(Charles F. Meyer) English Corpus Linguistics An

This page intentionally left blank
English Corpus Linguistics

An Introduction
English Corpus Linguistics is a step-by-step guide to creating and analyzing
linguistic corpora. It begins with a discussion of the role that corpus linguistics
plays in linguistic theory, demonstrating that corpora have proven to be very
useful resources for linguists who believe that their theories and descriptions
of English should be based on real, rather than contrived, data. Charles F.
Meyer goes on to describe how to plan the creation of a corpus, how to collect
and computerize data for inclusion in a corpus, how to annotate the data that
are collected, and how to conduct a corpus analysis of a completed corpus.
The book concludes with an overview of the future challenges that corpus
linguists face to make both the creation and analysis of corpora much easier
undertakings than they currently are. Clearly organized and accessibly written,
this book will appeal to students of linguistics and English language.
cnniis i. mivin is Professor of Applied Linguistics at the University of
Massachusetts, Boston. He has published numerous books and articles on lin-
guistics, including Apposition in Contemporary English (Cambridge, 1992),
and The Verb in Contemporary English, co-edited with Bas Aarts (Cambridge,
1995). He is currently editor of the Journal of English Linguistics and former
co-ordinator of the International Corpus of English (ICE).
s1toi is i ioii sn iotoi
The aim of this series is to provide a framework for original work on the English
language. All are based securely on empirical research, and represent theoretical and
descriptive contributions to our knowledge of national varieties of English, both
written and spoken. The series will cover a broad range of topics in English grammar,
vocabulary, discourse, and pragmatics, and is aimed at an international readership.
Already published
Christian Mair
Innitival complement clauses in English: a study of syntax in discourse
Charles F. Meyer
Apposition in contemporary English
Jan Firbas
Functional sentence perpective in written and spoken communication
Izchak M. Schlesinger
Cognitive space and linguistic case
Katie Wales
Personal pronouns in present-day English
Laura Wright
The development of standard English 13001800: theories, descriptions, conicts
s1toi is i ioii sn iotoi
Editorial Board
Bas Aarts, John Algeo, Susan Fitzmaurice,
Richard Hogg, Merja Kyt o, Charles Meyer
An Introduction
An Introduction
CHARLES F. MEYER
University of Massachusetts at Boston
iuniisuio n\ rui iiiss s\xoicari oi rui uxiviisir\ oi caxniioci
The Pitt Building, Trumpington Street, Cambridge, United Kingdom
caxniioci uxiviisir\ iiiss
The Edinburgh Building, Cambridge CB2 2RU, UK
40 West 20th Street, New York, NY 10011-4211, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
Ruiz de Alarcn 13, 28014 Madrid, Spain
Dock House, The Waterfront, Cape Town 8001, South Africa
http://www.cambridge.org
First published in printed format
ISBN 0-521-80879-0 hardback
ISBN 0-521-00490-X paperback
ISBN 0-511-02945-4 eBook
Charles F. Meyer 2004
2002
(Adobe Reader)
To Libby and Freddie

Contents
Preface page xi
1 Corpus analysis and linguistic theory 1
2 Planning the construction of a corpus 30
3 Collecting and computerizing data 55
4 Annotating a corpus 81
5 Analyzing a corpus 100
6 Future prospects in corpus linguistics 138
Appendix 1 Corpus resources 142
Appendix 2 Concordancing programs 151
References 153
Index 162
ix
Preface
When someone is referred to as a corpus linguist, it is tempting
to think of this individual as studying language within a particular linguistic
paradigm, corpus linguistics, on par with other paradigms within linguistics,
such as sociolinguistics or psycholinguistics. However, if the types of linguistic
analyses that corpus linguists conduct are examined, it becomes quite evident
that corpus linguistics is more a way of doing linguistics, a methodologi-
cal basis for pursuing linguistic research (Leech 1992: 105), than a separate
paradigm within linguistics.
To understand why corpus linguistics is a methodology, it is rst of all neces-
sary to examine the main object of inquiry for the corpus linguist: the linguistic
corpus. Most corpus linguists conduct their analyses giving little thought as to
what a corpus actually is. But dening a corpus is a more interesting question
than one would think. A recent posting on the Corpora list inquired about the
availability of an online corpus of proverbs (Maniez 2000).
1
This message led
to an extensive discussion of how a corpus should be dened. Could something
as specic as a computerized collection of proverbs be considered a corpus, or
would the body of texts from which the proverbs were taken be a corpus and
the proverbs themselves the result of a corpus analysis of these texts?
The answer to this question depends crucially on how broadly one wishes to
dene a corpus. The Expert Advisory Group on Language Engineering Stan-
dards (EAGLES) denes a corpus quite generally, saying that it can poten-
tially contain any text type, including not only prose, newspapers, as well as
poetry, drama, etc., but also word lists, dictionaries, etc. (Corpus Encoding
Standard: http://www.cs.vassar.edu/CES/CES1-0.html). Accordingtothis de-
nition, a collection of proverbs would indeed constitute a corpus. However, most
linguists doing corpus analyses would probably prefer a more restricted de-
nition of corpus, one that acknowledged the broad range of interests among
individuals who use corpora in their research but that dened a corpus as some-
thing more than a collection of almost anything. For the purposes of this book,
then, a corpus will be considered a collection of texts or parts of texts upon
which some general linguistic analysis can be conducted. In other words, one
does not create a corpus of proverbs to study proverbs, or a corpus of relative
1
Appendix 1 contains further information on the various corpus resources discussed in this book:
Internet discussion lists such as Corpora as well as all the corpora described in this and
subsequent chapters.
xi
xii Preface
clauses to study relative clauses. Instead, one creates a corpus which others can
use to study proverbs or relative clauses.
If a corpus is dened as any collection of texts (or partial texts) used for
purposes of general linguistic analysis, then corpus linguistics has been with
us for some time. Otto Jespersens multi-volume A Modern English Grammar
on Historical Principles (190949) would not have been possible had it not
been based on a corpus representing the canon of English literature: thousands
of examples drawn from the works of authors such as Chaucer, Shakespeare,
Swift, and Austin that Jespersen used to illustrate the various linguistic struc-
tures he discusses. In recent times, a corpus has come to be regarded as a body
of text made available in computer-readable form for purposes of linguistic
analysis. The rst computer corpus ever created, the Brown Corpus, qualies
as a corpus because it contains a body of text one million words of edited
written American English made available in an electronic format (the ICAME
CD-ROM, 2nd edn.) that can be run on multiple computer platforms (Macin-
tosh, DOS/Windows, and Unix-based computers).
Modern-day corpora are of various types. The Brown Corpus is a balanced
corpus because it is divided into 2,000-word samples representing different
types (or genres) of written English, including press reportage, editorials, gov-
ernment documents, technical writing, and ction. The purpose of designing
this corpus in this manner is to permit both the systematic study of individ-
ual genres of written English and a comparison of the genres. In contrast,
the Penn Treebank is not a balanced corpus: instead of containing a range of
different genres of English, it consists of a heterogeneous collection of texts
(totalling approximately 4.9 million words) that includes a large selection of
Dow Jones newswire stories, the entire Brown Corpus, the ction of authors
such as Mark Twain, and a collection of radio transcripts (Marcus, Santorini, and
Marcinkiewicz 1993). In creating this corpus, there was no attempt to balance
the genres but simply to make available in computer-readable form a sizable
body of text for tagging and parsing.
The Brown Corpus and Penn Treebank differ so much in composition be-
cause they were created for very different uses. Balanced corpora like Brown
are of most value to individuals whose interests are primarily linguistic and who
want to use a corpus for purposes of linguistic description and analysis. For in-
stance, Collins (1991a) is a corpus study of modal verbs expressing necessity
and obligation (e.g. must meaning necessity in a sentence such as You must
do the work). In one part of this study, Collins (1991a) compared the relative
frequency of these modals in four genres of Australian English: press reportage,
conversation, learned prose, and parliamentary debates. Collins (1991a: 1523)
selected these genres because past research has shown them to be linguistically
quite different and therefore quite suitable for testing whether modals of neces-
sity and obligation are better suited to some contexts than others. Not only did
Collins (1991a) nd this to be the case, but he was able to explain the varying
frequency of the modals in the four genres he studied. The fewest instances of
Preface xiii
these modals were in the press reportage genre, a genre that is factual, [and]
non-speculative and that would therefore lack the communicative context that
would motivate the use of modals such as must or ought. In contrast, the conver-
sations that Collins (1991a) analyzed contained numerous modals of this type,
since when individuals converse, they are constantly expressing necessity and
obligation in their conversations with one another. To carry out studies such as
this, the corpus linguist needs a balanced and carefully created corpus to ensure
that comparisons across differing genres of English are valid.
In designing a corpus such as the Penn Treebank, however, size was a more
important consideration than balance. This corpus was created so that linguists
with more computationally based interests could conduct research in natural
language processing (NLP), an area of study that involves the computational
analysis of corpora often (though not exclusively) for purposes of modeling
human behavior and cognition. Researchers in this area have done consider-
able work in developing taggers and parsers: programs that can take text and
automatically determine the word class of each word in the text (noun, verb,
adjective, etc.) and the syntactic structure of the text (phrase structures, clause
types, sentence types, etc.). For these linguists, a large corpus (rather than a bal-
anced grouping of genres) is necessary to provide sufcient data for training
the tagger or parser to improve its accuracy.
Even though descriptive/theoretical linguists and computational linguists use
corpora for very different purposes, they share a common belief: that it is
important to base ones analysis of language on real data actual instances
of speech or writing rather than on data that are contrived or made-up.
In this sense, then, corpus linguistics is not a separate paradigm of linguistics
but rather a methodology. Collins (1991a) could very easily have based his
discussion of modals on examples he constructed himself, a common practice
in linguistics that grew out of the Chomskyan revolution of the 1950s and
1960s with its emphasis on introspection. However, Collins (1991a) felt that
his analysis would be more complete and accurate if it were based on a body of
real data. Likewise, the computational linguist attempting to develop a tagger
or parser could tag or parse a series of articially constructed sentences. But
anyone attempting this kind of enterprise knows that a tagger or parser needs
a huge collection of data to analyze if it is expected to achieve any kind of
accuracy.
Further evidence that corpus linguistics is a methodology can be found by
surveying the various types of corpora available and the types of linguistic
analyses conducted on them. The CHILDES Corpus contains transcriptions
of children speaking in various communicative situations and has been stud-
ied extensively by psycholinguists interested in child language acquisition
(MacWhinney 2000). The Helsinki Corpus contains various types of written
texts from earlier periods of English and has been used by historical linguists to
study the evolution of English (Rissanen 1992). The COLT Corpus (the Bergen
Corpus of London Teenage English) contains the speech of London teenagers
xiv Preface
and has been analyzed by sociolinguists interested in studying the language of
a particular age group (Stenstr om and Andersen 1996). In short, linguists of
various persuasions use corpora in their research, and are united in their belief
that ones linguistic analysis will benet from the analysis of real language.
If corpus linguistics is viewed as a methodology as a way of doing linguistic
analysis it becomes increasingly important that corpora are carefully created
so that those analyzing them can be sure that the results of their analyses will
be valid. If a corpus is haphazardly created, with little thought put into its com-
position, then any analysis based on the corpus will be severely compromised.
This book seeks to help corpus linguists understand the process of corpus cre-
ation and analysis by describing what exactly is involved in creating a corpus
and what one needs to do to analyze a corpus once it is created. If corpus lin-
guists understand the methodological assumptions underlying both the creation
and subsequent analysis of a corpus, not only will they be able to create better
corpora but they will be better able to judge whether the corpora they choose
to analyze are valid for the particular linguistic analysis they wish to conduct.
Although much of the discussion is relevant to the creation and analysis of any
kind of corpus in any language, this book pays special attention to these issues
as they apply to English language corpora.
To describe the process of corpus creation and analysis, I have divided this
book into chapters that focus on the relationship between empirical studies of
language and general linguistic theory, the considerations involved in the plan-
ning and creation of a corpus, the kinds of linguistic annotation that can be
added to corpora to facilitate their linguistic analysis, and the process involved
in analyzing a corpus once it has been created. In chapter 1 (Corpus analysis
and linguistic theory), I discuss the role that corpora play in descriptive lin-
guistic analysis and explore a controversy in modern-day linguistics that has
been simmering since the rise of generative grammar in the 1950s: the conict
between the descriptive linguist, who often uses a linguistic corpus to produce
descriptions of linguistic constructions, and the theoretical linguist, who stereo-
typically sits in his or her ofce contriving the sentences upon which some new
theoretical point about language will be based. In this chapter, I argue that the
corpus linguist and generative grammarian are often engaged in complemen-
tary, not contradictory areas of study: while the goals of the corpus linguist
and the generative grammarian are often different, there is an overlap between
the two disciplines and, in many cases, the ndings of the corpus linguist have
much to offer to the theoretical linguist. To illustrate how corpus analysis can
benet linguistic theory and description, I provide a sample analysis of ellip-
tical coordinations that I conducted, and then give an overview of some of the
corpora currently available and the types of linguistic analyses that they permit.
After discussing the role of corpus analysis in linguistics, in chapter 2
(Planning the construction of a corpus), I describe the various factors that
have to be considered before the actual compilation of a corpus is begun. I
discuss such considerations as how the corpus compiler determines the size of
Preface xv
a corpus, the types of texts that should be included in it, the number of samples
for each text type, and the length of each text sample. Once decisions such as
these are made, the actual creation of the corpus can begin, and in chapter 3
(Collecting and computerizing data), I provide advice on how a corpus can
be most efciently created. I discuss how to collect texts for inclusion in a cor-
pus (i.e. make recordings and locate suitable written material), keep accurate
records of the texts collected, obtain permission for written and spoken texts,
and encode the texts in electronic form(i.e. transcribe spoken texts and optically
scan printed material).
After a corpus has been created, its future use and analysis will be greatly
facilitated if certain kinds of information are added in the form of linguis-
tic annotation, the topic of chapter 4 (Annotating a corpus). In this chapter,
I describe three kinds of annotation, or markup, that can be inserted in cor-
pora: structural markup, which provides descriptive information about the
corpus, such as the boundaries of overlapping speech segments in spoken texts
or font changes in written texts; part-of-speech markup, which is inserted by
software that automatically assigns each word in a corpus a part-of-speech des-
ignation (e.g. proper noun, modal verb, preposition, etc.); and grammatical
markup, which is inserted by software that actually parses a corpus, identify-
ing structures larger than the word, such as prepositional phrases or subordinate
clauses.
While chapters 24 focus on the creation of a corpus, chapter 5 (Analyzing a
corpus) describes the process of analyzinga corpus. Inthis chapter, I conduct an
actual corpus analysis to illustrate the various methodological issues that must
be considered in any corpus analysis. I discuss how corpus analysts can best
determine whether the size of the corpus they plan to analyze is suitable for the
analysis being conducted, how analyses can be reliably conducted on different
corpora collected under different circumstances, what software is available for
assisting in the analysis of corpora, and once the analysis is completed, how
the results of the analysis can be subjected to statistical analysis. In the nal
chapter, chapter 6 (Future prospects in corpus linguistics), I discuss where
corpus linguistics is headed as a discipline, given projected developments in
technology and the cost (in money and effort) it takes to create a corpus.
Although the approach I take in this book is relevant to the interests of a range
of different corpus linguists, my primary focus is on how balanced corpora can
be created and analyzed for purposes of descriptive linguistics analysis. For this
reason, some topics are treated in less detail than they would be by corpus lin-
guists withother interests. For instance, while the discussionof taggingandpars-
ing in chapter 4 refers to work in natural language processing done in this area,
I do not treat the topic of parsing in as much detail as a computational linguist
designing parsers would. Likewise, in the discussion of statistics in chapter 5,
there are many more statistical tests than I discuss that could have been cov-
ered. But the audience for whom these and other chapters were intended
linguists interested in creating and analyzing corpora have more limited
xvi Preface
interests in these areas. As a consequence, the areas are discussed in less detail,
and more attention is given to actual linguistic analyses of corpora.
