
as its scope. It can only do so if the relative clause is an adjunct to marker (which I prefer)
the DP /Simon’s only bedtime story/.
Likewise one may wonder about the place of restrictive relative
clauses. It is clear that they can be neither adjuncts to DP nor ad-
juncts to D0 (because then /the only/ cannot take scope over them).
Restrictive relative clauses are therefore adjuncts of either a N0 or a
NP.

Problems with Structural Explanations

So far we have seen that differences in interpretation can be explained
by differences in structure. The next example is not of the same kind—
at least at first sight. This example has to do with quantification.
Suppose that the professors complain about office space and the ad-
ministrator says:

Every professor has an office. (18.10)

He might be uttering a true sentence even if there is a single office that

is assigned to all professors. If this is to be understood as a remark
about how generous the university is, then it is probably just short for:

Every professor has his own office. (18.11)

In which case it would be clear that each professor has a different

office. The first reading is semantically stronger than the second. For,
 semantics iv: scope 219

if there is a single office and it is assigned to every professor, then every

professor has an office, albeit the same one. We can use ‘stilted talk’
to make the difference clearer:

There is an office such that it (18.12)

is assigned to every professor.
Every professor is such that an office (18.13)
is assigned to him.

In the first sentence, /there is an office/ takes scope over /every

professor/. In the second sentence /every professor/ takes scope
over /an office/.
Returning to the original sentence (18.10), however, we have diffi-
culties assigning different scope relations to the quantifier: clearly, syn-
tactically /every professor/ takes /an office in its scope in both
readings. This problem has occupied syntacticians for a long time.
The conclusion they came up with is that the mechanism that gets the
different readings is syntactic, but that the derivation at some point
puts the object into c-commanding position over the subject.
There are other examples that are not easily accounted for by syn-
tactic means. Let’s look at an example:

John searches for the holy grail. (18.14)

There are at least three ways of reading this sentence. Under one
interpretation, it means that there is something that John searches
for, and he thinks that is the holy grail. Under another interpretation,
it means that John is looking for something, which might or might
not be the holy grail, but which John believes to be the holy grail.
Under the third interpretation, it is unclear whether there is a holy
grail at all, but John believes that there is, and he is searching for it.
This particular case is interesting because people are divided over the
issue whether or not the holy grail existed (or exists, for that matter).
Additionally, it is not clear what it actually was. So, John might find
it and not know that he has found it. We may paraphrase the readings
as follows:

There is the holy grail and John (18.15)

searches for it as that.
There is something and John (18.16)
searches for it as the holy grail.
John is searching for something (18.17)
as the holy grail.

Here, the meaning difference is brought out as a syntactic scope differ-

ence. The first is the strongest sentence: it implies that both speaker
and John identify a specific object as the holy grail. The second is
somewhat weaker: according to it, there is something of which only
John believes that it is the holy grail. The third is the weakest: only
220 an open introduction to linguistics

John believes that there is such a thing as the holy grail, but it is not
specified whether it actually exists.
Such differences in meaning are quite real. There are people who
do in fact search for the holy grail. Some of them see in it the same
magical object as which it is described in the epics, while others do not
think the object has or ever had the powers associated to it. Rather,
for them the importance is that the knights of King Arthur’s court
possessed or tried to possess it. So, they regard the medieval stories as
myths, that nevertheless tell of something real (as Schliemann believed
that Homer’s Iliad may not have been about the Greeks and their gods,
but at least about the fall of a real city, Troy, one that he later found).
Now, whether or not these differences can be related back to differ-
ences in structures corresponding to (18.14) remains to be seen, since
such a claim is difficult to substantiate.

Study Questions
1. Syntactic, structural differences are a main cause of different inter-
pretations, but they are not the only one. Can you think of some
other causes and give examples?

2. Consider the sentence /I saw the man on the hill with the telescope./.
What are the three possible interpretations of this sentence? Are
these caused by syntactic differences? Hint: think about the position of the
PP /with the telescope/.
3. We have considered three different interpretations of sentence (18.14),
in sentences (18.15)-(18.17). The differences between these inter-
pretations are somewhat difficult to express clearly. Try to give a
clearer expression of the different interpretations, using the notions
of intension and extension (see page 191).

4. A Negative Polarity Item (NPI) is a lexical item that can only occur
in negative contexts. Using this fact and what you already know
about scope, c-commanding and (non-)reflexive pronouns, answer
the following questions:

I. *He likes pie at all.

II. He does not like pie at all.
III. Every baker1 likes himself1 .
IV. *Every baker1 likes him1 .
V. Less than three bakers like pie at all.
VI. *More than three bakers like pie at all.

a. Why is example I ungrammatical, while II is fine?

b. Why is example III grammatical, while IV is ungrammatical?
c. What do examples I and IV have in common?
d. Syntactically, what is /Less than three/?
e. Semantically, what feature does /Less than three/ have that
allows it to be used with at all?
19
Semantics V: Cross-Categorial Parallelism

There is an important distinction in the study of noun deno-

tations between count nouns and mass nouns. An equally im-
portant distinction is between processes and achievements. It is
possible to show that the division inside the class of nouns and
the division inside the class of verbs are quite similar.

We have said that nouns denote objects and that verbs denote events.
It has been observed, however, that some categorisations that have
been made in the domain of objects carry over to events and vice
versa. They ultimately relate to an underlying structure that is similar
in both of them.

Mass and Count Nouns

A particular example of this is the distinction between mass and count
nouns. A noun is said to be a count noun if what it refers to in the
singular is a single object that cannot be seen as consisting of parts
that are also objects denoted by this noun; for example, ‘bus’ is a
count noun. In the singular it denotes a thing that cannot be seen
as consisting of two or more buses. It has parts, for sure, such as a
motor, several seats, windows and so on. But there is no part of it
which is again a bus. We say that the bus is an integrated whole
with respect to being a bus. Even though some parts of it may not be
needed for it to be a bus, they do not by themselves constitute another
bus. Ultimately, the notion of integrated whole is a way of looking at
things: a single train, for example, may consist of two engines in front
of several dozens of wagons. It may be that it has even been obtained
by fusing together two trains. However, that train is not seen as two
trains: it is seen as an integrated whole. That is why things are not
as clear cut as we might like them to be. Although we are mostly
in agreement as to whether a particular object is a train or not, or
whether it is two trains, an abstract definition of an integrated whole
is hard to give.
The treatment of count nouns in formal semantics has therefore
been as follows. A noun denotes a certain set of things, the integrated
wholes of the kind. The word /mouse/ denotes, for example, the set of
all mice in this world; call that set M . A particular object is a mouse
if and only if it is in M . Groups of mice are subsets of M . Therefore,
222 an open introduction to linguistics

the plural /mice/ denotes the set of all subsets of M . (Maybe just the
subsets that contain more that one element, but that is not important
here.) Let us then say this: /A is a B/ is true if what /A/ denotes
is an element of the denotation of /B/. Then /Paul is a mouse./ is
true if and only if Paul is a member of M , that is, if Paul is a mouse.
We shall say the same about /A are B/: it is true if and only if what
/A/ denotes is in /B/. Therefore /Paul and Doug are mice./ is true
if and only if the denotation of /Paul and Doug/ is in the set of all
subsets of M , that is, if it is a subset of M . The denotation of /Paul
and Doug/ is a subset of M exactly if both Paul and Doug are mice.
Let us return to the issue of dividing objects of a kind. It seems
clear that the division into smaller and smaller units must stop at some
point. A train cannot be divided into smaller and smaller trains. At
some point, in fact very soon, it stops to be a train. This does not
apply to some other things, e.g. water. Although science tells us that
things are otherwise, in everyday life water is divisible to any degree
we like. We can take an arbitrary quantity of water and divide it as
we please—the parts will still be water. Thus, /water/ is not a count
noun; it is instead called a mass noun.
One problem remains, though. We have not talked about mass
things, but instead we have consistently talked about mass or count
nouns. We said, for example, that /water/ is a mass noun, not that
the substance water itself is a mass substance. In this case, it is easy to
confuse the two. But there are occasions where the two are different.
The word /furniture/, for example, turns out to be a mass noun,
but is not a mass substance, since what it denotes clearly cannot be
indefinitely divided into parts that are also called furniture.
But how do we know that something is a mass noun if we cannot
ultimately rely on our intuitions about the world? There are a number
of tests that establish this. First, mass nouns do not occur in the plu-
ral. We do not find ∗ /furnitures/ or ∗ /courages/. On the surface,
/waters/ seems to be an exception. However, in this case, /water/
is used in a different sense, where it denotes clearly defined patches of
water, such as rivers or lakes. A better test is this one: mass nouns
freely occur with so-called classifiers, while count nouns do not, for
example:
a glass of water (19.1)
a piece of furniture (19.2)
∗ ∗
a glass/ piece of bus (19.3)
Notice that one does not say ∗ /a glass of furniture/, nor a ∗ /piece
of water/. The kind of classifier that goes with a mass noun depends
on what it actually denotes. Some can be used across the board, like
/a lot/. Notice that classifiers must be used without the indefinite
determiner of the noun. So, with respect to (19.3), we can say /a
piece of a bus/ but then /piece/ is no longer a classifier.
There is an obvious distinction between whether or not objects of
a kind are divisible into objects of the same kind, i.e. whether they
are mass substances, and whether a noun denoting these objects in a
 semantics v: cross-categorial parallelism 223

language is called a mass noun. These concepts must be kept separate.

To a certain degree, languages exercise their freedom of seeing the
world in a different light. Some nouns are mass nouns when it comes
to using the tests, but everybody knows the kind they denote is not
divisible; an example is /furniture/. A somewhat less clear case is
/hair/.
Notice furthermore that whether or not things of a kind can be
divided into things of the same kind can vary according to your world
view, while the mass/count noun distinction is fixed in the language.
This is clearest when we look at /water/. The scientific world view
entails that the process of division must, at some point, come to a
halt. Eventually we have a single molecule of water, and here the
division must stop. Yet, our own experience is different. As humans
we experience water as arbitrarily divisible. The language we speak
has been formed not with a scientific world view in mind; there is no
authority that declares that from now on /water/ must cease to be
used as a mass noun.

Telic and Atelic Verbs

Instead of nouns, let’s look at verbs. Verbs denote events, as we have
said, and its a useful analogy to see events as a kind of film. We may
picture them as a sequence of scenes, lined up like birds on a telephone
cable. For example, scene 1 may have Paul 10 feet away from some
squirrel; scene 2 sees Paul being 8 feet away from the squirrel; scene
3 sees Paul just 6 feet away from the squirrel; and so on, until he is
finally right next to it, ready to eat it. Assume that the squirrel is not
moving at all. This sequence can then be summarized by saying:

Paul is attacking the squirrel. (19.4)

Similarly, in scene 1 someone is sitting over a blank piece of paper. In

scene 2, he has drawn a small line, in scene 3 a fragment of a circle.
Scene by scene, this fragment of a circle grows, until in the last scene
you see a complete circle. You may say:

He has drawn a circle. (19.5)

While you are watching the film, you can say:

He is drawing. (19.6)

Or you can say:

He is spreading ink on the paper. (19.7)

These are all legitimate ways of describing what is or was going on.
Unfortunately, the director has decided to cut a few of the scenes at the
end. So we now have a different film. Now we are not in a position to
truthfully utter (19.5) on the basis of the film any more. This is because
even though what that person began to draw looks like the beginning
of a circle, that circle may actually never have been completed, or
224 an open introduction to linguistics

may have ended up as something completely different. Notice that the

situation is quite like the one where we stop watching the film: we are
witnessing part of the story and can guess what the rest is like, but we
do not know whether the circle gets completed. However, (19.6) and
(19.7) are still fine, regardless of the ending. Even if the director cuts
parts of the beginning, still (19.6) and (19.7) are fine. No matter how
short the film is and no matter what really happens thereafter: the
description is valid.
This is the same situation as before with the nouns. Certain de-
scriptions can be applied also to subdivisions of the film, others can-
not. Those events that can be divided are called atelic; the others
being telic.
Note that by definition, a telic event is one that ends in a certain
state, without which the event would not be the same. In other words:
if we cut out parts of the beginning, that would not hurt. But if we
could out parts of the end, that would make a difference. An example
is the following.

John went to the station. (19.8)

Here, it does not matter where John started out from, as long it was
somewhere away from the station. We can cut parts of the beginning
of the film, and it will still be a film about John’s going to the station.
Telic events are directed towards a goal (that gave them their name;
in Ancient Greek, ‘telos’ meant ‘goal’.). As it appears, the majority of
non-divisible events are telic. A different one is:

John wrote a novel. (19.9)

Here, cutting out parts of the film anywhere will result in making
(19.9) false, because John did not write the novel in its entirety.
Now, how do we test for divisibility (= atelicity)? The standard test
is to see whether /for an hour/ is appropriate as opposed to /in an
hour/. Divisible events can occur with /for an hour/, but not with
/in an hour/. With indivisible events it is the other way around:

John wrote a novel in an hour. (19.10)

∗
John wrote a novel for an hour. (19.11)
∗
John was writing in an hour. (19.12)
John was writing for an hour. (19.13)

So, divisible events can be distinguished from non-divisible events.

However, let us see if we can make the parallel even closer. We have
said that mass terms are divisible. In addition, they are non-countable:
if you pour a little water into your glass, and then again a little bit,
you will still have water; you cannot say you have two water(s). You
can only say this if, say, you have two glasses of water, so the bits of
water are separate. (Actually, with water this still sounds odd, but
water in the sense of rivers allow this use.) Also, it does not make
sense to divide your portion of water in some way and say that you
have two pieces of water in your glass.
 semantics v: cross-categorial parallelism 225

Likewise, suppose that John is running from 1 to 2pm and from

2pm to 3pm and did not stop in between—we would not say that he
ran twice. The process of running stretches along the longest interval
of time it possibly can. There is only one process, just as there is only
one mass of water in your glass, without any boundary. You could say
that the mass of water, too expands its boundaries to the largest extent
it can: up until where the glass defines the boundary of water. Only if
you put another glass of water next to it, you could say that there are
two glasses of water. And if John is running from 1 to 2pm and then
from 3 to 4pm, he ran twice; there are now two processes of running,
because he did stop in between. The existence of a boundary between
things or events determines whether we have one, two, or several of
them.
We must distinguish between the event type denoted by a verb and
the event type denoted by the sentence as a whole. Take the verb
/drinking/. This denotes a process, and is therefore atelic:

∗
Alex was drinking in an hour. (19.14)
Alex was drinking for an hour. (19.15)

When combined with a mass noun it remains atelic, but when com-
bined with a count noun (or a non-divisible substance) it is telic.