There are many people without whose advice and support this book would
not have been possible. I amvery grateful to Bill Kretzschmar, who encouraged
me to write this book and who has offered many helpful comments on many
sections. Merja Kyt o, series editor for Studies in English Language, read the
entire manuscript and provided feedback that has improved the book immensely.
Two anonymous readers for Cambridge University Press read several draft
chapters and gave me numerous comments that both strengthened the draft
chapters and offered suggestions for completing the additional chapters I needed
to write. Andrew Winnard, senior acquisitions editor at Cambridge University
Press, provided expert guidance in taking the book through the review process.
Others have given me very useful comments on individual chapters: Bas Aarts
(chapter 1), Eric Atwell (chapter 4), Gerald Nelson (chapter 4), Robert Sigley
(chapter 5), and Atro Voutilainen (chapter 4). Finally, I owe an extreme debt
of gratitude both to my wife, Elizabeth Fay, who offered constant support,
love, and encouragement during the years I spent writing this book, and to my
son, Frederick Meyer, who at age three doesnt fully understand what corpus
linguistics is but who has tried to be patient when I retreated to my study to
sneak a few minutes to write this book.
1 Corpus analysis and linguistic theory
When the rst computer corpus, the Brown Corpus, was being cre-
ated in the early 1960s, generative grammar dominated linguistics, and there
was little tolerance for approaches to linguistic study that did not adhere to
what generative grammarians deemed acceptable linguistic practice. As a con-
sequence, even though the creators of the Brown Corpus, W. Nelson Francis
and Henry Ku cera, are now regarded as pioneers and visionaries in the corpus
linguistics community, in the 1960s their efforts to create a machine-readable
corpus of English were not warmly accepted by many members of the linguistic
community. W. Nelson Francis (1992: 28) tells the story of a leading genera-
tive grammarian of the time characterizing the creation of the Brown Corpus
as a useless and foolhardy enterprise because the only legitimate source
of grammatical knowledge about a language was the intuitions of the native
speaker, which could not be obtained from a corpus. Although some linguists
still hold to this belief, linguists of all persuasions are now far more open to
the idea of using linguistic corpora for both descriptive and theoretical studies
of language. Moreover, the division and divisiveness that has characterized the
relationship between the corpus linguist and the generative grammarian rests
on a false assumption: that all corpus linguists are descriptivists, interested only
in counting and categorizing constructions occurring in a corpus, and that all
generative grammarians are theoreticians unconcerned with the data on which
their theories are based. Many corpus linguists are actively engaged in issues
of language theory, and many generative grammarians have shown an increas-
ing concern for the data upon which their theories are based, even though data
collection remains at best a marginal concern in modern generative theory.
To explain why corpus linguistics and generative grammar have had such
an uneasy relationship, and to explore the role of corpus analysis in linguistic
theory, this chapter rst discusses the goals of generative grammar and the three
types of adequacy (observational, descriptive, and explanatory) that Chomsky
claims linguistic descriptions can meet. Investigating these three types of ade-
quacy reveals the source of the conict between the generative grammarian and
the corpus linguist: while the generative grammarian strives for explanatory
adequacy (the highest level of adequacy, according to Chomsky), the corpus
linguist aims for descriptive adequacy (a lower level of adequacy), and it is ar-
guable whether explanatory adequacy is even achievable through corpus anal-
ysis. However, even though generative grammarians and corpus linguists have
1
different goals, it is wrong to assume that the analysis of corpora has nothing to
contribute to linguistic theory: corpora can be invaluable resources for testing
out linguistic hypotheses based on more functionally based theories of gram-
mar, i.e. theories of language more interested in exploring language as a tool
of communication. And the diversity of text types in modern corpora makes
such investigations quite possible, a point illustrated in the middle section of the
chapter, where a functional analysis of coordination ellipsis is presented that
is based on various genres of the Brown Corpus and the International Corpus
of English. Although corpora are ideal for functionally based analyses of lan-
guage, they have other uses as well, and the nal section of the chapter provides
a general survey of the types of linguistic analyses that corpora can help the
linguist conduct and the corpora available to carry out these analyses.
1.1 Linguistic theory and description
Chomsky has stated in a number of sources that there are three levels
of adequacy upon which grammatical descriptions and linguistic theories can
be evaluated: observational adequacy, descriptive adequacy, and explanatory
adequacy.
If a theory or description achieves observational adequacy, it is able to de-
scribe which sentences in a language are grammatically well formed. Such a
description would note that in English while a sentence such as He studied for
the exam is grammatical, a sentence such as *studied for the exam is not. To
achieve descriptive adequacy (a higher level of adequacy), the description or
theory must not only describe whether individual sentences are well formed but
in addition specify the abstract grammatical properties making the sentences
well formed. Applied to the previous sentences, a description at this level would
note that sentences in English require an explicit subject. Hence, *studied for
the exam is ungrammatical and He studied for the exam is grammatical. The
highest level of adequacy is explanatory adequacy, which is achieved when the
description or theory not only reaches descriptive adequacy but does so using
abstract principles which can be applied beyond the language being considered
and become a part of Universal Grammar. At this level of adequacy, one would
describe the inability of English to omit subject pronouns as a consequence of
the fact that, unlike Spanish or Japanese, English is not a language which per-
mits pro-drop, i.e. the omission of a subject pronoun that is recoverable from
the context or deducible from inections on the verb marking the case, gender,
or number of the subject.
Within Chomskys theory of principles and parameters, pro-drop is a conse-
quence of the null-subject parameter (Haegeman 1991: 1720). This parame-
ter is one of many which make up universal grammar, and as speakers acquire a
language, the manner in which they set the parameters of universal grammar is
determined by the norms of the language they are acquiring. Speakers acquiring
1.1 Linguistic theory and description 3
English would set the null-subject parameter to negative, since English does not
permit pro-drop; speakers of Italian, on the other hand, would set the parameter
to positive, since Italian permits pro-drop (Haegeman 1991: 18).
Because generative grammar has placed so much emphasis on universal
grammar, explanatory adequacy has always been a high priority in generative
grammar, often at the expense of descriptive adequacy: there has never been
much emphasis in generative grammar in ensuring that the data upon which
analyses are based are representative of the language being discussed, and with
the notion of the ideal speaker/hearer rmly entrenched in generative grammar,
there has been little concern for variation in a language, which traditionally
has been given no consideration in the construction of generative theories of
language. This trend has become especially evident in the most recent theory
of generative grammar: minimalist theory.
In minimalist theory, a distinction is made between those elements of a lan-
guage that are part of the core and those that are part of the periphery. The
core is comprised of pure instantiations of UG and the periphery marked
exceptions that are a consequence of historical accident, dialect mixture, per-
sonal idiosyncracies, and the like (Chomsky 1995: 1920). Because variation
is limited to nonsubstantive elements of the lexicon and general properties of
lexical items (Chomsky 1995: 170), those elements belonging to the periphery
of a language are not considered in minimalist theory; only those elements that
are part of the core are deemed relevant for purposes of theory construction.
This idealized view of language is taken because the goal of minimalist theory
is a theory of the initial state, that is, a theory of what humans know about
language in advance of experience (Chomsky 1995: 4) before they encounter
the real world of the language they are acquiring and the complexity of structure
that it will undoubtedly exhibit.
This complexity of structure, however, is precisely what the corpus linguist
is interested in studying. Unlike generative grammarians, corpus linguists see
complexity and variation as inherent in language, and in their discussions of
language, they place a very high priority on descriptive adequacy, not explana-
tory adequacy. Consequently, corpus linguists are very skeptical of the highly
abstract and decontextualized discussions of language promoted by generative
grammarians, largely because such discussions are too far removed from ac-
tual language usage. Chafe (1994: 21) sums up the disillusionment that corpus
linguists have with purely formalist approaches to language study, noting that
they exclude observations rather than . . . embrace ever more of them and that
they rely too heavily on notational devices designed to account for only those
aspects of reality that fall within their purview, ignoring the remaining richness
which also cries out for understanding. The corpus linguist embraces complex-
ity; the generative grammarian pushes it aside, seeking an ever more restrictive
view of language.
Because the generative grammarian and corpus linguist have such very dif-
ferent views of what constitutes an adequate linguistic description, it is clear
why these two groups of linguists have had such a difcult time communicating
and valuing each others work. As Fillmore (1992: 35) jokes, when the cor-
pus linguist asks the theoretician (or armchair linguist) Why should I think
that what you tell me is true?, the generative grammarian replies back Why
should I think that what you tell me is interesting? (emphasis added). Of pri-
mary concern to the corpus linguist is an accurate description of language; of
importance to the generative grammarian is a theoretical discussion of language
that advances our knowledge of universal grammar.
Even though the corpus linguist places a high priority on descriptive ade-
quacy, it is a mistake to assume that the analysis of corpora has nothing to offer
to generative theory in particular or to theorizing about language in general. The
main argument against the use of corpora in generative grammar, Leech (1992)
observes, is that the information they yield is biased more towards performance
than competence and is overly descriptive rather than theoretical. However,
Leech (1992: 108) argues that this characterization is overstated: the distinction
between competence and performance is not as great as is often claimed, since
the latter is the product of the former. Consequently, what one discovers in a
corpus can be used as the basis for whatever theoretical issue one is exploring.
In addition, all of the criteria applied to scientic endeavors can be satised in a
corpus study, since corpora are excellent sources for verifying the falsiability,
completeness, simplicity, strength, and objectivity of any linguistic hypothesis
(Leech 1992: 11213).
Despite Leechs claims, it is unlikely that corpora will ever be used very
widely by generative grammarians, even though some generative discussions
of language have been based on corpora and have demonstrated their potential
for advancing generative theory. Working within the framework of government
and binding theory (the theory of generative grammar preceding minimalist
theory), Aarts (1992) used sections of the corpus housed at the Survey of
English Usage at University College London to analyze small clauses in
English, constructions like her happy in the sentence I wanted her happy that
can be expanded into a clausal unit (She is happy). By using the London Corpus,
Aarts (1992) was not onlyable toprovide a complete descriptionof small clauses
in English but to resolve certain controversies regarding small clauses, such as
establishing the fact that they are independent syntactic units rather than simply
two phrases, the rst functioning as direct object and the second as complement
of the object.
Haegeman(1987) employedgovernment andbindingtheorytoanalyze empty
categories (i.e. positions in a clause where some element is missing) in a specic
genre of English: recipe language. While Haegemans investigation is not based
on data from any currently available corpus, her analysis uses the type of data
quite commonly found in corpora. Haegeman (1987) makes the very interest-
ing claim that parametric variation (such as whether or not a language exhibits
pro-drop) does not simply distinguish individual languages from one another
but can be used to characterize regional, social, or register variation within a
1.2 Corpora in functional descriptions of language 5
particular language. She looks specically at examples from the genre (or reg-
ister) of recipe language that contain missing objects (marked by the letters [a],
[b], etc. in the example below):
(1) Skin and bone chicken, and cut [a] into thin slices. Place [b] in bowl with mushrooms.
Pur ee remaining ingredients in blender, and pour [c] over chicken and mushrooms.
Combine [d] and chill [e] well before serving. (Haegeman 1987: 2367)
Government and binding theory, Haegeman (1987: 238) observes, recognizes
four types of empty categories, and after analyzing a variety of different exam-
ples of recipe language, Haegeman concludes that this genre contains one type
of empty category, wh-traces, not found in the core grammar of English (i.e. in
other genres or regional and social varieties of English).
What distinguishes Haegemans (1987) study from most other work in gen-
erative grammar is that she demonstrates that theoretical insights into universal
grammar can be obtained by investigating the periphery of a language as well
as the core. And since many corpora contain samples of various genres within a
language, they are very well suited to the type of analysis that Haegeman (1987)
has conducted. Unfortunately, given the emphasis in generative grammar on in-
vestigations of the core of a language (especially as reected in Chomskys
recent work in minimalism), corpora will probably never have much of a role
in generative grammar. For this reason, corpora are much better suited to func-
tional analyses of language: analyses that are focused not simply on providing
a formal description of language but on describing the use of language as a
communicative tool.
1.2 Corpora in functional descriptions of language
Even though there are numerous functional theories of language, all
have a similar objective: to demonstrate how speakers and writers use language
to achieve various communicative goals.
1
Because functionalists are interested in language as a communicative tool,
they approach the study of language from a markedly different perspective than
the generative grammarian. As formalists, generative grammarians are pri-
marily interested in describing the form of linguistic constructions and using
these descriptions to make more general claims about Universal Grammar. For
instance, in describing the relationship between I made mistakes, a sentence
in the active voice, and its passive equivalent, Mistakes were made by me, a
generative grammarian would be interested not just in the structural changes
in word order between actives and passives in English but in making more
general claims about the movement of constituents in natural language. Conse-
quently, the movement of noun phrases in English actives and passives is part
1
Newmeyer (1998: 1318) provides an overview of the approaches to language study that various
functional theories of language take.
of a more general process termed NP [noun phrase] movement (Haegeman
1991: 2703). A functionalist, on the other hand, would be more interested in
the communicative potential of actives and passives in English. And to study
this potential, the functionalist would investigate the linguistic and social con-
texts favoring or disfavoring the use of, say, a passive rather than an active
construction. A politician embroiled in a scandal, for instance, might choose to
utter the agentless passive Mistakes were made rather than I made mistakes or
Mistakes were made by me because the agentless passive allows the politician
to admit that something went wrong but at the same time to evade responsibil-
ity for the wrong-doing by being quite imprecise about exactly who made the
mistakes.
Because corpora consist of texts (or parts of texts), they enable linguists to
contextualize their analyses of language; consequently, corpora are very well
suited to more functionally based discussions of language. To illustrate how
corpora can facilitate functional discussions of language, this section contains
an extended discussion of a functional analysis of elliptical coordinations in
English based on sections of the Brown Corpus and the American component of
the International Corpus of English (ICE). The goal of the analysis (described
in detail in Meyer 1995) was not simply to describe the form of elliptical
coordinations in speech and writing but to explain why certain types of elliptical
coordinations are more common than others, why elliptical coordinations occur
less frequently in speech than in writing, and why certain types of elliptical
coordinations are favored more in some written genres than others.
The study was based on a 96,000-word corpus containing equal proportions
of different types of speech and writing: spontaneous dialogues, legal cross
examinations, press reportage, belles lettres, learned prose, government doc-
uments, and ction. These genres were chosen because they are known to be
linguistically quite different and to have differing functional needs. Govern-
ment documents, for instance, are highly impersonal. Consequently, they are
likely to contain linguistic constructions (such as agentless passives) that are
associated with impersonality. Spontaneous dialogues, on the other hand, are
much more personal, and will therefore contain linguistic constructions (such
as the personal pronouns I and we) advancing an entirely opposite communica-
tive goal. By studying genres with differing functional needs, one can take a
particular linguistic construction (such as an elliptical coordination), determine
whether it has varying frequencies and uses in different genres, and then use
this information to determine why such distributions exist and to isolate the
function (or communicative potential) of the construction.
Inanelliptical coordination, some element is left out that is recoverable within
the clause in which the ellipsis occurs. In the sentence I wrote the introduction
and John the conclusion the verb wrote is ellipted in the second clause under
identity with the same verb in the rst clause. There are various ways to describe
the different types of ellipsis occurring in English and other languages. Sanders
(1977) uses alphabetic characters to identify the six different positions in which
ellipsis can occur, ranging from the rst position in the rst clause (position A)
to the last position in the second clause (position F):
A B C & D E F
Although there is disagreement about precisely which positions permit ellipsis
in English, most would agree that English allows ellipsis in positions C, D, and
E. Example (2) illustrates C-Ellipsis: ellipsis of a constituent at the end of the
rst clause (marked by brackets) that is identical to a constituent (placed in
italics) at the end of the second clause.