∗
Alex drank water in an hour. (19.16)
Alex drank water for an hour. (19.17)
Alex drank a beer in an hour. (19.18)
∗
Alex drank a beer for an hour. (19.19)

Likewise, a series of individual telic events can make up a higher level

atelic event:

∗
Alex was drinking beers in an hour. (19.20)
Alex was drinking beers for an hour. (19.21)

Putting It All Together

Having taken a closer look at phonology, morphology, syntax and se-
mantics, we shall revisit the big picture of the first chapter. We said
that there is one operation, called ‘merge’ and that it operates on all
of these four levels at the same time. However, we had to make con-
cessions in the way we construe the levels themselves. For example, we
argued that the English past tense marker was [d], but that it some-
times gets modified in a predictable way to [t] or [@d]. Thus we were
led to posit two levels: a deep phonological level and a surface phono-
logical level. Likewise we have posited a deep syntactic level and its
corresponding surface level (after movement has taken place), and also
a deep and surface morphological level.
This leaves us with a problem if we want to apply the ‘merge’ opera-
tion to all levels at the same time, since we can apply the morphological
rules only after we have the surface syntactical representation, because
226 an open introduction to linguistics

the latter reorders the lexical elements. Likewise, the deep phonologi-
cal representation becomes apparent only after we have computed the
surface morphological form. Thus, the parallel model should be re-
placed by a sequential model, which has been advocated for by Igor
Mel’čuk (in his Meaning-to-Text theory). In this model, the levels are
not parallel, but rather they are ordered sequentially. Thus, when we
want to say something, we do this by first organising the semantic
representation, then the words on the basis of that representation and
the syntactic rules, then the morphological representation on the basis
of the lexical, and the phonological on the basis of the morphologi-
cal representation. Listening and understanding involves the reversed
sequence.
An alternative paradigm is that of generative grammar. Generative
grammar assumes a generative process, which is essentially indepen-
dent of all levels. It runs by itself, but it interfaces with the phonology
and the meaning at certain points. The transformations are not taken
to be operations that are actually executed, but rather as ways to or-
ganize syntactic (and linguistic) knowledge. This makes the empirical
assessment of this theory very difficult, because it is difficult to say
what sort of evidence is evidence for or against this model.
Another option is to use other models of syntax, for example ones
that try to eliminate the distinction between deep and surface struc-
ture. In Generalized Phrase Structure Grammar (GPSG), for example,
the question words are generated directly in sentence-initial position;
there simply is no underlying structure that first puts the object ad-
jacent to the verb, only to move it to the beginning of the sentence
later.
There are probably as many theories as there are linguists. But
even though the discussion surrounding the architecture of linguistics
has been very fashionable in the last decades, some things still remain
that are problematic for most theories.
Let’s just consider one very irritating fact. We have seen earlier
that Malay uses reduplication for the plural. This poses a number
of problems: first of all the plural sign has no substance: there is
no actual string that signals the plural (like the English /s/). What
signals the plural is the reduplication. This is essentially a function ρ
that turns a string x into that string concatenated with itself in the
following way:

ρ ( x ) : = xa - a x (19.22)

Thus, ρ(kerani) = kerani-kerani. This function does not care

whether the input string is an actual word of Malay. It could be
anything. But it is this function which captures the phonology of the
plural sign. This has the consequence that the phonological represen-
tation of signs must be very complicated if the story of parallelism is to
be upheld (recall the plural of /mouse/ with a floating segment). We
have to postulate signs whose phonology is not a string but a function
on strings. Unfortunately, no other theory can do better here if that
is what Malay is like. Thus, the door has to be opened: there is more
 semantics v: cross-categorial parallelism 227

to the operation of merge than concatenating strings if we also need

this additional function. If that is so, we can try the same for syntax;
there might be more to syntax than concatenating constituents. It has
been claimed that Chinese has a construction that duplicates entire
constituents. Even if that story is not exactly true, the news is irritat-
ing. It would mean that entire constituents are there just to convey a
single piece of meaning.
Japanese and mass/count [c] 3in

(1) Ringo o tabeta.

apple-acc ate
(I) ate an apple

• Japanese doesn’t have determiners

• Japanese seems to make no distinction between mass or count nouns

on the noun form itself

• Japanese does mark noun-type when counting units of things

1.5in

(2) Watashi-wa ringo-o ip-pun de tabeta.

I ate APPLE within one minute.
(3) Watashi-wa ringo-o ip-pun-kan de tabeta.
I ate APPLE for one minute.

• “in one minute” and “for one minute” can be used with the same
predicates in Japanese

• this then marks the telicity of the event

(4) Watashi-wa ringo-o ip-pun de tabeta.

I ate APPLE within one minute.
(5) Watashi-wa ringo-o ip-pun de tabete-shimatta.
I ate up APPLE entirely within one minute.
(6) *Watashi-wa ringo-o ip-pun de tabete-ita.
I was eating APPLE within one minute.
(7) Watashi-wa ringo-o ip-pun-kan de tabeta.
I ate APPLE for one minute.
(8) *Watashi-wa ringo-o ip-pun-kan de tabete-shimatta.
I ate up APPLE entirely for one minute.
(9) Watashi-wa ringo-o ip-pun-kan de tabete-ita.
I was eating APPLE for one minute.

• The Japanese example illustrates that the ideas of quantification

and cumulative reference are cross-linguistically relevant even when
they are coded with different linguistic structures

• Suggests these are cognitively basic ideas that are then language
specifically realized
228 an open introduction to linguistics

Study Questions
1. How do the terms mass noun, mass substance, count noun and
count substance relate to each other? Try to define each term.

2. What are telic and atelic verbs? How are these related to the
notions of event, process and state?

3. What is the reason for choosing a sequential model of the ‘merge’

operation over a parallel model?

4. For each of the following nouns, say whether they are mass or count
nouns and whether they refer to mass or count substances.

a. lego block
b. iron ore
c. jeans
d. cutlery
e. lake

5. Of each of the following events, determine whether they are telic or

atelic.

a. Harry helped students.

b. Harry painted a picture.
c. Harry flew.
d. Harry flew to Dumbledore.
e. Harry ate a cookie.

6. We have proposed deep and surface levels for phonology, morphol-

ogy, and syntax, but not for semantics. Why don’t we need a deep
and surface semantic level? Or do you think we do? If so, how
would you define the deep and surface semantic levels?
20
Pragmatics

Pragmatics is the study of meaning in context. You can say the same
sentence in different contexts, and the place in time in which you say
it, the people that you say it to, the information that both you and
the hearer share, all influence how the sentence is interpreted. Major
topics include topic structure, focus, speech acts, and presupposition
and conversational implicatures. In this short introduction to prag-
matics we’ll only cover the latter two topics, presuppositions and
conversational implicatures.

Presupposition

Presupposition theory is one of the most researched topics in pragmat-

ics. But what are presuppositions? How can you identify them in a
text and how should they be analyzed? Let’s first look at an example:

(1) John stopped smoking.

(2) Jayne emailed JOHN too.

Example (1) asserts that John, at the time of the utterance, is not a
smoker. But this sentence also communicates additional information.
It also communicates that John used to smoke at a time earlier than
utterance time. This second piece of information is called a presup-
position. In order to understand the asserted information we have to
first accept the presupposed information as ‘background’ information.
(It would be very odd to say that John stopped smoking if John had
never been a smoker.) Example (2) also has a presupposition. This
sentence asserts that Jayne communicated with John via electronic
mail, and presupposes that Jayne also emailed someone else. Note
that the phonetic focus on JOHN in this example helps remove some
ambiguity, and Jayne EMAILED John too would instead presuppose
that Jayne also did something else to John (but will still assert the
same thing as (2)).
Presupposed information is semantically different from asserted in-
formation, and we see this clearly when we examine how both types
of information behave when in the scope of logical operators. Logi-
cal operators are lexical items that communicate logical functions, like
negation. Consider the negated version of (1):
230 an open introduction to linguistics

(3) John didn’t stop smoking.

P: John used to smoke

Example (3) adds a negation to the sentence. Negation then makes

clear that what is asserted is that John is still a smoker. But the pre-
supposition does not change. Even if John didn’t stop smoking, we
still understand (3) as stating that he used to smoke. This is called
The Negation Test for Presupposition and for awhile semanticists
thought this was sufficient for identifying presuppositions. However,
for some constructions the negation test doesn’t give clear results or
cannot be naturally used. For example, we can negate (2), but the
negation and its scopal relation with the focused JOHN is not clear,
making the sentences awkward and the presupposition harder to iden-
tify.

(4) Jayne didn’t email JOHN too.

But luckily, the main feature of presupposition is not that it is un-

affected by negation, but that it is unaffected by logical operators in
general. This means that embedding sentences under if-then state-
ments, modal adverbs or modal verbs, etc. can all be used as tests
for presupposition, and for most sentences some of these will be nat-
ural. For this reason, a more thorough test has been introduced, the
S-Family of tests:

1. S (simple),

2. not-S (denial),

3. S? (yes-no question),

4. if-S (antecedent of conditional sentence, and

5. perhaps-S (possibility).

Take a sentence S. What inferences does S communicate? Inferences

that are presuppositions will persist under the different embeddings
(logical environments) listed above. Let’s try it for (1)

1. S (simple): John stopped smoking.

2. not-S (denial): John didn’t stop smoking.

3. S? (yes-no question): Did John stop smoking?

4. if-S (antecedent of conditional sentence): If John stopped smoking,

he won’t be taking so many breaks.

5. perhaps-S (possibility): Perhaps John stopped smoking.

You should verify for yourself that for each sentence in the S-family
of tests, the presuppositional inference (that John used to smoke at
a time previous to the utterance time) goes through, but that the
assertion (that John does not smoke now) gets modified in all but the
first statement.
 pragmatics 231

Presupposition Triggers

Where do presuppositions come from ? Most semanticists believe that

presuppositions are part of the meaning associated with certain lexical
items. These lexical items are then called presupposition triggers.
In the above examples, the change-of-state verb stop is the trigger.
Stopping something has a prerequisite associated with it that to be
felicitous, whatever is stopped must have first been going on. This
prerequisite is then the presupposition. Stop is a change-of-state verb,
that describes a change from one state to another. Other change-of-
state verbs include start, continue and quite. Change-of-state verbs as
a class actually seem to all presuppose their initial states, so they are
all presupposition triggers. Consider the following examples and
There are many other sets of triggers. For example, possessives verify for yourself that the change-of-
state verbs are triggers:
trigger the presupposition of the existence of the thing possessed (see
(6), (9))): • Ed started jogging.
• Ann continues to bike to work.
(5) Jack’s son is bald.
• Bo quit complaining.
(6) P: Jack has a son
(7) P: Jack exists.
(8) The Bernoulliborg’s rotating door is broken.
(9) P:The Bernoulliborg has a rotating door.
(10) P: Bernoulliborg exists.

Cognitive factive verbs and factive adjective presuppose their com-

plements (e.g. (12), (16)), it-clefts presuppose the information in the
relative clause ((20)), definite descriptions presuppose the existence of
their referents (e.g. (22)), and proper names presuppose the existence
of their referent ((7), (10), (13), (14), (17) (18), (21)):

(11) John realized Sara was the best programmer in the group.
(12) P: Sara is the best programmer in the group.
(13) P: John exists.
(14) P: Sara exists.
(15) Sara was surprised John noticed her.
(16) P: John noticed Sara
(17) P: John exists.
(18) P: Sara exists.
(19) It was Jack who found the hidden village.
(20) P: someone found the hidden village
(21) P: Jack exists.
(22) P: there is a hidden village

For all of the above examples, you can check if the presupposition is
as claimed by applying tests from the S-family of tests.
232 an open introduction to linguistics

Trigger type Triggers

Proper names John, Mary, Berlin, Groningen
Change-of-state verbs stop, quit, begin, start, continue
Factive verbs know, realized, discover
Factive adjectives be surprised that, be happy that, etc.
It-clefts it was X who (someone) did something
Definite Descriptions the King of Sweden, the cafeteria, the desks

Presuppositional Projection

Because presuppositions function as a kind of background information,

we interpret them almost as if we’ve already been told them. This make
the use of presupposition alongside assertion a very efficient way of
communicating compared to simply using assertion only. For example,
we could paraphrase presuppositional sentences in the following way:

(23) John stopped smoking.

(24) John used to smoke, and he stopped doing it.
(25) It was John who borrowed my boat without asking.
(26) Someone borrowed my boat without asking and it was John
who did that.
(27) The students loved the book.
(28) There are some students and there’s a book. They love it.

Based on their behavior, two main theories of presupposition have

been developed. Irene Heim developed a theory of presupposition
withing context change semantics (CCS). When a presupposition is en-
countered, the hearer should update their context so that it includes
the presupposition. This means they interpret a sentence like (25)
as if they had heard (26). This is process is called accommoda-
tion. The other theory of presupposition is the Presuppositions as
Anaphors theory, developed by Rob van der Sandt and Bart Geurts in
Nijmegen. They instead argue that presuppositions are just another
type of anaphoric expression, not unlike a pronoun or do so. Unlike
pronouns however, presuppositions carry quite a bit of information.
This makes it possible for speakers to accommodate them even if they
did not know the information before hand. The argue that the infor-
mation that is presupposed get projected to the global context, as if
it it had been said before. But presuppositional material will not be
projected to this context if it (1) is not informative or (2) if it leads
to a contradiction. This then explains why sometimes presuppositions
do not project, e.g.

(29) John doesn’t have a son. But if John had a son, his son is
bald.

In Example (29), his son presupposes John has a son. But if we

 pragmatics 233

project this to the main context, we get something that is epistemically

incoherent (the belief system of the speaker would not be consistent),
e.g.

(30) ??John has a son. John doesn’t have a son. But if John had a
son, his son is bald.

So we cannot accommodate the presuppositional information in the

main context. We then look at each level of embedding. This sentence Projecting gold blocks, not presuppo-
actually contains another level of embedding: the if -clause creates a sitions
context:

(31) ??John doesn’t have a son. But if John has a son & if John had
a son, his son is bald.

This is slightly contradictory, because “ if John had a son” with the

subjunctive verb form had suggests that John doesn’t have a son. In-
stead we have to accommodate in the local clause where the presup-
position was presupposed:

(32) John doesn’t have a son. But if John had a son, John has a
son and he is bald.

This is the most straightforward interpretation of (29); his son cannot

be bald if there is no son.
Presuppositions that do not project, are said to be cancelled. Can-
celling a presupposition usually only happens when they are triggered
under and embedding. When an unembedded presupposition is can-
celled, it has the same awkwardness as when an entailment is cancelled.
Both types of cancellations are clearly infelicitous:

(33) ??Dan went to the Groningen castle, but there is no Groningen

castle.
(34) ??Dan went to the Groningen castle, but he didn’t go anywhere.

Now let’s look at another type meaning that sentences can commu-
nicate that goes beyond the simple semantic meaning, but that is
based on different principles and behaves differently than presuppo-
sition: conversational implicatures.

Conversational Implicatures – Beyond literal meaning

Sentences have basic meanings, and this is what semantics studies. But
in actual communicative settings we can use sentences to mean more
than their basic semantic meaning. This is because our expectations
about how the world works, in particular how we expect people to
behave to each other, allows us to get a richer interpretation from
what people say. Much of our communication is therefore not literal,
e.g. someone who says I could eat a horse doesn’t actually mean that
the speaker is able to eat a large animal.
The first linguist to look at this systematically was Paul Grice. He
234 an open introduction to linguistics

developed a theory of how semantic meaning is enriched by the fact

that we interact and cooperate with each other, developing what he
called the Cooperative Principle.

• The Cooperative Principle Make your contribution such as is

required, at the stage at which it occurs, by the accepted purpose
or direction of the talk exchange in which you are engaged.

Based on this overriding principle of how humans should converse,

he then developed more specific rules that govern how much we should
say, the manner in which it should be said etc. Examples abound
where the meaning of an expression is either different from what is
expected, or the literal meaning is enriched. Grice’s contribution was
that he identified classes of such enriched meaning, all derived from a
hearer’s expectation that the speaker will follow conventions that are
derived from the Cooperative Principle. These rules are called maxims
and there are four: (1) The Maxim of Quality, (2) The Maxim of Rele-
vance, (3) The Maxim of Quantity and (4) The Maxim of Manner. It’s
not always easy to recognize the effect of the Cooperative Principle in
action. When the Maxim’s are followed, conversations are unremark-
able. That’s why the Maxim’s are often illustrated with cases where
they are broken, which can also have a communicative effect.

Maxim of Quality

• Do not say what you believe to be false.

• Do not say that for which you lack adequate evidence.

The Maxim of Quality is a good rule to follow for successful com-

munication. Still, speakers can also break the maxim of Quality for
special communicative effects. Example (35) shows someone break-
ing the Maxim of Quality, by saying something that is clearly false.
Speaker B seems to not want to say how old they are. That is the im-
plication that flounting the maxim seems to communicate. In example
(36), Speaker B might not know the way, but in order to be polite, and
to come across as helpful, may simply agree with what Speaker A has
suggested. In this way speaker B follows the maxi of Q, but without
being very helpful.

(35) speaker a: How old are you?

speaker b: I’m 1000 years old.
(36) speaker a: Excuse me, should I turn left here to get to
Main Street?
speaker b: Yes dear, that’s right.

Maxim of Relevance

• Be relevant.
 pragmatics 235

It has been argued that all of Grice’s maxim’s could be subsumed

under the maxim of relevance, and in fact, an entire theory of linguistic
meaning, Relevance Theory, has been developed around this maxim.
Being relevant, and expecting others to be relevant, does seem to be a
characteristic of rational, normal conversation. For example:

(37) child: Can I have 5 Euros?

mom: My purse is in the bedroom.

In example (37), the mother’s contribution does not seem to answer

the question posed by the child on the surface. But clearly we all
understand that the answer is affirmative. That’s because the next
step in agreeing to give someone money is to actually give it, which
the Mom facilitates by giving instructions on where to find the money.
In this way, the maxim of relevance can be used to explain why seem-
ingly unrelated statements are perfectly reasonable responses to a ques-
tion. Speakers and hearers assume relevance, so they will try to find
an interpretation that makes the response reasonable in the context.
Here are some more examples of experimental items that were used in
a study investigating children’s acquisition of conversational implica-
tures (Surian et al. 2010). In each case, Tom’s response fails to follow
the maxim of relevance, while Jane’s response does follow Relevance:

(38) lucy: Where do you live?

*tom: I live on the moon.
jane: I live in London.
(39) lucy: Do you have any brothers?
*tom: Yes, I have 500 brothers.
jane: Yes, I have two brothers.
(40) lucy: What’s your favourite programme on television?
tom: My favourite is cartoons.
*jane: My favourite is sandwiches.

Maxim of Quantity

• Make your contribution as informative as is required (for the current

purposes of the exchange).

• Do not make your contribution more informative than is required.

The Maxim of Quantity is one of the most studied maxims.

(41) speaker a: Did you do all your homework?

speaker b: I did math.
Implicature: Speaker did not do all the homework.
(42) speaker a: Did you get Harrison Ford’s autograph?
speaker b: I got Garrik Hagon’s.
Implicature= Speaker B did not get Harrison Ford’s auto-
graph.
236 an open introduction to linguistics

Both (41) and (42) illustrate how we can derive additional meaning be-
yond what is actually said by considering how the speaker’s seemingly
indirect response actually might be very informative if the speaker is
following the Maxim of Quantity.
Below are some more examples from the acquisition study of Surian
et al.

(43) lucy: What did you eat for lunch ?

*tom: Some food.
jane: Pizza.
(44) lucy: What did you see at the cinema last night?
tom: Snow White.
*jane: A film.
(45) lucy: What did you have for breakfast?
tom: I had cornflakes, and then a boiled egg and toast.
*jane: A hard boiled egg cooked in hot water in a sauce
pan.

Examples (43) and (44) both illustrate the Quantity maxim’s re-
quirement that speakers should be informative. Example (45) illus-
trates the Quantity’s maxim’s requirement that speaker’s avoid re-
dundancy. Prolixity:
Some quantity implicatures are so frequent that many researchers the fact of using too many words and
therefore being boring or difficult to
believe they are no longer calculated on the fly by adults, but their read or listen to: Despite all its ab-
meanings have become conventional. The most well-studied quantity surd prolixity, this is one of the great
books in the English language. (From
implicatures is the some-not all implicature. This is the implicature Cambridge Dictionary Online)
calculated on the basis of the quantity maxim whenever the words
some is used. Consider:

(46) Some of the students passed the exam.

(47) Some windows have lights on.

Anyone who has taken an introductory logic course knows that some
semantically seems to just mean at least one. But often in normal
conversation we instead interpret it as meaning ‘not all’, even though
it is compatible with all. This interpretation is analyzed as being the
result of a conversational implicature based on the Maxim of Quantity:
A teacher who utters (46) would say that all the students passed the
exam if that was the case (why not share this amazing success!). Sim-
ilarly, if all the windows have lights on, the speaker should say this. If
they do not, it conversationally implies that not all windows had lights
on.
Maxim of Manner

• Be clear.

• Avoid obscurity of expression.

• Avoid ambiguity.

• Be brief (avoid unnecessary prolixity).

 pragmatics 237

• Be orderly.

(48) mom: Which do you like, tea or milk?

child: Maybe tea or maybe milk.

Why the child’s response seems awkward (and non-adult-like) can be

explained by appealing to the Maxim of Manner. The child is not
avoiding ambiguity. When it seems that a competent speaker is actu-
ally intentionally breaking the Manner maxim, we should look further
for hidden meaning. We can see this very clearly in a classic example
of the Maxim of Manner from Grice:

(49) Dr. Z is writing a letter of recommendation for a Ph.d. student

for a position in a Philosophy Department. His letter says:
Dear Sir or Madam, Dr. X’s command of English is excellent,
and he is always very punctual. Yours truly, Dr. Z.
Implicature= Don’t hire Dr. X.

Here’s another example, with an extensive analysis, taken from

Chris Pott’s notes on implicature calculation. It shows step by step
how the reasoning process should go in interpretation.

Pott’s Notes on Implicature Calculation

For the next example, I believe we need to supplement Grice

with the following principle: Definition 3 (The division of prag-
matic labor; Horn 1984; Levinson 2000). Normal events are
reported with normal language. Unusual events are reported
with unusual language.

(50) speaker: To show that she is pleased, Sue con-

tracts her zygomatic major muscle and her orbic-
ularis oculi muscle.
Implicature : Sues expressions of happiness are cold,
clinical, and robotic.

1. Assume the speaker is cooperative.

2. Assume scientific language is associated with being cold and

clinical.

3. There is a shorter, less obscure form, smiles, competing with

contracts her zygomatic major muscle and her orbicularis
oculi muscle.

4. By the Levinson/Horn heuristic def. 3, Sue’s smiles must

be unusual.

5. By 2. (and a theory of connotations!), her smiles are unusual

in being cold and clinical. Comments The implicature is
highly dependent upon contextual assumptions, and it leans
heavily on cooperativity.
238 an open introduction to linguistics

For example, if the speaker is known to be cold and clinical

himself, then we do not draw the implicature, because premise
(1.) is false in the relevant sense. Similarly, if the context is
that of an anatomy class, then the competition in (3.) breaks
down.

You may have already noticed that sometimes it seems as if the

same example can be explained by more than one of the maxims. For
example, in example (48), both manner and quantity could be argued
to be playing a role. This is a problem with Grice’s system that was
noticed early on. For example, consider what Sadock (1978) has to
say about this:

Sadock (1978) on Implicature Calculation

First of all, the Cooperative Principles with its maxims of

Quality, Quantity, Relation and Manner is so vague that
almost anything can be “worked out” on the basis of almost
any meaning. Let me briefly rehash the maxims in an effort to
prove my point:

1. The maxim of Quantity has two parts that require the coop-
erative speaker to say as much but no more than is required
for his particular purposes in the “talk exchange”

2. The maxim of Quality also has two parts, and demands that
the speaker say only what he believes to be true and on that
for which he has sufficient evidence;

3. The maxim of Relation (relevance) urges the speaker to

make hist contribution relevant; and

4. The maxim of Manner cautions the speaker to be methodical

to avoid ambiguity, prolixity, and obscurity.

So powerful is each of the maxims that at times they vie for the
privilege of explaining the same facts. It is not clear to me, for
example, how one could both be relevant and say either less or
more than is required. In what way is the avoidance of prolixity
different from saying only as much as is required? If a particular
contribution is obscure does it not also lack relevance? While
it is perhaps possible to eliminate some of the redundancy in
the maxims, I feel that the extreme power of the system is in
fact an unavoidable characteristic.

How do you identify conversational implicatures? Well one

feature of an implicature is that because it is a non-logical inference,
it can be cancelled or suspended without creating a contradiction. So
one test to see if you are dealing with an implicature is to try to correct
it with actually and see if it is felicitous. Below you can see that for
each of the four maxims, we can suspend the implicature with actually.
 pragmatics 239

Compare the following cases of an entailment ((51)), a presupposition

((52)) and an implicature of each type, with actually:

(51) John and Mary went to the party. ?? Actually, John didn’t
go.
(52) John stopped smoking for stoptober. ?? Actually, he never
smoked.
(53) I’m 1000 years old. Actually, I’m just kidding.
(54) lucy: Do you have a car ?
tom: I have a really fast bike. Actually, I also have a car.
(55) Some of the students passed the exam. Actually, all of them.
(56) Dr.Z is almost always on time. Actually, he’s always on time.
I was just kidding.

Example (55) is clearly easy to correct without concluding the the

speaker is epistemically inconsistent. The same can be said for the
other implicatures. This cannot be said of the entailment or presup-
positional examples, where it seems that the speaker doesn’t know
what they are talking about.

Language Technology Applications of Grice’s Maxims

In Language Technology there has been some recent interest in using

Grice’s maxims as a way to evaluate conversational dialogue systems.
In a student paper at RANLP, Jwalapuram (2017) reported on research
that evaluated popular conversational chatbots by asking humans to
judge them on how well the responses correlated with Gricean maxims.
They operationalized the maxims by asking participants to answer
questions like the following for sample dialogues from different systems:
Instructions given to annotators in Jwalapuram

After presenting the user with the example dialog, the user is
asked to rate the system performance on a Likert scale from 1
to 5 for 4 questions:

1. Is the system as informative as it can be and gives as much

information as is needed and no more? (Maxim of Quantity)

2. Is the system truthful or does it give information that is

false, to the best of your world knowledge? (Maxim of Qual-
ity)

3. Are the system responses relevant, i.e., does the system re-
spond with things pertinent to the discussion? (Maxim of
Relation)

4. Are the system responses clear, orderly and without obscu-

rity and ambiguity? (Maxim of Manner)
240 an open introduction to linguistics

It’s helpful to see some example dialogues and how they were rated.
Here are two sample dialogues, with an excerpt from the paper:

Dialogues from Jwalapuram 2017

We discuss the scores of Dialogue 5 and Dialogue 6 in a

little more detail. D6 is from a train timetable information
agent that uses explicit confirmation and was found to have a
high transaction success rate (Danieli and Gerbino, 1995) as
opposed to D5’s more efficient implicit confirmation (Walker
et al., 1997).

Dialog 5 (sample):

User: I want to go from Torino to Roma

System: Approximately what time
of day would you like to travel?
User: What are the options?
System: Morning or evening.

Dialog 6 (sample):

System: Do you want to leave from Trento? Yes or No?

User: No.
System: Where do you want to leave from?
User: I want to leave from Torino.
System: Do you want to leave from Torino? Yes or No?
User: Yes.

D5 performs better than D6 in all four counts . . . despite accom-

plishing the same task. There is an especially high difference in
relevance and manner ratings (statistically significant p-values
of <0.0001 and 0.0063 respectively). This indicates that users
do not like explicit confirmations despite them having a higher
task success rate (as implicit confirmation may be more likely
to generate errors or repair dialog) (Danieli and Gerbino, 1995).

Summary
There is more to meaning than the semantic meaning we’ve examined
until now. Words and expressions carry with them prerequisites on
the context in which they are used. One class of these prerequisites is
called presuppositions, a class of inference that distinguishes itself
from other information by being unaffected by logical operators. An-
other type of enriched meaning is that of conversational implicatures.
This meaning is derived from our expectation of what ideal human
communication should be like. Paul Grice identified features of suc-
cessful, normal communication, calling it the cooperative principles.
He further specified the cooperative principle into four maxims. Using
 pragmatics 241

these maxims, we can understand why sometimes the literal meaning of

an utterance is enriched. These ‘enrichments’ are called conversational
implicatures. This is not just useful for analyzing everyday commu-
nication, but it also could have applications in language technology,
such as in providing a method to evaluate dialogue systems.

Study Questions
Presupposition, Entailment or Conversational Implicature
For all of the sentence pairs (or conversation-sentence pairs) below:

1. Say whether the second sentence (the one labelled (ii)) is an entail-
ment, a presupposition or a conversational implicature of the first
sentence or conversation (i);

2. Explain how you know this (based on what tests, e.g. negation,
deniability).

Question a. already has an example answer.

1. John stopped smoking.

2. John used to smoke.

Answer: Sentence (i) presupposes sentence (ii). You can see this by
negating (i) and seeing that (ii) would still hold.

1. a: Did Jill bike from Groningen to Brussels?

b: She biked to Breda.

2. Jill didn’t bike all the way to Brussels.

1. Some of Harry’s friends are vegetarians.

2. Not all of Harry’s friends are vegetarians.

1. Harry didn’t discover that his friend eats meat.

2. Harry has a friend.

1. The Doctor has a sonic screwdriver and the Doctor has psychic
paper.

2. The Doctor has psychic paper.

1. a: Did Vincent do well in his programming course?

b: He really enjoyed it.