(2) The author wrote [ ] and the copy-editor revised the introduction to the book.
Examples (3) and (4) illustrate D- and E-Ellipsis: ellipsis of, respectively, the
rst and second parts of the second clause.
(3) The students completed their course work and [ ] left for summer vacation.
(4) Sally likes sh, and her mother [ ] hamburgers.
The rst step in studying the functional potential of elliptical coordinations
in English was to obtain frequency counts of the three types of elliptical coordi-
nations in the samples of the corpus and to explain the frequency distributions
found. Of the three types of ellipsis in English, D-Ellipsis was the most frequent,
accounting for 86 percent of the elliptical coordinations identied in the corpus.
In contrast, both C- and E-Ellipsis were very rare, occurring in, respectively,
only 2 percent and 5.5 percent of the elliptical coordinations.
2
These frequency
distributions are identical to those found by Sanders (1977) in a survey he con-
ducted of the frequency of ellipsis types in a variety of different languages.
For instance, Sanders (1977) found that while all of the languages of the world
allow D-Ellipsis, far fewer permit C-Ellipsis.
To explain typological distributions such as this, Sanders (1977) invokes two
psycholinguistic constraints: the suspense effect (as Greenbaum and Meyer
1982 label it) and the serial position effect. Briey, the suspense effect predicts
that ellipsis will be relatively undesirable if the site of ellipsis precedes the
antecedent of ellipsis, since the suspense created by the anticipation of the
ellipted item places a processing burden on the hearer or reader. C-Ellipsis
is therefore a relatively undesirable type of ellipsis because the antecedent of
ellipsis (the introduction to the book in example 2) comes after the ellipsis
in position C at the end of the rst clause. D- and E-Ellipsis, on the other
hand, are more desirable than C-Ellipsis because neither ellipsis type violates
the suspense effect: for both types of ellipsis, the antecedent of ellipsis occurs
in the rst clause (position A for D-Ellipsis and position B for E-Ellipsis) in
positions prior to ellipsis in the D- and E-positions in the second clause.
2
The remaining 6.5 percent of elliptical coordinations consisted of constructions exhibiting more
than one type of ellipsis and therefore no tendency towards any one type of ellipsis. For example,
the example below contains both C- and D-Ellipsis: ellipsis of the direct object in the rst clause
and subject of the second clause.
(i) We
1
tried out [ ]
2
and [ ]
1
then decided to buy the car
2
.
Table 1.1 The favorability of C-, D-, and E- Ellipsis
Ellipsis type Suspense effect Serial position effect
D-Ellipsis F F
E-Ellipsis F L
C-Ellipsis L L
F= favorable
L= less favorable
The serial position effect is based on research demonstrating that when given
memory tests, subjects will remember items placed in certain positions in a
series better than other positions. For instance, subjects will recall items placed
rst in a series more readily and accurately than items placed in the middle of
a series. The results of serial learning experiments can be applied to the six po-
sitions in a coordinated construction (AF) and make predictions about which
antecedent positions will be most or least conducive to memory retention and
thus favor or inhibit ellipsis. Position A, the antecedent position for D-Ellipsis
(see example 3), is the position most favorable for memory retention. Conse-
quently, D-Ellipsis will be the most desirable type of ellipsis according to the
serial position effect. The next most favorable position for memory is position
B, the antecedent position for E-Ellipsis, making this type of ellipsis slightly
less desirable than D-Ellipsis. And increasingly less desirable for memory re-
tention is the F-position, the antecedent position for C-Ellipsis, resulting in this
type of ellipsis being the least desirable type of ellipsis in English.
Working together, the Suspense and Serial Position Effects make predic-
tions about the desirability of ellipsis in English, predictions that match exactly
the frequency distributions of elliptical coordinations found in the corpora.
Table 1.1 lists the three types of ellipsis in English and the extent to which
they favorably or unfavorably satisfy the suspense and serial position effects.
D-Ellipsis quite favorably satises both the suspense and serial position effects,
a fact offering an explanation of why D-Ellipsis was the most frequent type
of ellipsis in the corpus. While E-Ellipsis satises the suspense effect, it less
favorably satises the serial position effect, accounting for its less frequent oc-
currence in the corpus than D-Ellipsis. However, E-Ellipsis was more frequent
than C-Ellipsis, a type of ellipsis that satises neither the suspense nor the
serial position effect and was therefore the least frequent type of ellipsis in the
corpus.
While the suspense and serial position effects make general predictions about
the favorability or unfavorability of the three ellipsis types in English, they fail to
explain the differing distributions of elliptical coordinations in speech and writ-
ing and in the various genres of the corpora. In speech, of the constructions in
which ellipsis was possible, only 40 percent contained ellipsis, with the remain-
ing60percent containingthe full unreducedform. Inwriting, incontrast, ellipsis
was much more common: 73 percent of the constructions in which ellipsis was
possible contained ellipsis, with only 27 percent containing the full unreduced
form. To explain these frequency differences, it is necessary to investigate why
repetition (rather than ellipsis) is more necessary in speech than in writing.
The role of repetition in speech is discussed extensively by Tannen (1989:
4753), who offers a number of reasons why a construction such as (5) below
(taken from a sample of speech in the American component of ICE) is more
likely to occur in speech than in writing.
(5) Yeah so we got that and we got knockers and we got bratwurst and we got <unintel-
ligible>wurst or kranzwurst or something I dont know. (ICE-USA-S1A-016)
In (5), there are four repetitions of a subject and verb (we got) in the D-position
that could have been ellipted rather than repeated. But in this construction, repe-
titionserves a number of useful purposes quite unique tospeech. First, as Tannen
(1989: 48) observes, the repetition allows the speaker to continue the owof the
discourse in a more efcient, less energy-draining way by enabling him/her to
continue speaking without worrying about editing what is being said and getting
rid of redundancies, a task that would greatly slow down the pace of speech. At
the same time, repetition is benecial to the hearer by providing semantically
less dense discourse (p. 49), that is, discourse containing an abundance of old
rather than new information. Moreover, repetition can create parallel structures
(as it does in example 5), and as many researchers have noted, parallelism is a
very common device for enhancing the cohesiveness of a discourse.
In addition to having a different distribution in speech and writing, elliptical
coordinations also had different distributions in the various genres of writing
that were investigated. If the genres of ction and government documents are
compared, very different patterns of ellipsis can be found. In ction, D-Ellipsis
constituted 98 percent of the instances of ellipsis that were found. In govern-
ment documents, on the other hand, D-Ellipsis made up only 74 percent of the
instances of ellipsis, with the remaining 26 percent of examples almost evenly
divided between C-Ellipsis and E-Ellipsis.
The high incidence of D-Ellipsis in ction can be explained by the fact that
ctionis largelynarration, andnarrative action, as Labov(1972: 376) has shown,
is largely carried forth in coordinate sentences. These sentences will often have
as subjects the names of characters involved in the narrative action, and as these
names are repeated, they will become candidates for D-Ellipsis. For instance,
in example (6) below (which was taken from a sample of ction in the Brown
Corpus), the second sentence (containing two coordinated clauses) begins with
reference to a male character (He) at the start of the rst clause, a reference that
is repeated at the start of the second clause, leading to D-Ellipsis rather than
repetition of the subject. Likewise, the last two sentences (which also consist
of coordinated clauses) begin with references to another character (Virginia
initially and then She), which are repeated and ellipted in the D-positions of
subsequent clauses.
(6) The days seemed short, perhaps because his routine was, each day, almost the same.
He rose late and [ ] went down in his bathrobe and slippers to have breakfast either
alone or with Rachel. Virginia treated him with attention and [ ] tried to tempt his
appetite with special food: biscuits, cookies, candies the result of devoted hours
in the tiled kitchen. She would hover over him and, looking like her brother, [ ]
anxiously watch the progress of Scottys fork or spoon. (K01 61080)
Although the government documents in the corpus contained numerous ex-
amples of D-Ellipsis, they contained many more examples of C-Ellipsis than the
samples of ction did. One reason that C-Ellipsis occurred more frequently in
government documents is that this type of construction has a function well suited
to government documents. As Biber (1988) has noted, the genre in which gov-
ernment documents can be found, ofcial documents, has a strong emphasis on
information, almost no concern for interpersonal or affective content (p. 131),
and a tendency towards highly explicit, text-internal reference (p. 142).
Instances of C-Ellipsis quite effectively help government documents achieve
these communicative goals. First of all, because government documents are
so focused on content or meaning, they are able to tolerate the stylistic awk-
wardness of constructions containing C-Ellipsis. In example (7) below (taken
froma government document in the Brown Corpus), there is a very pronounced
intonation pattern created by the C-Ellipsis, resulting in pauses at the site of
ellipsis and just prior to the ellipted construction that give the sentence a rather
abrupt and awkward intonation pattern.
(7) Each applicant is required to own [ ] or have sufcient interest in the property to be
explored. (H01 198090)
This awkwardness is tolerated in government documents because of the over-
riding concern in this genre for accuracy and explicitness. An alternative way
to word (7) would be to not ellipt the noun phrase in the C-position but instead
to pronominalize it at the end of the second clause:
(8) Each applicant is required to own the property to be explored or have sufcient
interest in it.
However, even though this wording results in no confusion in this example, in
general when a third-person pronoun is introduced into a discourse, there is
the potential that its reference will be ambiguous. If, in the case of (7), ellipsis
is used instead of pronominalization, there is no chance of ambiguity, since
the constraints for ellipsis in English dictate that there be only one source for
the ellipsis in this sentence (the noun phrase the property to be explored in the
second clause). Consequently, through ellipsis rather than pronominalization,
the communicative goal of explicitness in government documents is achieved.
The discussion of coordination ellipsis in this section provides further ev-
idence that corpus-based analyses can achieve explanatory adequacy: the
results of the study establish a direct relationship between the frequency of
the various types of elliptical coordinations across the languages of the world
1.3 Corpus-based research in linguistics 11
and their overall frequency in English. More importantly, however, the analysis
provides principled functional explanations for these frequency distributions
in English: certain kinds of elliptical coordinations place processing burdens
on the hearer/reader, thus making their overall frequency less common; at the
same time, the less common constructions are sometimes necessary because
they are communicatively necessary in certain contexts (e.g. the need in gov-
ernment documents to use a rare type of ellipsis, C-ellipsis, because this kind of
construction prevents potential ambiguity that might occur with an alternative
full-form containing a third-person pronoun).
Although not all corpus studies are as explicitly functional as the study of
coordination ellipsis in this section, all corpus-based research is functional in
the sense that it is grounded in the belief that linguistic analysis will benet
if it is based on real language used in real contexts. And as the next section
will demonstrate, this methodological principle has inuenced how research is
conducted in numerous linguistic disciplines.
1.3 Corpus-based research in linguistics
Linguists of all persuasions have discovered that corpora can be very
useful resources for pursuing various research agendas. For instance, many
lexicographers have found that they can more effectively create dictionaries
by studying word usage in very large linguistic corpora. Much current work
in historical linguistics is now based on corpora containing texts taken from
earlier periods of English, corpora that permit a more systematic study of the
evolution of English and that enable historical linguists to investigate issues that
have currency in modern linguistics, such as the effects of gender on language
usage in earlier periods of English. Corpora have been introduced into other
linguistic disciplines as well, and have succeeded in opening up new areas of
research or bringing new insights to traditional research questions. To illustrate
how corpora have affected research in linguistics, the remainder of this chapter
provides an overview of the various kinds of corpus-based research now being
conducted in various linguistic disciplines.
3
1.3.1 Grammatical studies of specic linguistic constructions
The study of coordination ellipsis in the previous section illustrated
a very common use of corpora: to provide a detailed study of a particular
grammatical construction that yields linguistic information on the construction,
3
The following sections do not provide an exhaustive listing of the research conducted in the
various areas of linguistics that are discussed. For a comprehensive survey of corpus-based re-
search, see either Bengt Altenbergs online bibliography: 1989: http://www.hd.uib.no/icame/
icame-bib2.txt; 19908: http://www.hd.uib.no/icame/icame-bib3.htm; or Michael Barlows:
http://www.ruf.rice.edu/barlow/refn.html.
such as the various forms it has, its overall frequency, the particular contexts in
which it occurs (e.g. speech rather than writing, ction rather than spontaneous
dialogues, and so forth), and its communicative potential.
Corpus-based research of this nature has focused on the use and structure of
many different kinds of grammatical constructions, such as appositives in con-
temporary English (Meyer 1992) and earlier periods of the language (Pahta and
Nevanlinna 1997); clefts and pseudo-clefts (Collins 1991b); innitival comple-
ment clauses (Mair 1990); past and perfective verb forms in various periods of
English (Elsness 1997); the modals can/may and shall/will in early American
English (Kyt o 1991); and negation (Tottie 1991) (see the ICAME Bibliography
for additional studies).
To investigate the use and structure of a grammatical construction, most
have found it more protable to investigate constructions that occur relatively
frequently, since if a construction occurs too infrequently, it is often hard to
make strong generalizations about its form and usage. For instance, in the
discussion of coordination ellipsis in the previous section, the infrequent occur-
rence of instances of E-Ellipsis (e.g. Joes a vegetarian, and Sally a carnivore)
helped make the theoretical point that if a particular grammatical construc-
tion occurs rarely in the worlds languages, in those languages in which it
does occur, it will have a very infrequent usage. At the same time, the lack of
many examples of E-Ellipsis made it difcult to make strong generalizations
about the usage of this construction in English. In many respects, this prob-
lem is a consequence of the relatively small corpus upon which the study of
coordination ellipsis was based. For this reason, to study some linguistic con-
structions, it will often be necessary to study very large corpora: the British
National Corpus, for instance (at 100 million words in length), rather than the
Brown Corpus (at one million words in length). However, for those construc-
tions that do occur frequently, even a relatively small corpus can yield reliable
and valid information. To illustrate this point, it is instructive to compare two
studies of modal verbs in English Coates (1983) and Mindt (1995) whose
results are similar, even though the studies are based on very different sized
corpora.
Coates (1983) was one of the earlier corpus studies of modals and was based
on two corpora totaling 1,725,000 words: the Lancaster Corpus (a precursor to
the LOB Corpus of written British English) and sections of the London Corpus
containing speech, letters, and diaries. Coates (1983) used these two corpora to
describe the different distributions of modals in writing and speech and the more
frequent meanings associated with the individual modals. Mindts (1995) study
of modals was based on a much larger group of corpora that together totaled
80 million words of speech and writing: the Brown and LOB corpora, sections
of the LondonLund Corpus containing surreptitiously recorded speech, the
LongmanLancaster Corpus, and CD-ROMS containing newspaper articles
from The Times and the Independent. Mindt (1995) used these corpora not only
to study the form and meaning of modals but to provide a comprehensive view
of the verb phrase in English based on the approximately 30,000 verb phrases
he identied in his corpora.
Although the size of Coates (1983) and Mindts (1995) corpora is drastically
different, many of their results are strikingly similar. Both studies found a
more frequent occurrence of modals in speech than in writing. Although the
rankings are different, both studies found that will, can, and would were the most
frequently occurring modals in speech, and that will and would were the most
frequent modals in writing. Certain modals tended to occur most frequently
in one medium rather than the other: may in writing more often than speech;
shall more often in speech than in writing. Even though both studies contain
frequency information on the meanings of modals, it is difcult to make direct
comparisons: the two studies used different categories to classify the meanings
of modals, and Coates (1983) calculated frequencies based only on an analysis
of one of her corpora (the London Corpus), biasing her results more towards
speech and certain kinds of unprinted material. Nevertheless, the results that
can be compared illustrate that frequently occurring grammatical constructions
can be reliably studied in relatively small corpora.