2. Vincent didn’t do well in his programming course.

1. At least three students own an electric bike.

2. At least three students own a bike.

242 an open introduction to linguistics

1. Rose realized that time travel is possible.

2. Time travel is possible.

1. Last year the AI department started a matching program.

2. There was no matching program at the AI department before last

year.
21
Language Acquisition

Children’s knowledge of language

Learning sounds
Newborn children are immediately able to cry. Interestingly, it’s clear
both to them and us that crying is not a meaningful speech sound.
Actually, newborns seem to know without experience that certain dis-
tinctions are not important in human languages. For example, babies
only a few days old do not respond differently if the speaker is male,
female, young, old or if the articulation is louder or softer. Instead
they seem to immediately focus on linguistically relevant contrasts. Two paradigms for studying the lin-
Soon after birth, infants start practicing making speech sounds, us- guistic development of babies are HAS
ing both passive and active articulators. Initially most sounds are (High Amplitude Sucking) and Pref-
erential Looking. Watch the following
vowels, because babies are unable to control the pressure, closing and video about how babies learning
opening necessary to produce plosives. Gradually they gain more con- English develop their recognition
trol, and will begin ‘babbling’, that is: producing all speech sounds of English and non-English speech
sounds. https://www.youtube.com/
present in the world’s languages. At this stage children are also able watch?v=WvM5bqUsbu8
to hear contrasts between sounds that are not present in their own
language. For example, English doesn’t have unaspirated intial conso-
nants, and a native English speaking adult does not hear this contrast
if you play audio files for them. However, Hindi, and many other
languages do have this contrast. Experimental studies of very young
babies perception (under six months) at the universal babbling stage
have found that even children being raised in an English speaking en-
vironment are able to hear the Hindi distinction.
Eventually, at around 9-11 months, babbling becomes language spe-
cific, and children only produce sounds that are present in their native
language. . At this point, children also lose their ability to distinguish Please note that everyone should
between all the sound contrasts in the worlds languages. They instead watch the TED talk from Dr. Patri-
cia Kuhl about comprehension at this
become specialized in hearing relevant contrasts in their own language. stage of language learning.
This changes has an advantage: hearing only relevant distinctions (e.g.
ones that lead to minimal pairs in your language) makes it easier to
learn words.
At about 12 months, children begin to say their first words. Often
the first words is mama or papa. Please note, that both these words
use bilabials. Cross-linguistically, words for mothers and fathers often
are made up of sounds that are easy to see visually, and this seems to
244 an open introduction to linguistics

be related to the fact that the child will be able to see how the parent
produces these words, allowing them to mirror their own production
on the parents. (But do we say mama for mother because that is
just an arbitrary word we associated with mothers, or do we choose to
let mama mean mother because these are sounds that the child (can)
produce?)

Vocalization Type Age

Crying 0;0
Cooing 2 months
Gooing 4 months
Babbling-universal 6-9 months
Babbling-specific 9-11 months
One-word (holophrastic) 12 months
Two-word 18-24 months
Telegraphic 24-30 months
“Later development” 30+ months

Learning words
Children are amazingly good at learning words once they start. Usu-
ally a handful of words are produced around 12 months, but later
children learn multiple new words a day. Keep in mind as well that
when learning words, children are also learning concepts, and how to
categorize the world. Word learning is much more than simply learning
to name things. First, getting the concepts right is sometimes hard.
Second, initially children do not know that many words, yet they still
attempt to communicate, which leads to a number of ‘shortcomings’
(errors?) in their production of words.
For both these reasons, cross-linguistically children make two com-
mon errors in learning the reference of words. Overextension errors
are when a child applies a word too widely, calling all men pappa, or
all meals breakfast, for example. Underextension errors are when chil-
dren interpret a word too specifically, such as only calling chocolate
cupcakes ‘cupcake’ or only their own car is a car.
Even thought very young children know very few words, they are
very adept at making them work for them. At this age children often From Vocabulary development in
use a single word to mean an entire sentence. So cookie can mean Give children with Down syndrome: Longi-
tudinal cross-sectional data (Zampini
me a cookie or Here is your cookie. This is called the holophrastic phase & D’Odorico, 2013)
in word learning, where a single word can be interpreted as a whole Comparison of vocabulary size be-
tween 18-30 months in Typically
(holo-) sentence (phrastic). Developing children in different per-
centiles, compared to children with
Age Active, productive vocabulary Down’s syndrome in the 50 percentile.
15 months 10 words
18 months 15 words
18+ 30-50 words in one month

The average 18-month-old recognizes about 260 words, whereas

preschool-aged children recognize between 1,000 and 10,000 words (Fen-
son et al., 2007; Shipley & McAfee, 2015). Age Notation in Acquisition Re-
search
In language acquisition researcher, age
in years is followed by a semi-colon,
followed by the number of months.
Thus age 3 is written as 3;0, 6 months
is written as 0;6.
 language acquisition 245

At age 5;0 children seem to be able to learn an average 9-12 words a

day (10,000/(365 x 3)) Children seem to be able to learn so many words
quickly because they are able to do something called ‘fast-mapping’,
young children need only one exposure to add a new word to their
vocabulary This is remarkably fast. We know from other research
that adult L2 (second language) learners need 20+ exposures to learn
to recognize a new word. Experiments that look at children’s fast-
mapping ability usually test to see how well they can remember the
name of novel words that are pointed out to them in an uncluttered
environment, when they are explicitly taught the word and its mean-
ing. But how do children learn the words of their language in a natural
context? Word learning, in a natural context, is actually not an easy
task. This was already recognized by Quine (1960) with his example
of “gavagai”: Here’s a paraphrase of his example (from Samualsen and
McMurry, 2018):

Imagine you are a field linguist studying a community whose

language you do not know. You go hunting with a group of
tribesmen and see a rabbit hop past. One of the tribesmen shouts
“gavagai.” How to you determine what this new word means?
It could be “rabbit” but it could also be “hopping,” “fluffy,”
“dinner,” “get it!” or a host of other things. You can watch a short video about the
gavagai problem from the perspective
of the philosophy of language here:
How do children learn even simpler words, like concrete nouns?
https://www.youtube.com/watch?v=
There are actually two reigning theories currently. The first argues 0YY4qxkqm3o
that children are able to infer the meaning of words cross-situationally,
by remembering features of each situation in which a word is used,
and keeping a record of which words were heard in which contexts.
For examples, at breakfast a child hears the word milk, when there
are many things on the table, including milk. The association between
the word milk and all as yet unknown things at the table will be
strengthened. Later, when in the grocery store, father says milk, the
association between the word milk, and all other items present, will
be strengthened as well. Eventually, milk will become the strongest
associated word with milk. See an example with animals in the margin.

The alternative theory, advanced by John Trueswell and Laila Gleit-

man, is that children simply guess what the referent of a new word is
from among the things that are present in their environment when
they heard the word, excluding of course objects that already have a
name in their vocabulary. This theory is called Propose-but-Verify and
it predicts that word learning works roughly like this. At breakfast,
An example of how cross-situational
the word milk is used. The child guesses it means knife. Later at the vocabulary learning works
grocery store, the father says milk again. Because there is no knife
present, the child realizes their hypothesis about the meaning of milk
was wrong, and guesses again from among the objects present at the
grocery store. The child does not store information about what other
objects were present at either occasion, only storing the link that they
guess between milk and an object. Eventually, the child will guess
right.
246 an open introduction to linguistics

Both these theories of word meaning make different predictions.

Cross-situational word learning predicts gradual acquisition of words,
where two words may, for awhile, be in competition with each other.
Accuracy will slowly increase, but there will not be great jumps in ac-
curacy. Propose-but-verify, on the other hand, predicts that for some
words, children will guess right immediately, getting those words right
all the time from the first encounter. For other words, they will guess
wrong, and when they realize this, because the theory simply predicts
guessing, they will simply randomly choose among existing unclaimed
meanings. This means that after an incorrect guess, children will show
chance performance if they are using the Propose-but-Verify method.
Both theories have been researched with novel word learning exper-
iments both with adults and children, and support for both theories
has been shown. Actually, very simple computer simulations also show
that both theories will eventual lead to word learning if given sufficient
exposures and time. But researchers are interested in which theory
seems to more accurately characterize human word learning. Results
suggest that children are initially too good at word learning if cross-
situational learning is being used, because they are correct after only
1 or 2 exposures, which would be insufficient time for a strong asso-
ciation to have been created. Work by Trueswell has also shown that
both adults and children show chance performance after a wrong guess,
which is consistent with Propose-but-Verify. However, recent work by
Roembke and MacMurry found evidence that both cross-situational
learning and Propose-but-Verify may may be involved. They found
that in an experiment with many exposures, adults showed slightly
better than chance guessing, which is not explained by the Propose-
but-Verify proposal.
Researchers also study how children learn more complex words. In
fact, children also seem to be able to learn quite complex lexical in-
formation even from simply hearing others use these words. Consider
this description of an experiment on transitive and intransitive verb
learning described by Swigly (2010):

By the age of 2, children reveal long-term storage of syntactic

information about a word, even when the initial presentation
event provides essentially no additional semantic content (Yuan
& Fisher, 2009). In that study, 28 month olds watched a film
of two women talking about unpictured events (“Hey . . . Jim
is gonna blick the cat!” . . . “Really?”. . .) consistently
using transitive sentences (e.g., “blick the cat”) or intransitive
sentences (“she was blicking”). One or two days later, children
were told to “find blicking” while viewing pairs of scenes. One
scene involved a novel activity appropriate for a transitive verb
(two participants), and one scene was appropriate for an intransi-
tive verb (one participant). Children’s looking patterns depended
upon their prior exposure to the word. If they had heard about
something getting blicked, they appropriately fixated the transi-
tive event more; if they had heard about someone blicking, they
 language acquisition 247

fixated the intransitive event more. (From: Swigly, 2010)

From this example we can see that children can learn quite com-
plex information about a word, including whether it is a transitive or
intransitive verb, just by hearing it used. This example also shows
that children seem to really be able to ‘fastmap’ verbs even when they
are not provided with any visual cues, retaining that knowledge across
a time span of a day or two between the initial presentation of the
verbs and the testing(!). This clearly shows that small children seem
to be exceptionally good at learning new words and word meanings.
But how do they move on from the one word stage to more complex
utterances?

Learning to combine words

After the one-word stage, where a single word like Up! might be used
by a child to mean Pick me up!, we come to the two-word stage. With
two or more words we can start to look for evidence that children know
something about syntactic structure. One characteristic of children’s
speech at the 2+ word stage is that it is often telegraphic, i.e. children
tend to remove all function words. So instead of saying “Give me the
book” children say “Give book”, leaving most function words out.
When we start to talk about syntax, a major question to investigate
is whether or not children understand that language has a hierarchical
structure. Matthei (1982) did an experimental study investigating
this question. 35 children between the ages of 3;9-6;30 (19 girls) were
trained in the meaning of the word second and how to count objects
from left to right. In the experiment, they were shown sequences of red
and green balls. There were two types of experimental items: neutral
arrays and biased arrays. See below:

Children were then asked questions like Point to the second green
ball. Note that the NP here, second green ball, is analyzed as the
feature of being second applying to the feature of being a green ball,
because second scopes over green ball in a hierarchical tree structure
(see margin example). However if children interpret it as having a flat,
non-hierarchical structure, just as in e.g. the big red ball, then being
second and being green will both be interpreted as features of the ball,
248 an open introduction to linguistics

as if they were two intersective adjectives taking scope over the ball
only.
For the neutral arrays children and adults correct identified the ball.
(See the arrows). However for the biased array, 54.5 % of responses
were incorrect, with children incorrectly choosing the second ball that
was green. This suggests that the children interpreted second green ball
as the ball that is second and is green, i.e. with a flat, non-hierarchical
structure. Because in the neutral array the second ball was red, this
interpretation was not possible.
These results were taken as evidence that children do not under-
stand hierarchical structure. However, children were correct almost
half the time, so this suggests that the error was not really systematic.
For this reason Matthei investigated whether children’s errors could
be explained by the complexity of the syntax used in the sentences
being too complex. Children had to interpret a noun phrase with two
modifiers. If the syntax of the DPs could be simplified they might
Structures assumed in syntax at the
perform better. They did this in Experiment 2, using a single modifier time of Matthei’s experiment. Note
and two different types of objects. You can see an example of a biased the three branches for flat structures.)
experimental item used in Experiment 2 below.

In this experiment, children were asked to Point to the second ball.

The expectation was that if children made errors in the previous ex-
periment because of the double prenominal modifier, in this task they
would do much better. Unexpectedly, children still made errors 36%
of the time, choosing the first red ball that appeared in the second
position. This result strengthens the conclusion that children strug-
gle with hierarchical structure, and that hierarchical structure is not
innate.
This isn’t the end of the story however. In 1984 Hamburger and
A cartoon example of motherese.
Crain also redid the experiment, but with a small change in the method- A more modern term is caretaker
ology. Hamburger and Crain were concerned that children’s non-adult speech. Caretaker speech is charac-
terized by exaggerated intonation,
like responses might be caused by incremental processing, so they first
extra loudness, slower tempo and
let children hear the entire sentence, before letting them see the stim- simpler syntactic structures, among
uli. In their experiment, the error rate dropped to 14%. This result, other features. But the motherese
hypothesis refers not to the idea that
and later results, showed that children do know that noun phrases can children learn language from this
have hierarchical structures. special way of talking to them, but to
the more general idea that children
Hamburger and Crain’s results strongly suggest that Matthei’s chil-
learn their language from some kind
dren were simply beginning with pointing as soon as they heard the of teacher-like interaction with their
word second, without waiting. caretakers.

We should note that Crain is a strong nativist: he belongs to the

group of people that believe that children are born with complete,
adult grammatical understanding. Any errors made by chidren are
simply a result of the experimenter creating an experiment that doesn’t
make sense to a child. For example, Crain believes it’s quite strange
for a child to be asked whether or not a sentence matches a picture,
or whether or not you can say a certain phrase. He argues that exper-
iments with children should not use these methods. So this result is
 language acquisition 249

exactly what Crain and colleagues would expect: children gave non-
adult answers because the way in which they were tested was not
appropriate for them.

The Logical Problem of Language Acquisition

How do children become proficient adults speakers? This is one of

the major research questions in language acquisition. We know that
children’s comprehension and production seem to differ from adult’s,
but how do they then reach the adult state? Below, four major pro-
posals are presented: (1) Motherese, (2) Behaviorism, (3) Analogical
Learning and (4) Innateness (Nativism).