1.3.2 Reference grammars
While it is quite common to use corpora to investigate a single gram-
matical construction in detail, it is also possible to use corpora to obtain informa-
tion on the structure and usage of many different grammatical constructions and
to use this information as the basis for writing a reference grammar of English.
As was noted in the Preface, there is a long tradition in English studies,
dating back to the nineteenth and early twentieth centuries, to use some kind
of corpus as the basis for writing a reference grammar of English, a tradition
followed by grammarians such as Jespersen (190949) or Curme (1947), who
based their grammars on written material taken from the works of eminent
Englishwriters. Manymodern-dayreference grammars followthis tradition, but
instead of using the kinds of written texts that Jespersen and Curme used, have
based their discussions of grammar on commonly available corpora of written
and spoken English. One of the rst major reference works to use corpora
were the two grammars written by Quirk, Greenbaum, Leech, and Svartvik:
AGrammar of Contemporary English(1972) andAComprehensive Grammar of
the English Language (1985). In many sections of these grammars, discussions
of grammatical constructions were informed by analyses of the London Corpus.
For instance, Quirket al.s (1985: 1351) descriptionof the nounphrase concludes
with a table presenting frequency information on the distribution of simple and
complex noun phrases in various genres of the London Corpus. In this table,
it is pointed out that in prose ction and informal spoken English, a sentence
with the structure of (9) would be the norm: the subject contains a simple noun
phrase (the pronoun he) and the object a complex noun phrase consisting of a
head noun (guy) followed by a relative clause (who is supposed to have left).
(9) Hes the guy who is supposed to have left (ICE-GB S1A-008 266)
In scientic writing, in contrast, this distribution of simple and complex noun
phrases was not as common. That is, in scientic writing, there was a greater
tendency to nd complex noun phrases in subject positions. Thus, in scientic
writing, it was not uncommon to nd sentences such as (10), a sentence in which
a very complex noun phrase containing a head (those) followed by a relative
clause (who have . . . ) occurs in subject position of the sentence:
(10) Even those who have argued that established, traditional religions present a major
hegemonic force can recognize their potential for developing an internal plural-
ism. (ICE-GB:W2A-012 40)
Information of this nature is included in the Quirk et al. grammars because one
of the principles underlying these grammars is that a complete description of
English entails information not just on the form of grammatical constructions
but on their use as well.
More recent reference grammars have relied even more heavily on corpora.
Like the Quirk et al. grammars, these grammars use corpora to provide informa-
tion on the form and use of grammatical constructions, but additionally contain
extensive numbers of examples from corpora to illustrate the grammatical con-
structions under discussion. Greenbaums Oxford English Grammar (1996) is
based almost entirely on grammatical information extracted from the British
Component of the International Corpus of English (ICE-GB). The Collins
COBUILD Project has created a series of reference grammars for learners of
English that contains examples drawn from Bank of English Corpus (Sinclair
1987). Biber et al.s Longman Grammar of Spoken and Written English (1999)
is based on the Longman Spoken and Written English Corpus, a corpus that is
approximately 40 million words in length and contains samples of spoken and
written British and American English. This grammar provides extensive infor-
mation not just on the form of various English structures but on their frequency
and usage in various genres of spoken and written English.
1.3.3 Lexicography
While studies of grammatical constructions can be reliably conducted
on corpora of varying length, to obtain valid information on vocabulary items,
it is necessary to analyze corpora that are very large. To understand why this
is the case, one need only investigate the frequency patterns of vocabulary
in shorter corpora, such as the one-million-word LOB Corpus. In the LOB
Corpus, the ve most frequent lexical items are the function words the, of,
and, to, and a. The ve least frequent lexical items are not ve single words
but rather hundreds of different words that occur from ten to fteen times
each in the corpus. These words include numerous proper nouns as well as
miscellaneous content words such as alloy, beef, and bout. These frequencies
illustrate a simple fact about English vocabulary (or, for that matter, vocabulary
patterns in any language): a relatively small number of words (function words)
will occur with great frequency; a relatively large number of words (content
words) will occur far less frequently. Obviously, if the goal of lexical analysis
is to create a dictionary, the examination of a small corpus will not give the
lexicographer complete information concerning the range of vocabulary that
exists inEnglishandthe varyingmeanings that these vocabularyitems will have.
Because a traditional linguistic corpus, such as the LOB Corpus, is a mere
snapshot of the language at a certain point in time (Ooi 1998: 55), some have
argued that the only reliable way to study lexical items is to use what is termed
a monitor corpus, that is, a large corpus that is not static and xed but that
is constantly being updated to reect the fact that new words and meanings are
always being added to English. This is the philosophy of the Collins COBUILD
Project at Birmingham University in England, which has produced a number
of dictionaries based on two monitor corpora: the Birmingham Corpus and the
Bank of English Corpus. The Birmingham Corpus was created in the 1980s
(cf. Renouf 1987 and Sinclair 1987), and while its size was considered large at
the time (20 million words), it would nowbe considered fairly small, particularly
for the study of vocabulary items. For this reason, the Birmingham Corpus has
been superseded by the Bank of English Corpus, which as of October 2000
totaled 415 million words.
The Bank of English Corpus has many potential uses, but it was designed
primarily to help in the creation of dictionaries. Sections of the corpus were
used as the basis of the BBC English Dictionary, a dictionary that was intended
to reect the type of vocabulary used in news broadcasts such as those on the
BBC (Sinclair 1992). Consequently, the vocabulary included in the dictionary
was based on sections of the Bank of English Corpus containing transcriptions
of broadcasts on the BBC (70 million words) and on National Public Radio
in Washington, DC (10 million words). The Bank of English Corpus was also
used as the basis for a more general purpose dictionary, the Collins COBUILD
English Dictionary, and a range of other dictionaries on such topics as idioms
and phrasal verbs. Other projects have used similar corpora for other types
of dictionaries. The Cambridge Language Survey has developed two corpora,
the Cambridge International Corpus and the Cambridge Learners Corpus, to
assist in the writing of a number of dictionaries, including the Cambridge
International Dictionary of English. Longman publishers assembled a large
corpus of spoken and written American English to serve as the basis of the
Longman Dictionary of American English, and used the British National Corpus
as the basis of the Longman Dictionary of Contemporary English.
To understand why dictionaries are increasingly being based on corpora, it
is instructive to review precisely how corpora, and the software designed to
analyze them, can not only automate the process of creating a dictionary but
also improve the information contained in the dictionary. A typical dictionary,
as Landau (1984: 76f.) observes, provides its users with various kinds of infor-
mation about words: their meaning, pronunciation, etymology, part of speech,
and status (e.g. whether the word is considered colloquial or non-standard).
In addition, dictionaries will contain a series of example sentences to illustrate
in a meaningful context the various meanings that a given word has.
Prior to the introduction of computer corpora in lexicography, all of this infor-
mation had to be collected manually. As a consequence, it took years to create
a dictionary. For instance, the most comprehensive dictionary of English, the
Oxford English Dictionary (originally entitled New English Dictionary), took
fty years to complete, largely because of the many stages of production that the
dictionary went through. Landau (1984: 69) notes that the 5 million citations in-
cluded in the OEDhad to be painstakingly collected . . . subsorted . . . analyzed
by assistant editors and dened, with representative citations chosen for inclu-
sion; and checked and redened by [James A. H.] Murray [main editor of the
OED] or one of the other supervising editors. Of course, less ambitious dictio-
naries than the OEDtook less time to create, but still the creation of a dictionary
is a lengthy and arduous process.
Because so much text is now available in computer-readable form, many
stages of dictionary creation can be automated. Using a relatively inexpensive
piece of software called a concordancing program (cf. section 5.3.2), the lexi-
cographer can go through the stages of dictionary production described above,
and instead of spending hours and weeks obtaining information on words, can
obtain this information automatically from a computerized corpus. In a matter
of seconds, a concordancing program can count the frequency of words in a
corpus and rank them from most frequent to least frequent. In addition, some
concordancing programs can detect prexes and sufxes and irregular forms
and sort words by lemmas: words such as runs, running, and ran will not
be counted as separate entries but rather as variable forms of the lemma run.
To study the meanings of individual words, the lexicographer can have a word
displayed in KWIC (key word in context) format, and easily view the varying
contexts in which a word occurs and the meanings it has in these contexts. And
if the lexicographer desires a copy of the sentence in which a word occurs, it can
be automatically extracted from the text and stored in a le, making obsolete
the handwritten citation slip stored in a ling cabinet. If each word in a corpus
has been tagged (i.e. assigned a tag designating its word class; cf. section 4.3),
the part of speech of each word can be automatically determined. In short, the
computer corpus and associated software have completely revolutionized the
creation of dictionaries.
In addition to making the process of creating a dictionary easier, corpora
can improve the kinds of information about words contained in dictionaries,
and address some of the deciencies inherent in many dictionaries. One of the
criticisms of the OED, Landau (1984: 71) notes, is that it contains relatively
little information on scientic vocabulary. But as the BBC English Dictionary
illustrates, if a truly representative corpus of a given kind of English is created
(in this case, broadcast English), it becomes quite possible to produce a dictio-
nary of any type of English (cf. section 2.5 for a discussion of representativeness
in corpus design). And with the vast amount of scientic English available in
computerized form, it would now be relatively easy to create a dictionary of
scientic English that is corpus-based.
Dictionaries have also been criticized for the unscientic manner in which
they dene words, a shortcoming that is obviously a consequence of the fact that
many of the more traditional dictionaries were created during times when well-
dened theories of lexical meaning did not exist. But this situation is changing
as semanticists turn to corpora to develop theories of lexical meaning based
on the use of words in real contexts. Working within the theory of frame
semantics, Fillmore (1992: 3945) analyzed the meaning of the word risk in a
25-million-word corpus of written English created by the American Publishing
House for the Blind. Fillmore (1992: 40) began his analysis of risk in this
corpus working from the assumption that all uses of risk t into a general
frame of meaning that there is a probability, greater than zero and less than
one, that something bad will happen to someone or something. Within this
general frame were three frame elements, i.e. differing variations on the main
meaning of risk, depending upon whether the risk is not caused by someones
action (e.g. if you stay here you risk getting shot), whether the risk is due in
some part to what is termed the Protagonists Deed (e.g. I had no idea when
I stepped into that bar that I was risking my life), or whether the risk results
from the Protagonists decision to perform the Deed (e.g. I know I might lose
everything, but what the hell, Im going to risk this weeks wages on my favorite
horse) (Fillmore 1992: 412).
In a survey of ten monolingual dictionaries, Fillmore (1992: 3940) found
great variation in the meanings of risk that were listed, with only two dictionaries
distinguishing the three meanings of risk. In his examination of the 25-million-
word corpus he was working with, Fillmore (1992) found that of 1,743 instances
of risk he identied, most had one of the three meanings. However, there were
some examples that did not t into the risk frame, and it is these examples that
Fillmore (1992: 43) nds signicant, since without having examined a corpus,
we would not have thought of them on our own. Fillmores (1992) analysis
of the various meanings of the word risk in a corpus effectively illustrates the
value of basing a dictionary on actual uses of a particular word. As Fillmore
(1992: 39) correctly observes, the citation slips the lexicographers observed
were largely limited to examples that somebody happened to notice . . . But
by consulting a corpus, the lexicographer can be more condent that the results
obtained more accurately reect the actual meaning of a particular word.
1.3.4 Language variation
Much of the corpus-based research discussed so far in this section has
described the use of either grammatical constructions or lexical items in some
kindof context: speechvs. writing, or scientic writingvs. broadcast journalism.
The reasons these kinds of studies are so common is that modern-day corpora,
fromtheir inception, have been purposely designed to permit the study of what is
termedgenre variation, i.e. howlanguage usage varies accordingtothe context
in which it occurs. The rst computer corpus, the Brown Corpus, contained
various kinds of writing, and this corpus design has inuenced the composition
of most balanced corpora created since then.
Because corpus linguists have focused primarily on genre variation, they have
a somewhat different conception of language variation than sociolinguists do. In
sociolinguistics, the primary focus is howvarious sociolinguistic variables, such
as age, gender, and social class, affect the way that individuals use language.
One reason that there are not more corpora for studying this kind of variation is
that it is tremendously difcult to collect samples of speech, for instance, that are
balanced for gender, age, and ethnicity (a point that is discussed in greater detail
in section 2.5). Moreover, once such a corpus is created, it is less straightforward
to study sociolinguistic variables than it is to study genre variation. To study
press reportage, for instance, it is only necessary to take from a given corpus
all samples of press reportage, and to study within this subcorpus whatever one
wishes to focus on. To study variation by gender in, say, spontaneous dialogues,
on the other hand, it becomes necessary to extract froma series of conversations
in a corpus what is spoken by males as opposed to females a much more
complicated undertaking, since a given conversation may consist of speaker
turns by males and females distributed randomly throughout a conversation,
and separating out who is speaking when is neither a simple nor straightforward
computational task. Additionally, the analyst might want to consider not just
which utterances are spoken by males and females but whether an individual
is speaking to a male or female, since research has shown that how a male or
female speaks is very dependent upon the gender of the individual to whom
they are speaking.
But despite the complications that studying linguistic variables poses, de-
signers of some recent corpora have made more concerted efforts to create
corpora that are balanced for such variables as age and gender, and that are
set up in a way that information on these variables can be extracted by various
kinds of software programs. Prior to the collection of spontaneous dialogues in
the British National Corpus, calculations were made to ensure that the speech
to be collected was drawn from a sample of speakers balanced by gender, age,
social class, and dialect region. Included within the spoken part of the BNC is
a subcorpus known as the Corpus of London Teenage English (COLT). This
part of the corpus contains a valid sampling of the English spoken by teenagers
from various socioeconomic classes living in different boroughs of London. To
enable the study of sociolinguistic variables in the spoken part of the BNC, each
conversation contains a le header (cf. section 4.1), a statement at the start of
the sample providing such information as the age and gender of each speaker in
a conversation. Asoftware program, Sara, was designed to read the headers and
do various analyses of the corpus based on a pre-specied selection of sociolin-
guistic variables. Using Sara, Aston and Burnard (1998: 11723) demonstrate
how a query can be constructed to determine whether the adjective lovely is, as
many have suggested, used more frequently by females than males. After using
Sara to count the number of instances of lovely spoken by males and females,
they conrmed this hypothesis to be true.
Other corpora have been designed to permit the study of sociolinguistic
variables as well. In the British component of the International Corpus of
English (ICE-GB), ethnographic information on speakers and writers is stored
in a database, and a text analysis program designed to analyze the corpus,
ICECUP (cf. section 5.3.2), can draw upon information in this database to re-
strict searches. Even though ICE-GB is not a balanced corpus it contains the
speech and writing of more males than females a search of lovely reveals the
same usage trend for this word that was found in the BNC.
Of course, programs such as Sara and ICECUP have their limitations. In
calculatinghowfrequentlymales andfemales use lovely, bothprograms canonly
count the number of times a male or female speaker uses this expression; neither
program can produce gures that, for instance, could help determine whether
females use the word more commonly when speaking with other females than
males. And both programs depend heavily on how accurately and completely
sociolinguistic variables have been annotated, and whether the corpora being
analyzed provide a representative sample of the variables. In using Sara to gather
dialectal information from the BNC, the analyst would want to spot check the
ethnographic information on individuals included in the corpus to ensure that
this information accurately reects the dialect group in which the individuals are
classied. Even if this is done, however, it is important to realize that individuals
will style-shift: they may speak in a regional dialect to some individuals but
use a more standard form of the language with others. In studying variation by
gender in ICE-GB, the analyst will want to reviewthe results with caution, since
this corpus does not contain a balanced sample of males and females. Software
such as Sara or ICECUP may automate linguistic analyses, but it cannot deal
with the complexity inherent in the classication of sociolinguistic variables.
Therefore, it is important to view the results generated by such programs with
a degree of caution.
Although traditionally designed corpora such as Brown or LOB might seem
decient because they do not easily permit the study of sociolinguistic vari-
ables, this deciency has been more than compensated for by the important
information on genre variation that these corpora have yielded. Bibers (1988)
study of the linguistic differences between speech and writing effectively il-
lustrates the potential that corpora have for yielding signicant insights into
the structure of different written and spoken genres of English. Using the LOB
Corpus of writing and the LondonLund Corpus of speech, Biber (1988) was
able to show that contrary to the claims of many, there is no strict division
between speech and writing but rather that there exists a continuum between
the two: certain written genres (such as ction) contain linguistic structures
typically associated with speech, whereas certain spoken genres (such as pre-
pared speeches) contain structures more commonly associated with writing. To
reach this conclusion, Biber (1988) rst used a statistical test, factor analysis
(explained in section 5.4.1), to determine which linguistic constructions tended
to co-occur in particular texts. Biber (1988: 13) was interested in grammati-
cal co-occurrences because he believes that strong co-occurrence patterns of
linguistic features mark underlying functional dimensions; that is, that if pas-
sives and conjuncts (e.g. therefore or nevertheless) occur together, for instance,
then there is some functional motivation for this co-occurrence. The functional
motivations that Biber (1988) discovered led him to posit a series of textual
dimensions. Passives and conjuncts are markers of abstract uses of language,
Biber (1988: 1514) maintains, and he places them on a dimension he terms
Abstract versus Non-Abstract Information. High on this dimension are two
types of written texts that are dense in linguistic constructions that are markers
of abstract language: academic prose and ofcial documents. Low on the di-
mension are two types of spoken texts that contain relatively few abstractions:
face-to-face conversations and telephone conversations. However, Biber (1988)
also found that certain kinds of written texts (e.g. romantic ction) were low on
the dimension, and certain kinds of spoken texts (e.g. prepared speeches) were
higher on the dimension. Findings such as this led Biber (1988) to conclude
that there is no absolute difference between speech and writing. More recently,
Biber (1995) has extended this methodology to study genre analysis in corpora
of languages other than English, and in corpora containing texts from earlier
periods of English (Biber and Burges 2000).
1.3.5 Historical linguistics
The study of language variation is often viewed as an enterprise con-
ducted only on corpora of Modern English. However, there exist a number of
historical corpora corpora containing samples of writing representing earlier
dialects and periods of English that can be used to study not only language
variation in earlier periods of English but changes in the language from the past
to the present.
Much of the interest in studying historical corpora stems from the creation
of the Helsinki Corpus, a 1.5-million-word corpus of English containing texts
from the Old English period (beginning in the eighth century) through the early
ModernEnglishperiod(the rst part of the eighteenthcentury). Texts fromthese
periods are further grouped into subperiods (ranging from70100 years) to pro-
vide what Rissanen (1992: 189) terms a chronological ladder of development,
that is, a grouping of texts from a specic period of time that can be compared
with other chronological groupings of texts to study major periods of linguistic
development within the English language. In addition to covering various peri-
ods of time, the texts in the Helsinki Corpus represent various dialect regions in
Englandanddifferent genres (e.g. law, homilies andsermons, ction, andletters;
for full details, see Kyt o 1996); and it contains sociolinguistic information on
authors (e.g. age, gender) for texts written from the Middle English period on-
wards, since prior to this period sociolinguistic information is too inconsistent
to form a basis for any socio-historical considerations (Rissanen 1992: 193).
To ll gaps in the Helsinki Corpus, ARCHER (A Representative Corpus
of English Historical Registers) was created (cf. Biber et al. 1994; Biber and
Burges 2000). ARCHER is 1.7 million words in length and contains American
as well as British English. It covers the years 16501990, with texts divided
into fty-year subgroups. Like the Helsinki Corpus, it contains various genres
of English, representing not just formal exposition (e.g. scientic writing) but
personal letters and diaries and more spoken-based genres, such as sermons
and ctional dialogue.
Although FLOB (the Freiburg LOB Corpus of British English) and FROWN
(the Freiburg Brown Corpus of American English) are not historical corpora in
the sense that the Helsinki and ARCHER corpora are, they do permit the study
of changes in British and American English between 1961 and 1991. FLOBand
FROWNreplicate the LOBand Brown corpora, respectively, but with texts pub-
lished in the year 1991. Thus, FLOBand FROWNallowfor studies of linguistic
change in British and American English over a period of thirty years. Although
thirty years is not a long time for language to change, studies of FLOB and
FROWN have documented changes in the language during this period (cf., for
instance, Mair 1995). Consequently, FLOBandFROWNare synchronic corpora
on the one hand, providing resources for describing Modern English, but di-
achronic corpora on the other when compared with the LOBand Brown corpora.
The two main historical corpora, the Helsinki and ARCHER corpora, are
what Rissanen (2000) terms multi-purpose general corpora because they con-
tain many different texts and text fragments covering various periods of English.
There are, however, more focusedhistorical corpora coveringspecic works, au-
thors, genres, or periods. These corpora include electronic versions of Beowulf
(The Electronic Beowulf; cf. Prescott 1997andhttp://www.uky.edu/kiernan/
eBeowulf/guide.htm); the works of Chaucer and other Middle English writers
(the Corpus of Middle English Prose and Verse; cf. http://www.hti.umich.edu/
english/mideng/); collections of early English letters (the Corpus of Early
English Correspondence; cf. Nevalainen and Raumolin-Brunberg 1996); and
early Modern English tracts (the Lampeter Corpus; cf. Schmied and Claridge
1997). As Rissanens (2000) discussion of these and other historical corpora
indicates, the creation of historical corpora has become as active an enterprise
as the creation of corpora of Modern English.
Historical corpora have greatly enhanced our ability to study the linguistic
development of English: such corpora allow corpus linguists not only to study
systematically the development of particular grammatical categories in English
but to gain insights into how genres in earlier periods differed linguistically and
how sociolinguistic variables such as gender affected language usage.
Finegan and Bibers (1995) study of the alternation of that and zero in con-
structions such as I know(that) this is true provides a good example of the types
of genre variation that can be studied in historical corpora. Finegan and Biber
(1995) studied the alternation of that and zero in various genres of ARCHER
and the Helsinki Corpus. In Modern English, this alternation is a marker of for-
mal vs. informal style. As Greenbaum, Nelson and Weizmans (1996: 84) study
of speech and writing demonstrated, that tended to be more common in formal
styles, such as academic writing and monologues, and much less common in
informal styles, such as conversations and personal letters. This trend, however,
was only partially conrmed by Finegan and Biber (1995): although the most
colloquial written genre that they investigated, personal letters, contained the
most instances of zero (at least in some periods), this genre (and the two more
written-based genres, sermons and medicine) showed that historically in written
genres of English, that has been preferred over zero in complement clauses, and
that the trendtowards that has increasedinall genres since 1750. Of course, these
results are tentative, as Finegan and Biber (1995) correctly observe. Neverthe-
less, they demonstrate the potential that historical corpora have for documenting
changes in the language in both written and speech-based genres of English.
The study of sociolinguistic variables in historical corpora has provided nu-
merous insights into the use of language in earlier periods, particularly in the
area of gender differences. Nevalainen (2000) used the Corpus of Early English
Correspondence to explore a hypothesis concerning gender that is popular in
sociolinguistics, namely that rather than simply favour[ing] prestige forms,
females in fact create them (emphases in original). Nevalainen (2000) chose
personal letters as the basis of the study because they are the best kind of genre
from earlier periods of English to study the day-to-day interaction between
people, the kind of interaction in which gender differences are likely to emerge.
In her corpus, Nevalainen (2000) analyzed gender differences in the use of three
different linguistic constructions: the replacement in subject positions of ye by
you, the replacement of the third-person-singular verb sufx-(e)th by-(e)s, and
the loss of multiple negation and subsequent increase in the use of non-
assertive indenites such as any or ever. While Nevalainen (2000) found that the
rst two changes were led by female writers, the loss of double negation was
promoted by males, a change that Nevalainen (2000) attributes to the two
sexes differential access to education in general and to professional special-
izations in particular. These mixed results lead Nevalainen (2000) to conclude
that womens role in language change was different in the sixteenth century than
it is in the twentieth century, and that further information is needed to determine
precisely when women came to generally promote linguistic change.
1.3.6 Contrastive analysis and translation theory
Corpora, inparticular parallel corpora, have beencreatedtofacilitate
contrastive analyses of English and other languages, advance developments in
translation theory, and enhance foreign language teaching.
One of the earlier parallel corpora, the EnglishNorwegian Parallel Corpus,
provides a good illustration of how parallel corpora have been designed and of
the particular kinds of linguistic analyses that can be conducted on them. The
EnglishNorwegianParallel Corpus contains samples of EnglishandNorwegian
ction and non-ction that are 10,00015,000 words in length (Johansson and
Ebeling 1996). This corpus was designed to permit three kinds of studies. First,
because the corpus contains samples of similar types of ction and non-ction
written originally in English or Norwegian, it can be used to study genre vari-
ation between English and Norwegian: how English ction, for instance, is
similar to or different than Norwegian ction. Second, each sample has been
translated from the original language into the other language. Consequently,
the corpus allows for the study of how English is typically translated into
Norwegian, or Norwegian into English. Finally, the corpus can be used to com-
pare the structure of, for instance, ction written originally in Norwegian with
ction translated into Norwegian, or non-ction written originally in English
with non-ction translated into English.
To analyze the corpus, software has been developed to align sentences and to
conduct various kinds of searches. For instance, one can search for an English
expression such as as it were and receive a list of every sentence containing
this expression as well as a sentence containing the Norwegian translation of
the expression:
(11) <s id=FW1.4.s153 corresp=FW1T.4.s154>She took a swig of whiskey and tried to
relocate herself, as it were, in her own skin.</s>
<s id=FW1T.4.s154 corresp=FW1.4.s153>Hun tok seg en god slurk med whisky,
og prvde agjennne seg selv liksom, i sitt eget skinn.</s> (Johansson and Ebeling
1996: 6)
This kind of display permits a range of different contrastive studies to be con-
ducted on the corpus. Hasselg ard (1997) studied sentence openings: the dif-
ferent grammatical constructions that English and Norwegian use to begin sen-
tences. She found that in the original English and Norwegian samples there
were distinct differences between the elements placed in the opening section of
a sentence (e.g. Norwegian more frequently places direct objects at the begin-
ning of a sentence than English does). However, in the translations, there were
fewer differences, with the source language having a greater effect on the
syntax of the target language (Hasselg ard 1997: 18).
This inuence that Hasselg ard (1997) discovered in her study is commonly
referred to as translationese (Johansson and Hoand 1993), i.e. features of
the translated text more characteristic of the source language than the language
the sentence is being translated into. Schmied and Sch afer (1996) describe
a number of examples of translationese in a parallel corpus of English and
German, the Chemnitz Corpus. Theynote that translationese caninvolve not just
particular syntactic and morphological differences but pragmatic differences as
well. For instance, in German, it is very common to express tentativeness with
adverbials or modal particles. However, Schmied and Sch afer (1996: 50) found
a number of examples in their corpus tending more towards the English norm
of using in the case of (12) a lexical modal structure in the German translation
rather than the more natural modal particle (given in brackets at the end of the
example).
(12) In the less developed areas it has tended to widen. (Commission of the European
Union, 1993: 10)
[ . . . ] in den weniger entwickelten Teilen der Gemeinschaft bestand sogar die
Tendenz zu einer Vergr osserung def Kluft. [natural language: hat sich die Kluft
eher vergr ossert] (Kommission der Europ aischen Gemeinschaften, 1993: 10)
Schmied and Sch afer (1996: 52) observe that information on translation such
as this can be used to train translators or help create bilingual dictionaries.
Because of the increased popularity of parallel corpora, there now exists
general-purpose software, such as ParaConc (Barlow 1999), that can be used
to align sentences in any two or more languages and to aid researchers in
conducting the kinds of studies detailed in this section.
1.3.7 Natural language processing
The primary emphasis in this chapter has been on how corpora can be
used to conduct linguistic analyses of English. However, there are many corpus
linguists whose interests are more computational than linguistic. These linguists
have created and used corpora to conduct research in an area of computational
linguistics known as natural language processing (NLP). For instance, the
North American Chapter of the Association for Computational Linguistics reg-
ularly has workshops and special sessions at which computational linguists in
NLP discuss the use of corpora to advance research in such areas as tagging,
parsing, information retrieval, and the development of speech recognition sys-
tems. One recent special session at the 2000 conference in Seattle, Washington,
on Large Corpus Annotation and Software Standards: Towards an American
National Corpus provided a forumfor the discussion of the necessary software
and annotation standards required for annotating very large corpora, such as
the American National Corpus, which is intended to be the American equiv-
alent of the British National Corpus (cf. chapter 4 for more details on corpus
annotation).
Because researchers in NLPhave their own distinct interests, the corpora they
use are designed differently than corpora such as Brown or LOB. This point is
illustrated by a survey of the corpora distributed by the Linguistic Data Con-
sortium (LDC), an organization working out of the University of Pennsylvania
(http://www.ldc.upenn.edu/). Many LDC corpora have been used to develop
speech recognition systems. The TIMITAcousticPhonetic Continuous Speech
Corpus contains digitized recordings of 630 speakers, representing various di-
alect regions in the United States, who were recorded reading a series of sen-
tences. Aligned with each speakers recording are phonetic as well as standard
orthographic transcriptions of each word a speaker utters. The Switchboard
Corpus contains digitized recordings of approximately 2,400 telephone con-
versations. The conversations are orthographically transcribed and the tran-
scriptions are linked to the audio (cf. Wheatley et al. 1992). The TIMIT and
Switchboard corpora are but two of the many LDC distributed corpora used to
improve speech recognition systems.
Other LDC corpora have been used to develop taggers and parsers and in-
formation retrieval systems. The Penn Treebank contains a heterogeneous col-
lection of spoken and written texts totaling 4.9 million words; sections of this
corpus have been tagged and parsed (cf. Marcus, Santorini, and Marcinkiewicz
1993 and http://www.cis.upenn.edu/treebank/cdrom2.html, and also sec-
tion 4.5 for a more detailed discussion of treebanks). The TIPSTER Corpus
contains a variety of written texts taken from such sources as the Wall Street
Journal and Associated Press newswire stories. Because the corpus is used
mainly to develop information retrieval systems, it is not tagged and parsed but
annotated with SGML tags, which are designed to reveal the structure of docu-
ments rather than their linguistic structure (cf. section 4.1 for more information
on SGML tags).
Corpora such as those described above have not been used that extensively
by corpus linguists doing descriptive studies of English, largely because many
of the corpora used by researchers in NLP are not suitable for many kinds of
linguistic analyses. In the abstract, the kinds of texts used in the Switchboard
Corpus telephone conversations would be very useful for linguistic analysis.
However, conversants whose speech is included in this corpus were instructed
beforehand to discuss a specic topic. Therefore, the speech in this corpus is
not natural (cf. section 3.2 for a discussion of natural speech). The Penn
Treebank is not a balanced corpus: a large percentage of it is press reportage.
Thus, the results of any linguistic analysis of this corpus will be limited to
making generalizations about a very small number of genres.