The Motherese hypothesis

One explanation of how children become proficient speakers of their

native language is that they learn it from their caregivers. This is
the motherese hypothesis. Of course, at some level this hypothesis is
most certainly true, because children do learn to speak the language
that they hear in their environment. But is it because adults are
actually teaching the children the language actively? Is there as certain
type of interaction with small babies and young children that is an
essential component to their acquisition process? The hypothesis was
empirically investigated by Peters (1983), who studied the features of
caregiver speech in three different cultures, and found big differences
between groups. The three groups were:

• White middle-class families in New England (Massachusetts and

other eastern coastal states)

• Black Mill worker families in Trackton, South Carolina

• Kaluli (Samoan) speakers in Papua New Guinea

For each group, Peters investigated a number of research questions

related to how adults interacted with children linguistically. Let’s go
through some of these questions in detail:
Do children have the right to participate in conversations?
Each of these cultures differed in how they interacted with infants and
small children. One major research question was: does the culture
believe that children have the right to participate in a conversation?
The white, middle-class families in New England included children in
conversations, and in fact encouraged children (even babies) to partic-
ipate. But among the mill workers in Trackton, children had to earn
the right to participate. For caretakers in Papua New Guinea, the
very question seemed odd: only as children become older and could
speak more fluently did they consider it possible to include them in
conversations.
Do caregivers talk to babies in a special way? In most west-
ern countries, there is a special manner in which we talk to children.
250 an open introduction to linguistics

We often use exaggerated intonation and exaggerated pronunciation.

This simplified and exaggerated form of English is called child-directed
speech (formally caregiver or motherese), and among the New England Note that New England refers to an
families studied this was a common way to speak to children. This was area on the north east coast of the
United States.
quite different from the other cultures. In Trackton, caregivers didn’t
talk to babies. In Papua New Guinea, caretakers spoke with babies
but they didn’t modify their language.
What do babies mean when they make early sounds? The cul-
tures also differed as to how they interpreted babies first vocalizations.
The families in New England believed that even babies were making
sounds that they could interpret, and readily ascribed meanings to the
babies’ attempts. But the families in Trackton and Samoa did not
interpret early vocalizations as meaningful.

How do caregivers react to babies’ sounds? For the white

middle-class English speakers, babies’ and small children’s sounds should
be repeated and paraphrased whenever possible. This is also a com-
mon feature of interactions with babies and small children: the mill
workers in Trackton instead responded that they could simply be ig-
nored, probably because they did not believe they were meaningful.
Caregivers in Samoa instead believed that children’s attempts should A speech therapist or language re-
be accepted, but modified or corrected. searcher can estimate a child’s
vocabulary by using a specialized
test, such as the Peabody Picture
Vocabulary Test. A certain number
All children from each speech community (New England, Trackton, of words selected from different
frequency sets in the vocabulary of
Papua New Guinea) become proficient speakers of the adult language. the language are depicted on cards.
They may have differed in the rate of acquisition, and quantity of The investigator then asks children
acquisition, but they all eventually become native speakers. This sug- to identify (or name) the objects. At
a certain point children are unable
gests that Motherese or even frequent interaction with children, is not to continue, and the frequency of
necessary for language acquisition. Even in cultures where caregivers words tested at that point can be
used to estimate active and passive
do not speak with babies directly, adults do talk about what the ba- vocabulary size.
bies are doing between themselves, and talk about other things in the
environment, exposing babies and young children to language in this
indirect way. This seems to be sufficient for language learning to be
stimulated.

Input does matter for vocabulary

However, this doesn’t mean that parental input is completely irrele- Source: Hart & Risley 1995
vant. The opposite seems to be true when we look at rate of acquisition
and vocabulary size, and how it later relates to school success.
Much research has shown that vocabulary size is one of the strongest
predictors of reading and school success. Children’s early environment,
and the quantity, quality and breadth of vocabulary and structures
that are child is exposed to has a major effect on their vocabulary
acquisition. Research has shown that the amount of time parents talk
Source: Hart & Risely 1995
with their children (Huttenlocher et al. 1991), the breadth of the
vocabulary used, (Weizeman and Snow 2001) and also the variation in
vocabulary used all has a positive effect (Huttenolocher et al. 2010).
 language acquisition 251

This relationship between children’s early vocabulary size and fu-

ture academics was discovered in research that looked at the effect of
socio-economic status (SES) on word learning, where major differences
have been attested. For example, Huttenlocher et al. (1991) estimate
that between the ages of 14-26 months children in the professional fam-
ilies studied were exposed to over 200,000 words each week, compared
to around only 60,000 words for children in low-SES families. This
extrapolates into a difference of around 32 million more words for the
children in the professional families in this time period. Similar results
were found in a study by Hart & Risley (1995) (see graphs), where
they estimated that by age four children in professional families had
heard 45 million words, children in working class families had heard
25 million and children in lower income families had only heard 13
million words. This difference in exposure correlates with the size of
the productive vocabulary (words children actively could use, not the
words they recognize) at age 4. Because this measure then correlates
directly with reading and school success, these differences are impor-
tant. Interventions to help boost very young children’s exposure to
more speech and more varied speech would be expected have a major
impact on their later school success.
These results from vocabulary learning thus show that the input
children are exposed to does have an effect on their linguistic develop-
ment and how it relates to academic success. This means that while
the strong view of the motherese hypothesis, that parents need to
‘teach’ their children to talk, is not correct because all children seem
to learn the language in the environment around them, but variations
in vocabulary size (and syntactic complexity) are related to exposure
and verbal interaction. So we can conclude that actually talking to
babies and toddlers does help them develop linguistically.

Behaviorist ideas about language acquisition B.F. Skinner, Behaviorist.

“A child acquired verbal behavior when relatively unpatterned

vocalizations, selectively reinforced, gradually assume forms which
produce appropriate consequences in a given verbal community”
(Skinner 1959:31)
If children don’t learn their language from their parents, how do they
learn it? In the 1950’s behavioral psychologists thought that they had
the answer. Earlier research showed that animals could be conditioned
Skinner box experiment in mod-
to associate a neutral stimulus with a response. For example, Pavlov’s
ern gaming (watch at home):
dog learned to associate a bell with food, which then caused the dog https://www.youtube.com/watch?v=
to salivate when it heard a bell. Psychologist B.F. Skinner believed tWtvrPTbQ_c Note that this is the
same principle that was proposed for
that almost all behavior could be explained as a learned response to language acquisition.
a given stimuli, including language learning. This was a very popular
account of human learning, including language learning, in the 50’s.
This would mean that children would learn language from the adults
around them, by being rewarded with correct responses and under-
standing, and by negatively associating incorrect responses or inter-
pretations, utilizing some type of reinforcement.
252 an open introduction to linguistics

• positive evidence, reinforcement – when we say something correctly

we get what we wanted

• negative – corrections when we say something wrong

So a child that says ‘milk’ when being given a glass of milk will
be praised by the caregiver for that utterance. The desire to get that
praise will increase the association between the word milk and the ob-
ject over time, and the child will learn that the vocalization of milk
refers to the white liquid. A child who mispronounces the word papa,
will react to the negative reinforcement that the father will not rec-
ognize and respond, because they didn’t understand what the child
said. This will encourage the child to improve their pronunciation.
For Skinner, and other behaviorists, the child is (compared to adults)
extremely reactive to how linguistic stimulus is reinforced.
One problem with this idea is that it’s hard to see how the learning
can be extended to links between stimuli and responses as the links
become more complex, and more subtle. Let’s examine how this would
then work for language learning.
Setting aside the challenge of extending the account to more com-
plex and abstract words, and giving a more detailed account of the
actual learning process, we can instead examine its correctness in a
different way. If this account is true, what predictions follow? If chil-
dren are learning to react to linguistic stimuli from examples in the
environment, via some stimulus-outcome system, then they:

• Will imitate what adults say

• Won’t say things they haven’t heard

• Will revise their model of language based on how adults correct

their mistakes

• Will become better speakers if adults praise their correct attempts

We can then examine the degree to which these predictions seem

to be true as a way to judge whether or not behaviorist theories can
account for language learning. However, even the very first prediction
seems to be problematic. Children, even at the stage where they can
say several words, do not seem to be able or willing to imitate what
adults say. Here are a number of examples taken from the Childes
database of child language speech:

parent: He’s going out

child: He go out

parent: That’s an old-time train.

child: Old-time train

parent: Adam, say what I say, “Where can I put them?”

adam: Where I can put them?
 language acquisition 253

In each case, we see that even when trying to imitate, children

produce only what they can produce spontaneously. Further, this isn’t
because they have problems with the pronunciation, but because they
seem to be following their own production strategies.
Another prediction of the behaviorist view of language acquisition
is that children should only produce things that they have heard. But
there is ample evidence that this is not the case. Consider the following
example from Dutch between the child Jan (2;11) and his father, where
Jan leaves out the verb, and in the next utterance leaves out the subject
(also using an incorrect verb form). Leaving out the verb or subject is
not possible in these positions in Dutch.

jan: Dis mijn boekje. (This my little-book.)

dad: Ja. (Yes.)
jan: Is beetje kapotgemaken. (Is little broken)

Consider also the following dialogue between Pien (4;10) and her mother,
where Pien uses an incorrect regularized form of the past-tense verb,
eette (‘*eated’), instead of heb opgegeten:

mom: Ik zag je vanmiddag ook zo braaf eten. Bloemkool en kaas...

(I saw you this afternoon eating so well. Cauliflower and cheese. . . )
pien: Ma maar ik lustte geen kaas, maar ik eette het toch op.
(Mom but I don’t like cheese, but I eated it all up anyway)

The children could not have learned these sequences or these forms
from an adult. In fact, we see systematic errors in children’s speech at
similar ages across languages, so this suggests that there is a blueprint
to the acquisition process. In fact, the regularized form of the past-
tense error, like the form that Pien made in the dialogue above, is
a process seen in most languages: cross-linguistically, children often
make errors in acquiring distinctions between regular and irregular
morphological forms. These examples all suggest that forms cannot A U-shaped learning curve that shows
accuracy for irregular forms in En-
be simply memorized from what children hear from adult speakers.
glish. At time 1, children have simply
Let’s look at the case of the English past-tense morpheme in more memorized all forms, and produce
detail. This example is very important not just in linguistics but also them correctly because of this. At
time 2, accuracy drops, because
in cognitive science. A large body of evidence suggests that children children begin to incorrectly apply a
don’t simply memorize the past-tense forms in English. past-tense formation rule meant for
regular forms. At time 3, children
First recall that we can distinguish between two different types of
have realized that irregular forms
past-tense forms: regular and irregular. In English, and other lan- differ from regular forms, and they
guages as well, young children often go through a phase where they reach the adult stage.

overgeneralize regular past tense endings, applying them incorrectly

to irregular forms. For example, English children say goed instead
of went, and singed instead of sang. What’s remarkable about these
overgeneralizations is that they occur after the child has been correctly
using the irregular forms.
Steven Pinker and other researchers have noticed that this seems
to be a case where a linguistic form or type is acquired in what can
be described as a U-shaped curve. In the initial acquisition, the child
memorizes all forms, regular or irregular. This very young child is not
254 an open introduction to linguistics

yet aware of the difference between regular and irregular forms. As

the child learns more words, they realize that there is some type of
pattern between the present form and the past form for some words.
This leads the child to infer a ‘rule’ that is roughly: add /@d/ to
the end of the present tense root to get the past-tense. Once the
child has learned this rule, they overapply it, which leads to incorrect
overgeneralizations. This thus means that children will say words like
goed even though they have never heard it in the input. At a certain
point, children realize that there are two types of verbs, and they again
start to produce forms in an adult-like manner.
How children infer the past-tense form rule (and whether this is
actually a rule or could be understood some other way), and what
information triggers them to realize there are two types of verbs, are
questions that are hotly debated. The main take-away though is that
this is evidence that children are not simply memorizing the forms
around them to learn language.
We also know that children say different things than simply what
they hear in their environment. A very famous study that showed
that children actually seemed to have learned rules for their native
language was done by Jean Berko in 1958. In this study, they showed
children pictures of novel animals, like the ones to the right. They then
asked children to fill in the plural form in a frame, e.g. This is a wug.
Now there is another one. There are two of them. There are two . . . .
Children correctly and consistently applied the same plural endings
that adults would. Because the names of the creatures were novel, the
children could not have simply memorized the form. Instead, their
answers suggest that they have already internalized a rule about how
to create plural forms from a singular form.
These examples make clear that language learning isn’t simply mem-
orizing forms. Another prediction of behaviorist ideas is that children
should be able to learn from corrections. Corpus studies of natural in-
teractions between children and adults show that children are seldom
corrected by their parents. When caretakers do correct children, it’s
usually pronunciation or facts, but not grammar. In fact, examples
suggest that children are very insensitive to corrections:

child: Nobody don’t like me

mother: No, say ‘Nobody likes me’
child: Nobody don’t like me.
(8 times the Mother tries to correct the child)
You can also watch a video of the
mother: Now, listen carefully, say ‘Nobody likes me’. wug tests at https://www.youtube.
com/watch?v=MgB2iMuEZAA
child: Oh! nobody don’t likes me!
The LingSpace also has a video ex-
plaining why a language isn’t just the
Here is a similar exchange with a Dutch child (from Frijn & De memorization of forms, and also ex-
Haan, 1994): plains the wug test: https://www.
youtube.com/watch?v=1Lb6phcDte4
eva: mag ik de bord? (can I have the(de) plate)
mom: het bord (the(het) plate)
eva: mag ik de bord? (can I have the(de) plate?)
 language acquisition 255

mom: nee, je zegt het bord (no, you say the(het) plate.)
eva: mag ik de bord? (can I have the(de) plate?)
mom: het bord (the(het) plate)
eva: het bord. mag ik nou de bord? (the(het) plate. Can I have
the(de) plate now?)

These examples, and many more, suggest that children are not sen-
sitive to explicit negative evidence. Given the infrequency of caretakers
correcting children, for correction to play a major role in language ac-
quisition we would have to assume that children are very sensitive to
implicit corrections, but that seems rather unlikely.

Can language be learned by analogy?

Another idea that has been proposed is that perhaps children learn lan-
guage by analogy, learning patterns and then expanding them, without
And you can see a (youngish) Chom-
the necessity to refer to hidden structures. However, this strategy also sky, and other people talking about
would predict errors that we actually do not find. Universal Grammar on youtube:
https://www.youtube.com/watch?v=
I painted a red barn X xfiHd6DyuTU

I painted a barn red. X

I saw a red barn. X
I saw a barn red. ×
If children are learning by analogy, we should expect children to say
things like I saw a barn red because if you can use a transitive verb and
color to paint something that color, why can’t you use other transitive
verbs in a similar way? But these, and other similar examples, are not
attested errors in child language. The proposal that analogy plays a
major role in the acquisition process is also unlikely.