But despite the limitations of the corpora created by researchers in NLP,
the tools they have developed taggers and parsers in particular have been
instrumental in creating grammatically analyzed corpora that are of great value
to descriptive linguists. For instance, work done by computational linguists
in developing the Nijmegen tagger and parser for the International Corpus of
English greatly facilitated the grammatical annotation of the British component
of ICE, and led to a fully tagged and parsed one-million-word corpus of spoken
and written English that, when used with a text retrieval program(ICECUP), can
extract fromICE-GBa wealth of grammatical information that in the past could
be obtained only by manual analysis (cf. section 5.3.2 for more information on
ICE-GB and ICECUP). The UCREL team at Lancaster University consists of
a group of descriptive and computational linguists who worked together not
only to create the British National Corpus but to develop the tagger (CLAWS)
that was used to tag the corpus (Garside and Smith 1997). Speech recognition
programs have progressed to the point where they can recognize with varying
degrees of accuracy not just carefully scripted monologues but broadcast speech
(Nguyen et al. 1999). As these programs improve, they may in the future be used
to at least partially automate the process of transcribing speech for inclusion in
a corpus (cf. sections 3.5 and 3.6 for more on the transcription of speech). As
these research initiatives illustrate, while their goals may be different, corpus
linguists of all persuasions have much to contribute to each others work.
1.3.8 Language acquisition
Although studies of language acquisition have always had an empirical
basis, researchers in the areas of rst- and second-language acquisition have
tended not to make publicly available the data upon which their studies were
based. However, this situation is changing and there now exist corpora suitable
for studying both rst- and second-language acquisition.
To facilitate the study of both rst- and second-language acquisition, the
CHILDES (Child Language Data Exchange) System was developed. This sys-
tem contains a corpus of transcriptions of children and adults learning rst
and second languages that are annotated in a specic format called CHAT
and that can be analyzed with a series of software programs called CLAN
(MacWhinney 1996: 2). Although much of the corpus is focused on English,
a considerable part of it represents nineteen additional languages as well, and
the speech that is included is taken from normally developing individuals as
well as those with language disorders such as aphasia or autism (MacWhinney
1996: 3). The CHILDEScorpus has generated an impressive amount of research
on language acquisition, ranging from studies of children learning Germanic
and Romance languages to studies of children who have language disabilities
(MacWhinney 2000: 421f.).
To study second-language acquisition, a number of researchers have begun
developing what are called learner corpora: corpora containing the speech or
writing of individuals learning English as a second or foreign language. One
of the larger corpora in this area is called the International Corpus of Learner
English (ICLE). ICLEis currently more than two million words in length (http://
www.tr.ucl.ac.be/FLTR/GERM/ETAN/CECL/introduction.html); it is divided
into 500-word essays written by students fromfourteen different linguistic back-
grounds learning English as a foreign language (Granger 1998: 10). Other
learner corpora include the Longmans Learner Corpus (http://www.longman-
elt.com/dictionaries/corpus/lclearn.html) and the Hong Kong University of Sci-
ence and Technology (HKUST) Learner Corpus (Milton and Freeman 1996).
The creation of learner corpora has led to a number of studies in the area of
what Granger (1998: 1214) terms contrastive interlanguage analysis (CIA):
the use of corpora to compare the interlanguage a native speaker of, for instance,
French develops when learning English, and to compare across language groups
the structure of the various interlanguages that individuals from different rst-
language backgrounds develop. Altenberg and Tapper (1998) employed this
type of contrastive interlanguage analysis in their study of the use of adverbial
connectors in the section of ICLE containing the writing of Swedish students
learning English (a subcomponent called SWICLE). Additionally, they com-
pared the use of adverbial connectors by Swedish students with native-speaker
usage. Although their conclusions are tentative, they found that overall the
Swedish learners used connectives similarly to the way that native speakers
do. Compared with French learners of English, however, the Swedish learners
underused connectives, setting them apart from other language groups, which
tend to overuse connectives. What led to this underuse, Altenberg and Tapper
(1998: 92) note, was the tendency of Swedish learners to substitute less formal
connectives for more formal connectives.
1.3.9 Language pedagogy
One consequence of the development of learner corpora is that re-
searchers are taking information from them to develop teaching strategies for
individuals learning English as a second or foreign language. After discovering
that Swedish learners of English overused informal connectives in their writing,
Altenberg and Tapper (1998: 92) concluded that the best strategy for teaching
them the use of connectives would be to expose them to a greater range of
registers and to a more extensive training in expository writing. Gillard and
Gadsby (1998) describe how they used the Longmans Learner Corpus (LLC)
to help create a dictionary, the Longman Essential Activator (1997), designed
specically for students learning English. One way they used the corpus was to
discover which errors students tended to make and then to use this information
to design special help boxes included in the dictionary. For instance, Gillard
and Gadsby (1998: 167) discovered many errors in the usage of the word peoples
in the LLC, and they used this information to provide special information on
the use of this word in its entry in the dictionary. The idea behind using learner
corpora to develop teaching strategies is that they give teachers an accurate
depiction of how their students are actually using the language, information
that can then be incorporated into textbooks and lesson plans.
In addition to using information from learner corpora to develop teaching
strategies for learners of English, many have advocated that students them-
selves study corpora to help them learn about English, a methodology known
as data-driven learning (Johns 1994 and Hadley 1997). This method of teach-
ing has students investigate a corpus of native-speaker speech or writing with
a concordancing program to give them real examples of language usage rather
than the contrived examples often found in grammar books. Kettemann (1995),
for instance, provides examples of how students can use a concordancing pro-
gram to investigate various kinds of grammatical distinctions in English, such
as the difference between the past and present perfect. Kettemann (1995: 4) ar-
gues that a concordancing programis an extremely powerful hypothesis testing
device that allows the student to conduct inductive explorations of grammatical
constructions on vast amounts of data. Because concordancing programs can
generate so much data, Gavioli (1997: 84) argues that students need to be taught
how to interpret the data that they discover, a process that she claims can lead
to an effective language-learning activity.
The notion of data-driven learning has been extended to involve the ultimate
corpus: the World Wide Web. The Lingua Center within the Division of English
as an International Language at the University of Illinois has developed a
methodologycalledthe Grammar Safari (http://deil.lang.uiuc.edu/web.pages/
grammarsafari.html). This methodology has students use search engines to nd
and then study grammatical constructions in various online resources, such as
novels, newspapers, and magazines. Meyer (1998) describes the actual process
that students can follow to nd grammatical constructions on the WWW. He
shows how a search engine, such as AltaVista, and a web browser, such as
Netscape, can be used to study the disputed usage between you and I and to
nd not only examples of the usage but discussions of the usage in the archives
of various e-mail discussion lists, such as the Linguist List and Ask a Linguist.
1.4 Conclusions
This chapter has demonstrated that corpora have numerous uses, rang-
ing from the theoretical to the practical, making them valuable resources for
descriptive, theoretical, and applied discussions of language. Because corpus
linguistics is a methodology, all linguists even generativists could in prin-
ciple use corpora in their studies of language. However, most generativists feel
that a corpus enables one to study performance, not competence; as a result,
they continue to use introspection as the primary source of their data. In many
other disciplines of linguistics, corpora have proven to be valuable resources:
they are used for creating dictionaries, studying language change and variation,
understanding the process of language acquisition, and improving foreign- and
second-language instruction. However, for corpora to be most effective, they
must be compiled in a manner that ensures that they will be most useful to those
who are potential users of them. This topic how to plan, create, and annotate
a corpus is the subject of the next three chapters of the book.
Study questions
1. Corpus linguistics is based on the principle that linguistic theory and descrip-
tion is best based on real rather than contrived data. What are the advantages
and disadvantages of obtaining data from a corpus rather than from intro-
spection?
2. Why have lexicographers in recent years been using corpora to create dictio-
naries? What kinds of information can lexicographers get from corpora that
they cannot get using more traditional approaches to dictionary creation?
Study questions 29
3. Why is it easier to study genre variation in corpora than it is to study socio-
linguistic variables, such as age and gender?
4. Why have language teachers started to use corpora to teach English as a
second or foreign language? What can a student get from a corpus that he
or she cannot get from a traditional book on grammar?
2 Planning the construction of a corpus
Before the texts to be included in a corpus are collected, annotated, and
analyzed, it is important to plan the construction of the corpus carefully: what
size it will be, what types of texts will be included in it, and what population
will be sampled to supply the texts that will comprise the corpus. Ultimately,
decisions concerning the composition of a corpus will be determined by the
planned uses of the corpus. If, for instance, the corpus is be used primarily for
grammatical analysis (e.g. the analysis of relative clauses or the structure of
noun phrases), it can consist simply of text excerpts rather than complete texts.
On the other hand, if the corpus is intended to permit the study of discourse
features, then it will have to contain complete texts.
Deciding howlengthy text samples within a corpus should be is but one of the
many methodological considerations that must be addressed before one begins
collecting data for inclusion in a corpus. To explore the process of planning a
corpus, this chapter will consider the methodological assumptions that guided
the compilation of the British National Corpus. Examining the British National
Corpus reveals how current corpus planners have overcome the methodolog-
ical deciencies of earlier corpora, and raises more general methodological
considerations that anyone planning to create a corpus must address.
2.1 The British National Corpus
At approximately 100 million words in length, the British National
Corpus (BNC) (see table 2.1) is one of the largest corpora ever created. Most
of the corpus (about 90 percent) consists of various types of written British
English, with the remainder (about 10 percent) comprised of different types of
spoken British English. Even though writing dominates in the BNC, the amount
of speech in the corpus is the most ever made available in a corpus.
1
In planning the collection of texts for the BNC, a number of decisions were
made beforehand:
1. Even though the corpus would contain both speech and writing, more writing
would be collected than speech.
1
Prior to the creation of the British National Corpus, the LondonLund Corpus contained the most
speech available in a corpus: approximately 500,000 words (see Svartvik 1990).
30
2.1 The British National Corpus 31
Table 2.1 The composition of the British National Corpus
(adapted from Aston and Burnard 1998: 2933 and
http://info.ox.ac.uk/bnc/what/balance.html)
Speech
Type Number of texts Number of words % of spoken corpus
Demographically 153 4,211,216 41%
sampled
Educational 144 1,265,318 12%
Business 136 1,321,844 13%
Institutional 241 1,345,694 13%
Leisure 187 1,459,419 14%
Unclassied 54 761,973 7%
Total 915 10,365,464 100%
Writing
Type Number of texts Number of words % of written corpus
Imaginative 625 19,664,309 22%
Natural science 144 3,752,659 4%
Applied science 364 7,369,290 8%
Social science 510 13,290,441 15%
World affairs 453 16,507,399 18%
Commerce 284 7,118,321 8%
Arts 259 7,253,846 8%
Belief & thought 146 3,053,672 3%
Leisure 374 9,990,080 11%
Unclassied 50 1,740,527 2%
Total 3,209 89,740,544 99%
2
2. A variety of different genres of speech and writing would be gathered for
inclusion in the corpus.
3. Each genre would be divided into text samples, and each sample would not
exceed 40,000 words in length.
4. A number of variables would be controlled for in the entire corpus, such as
the age and gender of speakers and writers.
5. For the written part of the corpus, a careful record of a variety of variables
would be kept, including when and where the texts were written or published;
whether they were books, articles, or manuscripts; and who their target
audience was.
6. For the demographically sampled spoken samples, texts would be collected
from individuals representing the major dialect regions of Great Britain and
the various social classes existing within these regions.
2
Because fractions were rounded up or down to the nearest whole number, cumulative frequencies
in this and subsequent tables do not always add up to exactly 100 percent.
Even though the creators of the BNC had a very denite plan for the composi-
tion of the corpus, it is important to realize, as Biber (1993: 256) observes, that
the creation of a corpus is a cyclical process, requiring constant re-evaluation
as the corpus is being compiled. Consequently, the compiler of a corpus should
be willing to change his or her initial corpus design if circumstances arise
requiring such changes to be made. For instance, in planning the part of the
BNC containing demographically sampled speech, it was initially thought that
obtaining samples of speech from 100 individuals would provide a sufciently
representative sample of this type of speech (Crowdy 1993: 259). Ultimately,
however, the speech of 124 individuals was required to create a corpus bal-
anced by age, gender, social class, and region of origin (Aston and Burnard
1998: 32).
While changes in corpus design are inevitable, it is crucial that a number
of decisions concerning the composition of a corpus are made prior to the
collection of texts for the corpus, and the remainder of this chapter considers
the factors that inuenced the decisions detailed above that the creators of the
BNC made, and the general methodological issues that these decisions raise.
2.2 The overall length of a corpus
Earlier corpora, such as Brown and LOB, were relatively short, primar-
ily because of the logistical difculties that computerizing a corpus created. For
instance, all of the written texts for the Brown Corpus had to be keyed in by hand,
a process requiring a tremendous amount of very tedious and time-consuming
typing. This process has been greatly eased with the invention of optical scan-
ners (cf. section 3.7), which can automatically computerize printed texts with
fairly high rates of accuracy. However, technology has not progressed to the
point where it can greatly expedite the collection and transcription of speech:
there is much work involved in going out and recording speech (cf. section 3.2),
and the transcription of spoken texts still has to be done manually with either
a cassette transcription machine or special software that can replay segments
of speech until the segments are accurately transcribed (cf. section 3.6). These
logistical realities explain why 90 percent of the BNCis writing and only 10 per-
cent speech.
To determine how long a corpus should be, it is rst of all important to
compare the resources that will be available to create it (e.g. funding, research
assistants, computing facilities) with the amount of time it will take to collect
texts for inclusion, computerize them, annotate them, and tag and parse them.
Chafe, Du Bois, and Thompson (1991: 701) calculate that it will require six
person-hours to transcribe one minute of speech for inclusion in the Santa
Barbara Corpus of Spoken American English, a corpus containing not just
orthographic transcriptions of speech but prosodic transcriptions as well (cf.
section 4.1). For the American component of ICE, it was calculated that it
would take eight hours of work to process a typical 2,000-word sample of
2.2 The overall length of a corpus 33
writing, and ten to twenty hours to process a 2,000-word spoken text (with
spontaneous dialogues containing numerous speech overlaps requiring more
time than carefully articulated monologues). Since the American component of
ICE contains 300 samples of speech and 200 samples of writing, completing
the entire corpus would take a full-time research assistant working forty hours
a week more than three years to complete. If one applies these calculations to
the BNC, which is one hundred times the length of the American component
of ICE, it becomes obvious that creating a corpus requires a large commitment
of resources. In fact, the creation of the BNC required the collaboration of a
number of universities and publishers in Great Britain, as well as considerable
nancial resources from funding agencies.
After determining how large a corpus ones resources will permit, the next
step is to consider howlong the corpus needs to be to permit the kinds of studies
one envisions for it. One reason that the BNC was so long was that one planned
use of the corpus was to create dictionaries, an enterprise requiring a much
larger database than is available in shorter corpora, such as Brown or LOB. On
the other hand, all of the regional components of the ICE Corpus are only one
million words in length, since the goal of the ICE project is not to permit in-
depth lexical studies but grammatical analyses of the different regional varieties
of English a task that can be accomplished in a shorter corpus.
In general, the lengthier the corpus, the better. However, it is possible to
estimate the required size of a corpus more precisely by using statistical for-
mulas that take the frequency with which linguistic constructions are likely
to occur in text samples and calculate how large the corpus will have to be
to study the distribution of the constructions validly. Biber (1993: 2534)
carried out such a study (based on information in Biber 1988) on 481 text
samples that occurred in twenty-three different genres of speech and writing.
He found that reliable information could be obtained on frequently occurring
linguistic items such as nouns in as few as 59.8 text samples. On the other
hand, infrequently occurring grammatical constructions such as conditional
clauses required a much larger number of text samples (1,190) for valid in-
formation to be obtained. Biber (1993: 254) concludes that the most con-
servative approach would be to base the ultimate size of a corpus on the
most widely varying feature: those linguistic constructions, such as condi-
tional clauses, that require the largest sample size for reliable studies to be
conducted. A corpus of 1,190 samples would therefore be 2,380,000 words in
length (if text samples were 2,000 words in length, the standard length in many
corpora).
3
Unfortunately, such calculations presuppose that one knows precisely what
linguistic constructions will be studied in a corpus. The ultimate length of a
3
S anchez and Cantos (1997) describe a formula they developed, based on a corpus of Spanish they
created, to predict the length of corpus necessary to provide a reliable and accurate representation
of Spanish vocabulary. However, because their calculations are based on the Spanish language,
they caution that future research (p. 265) needs to be conducted to determine whether the
formula can be applied to other languages, such as English, as well.
corpus might therefore be better determined not by focusing too intently on the
overall length of the corpus but by focusing more on the internal structure of the
corpus: the range of genres one wishes to include in the corpus, the length and
number of individual text samples required to validly represent the genres that
will make up the corpus, and the demographic characteristics of the individuals
whose speech and writing will be chosen for inclusion in the corpus.
2.3 The types of genres to include in a corpus
The BNC, as table 2.1 indicates, contains a diverse range of spoken
and written genres. A plurality of the spoken texts (41 percent) were demo-
graphically sampled: they consisted of a variety of spontaneous dialogues and
monologues recorded from individuals living in various parts of Great Britain.
4
The remainder of the spoken samples contained a fairly even sampling (12
14 percent) of monologues and dialogues organized by purpose. For instance,
those texts in the genre of education included not just classroom dialogues and
tutorials but lectures and news commentaries as well (Crowdy 1993: 263). A
plurality of the written texts (22 percent) were imaginative: they represented
various kinds of ctional and creative writing. Slightly less frequent were sam-
ples of writing from world affairs (18 percent), the social sciences (15 percent),
and a number of other written genres, such as the arts (8 percent) and the natural
sciences (4 percent).
If the BNC is compared with the International Corpus of English (ICE), it
turns out that while the two corpora contain the same range of genres, the genres
are much more specically delineated in the ICE Corpus (see table 2.2) than
they are in the BNC. For instance, in both corpora, 60 percent of the spoken
texts are dialogues and 40 percent are monologues. In the ICE Corpus, this
division is clearly reected in the types of genres making up the spoken part
of the corpus. In the BNC, on the other hand, dialogues and monologues are
interspersed among the various genres (e.g. business, leisure) making up the
spoken part of the corpus (Crowdy 1993: 263).
Likewise, the other spoken genres in the ICE Corpus (e.g. scripted and un-
scripted speeches) are included in the BNCbut within the general areas outlined
in table 2.1. In both corpora, there is a clear bias towards spontaneous dialogues:
in the ICE Corpus, 33 percent of the spoken texts consist of direct conversations
or telephone calls; in the BNC, 41 percent of the spoken texts are of this type
(although some of the texts in the category are spontaneous monologues as
well).
While the amount of writing in the BNC greatly exceeded the amount of
speech, just the opposite is true in the ICE Corpus: only 40 percent of the
texts are written. While creative (or imaginative) writing was the most common
4
The types of individuals whose speech is represented in the demographically sampled part of the
BNC are discussed in greater detail in section 2.5.
Table 2.2 Composition of the ICE (adapted from Greenbaum 1996a: 2930)
Speech
Type Number of texts Length % of spoken corpus
Dialogues 180 360,000 59%
Private 100 200,000 33%
direct conversations 90 180,000 30%
distanced conversations 10 20,000 3%
Public 80 160,000 26%
class lessons 20 40,000 7%
broadcast discussions 20 40,000 7%
broadcast interviews 10 20,000 3%
parliamentary debates 10 20,000 3%
legal cross-examinations 10 20,000 3%
business transactions 10 20,000 3%
Monologues 120 240,000 40%
Unscripted 70 140,000 23%
spontaneous commentaries 20 40,000 7%
speeches 30 60,000 10%
demonstrations 10 20,000 3%
legal presentations 10 20,000 3%
Scripted 50 100,000 17%
broadcast news 20 40,000 7%
broadcast talks 20 40,000 7%
speeches (not broadcast) 10 20,000 3%
Total 300 600,000 99%
Writing
Type Number of texts Length % of written corpus
Non-printed 50 100,000 26%
student untimed essays 10 20,000 5%
student examination essays 10 20,000 5%
social letters 15 30,000 8%
business letters 15 30,000 8%
Printed 150 300,000 75%
informational (learned): 40 80,000 20%
humanities, social sciences,
natural sciences, technology
informational (popular): 40 80,000 20%
humanities, social sciences,
natural sciences, technology
informational (reportage) 20 40,000 10%
instructional: administrative, 20 40,000 10%
regulatory, skills, hobbies
persuasive (press editorials) 10 20,000 5%
creative (novels, stories) 20 40,000 10%
Total 200 400,000 101%
type of writing in the BNC, in the ICE Corpus it is not as prominent. More
prominent were learned and popular examples of informational writing: writing
from the humanities, social and natural sciences, and technology (40 percent of
the written texts). These categories are also represented in the BNC, although
the BNC makes a distinction between the natural, applied, and social sciences
and, unlike the ICE Corpus, does not include equal numbers of texts in each
of these categories. The ICE Corpus also contains a fairly signicant number
(25 percent) of non-printed written genres (such as letters and student writing),
while only 510 percent of the BNC contains these types of texts.
To summarize, while there are differences in the composition of the ICE
Corpus and BNC, overall the two corpora represent similar genres of spoken
and written English. The selection of these genres raises an important method-
ological question: why these genres and not others?
This question can be answered by considering the two main types of corpora
that have been created, and the particular types of studies that can be carried out
on them. The ICE Corpus and BNC are multi-purpose corpora, that is, they are
intended to be used for a variety of different purposes, ranging from studies of
vocabulary, to studies of the differences between various national varieties of
English, to studies whose focus is grammatical analysis, to comparisons of the
various genres of English. For this reason, each of these corpora contains a broad
range of genres. But in striving for breadth of coverage, some compromises had
to be made in each corpus. For instance, while the spoken part of the ICE
Corpus contains legal cross-examinations and legal presentations, the written
part of the corpus contains no written legal English. Legal written English was
excluded from the ICE Corpus on the grounds that it is a highly specialized
type of English rmly grounded in a tradition dating back hundreds of years,
and thus does not truly represent English as it is written in the 1990s (the years
during which texts for the ICE Corpus were collected). The ICE Corpus also
contains two types of newspaper English: press reportage and press editorials.
However, as Biber (1988: 18096) notes, newspaper English is a diverse genre,
containing not just reportage and editorials but, for instance, feature writing as
well a type of newspaper English not represented in the ICE Corpus.
Because general-purpose corpora such as the ICE Corpus and BNC do not
always contain a full representation of a genre, it is nowbecoming quite common
to see special-purpose corpora being developed. For instance, the Michigan
Corpus of Academic Spoken English (MICASE) was created to study the type
of speech used by individuals conversing in an academic setting: class lectures,
class discussions, student presentations, tutoringsessions, dissertationdefenses,
and many other kinds of academic speech (Powell and Simpson 2001: 34
40). By restricting the scope of the corpus, energy can be directed towards
assembling a detailed collection of texts that fully represent the kind of academic
language one is likely to encounter at an American university. As of June 2001,
seventy-one texts (813,684 words) could be searched on the MICASE website:
http://www.hti.umich.edu/micase/.
2.3 The types of genres to include in a corpus 37
At the opposite extreme of special-purpose corpora like MICASE are those
which have specialized uses but not for genre studies. The Penn Treebank
consists of a heterogeneous collection of texts totalling 100 million words,
ranging from Dow Jones newswire articles to selections from the King James
version of the Bible (Marcus, Santorini, and Marcinkiewicz 1993). The reason
that a carefully selected range of genres is not important in this corpus is that
the corpus is not intended to permit genre studies but to train taggers and
parsers to analyze English: to present them with a sufcient amount of data so
that they can learn the structure of numerous constructions and thus produce
a more accurate analysis of the parts of speech and syntactic structures present
in English (cf. section 4.5 for more on parsers). And to accomplish this goal, all
that is important is that a considerable number of texts are available, and less
important are the genres from which the texts were taken.
Because of the wide availability of written and spoken material, it is relatively
easy to collect material for modern-day corpora such as the BNC and ICE; the
real work is in recording and transcribing spoken material (cf. section 3.2), for
instance, or obtaining copyright clearance for written material (cf. section 3.3).
With historical corpora, however, collecting texts from the various genres ex-
isting in earlier periods is a much more complicated undertaking.
In selecting genres for inclusion in historical corpora, the goals are similar
to those for modern-day corpora: to nd a range of genres representative of
the types of English that existed during various historical periods of English.
Consequently, there exist multi-purpose corpora, such as Helsinki, which con-
tains a range of different genres (sermons, travelogues, ction, drama, etc.;
cf. section 1.3.5 for more details), as well as specialized corpora, such as the
Corpus of Early English Correspondence, a corpus of letters written during
the Middle English period. In gathering material for corpora such as these, the
corpus compiler must deal with fact that many of the genres that existed in ear-
lier periods are either unavailable or difcult to nd. For instance, even though
spontaneous dialogues were as common and prominent in earlier periods as they
are today, there were no tape recorders around to record speech. Therefore, there
exists no direct record of speech. However, this does not mean that we cannot get
at least a sense of what speech was like in earlier periods. In her study of early
American English, Kyt o (1991: 29) assembled a range of written texts that are
second-hand representations of speech: various verbatim reports, such as the
proceedings from trials and meetings and depositions obtained from witnesses,
and transcriptions of sermons. Of course, it can never be known with any great
certainty exactly how close to the spoken word these kinds of written texts are.
Nevertheless, texts of this type give us the closest approximation of the speech
of earlier periods that we will ever be able to obtain.
In other situations, a given genre may exist but be underrepresented in a
given period. In his analysis of personal pronouns across certain periods in the
Helsinki Corpus, Rissanen (1992: 194) notes that certain genres were dif-
cult to compare across periods because they were poorly represented during
certain periods: histories in Old English and laws and correspondence in Middle
English. Correspondence in the formof personal letters is not available until the
fteenth century. Other types of genres exist in earlier periods but are dened
differently than they are in the modern period. The Helsinki Corpus includes
the genre of science in the Old English period, but the texts that are included
focus only on astronomy, a reection of the fact that science played a much
more minimal role in medieval culture than it does today.
2.4 The length of individual text samples to be included
in a corpus
Corpora vary in terms of the length of the individual text samples that
they contain. The Lampeter Corpus of Early Modern English Tracts consists
of complete rather than composite texts. In the Helsinki Corpus, text samples
range from 2,00010,000 words in length. Samples within the BNC are equally
varied but are as long as 40,000 words. This is considerably lengthier than earlier
corpora, such as Brown or LOB, which contained 2,000-word samples, and the
LondonLund Corpus, which contained 5,000-word samples. The ICE Corpus
follows in the tradition of Brown and LOB and contains 2,000-word samples.
5
Because most corpora contain relatively short stretches of text, they tend to
contain text fragments rather than complete texts. Ideally, it would be desirable
to include complete texts in corpora, since even if one is studying grammatical
constructions, it is most natural to study these constructions within the context
of a complete text rather than only part of that text. However, there are numerous
logistical obstacles that make the inclusion of complete texts in corpora nearly
impossible. For instance, many texts, such as books, are quite lengthy, and to
include a complete text in a corpus would not only take up a large part of the
corpus but require the corpus compiler to obtain permission to use not just a
text excerpt, a common practice, but an entire text, a very uncommon practice.
Experience with the creation of the ICE Corpus has shown that, in general, it
is quite difcult to obtain permission to use copyrighted material, even if only
part of a text is requested, and this difculty will only be compounded if one
seeks permission to use an entire text (cf. section 3.3).
Of course, just because only text samples are included in a corpus does not
mean that sections of texts ought to be randomly selected for inclusion in a
corpus. It is possible to take excerpts that themselves form a coherent unit. In
the American component of ICE (and many earlier corpora as well), sections of
spoken texts are included that form coherent conversations themselves: many
conversations consist of subsections that formcoherent units, and that have their
own beginnings, middles, and ends. Likewise, for written texts, one can include
5
All of these lengths are approximate. For instance, some texts in the ICE Corpus exceed 2,000
words to avoid truncating a text sample in mid-word or mid-paragraph.
2.4 The length of text samples 39
the rst 2,000 words of an article, which contains the introduction and part of
the body of the article, or one can take the middle of an article, which contains a
signicant amount of text developing the main point made in the article, or even
its end. Many samples in the ICECorpus also consist of composite texts: a series
of complete short texts that total 2,000 words in length. For instance, personal
letters are often less than 2,000 words, and a text sample can be comprised of
complete letters totalling 2,000 words. For both the spoken and written parts of
the corpus, not all samples are exactly 2,000 words: a sample is not broken off in
mid-sentence but at a point (often over or just under the 2,000-word limit) where
a natural break occurs. But even though it is possible to include coherent text
samples in a corpus, creators and users of corpora simply have to acknowledge
that corpora are not always suitable for many types of discourse studies, and
that those wishing to carry out such studies will simply have to assemble their
own corpora for their own personal use.
In including short samples from many different texts, corpus compilers are
assuming that it is better to include more texts frommany different speakers and
writers than fewer texts from a smaller number of speakers and writers. And
there is some evidence to suggest that this is the appropriate approach to take in
creating a corpus. Biber (1990 and 1993: 24852) conducted an experiment in
which he used the LOB and LondonLund Corpora to determine whether text
excerpts provide a valid representation of the structure of a particular genre.
He divided the LOB and LondonLund corpora into 1,000-word samples: he
took two 1,000-word samples fromeach 2,000-word written sample of the LOB
Corpus, and he divided each 5,000-word sample of speech from the London
LundCorpus intofour 1,000-wordsamples. In110of these 1,000-wordsamples,
Biber calculated the frequency of a range of different linguistic items, such as
nouns, prepositions, present- and past-tense verb forms, passives, and so forth.
He concluded that 1,000-word excerpts are lengthy enough to provide valid
and reliable information on the distribution of frequently occurring linguistic
items. That is, if one studied the distribution of prepositions in the rst thousand
words of a newspaper article totalling 10,000 words, for instance, studying
the distribution of prepositions in the entire article would not yield different
distributions: the law of diminishing returns is reached after 1,000 words. On
the other hand, Biber found that infrequently occurring linguistic items (such as
relative clauses) cannot be reliably studied in a short excerpt; longer excerpts
are required.
In addition to studying the distribution of word categories, such as nouns or
prepositions, Biber (1993: 250) calculated the frequency with which newwords
are added to a sample as the number of words in the sample increases. He found,
for instance, that humanities texts are more lexically diverse than technical
prose texts (p. 252), that is, that as a humanities text progresses, there is higher
likelihood that new words will be added as the length of the text increases than
there will be in a technical prose text. This is one reason why lexicographers
need such large corpora to study vocabulary trends, since so much vocabulary
(in particular, open-class items such as nouns and verbs) occurs so rarely. And
as more text is considered, there is a greater chance (particularly in humanities
texts) that new words will be encountered.
Bibers (1993) ndings would seem to suggest that corpus compilers ought
to greatly increase the length of text samples to permit the study of infrequently
occurring grammatical constructions and vocabulary. However, Biber (1993:
252) concludes just the opposite, arguing that Given a nite effort invested
in developing a corpus, broader linguistic representation can be achieved by
focusing on diversity across texts and text types rather than by focusing on
longer samples from within texts. In other words, corpus compilers should
strive to include more different kinds of texts in corpora rather than lengthier
text samples. Moreover, those using corpora to study infrequently occurring
grammatical constructions will need to go beyond currently existing corpora
and look at additional material on their own.
6
2.5 Determining the number of texts and range
of speakers and writers to include in a corpus
Related to the issue of how long text samples should be in a corpus
is precisely how many text samples are necessary to provide a representative
sampling of a genre, and what types of individuals ought to be selected to sup-
ply the speech and writing used to represent a genre. These two issues can be
approached from two perspectives: from a purely linguistic perspective, and
from the perspective of sampling methodology, a methodology developed by
social scientists to enable researchers to determine how many elements from
a population need to be selected to provide a valid representation of the popu-
lation being studied. For corpus linguists, this involves determining how many
text samples need to be included in a corpus to ensure that valid generalizations
can be made about a genre, and what range of individuals need to be selected
so that the text samples included in a corpus provide a valid representation of
the population supplying the texts.