The arguments for innateness

The major alternative account to motherese, behaviorism and analogy
as the driving force behind language acquisition is the idea that it is
instead in part the result of innate knowledge and preprogrammed pro-
cesses, including specialized language acquisition abilities. This view is
called the ‘Nativist’ view of language acquisition, and it assumes that
some abstract features of language are genetically encoded. The task
of the child is then simplified immensely, because they simply need
to use the input in their environment together with existing innate
structures to create a grammar. The original proponent of this idea
was Noam Chomsky. Now it is the accepted view among linguists and
some psychologists.
One of the arguments for innateness is what’s called The Poverty
of the Stimulus. The argument is that the amount of language
children are exposed to is too little to account for their acquisition.
Small children are not exposed to very many different constructions.
The input they receive contains production errors, so it is far from
ideal. Children are seldom given explicit negative evidence, which
makes it hard to determine what the grammar they are learning is. But
256 an open introduction to linguistics

despite all these potential difficulties, language acquisition proceeds

very quickly. Babies that are only a few days old already recognize
their mother tongue, and children at age 6 or 7 already seem to be
adult-like (Note however that 7 year old children do not yet have an
adult grammar, but this is usually only recognizable with controlled
experimental testing, so most adults will not notice that children are
interpreting and using expressions differently than they do).
We know that many aspects of language learning are biologically de-
termined, because (1) they are learned naturally (i.e. without teaching
or intervention), and (2) this learning is only possible during certain
critical periods (Lenneberg, 1967). For example, infants between 0;6
-0;9 still can discriminate between all sounds in the worlds languages,
but by 0;10 they have lost this ability (see the lecture by Dr. Kuhl). Check out the Ted talk: The linguistic
genius of babies, by Dr. Patricia
After age six some children may have trouble learning the pronunci-
Kuhl. https://www.ted.com/talks/
ation of a new language. After puberty, learning to pronounce a new patricia_kuhl_the_linguistic_
language like a native speaker is exceptional, and rarely is possible genius_of_babies?language=en

without specialized practice and training. After puberty, the ability

to learn to use complex morphology is also impaired. These biological
limits may be helpful from an evolutionary standpoint: we can always
identify who is a member of our ‘tribe’, because they can speak like
us. They are also all evidence that there is a biological component to
language acquisition.
Theoretically, is it strange to believe a cognitive skill is innate? We
know from other areas that physical changes can be innate, and these
are uncontroversial. But many researchers have resisted this idea, of-
ten with the complaint that if we cannot identify the genes that are
then responsible for language learning, and if we cannot explain exactly
how it proceeds, innateness is a non-starter. However, consider the
following quote from Chomsky where he shows that it’s quite strange
that many researchers are reluctant to consider the idea that language
might be innately controlled, but they accept this idea in other do-
mains without resistance:

If someone came along and said that a bird embryo is somehow

“trained” to grow wings, . . . and human babies are “trained” to
grow arms . . . people would just laugh. Everybody agrees (this)
. . . is due to heredity – the fact that humans develop arms rather
than wings. . . . even though embryologists lack anything like a
detailed understanding of how genes regulate embryological de-
velopment.

So if language, clearly a cognitive skill, is innate, what does that

mean? Clearly not all of language is innate: we have to learn the
vocabulary and structures of the language in our environment. But
how much is innate and what form does it take?
What would the innate structures, our genetic language endowment,
actually look like? Are there actual rules in our brain? Or even trees?
Are there X-bar structures in our brain (and if that’s true, why is it
not really easy to learn X-bar structures in an introductory linguistics
class?) LingSpace also has a video about
nativists and universal grammar:
https://www.youtube.com/watch?v=
MLNFGWJOXjA
 language acquisition 257

Even if you accept that some language knowledge must be innate,

teasing apart the innate from the learned is still tricky. One way to
do it is to look cross-linguistically. This is also why when discussing
syntax we often talk of universal grammar, because syntacticians as-
sume that all languages will have at some level of abstraction the same
architecture. Because the innate aspects of language would be present
in all languages, acquisition research benefits from research in linguis-
tic typology, where researchers compare the structures of different,
unrelated languages. When they find something that seems to be a
universal, it is often a good candidate for an innate structure. One
syntactic example would be X-bar theory. An example from semantics
is the semantic universal that all quantificational determiners are con-
servative. However, it could also be that conservative quantifiers are
easier to learn or verify, in which case there is no need to appeal to an
innateness explanation to account for the universality of conservativity.
And if there is innate knowledge, are there correlates in the learning
process (the interface between the innate and the environment)? Do
we use general learning mechanisms to learn our language, the same
ones we also use to learn in other domains, or do we have specialized
learning mechanisms available just for language learning? Chomsky
(1965) Katz (1966) McNeill (1966) were all early proponents of what
is called the LAD or the language acquisition device. This is the
companion to innate knowledge:a specialized learning mechanism in
the brain that aids the child in language acquisition. Think of this as
an abstract machine, that gets input in the form of sentences in e.g.
English, and its task is to learn the rules of these sentences in order to
be able to generate new ones. How will this device do this, if it only
receives the surface structure forms of language? The explanation is
that the device must already be preprogrammed with universals, since
learning from the input provided will be insufficient (i.e the Poverty
of the Stimulus). The ‘pre-programmed’ information are universal
structures in the form of the universal grammar common to all
languages. The device then only has to determine from the input
exactly what structure the language has. Note that given the previous
discussion of motherese, whatever form the LAD has, it doesn’t seem
to require simplified or even direct input to be successful.
The innateness account (also called the nativist account), makes
a number of predictions as well, predictions that are distinct from
the behaviorist predictions. For example, the account predicts that
children learning very different languages will initially assume similar
structures, until they are confronted with evidence that leads them to
change their grammars. There is some evidence of this. One exam-
ple comes from negated disjunction. Consider the contrast between
English and French from a paper by Nicolae (2017):

(1) Mary didn’t invite Lucy or John for dinner.

(2) Mary didn’t invite Lucy or she didn’t invite John for dinner.
(3) Neither Lucy nor John were invited to dinner by Mary
258 an open introduction to linguistics

(4) Marie n’a pas invité Léa ou Jean à dîner. Marie has not invited
Lea or Jean for dinner
(5) Mary didn’t invite Lucy or she didn’t invite John for dinner.
(6) ??Neither Lucy nor John were invited to dinner by Mary.

As you can see, the English statement (1) is ambiguous. The disjunc-
tion can either take narrow scope ((3)) or wide scope ((2)) over the
negation, leading to two possible readings. Other languages like Ger-
man and Korean also allow both scopal interpretations. However, in
French, only the interpretation where the disjunction has wide scope
over the negation is natural. An interpretation where the negation
gets narrow scope, e.g. (6), is highly degraded. The same pattern
is found in Turkish, Russian, Mandarin and Japanese. Strikingly, all
children begin with the assumption that their language allows both
scopal interpretations. Acquisitionists who believe in innate structures
therefore see this as evidence that children must have initial universal
assumptions, in their LAD, that all scopal relations are possible.
Further evidence comes from children’s errors with wh-questions.
One common error in young children is a failure to invert the subject
and verb. Consider the following example in English and Italian:

(7) Why are you here?

(8) *Why you are here?
(9) Perché tu sei qui? (Eng. Why you are here?)

English children produce incorrect questions like (8), similar to Italian

construction. But exposure to English structures should allow their
LAD to figure out that the language they are learning requires inver-
sion.
This all relates to another discussion point in nativist language ac-
quisition. Because children’s omissions and errors are similar, they
seem to be systematic. Because errors are not explicitly corrected,
learning from the input alone will be difficult in some situations. Pinker
(1984) and other see language learning as learning what set of princi-
ples and constraints are relevant for your language. The child’s task
is then to figure out what this set is, gradually developing a hypoth-
esis about the adult grammar as they get more input. Pinker (1984)
identified four likely situations, shown in the diagram below.

Situation (a) corresponds to the state of a child learning English

who never produces correct questions, always using Italian word or-
 language acquisition 259

der. Situation (b) this would correspond to a child who makes some
errors in question formation, but can form some questions correctly.
Situation (c), where the child’s hypothesis set H is a subset of the adult
target hypothesis T, is the case where the child correctly uses a subset
of target forms, never using an incorrect form. But this child does not
yet produce the full range of questions. Finally, in situation (d), H is a
superset of T. In this case the child uses some forms correctly, but then
also produces incorrect forms. Pinker argued that it would be impos-
sible to go from the superset situation (d) to the correct hypothesis,
because children do not receive or respond to negative evidence. This
is called the superset problem. But only negative evidence in the
form of corrections would inform at child in situation (d) that they are
incorrectly assuming a superset of the target grammar. Because this
is a logically possible situation, Pinker and others have argued that
the only way to explain how children always converge on the target
adult system is to assume that there is innate knowledge in the form of
constraints on the form of grammar that prevents children from ever
ending up in the superset situation in the first place. This view of
language also makes some additional predictions. If the innateness of
language is there in part to prevent children from ending up in a su-
perset grammar, then we should predict that it will encourage children
to always assume a subset grammar. From a subset grammar, positive
evidence in the form of what structures adults around the child use,
are sufficient to extend the child’s hypothesis until it is the same as the
target grammar. This then predicts that if there are variations within
forms cross-linguistically, then children will initially seem to assume
the form that is a subset of the other form, regardless of the correct
forms in the adult language in that particular language.

Recursion: innate but late?

Some researchers believe that humans ability to use recursion, in lan-

guage, is one of the main things that distinguishes us from other higher
animals. Recursion in language manifests when you have a constituent
containing a constituent of the same type. “Do the same operation
again (Merge), and do it upon the same structure” Roeper (2007).
This allows us to have infinite systems from finite rules. For example,
we can have possessives inside possessives:

(10) My mom’s cousin’s friend’s dog is a corgi.

What does recursion look like when sentential constituents are embed-
ded in other sentence constituents? consider the following examples: You can watch a video from the
hit television series ‘Friends’ where
(11) Joey knows Monica and Chandler are in love. they use multiple recursive em-
beddings here: https://youtu.be/
(12) Rachel knows that Joey knows that Monica and Chandler are LUN2YN0bOi8?t=1m14s
in love.
(13) Monica knows that Rachel knows that Joey knows that Monica
and Chandler are in love.
260 an open introduction to linguistics

Syntax will allow this to go on for ever; the only limits come from our
processing abilities.
Because recursion is such an important aspect of language, it is also
a candidate for innateness. So an important question to ask is: what
do children know about recursion? Interestingly, young children do not
seem to understand recursive structures. For example, in the Childes
database we find evidence of children’s comprehension difficulties (from
Roeper, 2011):

mother: what’s your . . . what’s your . . . what’s your cousin Arthur’s

Mummy’s name?
sarah: I don’t . . . your cousin?

Young children seem to interpret recursive structures as if they were

conjunctions in the initial stages. For example, in a study by Gentile
(cited in Roeper), children were shown pictures like the following:

After being introduced to cookie monster and his sister, the exper-
imenter asked children “Can you show me Cookie Monster’s sisters’
picture?”. The correct answer is of course picture A, but one third of
the children actually chose picture B, e.g. “cookie monster and sister”.
If very young children do not understand recursion, does this mean
that it cannot be innate? Not necessarily. Many biologically deter-
mined developmental milestones only appear later in children’s devel-
opment, e.g. molars, puberty. Language may also have innate aspects
that are not initially present. Further, to be feasible, a theory of lan-
guage acquisition must also incorporate realistic assumptions about
children’s executive functioning, e.g. working memory, ability to sup-
press distracting information, etc. This means that there may be as-
pects of language development like the development of the ability to
deal with recursive structures that are dependent on processing abili-
ties, and thus recursion may not be initially available because children
lack e.g. sufficient working memory.

Summary

Children learn to speak their native language, without teaching, and

they do it very quickly. Children are not taught language by their
parents (except the vocabulary) but they still learn. Early theories of
language acquisition have claimed that parents teach children (moth-
erese), that children learn from stimuli and responses (behaviorism), or
that children learn by analogy. But most linguists (and some psychol-
 language acquisition 261

ogists) now agree that a substantial component of language is innate,

but they argue about how much is innate and how it manifests.

Exercises
1. L1 acquisition and Nativism (8 pts)
Consider the following dialogue and answer the questions below.

mother: What do you have there?

child: A ball!

mother: It’s an apple. Can you give me the apple?

child: [ pushes the apple ] Bye, ball!

mother: What happened?

child: The ball goed to you!

mother: The apple went to me?

child: Yes, it goed to you!

(a) How do you explain the use of the word ‘ball’ to refer to ‘apple’?
(b) How do Nativists believe that children learn their native lan-
guages?
(c) Give two arguments that support the Nativist view of language
learning.
(d) How do non-Nativists believe that children learn their native
languages?
(e) Would non-Nativists use the dialogue above as proof to support
their view of how children learn their native languages? Explain
your answer.
(f) The speech of the child in the example above is also character-
istic of a particular stage of U-shaped learning. Explain what
U-shaped learning is, and in which stage of U-shaped learning
this child is.
(g) Is there any evidence that this dialogue is not authentic?

2. Stages in Acquisition

(a) What is an MLU?

(b) List the stages in L1 acquisition.
(c) What distinguishes the universal babbling stage from the spe-
cific babbling stage?
(d) What is the holophrastic stage? Which two types of semantic
generalizations are displayed at this stage? Give examples.
262 an open introduction to linguistics

3. Vocabulary acquisition
In early diary studies, where parent linguists recorded their own
children’s speech, we can see evidence of a common process in early
referent learning discussed in the chapter. The table below presents
the words that children used initially for the first referent, and then
what they used that word to refer to later in subsequent references:

Study Word First referent Usages

Moore, 1896 ‘bird’ sparrows cows, dogs, cats
4. Chamberlain & Chamberlain ‘mooi’ moon cakes, round shapes in books, letter ‘O’
Moore 1896 ‘fly’ fly specks of dirt, dust, own toes, crumbs of bread
Leopold, 1949 ‘wau-wau’ dogs toy dog, soft slippers, picture of old man in furs

(a) What is the type of error in applying words to meanings illus-

trated in the above table?
(b) Why do children make this error? Give two possible reasons.
(c) Do you see a pattern in the what each word can be used to
refer to? Explain what this pattern is, and explain why children
might use this strategy.

5. Errors in children’s speech

For each of the following dialogues, identify the type error in the
child’s speech an explain why researchers think it occurs.

(a) child: *Pappa drinked coffee.

(b) child: *Why you are standing outside?
(c) child: *Give drum. Big drum broken. That busy truck not
go.
(d) child: Nice bunny. Very nice bunny.
parent: Yes, but thats a cat. A nice cat.

6. The Acquisition of Hierarchical Structure

In the study by Matthei, they showed children arrays and asked
them to point to certain objects. In their second experiment in the
Chapter on Language Acquisition, an example experimental item is
shown with boxes and balls. However, the original study used an-
imals such as hippos and bears. Consider the following array that
is a replication of the original experiment items:

(a) What was the purpose of Matthei’s experiments? I.e. what was
he investigating ?
(b) Let’s say you are going to recreate Matthei’s experiment. What
kind of question would you ask children to answer given the array
of objects above?
 language acquisition 263

(c) What answer would you expect from adults who have full gram-
matical understanding? (identify the animal by letter).
(d) What answer would you expect many of the children to give
if they responded similarly to the children in Matthei’s study?
(Identify the animal by letter).
(e) What did Matthei conclude from his results?
(f) Can we be sure of this conclusion? Elaborate on any potential
alternative explanations.
22
Language Families and History of Languages

It is clear even to an untrained person that certain languages, say

Italian and Spanish, are somehow related. Careful analysis has
established relationships between languages beyond doubt. A
precursor language called Indo-European has been proposed as
the ancestor of about half of the languages spoken in Europe and
many more. The study of the history of language tries to answer
(at least partly) one of the fundamental questions of mankind:
where do we come from?

Today, linguistics focuses mainly on the mental state of speakers and

how they come to learn language. For a large part, this investigation
dismisses input that comes from an area of linguistics that was once
dominant: historical linguistics. Historical linguistics is the study of
the history and development of languages. The roots of historical
linguistics go as far back as the late 17th century when it was observed
that English, German, Dutch, as well as other languages, shared a lot
of common features, and it was quickly observed that one could try to
reconstruct a kind of common ancestor language that existed a long
time ago, called Germanic, from which these languages originated.
To get an idea of the kind of evidence used for this, let’s look at a few
words in these languages:

English Dutch German

bring brengen [brEN@n] bringen [bKıN@n]
sleep slapen [slap@n] schlafen [Slaf@n]
(22.1)
ship schip [sxıp] Schiff [Sıf]
sister zuster [zystEr] Schwester [SwEste@]
good goed [xud] gut [gut]

This list can be made a lot longer, and after doing so, it turns out
that the correspondences between languages are largely systematic.
It can be observed, for example, that word-initial [p] in Dutch and
English corresponds to German [pf], that initial [s] becomes [S] before
[t] and [p], and so on. This has lead to two things: the postulation of
a language (of which we have no record!), called Germanic, a set of
words together with morphology and syntax for this language, and a
set of rules which show how the language developed into its daughter
languages. In the present case, the fact that Dutch [p] corresponds to
266 an open introduction to linguistics

In the context of historical linguistics,

the star ∗ indicates that a word or
German [pf] is explained by the fact that Germanic (not to be confused
sound is reconstructed, not that the
with German) had a sound ∗ p. This sound developed word initially sound is illegitimate
into [p] in Dutch and into [pf] in German. This is called a sound law.
We may write it in the same way as we did in phonology:
∗
p→p/# Dutch (22.2)
∗
p → pf / # German (22.3)

Often, one simply writes ∗ p > pf for the sound change. The similarity
to phonological rules is not accidental; there, they were taken to mean
a sequential process, a development from a sound to another in an
environment. Here a similar interpretation is implied, only that the
time span in which this is supposed to have taken place is much longer,
approximately two thousand years!
The list of Germanic languages is long. Apart from the ones just
listed, Danish, Swedish, Norwegian, Faroese, Icelandic, Frisian and
Gothic also belong to that group. Gothic is interesting because it is
a language of which we only have written records; we do not know
exactly how it sounded.
As with the Germanic languages, similarities can be observed be-
tween French, Spanish, Italian, Romanian and Portuguese, too. In
fact, all these languages come from a language which we know very
well: Latin. The development of Latin into these languages is well
documented, compared to other language groups. This is important,
since it allows to assert the existence of a parent language and changes
with certainty, whereas in most cases the parent language has to be
constructed from the daughter languages.
This is so, for example, with Celtic, from which Gaelic, Irish,
Welsh, and Breton descended. Cornish and Manx are also Celtic, but
became extinct in the 19th and 20th century, respectively. Another This is not the end of the story for
Celtic language, Gaulish, was spoken in the whole of France, but it Manx and Cornish, however. In the
late 20th century, both langauges have
was completely superseded by Latin. The only records we have of been revived by language enthusiasts.
Gaulish are in names of people and places. For example, we know of Speakers of the ‘resurrected’ forms of
both languages number in the hun-
the Gaulish king Vercingetorix through the writings of Caesar. The dreds, and nowadays, Manx and Cor-
name is a Celtic name for sure. nish even have some native speakers
Throughout the 19th century, it became apparent that there were again!

similarities to be found not only between the languages just discussed,

but also between Germanic, Latin, Greek, Celtic, Sanskrit, Old Per-
sian, Armenian, Slavic, Lithuanian, and Tocharian. It was proposed
that all these languages (and their daughters, of course) descended
from a single language called Indo-European. When researchers
made excavations in Anatolia in the 1920s and found remains of the
Hittite language, they soon realised that Hittite, too, belongs to this
group of languages. In the last 200 years a lot of effort has been
made to reconstruct the sound structure, morphology and syntax of
Indo-European, and to find out about the culture, belief, and ancient
homeland of the Indo-Europeans.
The proposed time frame for the Indo-European language family
is roughly this: the Indo-European language is believed to have been
spoken up to the 3rd millennium BC. Some equate the Indo-Europeans
 language families and history of languages 267

with people that lived in the region of the Balkan and the Ukraine in
the 5th millennium BC, some believe they originate further north in
Russia, while others equate them with the Kurgan culture, 4400–2900
BC, in the south of Russia (near the Caspian sea). From there, they
are believed to have spread into the Indian subcontinent, Persia and
large parts of Europe. The first to arrive in central Europe and Britain
were the Celts who established a large empire only to be topped by
the Romans and later by the Germans.

What Indo-European Looked Like

The plosive consonants of Indo-European are believed to be these:

unaspirated aspirated
voiceless voiced voiceless voiced
velar ∗k ∗g ∗ kh ∗ gh
∗ ∗ ∗ k̂h ∗ ĝh
(22.4)
palatal k̂ ĝ
apico-dental ∗ t ∗d ∗ th ∗ dh

labial ∗p ∗b ∗ ph ∗ bh

Other people have proposed, instead of the palatals, a series of labiove-

lars (∗ ku , ∗ gu , ∗ ku h, ∗ gu h). This difference is irrelevant from an ab-
stract point of view, but it makes certain sound changes more likely.
Another set is ∗ y, ∗ w, ∗ r, ∗ l, ∗ m and ∗ n, which could be either
syllabic or non-syllabic. Syllabic ∗ y was roughly [i], non-syllabic ∗ y
was [j]. Likewise, syllabic ∗ w was [u], non-syllabic ∗ w was [w]. The
non-syllabic ∗ r was perhaps trilled and non-syllabic ∗ m and ∗ n were
like [m] and [n]. Syllabic ∗ l was realized as l, and similarly were m and
˚ ˚
n. The vowels were ∗ i (= syllabic ∗ y), ∗ e, ∗ a, ∗ o and ∗ u (= syllabic
˚
∗ w).

Here are two examples of roots and their correspondences in various

Indo-European languages:
∗ wlku os ‘wolf’. In Latin we find /lupus/, in Greek /lykos/, in
˚ ∗
Sanskrit /vr .kah ./, in Lithuanian /vil̃kas/, in Germanic /wulfaz/,
from which English /wolf/, and German /Wolf/ [wOlf].
∗ dekm ‘ten’. In Sanskrit /daśa/, Latin /decem/, pronounced [dEkEm]
˚
or even [dEk˜E], with nasalized vowel, in Greek /deka/, Germanic
∗ /tehun/, from which Gothic /taihun/, German zehn [tse:n], and

English /ten/.
Indo-European is believed to have had not only singular and plural
verb conjugations, but also a dual (for two people). The dual was lost
in Latin, but retained in Greek and Sanskrit. Let’s look at an example,
268 an open introduction to linguistics

the root of which is ∗ bher ‘to carry’:

Sanskrit Greek Latin
Sg 1 bhar-ā-mi pher-ō fer-ō
2 bhar-ā-si pher-eis fer-s
3 bhar-ā-ti pher-ei fer-t
Du 1 bhar-ā-vah. − −
(22.5)
2 bhar-ā-thah. pher-e-ton −
3 bhar-ā-tah. pher-e-ton −
Pl 1 bhar-ā-mah. pher-o-mes fer-i-mus
2 bhar-ā-tha pher-e-te fer-tis
3 bhar-ā-nti pher-o-nti fer-u-nt
To be exact, although Latin /fero/ has the same meaning, it is con-
sidered to belong to another inflectional paradigm, because it does not
have the vowel ‘i’. Note also that in Attic and Doric Greek, the 1st plu-
ral was /pheromen/. Thus, there has been a variation in the endings,
which we do not see here.
You might notice that the verb ∗ bher is also found in English, in the
verbs /bear/ and /bring/ (often, root vowels become weak, giving rise
in the case of ∗ e to a so-called ‘zero’-grade ∗ bhr, from which /bring/
˚
is derived).

How Do We Know?
The reconstruction of a language which is no longer spoken (‘extinct’)
is a difficult task. There are two main methods: comparison between
languages, and internal reconstruction. The latter is used in absence of
comparative evidence. In that case, one observes certain irregularities
in the language and proposes an explanation by referring to possible
earlier developments in the language.
It has been observed, for example, that irregular inflection is older
than regular inflection. For example, in English there are plurals in
/en/ (/oxen/, /vixen/) and plurals formed by vowel change (/women/,
/mice/). These are explained by internal reconstruction as reflecting
an earlier state of the language where the plural was formed by addition
of /en/ and vowel change. The plural /s/ was a later development.
Likewise, this method predicts that the comparative in English was
once formed using /er/ and /est/, but at some point was (partly) re-
placed by forms involving /more/ and /most/. In both cases, German
reflects the earlier stage of English.
Notice, however, that the proposed changes are based on present-
day English. How, then, can we make certain that we are right?
First and foremost, there are written documents. We have transla-
tions of the bible into numerous languages, including medieval Geor-
gian (a Caucasian language), Old English (King Alfred’s bible), and
Gothic, for example. Latin and Greek literature has been preserved
to this day thanks to the effort of thousands of monks in monaster-
ies (copying was a very honourable and time consuming task in those
days). Other languages that have been preserved are, among others,
Avestan, Sanskrit and Hittite (written mostly in cuneiform). Most
 language families and history of languages 269

other languages were written down only from the early middle ages
onwards, mostly in the form of biblical texts and legal documents.
Now, this provides us with the written languages, but it does not
tell us how they were spoken. In the case of Hittite, the problem is
very obvious: the writing system was totally different from ours and it
had to be discovered how it was to be read. For Sanskrit we know from
the writings of the linguists of those days (among which Pan.ini (500
BC) is probably one of the latest) how the language was spoken. This
is because they gave us explicit descriptions of how the sounds were
produced. For Latin and Greek matters are less easy. The Greeks, for
example, did not realize the distinction between voiced and voiceless
consonants (they knew they were different but could not say what the
difference was). In the middle ages all learned people spoke Latin, but
the Latin they spoke was very different from classical Latin, both in
vocabulary and pronunciation. By that time, people did not know how
things were pronounced in the classical times. So how come we know?
One answer is: through mistakes people made when writing Latin.
Inscriptions in Pompeii and other sites provide telling examples. One
specific example is that /m/ after vowels was either completely lost
or it just nasalized the preceding vowel. This can be inferred from
inscriptions where one finds /ponte/ in place of what should have
been /pontem/. People who made the mistakes simply could not hear
the difference. (Which is not to say that there is none; only that it
was too small to be noticeable.) Also, in verse dating from that time
the endings in vowel plus /m/ counted as non-existent for the metre.
(This in turn we know for sure because we know what the metre was.)
All this is strong evidence that in classical times the final /m/ was not
pronounced.
Another method is looking at borrowings into other languages. The
name /Caesar/ (and the title that derived from it) was borrowed into
many languages, and appears in the form of /Kaiser/ in German,
taken from Gothic /kaisar/, and /césar/ [seza:r] in French. So, at
the time the Goths borrowed the word the letter /c/ was pronounced
[k]. And since French descends from Latin we must conclude that the
[k]-pronunciation is older.
A third source is the alphabet itself. The Romans did not distin-
guish /v/ from /u/. They wrote /v/ regardless. This shows that the
two were not felt to be distinct. It is unlikely that /v/ was pronounced
[v] (as in English /vase/). Rather, it was originally a bilabial approxi-
mant (the non-syllabic ∗ w mentioned above), and became a labiodental
fricative only later.
A problem is caused by the fact that languages consist of tens
of thousands of words, but a lot of them are not indigenous words.
Many words that we use in the scientific context, for example, come
from Latin and/or Greek. Moreover, they may have been borrowed
from these languages at any moment in time. The linguistic terms
(/phoneme/, /lexeme/) is a telling example. These words have been
artificially created from Greek source words, so learned words like these
have to be discarded.
270 an open introduction to linguistics

Another problem is caused by the fact that words change their

meaning in addition to their form. An example is German /schlimm/
‘bad’, which originally meant ‘inclined’. Or the word /vergammeln/ ‘to
rot’, which originates from Scandinavian /gamall/ ‘old’. Another ex-
ample is English /but/, which derives from a spatial preposition that
is still found in Dutch /buiten/ ‘outside’. From there it took more
abstract meanings, until the spatial meaning was completely lost. If
meanings can change so dramatically, we have to be very cautious.
If meanings would be constant, we could easily track words back in
time; we would just have to look at words in related languages that
had the same meaning. But if the meaning can change too, what are
we to look out for? Linguists have put a lot of effort into determining
in which ways the meanings of words can go and which meanings are
more stable than others.

Two Examples Among Many

As an example of the beauty and danger of historical linguistics, let’s

look at the history of two words.
The first example shows that once we know the sound changes lan-
guages have gone through, resemblances become very striking. The
English word /choose/ has relatives in Dutch and German. If we look
at the verbs that these languages normally use, we might get disap-
pointed: the Dutch word is /kiezen/, and the German is /wählen/.
The relation between Dutch and the English is quite easy to see. First,
notice that the Dutch verb has a perfect participle /gekozen/ ‘cho-
sen’, which has the /o/ in place of the /ie/. The change from [i] to
[o], called Ablaut, is widely attested in the Indo-European languages.
Now, [k] often becomes [tS] before either [e] or [i] (like Latin [ke] be-
came [tSe] in Italian, often with loss of /e/ in pronunciation). This
change occurred in English only, not in Dutch. However, we still have
to see why Ablaut occurred in English. The Old English word in fact
was /ceōsan/. (When no star appears, that means we have written
records.) The pronunciation of /c/ changed and incorporated the /e/
and the infinitive ending got lost like with other verbs.
The German case, on the other hand, seems hopeless. In fact,
/wählen/ does not come from a related root. However, in German we
do find a verb /kiesen/ in similar meaning, although it is no longer
in use. Strangely, the perfect passive participle (PPP) of the verb
/auserkiesen/ (two prefixes added, /aus/ and /er/), is still in use:
/auserkoren/ ‘chosen’, even though the verb itself is no longer in use.
Notice that the ablaut is the same as in Dutch, which incidentally also
uses the circumfix /ge- -en/ as does German. Finally, in German
the PPP has /r/ in place of /s/ (which would be pronounced [z]).
The change from /s/ to /r/ in between vowels (called ‘rhotacism’) is
a popular change. (There are more examples. The English /was/ is
cognate to Dutch /was/, but German has /war/.)
Now, the root from which all this derives is believed to be Ger-
manic ∗ /keusa/ ‘to try out, choose’. Once we have progressed this far,
 language families and history of languages 271

other words come into sight: Latin /gustare/ ‘to taste’ (from which,
via French, English got /disgusting/), Greek /geuomai/ ‘I taste’,
Sanskrit /juśati/ ‘he likes’, Old Irish /do-goa/ ‘to choose’—they all
look like cognates to this one. Indeed, from all these words one has
reconstructed the Indo-European root ∗ /geus/ ‘to choose’. The Latin
word presents the zero-grade ∗ /gus/, and in West Germanic, we find Roots typically had /e/. The zero
grade is the version of the root that
/kuzi/ ‘vote, choice’.
has no /e/. Similarly, ∗ /dik/ is the
But do not think that historical linguistics is always this straight- zero grade of ∗ /deik/ ‘to show’.
forward. Sometimes it can be tricky and deceiving. For example,
French /choisir/ ‘to choose’ does not come from Latin /gustare/
(although it does come from the same Indo-European root)—there is
no way to explain this with known sound laws. Known sound laws
predict /gustare/ to develop into /goûter/, and this is what we find.
Instead /choisir/ was taken from—Gothic! Indeed, French has taken
loanwords from Germanic, being occupied/inhabited in large parts by
Germanic tribes. The name /France/ itself derives from the name of
a tribe: /Frankon/.
With respect to Dutch and German, the reconstruction is actually
easy, since the languages split around the 17th century. Certain di-
alects of North Germany still have [p] where others have [pf]. Often,
it happens that a certain word is not attested in all languages of a
certain group. For example, English /horse/ corresponds to Dutch
/paard/ and German /Pferd/, with same meaning. The sound laws
allow to assert that /paard/ and /Pferd/ descend from the same word,
but for English /horse/ this is highly implausible. There is a German
word, /Ross/ ‘horse’, and a Dutch /ros/, which sound quite simi-
lar, but they are far less frequent. What we have to explain first is
why the words (/horse/, /ros/, /Ross/) are different. Now, we are
lucky to have sources confirming that the word was sometimes spelt
as /hros/ and sometimes as /ros/ in Old German. In Icelandic, it
is still /hross/. The loss of /h/ before /r/ is attested in other cases,
like the word /ring/. But the change happened in English, too, and
the only reason why the /h/ was preserved in /horse/ is that the /r/
changed places with /o/.
Second, where did Dutch and German get their main words from?
Both words come from the same source, but it is not Germanic, it is
Medieval Latin /paraveredus/ ‘post-horse’ (200 - 900 AD), which in
turn is /para/ + /veredus/. /para/ comes from Greek (!) /para/
‘aside’, and /veredus/ is Celtic. It is composed from /ve/ ‘under’
and /reda/ ‘coach’. In Welsh, there is a word /gorwydd/ ‘horse’. So,
/paraveredus/ had a more special meaning: it was the horse that
was running on the side. (The coach had two horses, one on the
right and one on the left; the one on the left was saddled, and the
one on the right was the ‘side horse’.) English has a word /palfrey/,
which denotes a kind of riding horse. The Indo-European root that has
been reconstructed is ∗ ekw os, compare Latin /equus/, Greek /hippos/
and Sanskrit /aśvah ./. Had we not known that the Latin word is
/equus/, we would have had to guess from French /cheval/ and Span-
ish /caballo/.
272 an open introduction to linguistics

Thus, roots do not survive everywhere. Words get borrowed, changed,

returned and so on. There are not too many roots that are attested
in all languages, and those that are are very elementary in meaning:
mostly kinship terms, personal pronouns and numbers.

Other Language Families

In addition to Indo-European, there is another language family in Eu-
rope: the Uralic language family. The languages that belong to that
family are Finnish, Estonian, Sami, Hungarian and a number of lesser
known languages spoken in the west and north of Russia. The affilia-
tion of Hungarian is nowadays not disputed, but in the 19th century
it was believed to be related to Turkish. Unfortunately, the written
records of these languages are at most 1000 years old, and the sim-
ilarities are not always that great. Finnish, Estonian and Sami can
easily be seen to be related, but Hungarian is very much different.
This may have to do with the fact that it was under heavy influence
from Slavic neighbours, Turkish (Hungary was under Ottoman rule
for a long time) and Germanic (not the least through the Habsburg
monarchy and the Austro-Hungarian empire). Another cause for the
large difference is that Hungarian originates from the Ugric branch of
the Uralic language family, while Finnish and Estonian are part of the
(Baltic-)Finnic branch. When compared to the other Ugric languages,
Khanty and Mansi, the relationship become more obvious.
In addition to these two language families, there is one special case
in Europe: Basque. Although many theories have been proposed,
Basque is not commonly accepted to belong to any one language family.
As such, it is called a language isolate.
Some of the other recognized language families are: Semitic (in-
cluding Hebrew, Ethiopic, Amharic, Aramaic and Arabic), Altaic
(Turkish, Tatar, Tungusic and Mongolian), Dravidian (spoken in the
south of India: Tamil, Telugu, Kannada and Malayalam), Austrone-
sian (Malay, Indonesian, Malagassy, languages spoken on Macronesia,
Micronesia and Polynesia), Eskimo-Aleut (Inuit (= Eskimo), indige-
nous languages spoken in Canada, Greenland, Western Siberia and the
Aleut Islands). Sometimes languages are grouped together because it
is believed that they are related, but relationships are actually hard
to come by. This is the case with the Caucasian languages (Geor-
gian and many others), although three main groups of related lan-
guages have now been established within the Caucasian languages. It
is believed that the people who live there have not moved for several
millennia. This has given rise to a dazzling number of quite distinct
languages in a relatively small area in and around the southern part
of the Caucasian mountains.

Probing Deeper in Time

People have tried to probe deeper into the history of mankind. One
way to do this is to classify people along genetic relations, another
has been to establish larger groupings among the languages. Although
 language families and history of languages 273

genetic relationships need not coincide with language relationships, the

two are to a large degree correlated, since both languages and genes
are generally passed on by parents to their children. An exciting example of the combina-
Two hypotheses are being investigated, both starting from Indo- tion of these two methods is the (rela-
tively recent) proposal of the Dené-
European. Joseph Greenberg proposed a macro-family called Eurasi- Yeniseian language family. This
atic, which includes Indo-European, Uralic, Altaic, Eskimo-Aleut, Ko- proposal links languages spoken by
native peoples in Central Siberia,
rean, Japanese and Ainu (spoken in the north of Japan) and Chuk- Alaska, North-western Canada, and
chi-Kamchatkan. Alternatively, it has been suggested by Russian the Southern United States based on
linguists (Illyč-Svityč and Dolgopolsky) that there is an even larger both linguistic and genetic evidence.

macro-family called Nostratic, which originally was believed to in-

clude all Eurasiatic families, Afro-Asiatic (spanning north Africa in-
cluding Semitic), Dravidian and Kartvelian (part of Caucasian). Green-
berg did not reject the existence of Nostratic but wanted to put it even
farther back in history, as a language which developed among other
into Eurasiatic. Moreover, the position of modern Nostraticists has
come closer to that of Greenberg’s views. At a larger scale there are
believed to be twelve such macro-families in this world, all ultimately
coming from a single language ... Examples of words that are believed
to have been passed to us by the single ancestor language are /kuan/
‘dog’; /mano/ ‘man’; /mena/ ‘think (about)’; /mi(n)/ ‘what’; /Paq’wa/
‘water’; /tik/ ‘finger, one’; and /pal/ ‘two’ (see 1 ]Merritt Ruhlen and 1[

John D. Bengtson. Global etymologies. In Merritt Ruhlen, editor, On the Origin of

Languages, pages 277–336. Stanford University Press, Stanford, 1994,
for more)
This is all highly speculative, but the evidence for Eurasiatic (or even
Nostratic) is not as poor as one might think. Common elements in
Eurasiatic are for example: first person in m, second person in t/n.
Greenberg has collected a list of some 250 words or elements that are
believed to be of Eurasiatic origin.
Bibliography

Merritt Ruhlen and John D. Bengtson. Global etymologies. In Merritt

Ruhlen, editor, On the Origin of Languages, pages 277–336. Stanford
University Press, Stanford, 1994.
Index

Ablaut, 270 Cherokee, 20

adjective, 115 circumfix, 94
adposition, 126 class, 205
adverb, 115 classification system, 39
affix, 82, 94 classifier, 222
affricates, 13, 40 clause
agreement, 134, 135 subordinate, 146
subject-verb, 137 coarticulation, 32
agreement feature, 141 coda, 59
allomorph, 93 comparability, 149
ambisyllabicity, 60 comparative, 97
analogy, 255 compensatory lengthening, 174
anaphor, 165 complement, 125
Anglo-Saxon genitive, 137 complementizer, 115
answer ellipsis test, 108 compounding, 84
approximants, 14 consonants, 13
archiphoneme, 55 constituency tests, 107, 108
argument, 116 constituent, 105, 109
oblique, 121 context, 108
attribute, 36 Continuity of Constituents, 106
autosegmental phonology, 60 contrastive in anaologous envi-
AVS, 36 ronments, 31
inconsistent, 36 Conversational Implicatures, 229
Aymara, 20 cooperative principle, the, 234
coordination, 105
behaviorism, 251 coordination test, 105, 107
beneficiary, 212 count noun, 221
binarism, 42 cross-situational word learning,
broad transcription, 31 245

c-command, 149 dactylus, 65

C.A.E., 31 Danish, 20
candidate, 76 dative, 136, 137
optimal, 76 derivation, 84, 112
cardinal vowels, 21 determiner, 115
case, 135 definite, 134
category, 110, 112 indefinite, 135
lexical, 115 diphthong, 21
major, 115 diphthongs, 40
 INDEX 277

Dutch, 16 Kham, 14

ellipsis test, 108 language, 113

epiglottis, 12 string, 109
event, 211 language isolate, 272
atelic, 224 larynx, 12
telic, 224 lax, 38
experiencer, 212 level, 120
extrasyllabicity, 60 Levinson, 238
license, 2
fast-mapping, 245 local admissibility, 126
feature, 35, 36 logical problem of acquisition,
grammatical, 136 the, 249
Finnish, 20 loudness, 65
flounting, 234
foot, 65 manner, 13
frame, 82 manner, Maxim of, 236
French, 16 mass noun, 222
Maximise Onset, 63
gavagai, 245 Maxims, 234
gender, 135, 205 metalanguage, 181
feminine, 135 metathesis, 75
masculine, 135 metre
neuter, 135 iambic, 65
genitive, 137 trochaic, 65
German, 20 metrical stress, 65
glottis, 12 minimal pairs, 27
grammar, 111 morph, 93
Grice, Paul, 234 morpheme, 93
morphological change, 94
Haida, 20 motherese hypothesis, the, 249
haplology, 71
Hausa, 15 Narrow transcription, 31
head, 125, 146 nasals, 14
movement, 146 natural class, 39
height, 19 necessity, 210
Horn, 238 node, 149
nominative, 136
Iaai, 14 Non-Crossing, 106
Icelandic, 20 Norwegian, 20
immediate constituent, 106 Not-Too-Similar Principle, 73
inflection, 84 noun, 115
innateness, 255 nucleus, 59
integrated whole, 221 number, 133
intension, 210
IPA, 16 object, 116
it-cleft test, 108 direct, 116
indirect, 137
Japanese, 16 object language, 180
278 an open introduction to linguistics

obstruent, 70 Recover the Morpheme, 75

occurrence, 104 Recover Voicing, 76
constituent, 110 recursion, 259
onset, 59 relative clause, 217
legitimate, 64 (non-)restrictive, 218
Optimality Theory, 74 Relevance, Maxim of, 235
OT, 74 replacement derivation, 112
overlap, 104 representation
surface, 75
part-of-speech, 91 syntactic, 133
passivazation test, 108 underlying, 75
person, 136 rhyme, 59
pharynx, 12 root, 56, 84
phoneme, 17, 27 rounding, 19
phonemic contrast, 29 rule
phonemic inventory, 28 context-free, 111
phonetics, 11 domain, 60
phonological representation Russian, 20
deep, 54
surface, 54 Sadock, 238
phrase, 115, 120 Sandhi, 49
Pinker, Steven, 258 Sapir-Whorf-Thesis, 205
place, 19 schwa, 20
plosives, 13 selectional feature, 141
pluperfect, 209 similar consonant, 71
position, 13 sonoricity hierarchy, 61
positive, 97 sonoricity peak, 61
postposition, 126 sonority, 18
Potts, Chris, 238 sound law, 266
precedence, 104, 106 specifier, 125
prefix, 94 spectogram, 22
preposition, 115 SR, 75
Presupposition, 229 start symbol, 111
pro-form, 126 state, 211
process, 211 stress, 65
projection, 120 primary, 65
intermediate, 120 secondary, 65
projection, presuppositional, 232 string, 109
pronoun, 136 subject, 116
propose-but-verify, 246 substitution test, 105, 107
suffix, 94
Quality, Maxim of, 234 superlative, 97
Quantity, Maxim of, 235 superset problem, the, 259
Swedish, 20
ranking, 76 syllabification, 62
realization rule, 39 syllable
reciprocal, 165 antepenultimate, 65
Recover Adjacency, 76 closed, 59
Recover Obstruency, 76 heavy, 66
 INDEX 279

open, 59 Xhosa, 15
penultimate, 65
weight, 66 Yiddish, 20
Syllable Equivalence, 76
Syllable Structure I, 59 zero grade, 271
Syllable Structure II, 61 Zulu, 15
syntactic representation
deep, 144
surface, 144

tense, 115
tenseness, 20
tension, 38
theme, 212
tier, 60
topicalisation, 143
trace, 171
trachea, 12
transfix, 95
transformation, 144
tree, 107
root, 107
two word stage, 247

UR, 75

V2, 146
value, 36
value range, 36
verb, 115
distributive, 207
transitive, 116
verb second, 146
vocal cords, 12
Voice Agreement Principle, 70
voiced, 13
voiceless, 13
voicing, 13
VOT, 22, 28
vowel chart, 19
vowels, 18

Wh-Movement, 145, 146

wh-word, 145
word, 115
word class, 81
Words are Constituents, 106
world, 210