There are linguistic considerations that need to be taken into account in deter-
mining the number of samples of a genre to include in a corpus, considerations
that are quite independent of general sampling issues. If the number of sam-
ples included in the various genres of the BNC and ICE Corpus are surveyed,
it is immediately obvious that both of these corpora place a high value on
spontaneous dialogues, and thus contain more samples of this type of speech
than, say, scripted broadcast news reports. This bias is a simple reection of
the fact that those creating the BNC and ICE Corpus felt that spontaneous di-
alogues are a very important type of spoken English and should therefore be
6
It is also possible to use elicitation tests to study infrequently occurring grammatical items tests
that ask native speakers to comment directly on particular linguistic items. See Greenbaum(1973)
for a discussion of how elicitation tests can be used to supplement corpus studies.
2.5 Determining the number of texts and range of speakers and writers 41
amply represented. The reason for this sentiment is obvious: while only a small
segment of the speakers of English create scripted broadcast news reports, all
speakers of English engage in spontaneous dialogues.
Although it is quite easy to determine the relative importance of sponta-
neous dialogues in English, it is far more difcult to go through every potential
genre to be included in a corpus and rank its relative importance and frequency
to determine how much of the genre should be included in the corpus. And
if one did take a purely proportional approach to creating a corpus, Biber
(1993: 247) notes, the resultant corpus might contain roughly 90% conversa-
tion and 3% letters and notes, with the remaining 7% divided among registers
such as press reportage, popular magazines, academic prose, ction, lectures,
news broadcasts, and unpublished writings since these gures provide a rough
estimate of the actual percentages of individuals that create texts in each of
these genres. To determine how much of a given genre needs to be included
in a corpus, Biber (1993) argues that it is more desirable to focus only on
linguistic considerations, specically how much internal variation there is in
the genre. As Biber (1988: 170f.) has demonstrated in his pioneering work on
genre studies, some genres have more internal variation than others and, con-
sequently, more examples of these genres need to be included in a corpus to
ensure that the genres are adequately represented. For instance, Biber (1988:
178) observes that even though the genres of ofcial documents and academic
prose contain many subgenres, the subgenres in ofcial documents (e.g. gov-
ernment reports, business reports, and university documents) are much more
linguistically similar than the subgenres of academic prose (e.g. the natural and
social sciences, medicine, and the humanities). That is, if the use of a con-
struction such as the passive is investigated in the various subgenres of ofcial
documents and academic prose, there will be less variation in the use of the
passive in the ofcial documents than in the academic prose. This means that a
corpus containing ofcial documents and academic prose will need to include
more academic prose to adequately represent the amount of variation present
in this genre.
In general, Bibers (1988) study indicates that a high rate of internal varia-
tion occurs not just in academic prose but in spontaneous conversation (even
though there are no clearly delineated subgenres in spontaneous conversation),
spontaneous speeches, journalistic English (though Biber analyzed only press
reportage and editorials), general ction, and professional letters. Less inter-
nal variation occurs in ofcial documents (as described above), science ction,
scripted speeches, and personal letters.
Because the BNC is a very lengthy corpus, it provides a sufcient number
of samples of genres to enable generalizations to be made about the genres.
However, with the much shorter ICE Corpus (and with other million-word
corpora, such as Brown and LOB, as well), it is an open question whether the
forty 2,000-word samples of academic prose contained in the ICE Corpus, for
instance, are enough to adequately represent this genre. And given the range
of variation that Biber (1988) documents in academic prose, forty samples are
probably not enough. Does this mean, then, that million-word corpora are not
useful for genre studies?
The answer is no: while these corpora are too short for some studies, for fre-
quentlyoccurringgrammatical constructions theyare quite adequate for making
generalizations about a genre. For instance, Meyer (1992) used 120,000-word
sections of the Brown, the LondonLund, and the Survey of English Usage
(SEU) Corpora to study the usage of appositions in four genres: spontaneous
conversation, press reportage, academic prose, and ction. In analyzing only
twenty 2,000-word samples from the press genre of the Brown Corpus, he
found ninety-seven examples of appositions of the type political neophyte Steve
Forbes in which the rst unit lacks a determiner and occurs before a proper noun
(p. 12); in eight 5,000-word samples from the press genre of the SEU Corpus,
he found only thirty-one examples of such appositions. These ndings allowed
Meyer (1992) to claim quite condently that not only were such appositions
conned to the press genre (the other genres contained virtually no examples
of this type of apposition) but they were more common in American English
than in British English.
Studying the amount of internal linguistic variation in a genre is one way
to determine how many samples of a genre need to be included in a corpus;
applying the principles of sampling methodology is another way. Although
sampling methodology can be used to determine the number of samples needed
to represent a genre, this methodology, as was demonstrated earlier in this
section, is not all that useful for this purpose. It is best used to select the
individuals whose speech and writing will be included in a corpus.
Social scientists have developed a sophisticated methodology based on math-
ematical principles that enables a researcher to determine howmany elements
from a sampling frame need to be selected to produce a representative and
therefore valid sample. A sampling frame is determined by identifying a spe-
cic population that one wishes to make generalizations about. For instance, in
creating the Tampere Corpus, Norri and Kyt o (1996) decided that their sampling
frame would not be just scientic writing but different types of scientic writ-
ing written at varying levels of technicality. After deciding what their sampling
frame was, Norri and Kyt o (1996) then had to determine which elements from
this frame they needed to include in their corpus to produce a representative
sampling of scientic writing, that is, enough writing from the social sciences,
for instance, so that someone using the corpus could analyze the writing and
be able to claim validly that their results were true not just about the specic
examples of social science writing included in the Tampere Corpus but all so-
cial science writing. Obviously, Norri and Kyt o (1996) could not include all
examples of social science writing published in, say, a given year; they had to
therefore narrow the range of texts that they selected for inclusion.
Social scientists have developed mathematical formulas that enable a re-
searcher to calculate the number of samples they will need to take from a
sampling frame to produce a representative sample of the frame. Kretzschmar,
2.5 Determining the number of texts and range of speakers and writers 43
Meyer, and Ingegneri (1997) used one of Kaltons (1983: 82) formulas to cal-
culate the number of books published in 1992 that would have to be sampled to
create a representative sample. Bowkers Publishers Weekly lists 49,276 books
as having been published in the United States in 1992. Depending on the level
of condence desired, samples from 2,1682,289 books would have to be in-
cluded in the corpus to produce a representative sample. If each sample from
each book was 2,000 words in length, the corpus of books would be between
4,336,000 and 4,578,000 words in length. Kaltons (1983) formula could be
applied to any sampling frame that a corpus compiler identies, provided that
the size of the frame can be precisely determined.
Sampling methodology can also be used to select the particular individuals
whose speech and writing will be included in a corpus. For instance, in plan-
ning the collection of demographically sampled speech for the BNC, random
location sampling procedures were used to select individuals whose speech
would be recorded (Crowdy 1993: 259). Great Britain was divided into twelve
regions. Within these regions, thirty sampling points were selected: locations
at which recordings would be made and that were selected based on their
ACORN prole . . . (A Classication of Regional Neighbourhoods) (Crowdy
1993: 260). This prole provides demographic information about the types of
people likely to live in certain regions of Great Britain, and the prole helped
creators of the BNC select speakers of various social classes to be recorded.
In selecting potential speakers, creators of the BNC also controlled for other
variables, such as age and gender.
In using sampling methodology to select texts and speakers and writers for
inclusion in a corpus, a researcher can employ two general types of sampling:
probability sampling and non-probability sampling (Kalton 1983). In probabil-
ity sampling (employed above in selecting speakers for the BNC), the researcher
very carefully pre-selects the population to be sampled, using statistical formu-
las and other demographic information to ensure that the number and type of
people being surveyed are truly representative. In non-probability sampling, on
the other hand, this careful pre-selection process is not employed. For instance,
one can select the population to be surveyed through the process of haphazard,
convenience, or accidental sampling (Kalton 1983: 90); that is, one samples in-
dividuals who happen to be available. Alternatively, one can employ judgment,
or purposive, or expert choice sampling (Kalton 1983:91); that is, one decides
before-hand who would be best qualied to be sampled (e.g. native rather than
non-native speakers of a language, educated vs. non-educated speakers of a
language, etc.). Finally, one can employ quota sampling (Kalton 1983: 91),
and sample certain percentages of certain populations. For instance, one could
create a corpus of American English by including in it samples reecting ac-
tual percentages of ethnic groups residing in the United States (e.g. 10 percent
African Americans, 15 percent Hispanic Americans, etc.).
Although probability sampling is the most reliable type of sampling, lead-
ing to the least amount of bias, for the corpus linguist this kind of sampling
presents considerable logistical challenges. The mathematical formulas used
in probability sampling often produce very large sample sizes, as the example
with books cited above illustrated. Moreover, to utilize the sampling techniques
employed by creators of the BNC in the United States would require consider-
able resources and funding, given the size of the United States and the number
of speakers of American English. Consequently, it is quite common for corpus
linguists to use non-probability sampling techniques in compiling their corpora.
In creating the Brown Corpus, for instance, a type of judgement sampling
was used. That is, prior to the creation to the Brown Corpus, it was decided
that the writing to be included in the corpus would be randomly selected from
collections of edited writing at four locations:
(a) for newspapers, the microlm les of the New York Public Library;
(b) for detective and romantic ction, the holdings of the Providence Athen-
aeum, a private library;
(c) for various ephemeral and popular periodical material, the stock of a large
secondhand magazine store in New York City;
(d) for everything else, the holdings of the Brown University Library. (Quoted
from Francis 1979: 195)
Other corpora have employed other non-probability sampling techniques. The
American component of the International Corpus of English used a combina-
tion of judgement and convenience sampling: every effort was made to
collect speech and writing from a balanced group of constituencies (e.g. equal
numbers of males and females), but ultimately what was nally included in
the corpus was a consequence of whose speech or writing could be most eas-
ily obtained. For instance, much ction is published in the form of novels or
short stories. However, many publishers require royalty payments from those
seeking reproduction rights, a charge that can sometimes involve hundreds of
dollars. Consequently, most of the ction in the American component of ICE
consists of unpublished samples of ction taken from the Internet: ction for
which permission can usually be obtained for no cost and which is available in
computerized form. The Corpus of Spoken American English has employed a
variation of quota sampling, making every effort to collect samples of speech
from a representative sampling of men and women, and various ethnicities and
regions of the United States.
The discussion thus far in this chapter has focused on the composition of the
BNC to identify the most important factors that need to be considered prior to
collecting texts for inclusion in a corpus. This discussion has demonstrated that:
1. Lengthier corpora are better than shorter corpora. However, even more im-
portant than the sheer length of a corpus is the range of genres included
within it.
2. The range of genres to be included in a corpus is determined by whether it
will be a multi-purpose corpus (a corpus intended to have wide uses) or a
special-purpose corpus (a corpus intended for more specic uses, such as
the analysis of a particular genre like scientic writing).
2.6 The time-frame for selecting speakers and texts 45
3. It is more practical to include text fragments in a corpus rather than complete
texts. These fragments can be as short as 2,000 words, especially if the focus
of study is frequently occurring grammatical constructions.
4. The number of samples of a genre needed to ensure valid studies of the genre
is best determined by how much internal variation exists in the genre: the
more variation, the more samples needed.
5. Probability sampling techniques can also be used to determine the number
of samples of a genre necessary for inclusion in a corpus. However, such
techniques are better used for selecting the individuals from whom text
samples will be taken.
6. If it is not possible to select individuals using probability sampling tech-
niques, then non-probability sampling techniques can be used, provided that
a number of variables are considered and controlled for as well as possible.
These variables will be the focus of the remainder of this chapter.
2.6 The time-frame for selecting speakers and texts
Most corpora contain samples of speech or writing that have been writ-
ten or recorded within a specic time-frame. Synchronic corpora (i.e. corpora
containing samples of English as it is presently spoken and written) contain
texts created within a relatively narrow time-frame. Although most of the writ-
ten texts used in the BNC were written or published between 1975 and 1993, a
few texts date back as far as 1960 (Aston and Burnard 1998: 30). The Brown
and LOB corpora contain texts published in 1961. The ICE Corpus contains
texts spoken or written (or published) in the years 1990-present.
In creating a synchronic corpus, the corpus compiler wants to be sure that
the time-frame is narrow enough to provide an accurate view of contemporary
English undisturbed by language change. However, linguists disagree about
whether purely synchronic studies are even possible: new words, for instance,
come into the language every day, indicating that language change is a con-
stant process. Moreover, even grammatical constructions can change subtly in
a rather short period of time. For instance, Mair (1995) analyzed the types of
verb complements that followed the verb help in parts of the press sections
of the LOB Corpus (comprised of texts published in 1961) and the FLOB
(FreiburgLancasterOsloBergen) Corpus, a corpus comparable to LOB but
containing texts published thirty years later in 1991. Mair (1995) found some
differences in the complementation patterns, even over a period as short as
thirty years. While help with the innitive to was the norm in the 1961 LOB
Corpus (e.g. I will help you to lift the books), this construction has been re-
placed in the 1991 FLOB Corpus by the bare innitive (e.g. I will help you lift
the books).
Mairs (1995) study indicates that language change can occur over a relatively
short period of time. Therefore, if one wishes to create a truly synchronic corpus,
then the texts included in the corpus should reect as narrow a time-frame
as possible. Although this time-frame does not need to be as narrow as the
one-year frame in Brown and LOB, a time-frame of ve to ten years seems
reasonable.
With diachronic corpora (i.e. corpora used to study historical periods of
English), the time-frame for texts is somewhat easier to determine, since the
various historical periods of English are fairly well dened. However, compli-
cations can still arise. For instance, Rissanen (1992: 189) notes that one goal
of a diachronic corpus is to give the student an opportunity to map and com-
pare variant elds or variant paradigms in successive diachronic stages in the
past. To enable this kind of study in the Helsinki Corpus, Rissanen (1992)
remarks that the Old and Early Middle English sections of the corpus were
divided into 100-year subperiods, the Late Middle and Early Modern English
periods into seventy- to eighty-year periods. The lengths of subperiods were
shorter in the later periods of the corpus to include the crucial decades of
the gradual formation of the fteenth-century Chancery standard within one
and the same period (Rissanen 1992: 189). The process that was used to de-
termine the time-frames included in the Helsinki Corpus indicates that it is
important in diachronic corpora not to just cover predetermined historical pe-
riods of English but to think through how signicant events occurring during
those periods can be best covered in the particular diachronic corpus being
created.
2.7 Sampling native vs. non-native speakers of English
If one is setting out to create a corpus of American English, it is best
to include only native speakers of American English in the corpus, since the
United States is a country in which English is predominantly a language spoken
by native speakers of English. On the other hand, if the goal is to create a corpus
of Nigerian English, it makes little sense to include in the corpus native speakers
of English residing in Nigeria, largely because Nigerian English is a variety of
English spoken primarily as a second (or additional) language. As these two
types of English illustrate, the point is not simply whether one obtains texts from
native or non-native speakers but rather that the texts selected for inclusion are
obtained from individuals who accurately reect actual users of the particular
language variety that will make up the corpus.
Prior to selecting texts for inclusion in a corpus, it is crucial to establish
criteria to be used for selecting the individuals whose speech or writing will be
included in the corpus. Since the American component of ICE was to include
native speakers of AmericanEnglish, criteria were establishedpreciselydening
a native speaker of American English. Although language theorists will differ in
their denitions of a native speaker of a language, for the American component
of ICE, a native speaker of American English was dened as someone who
2.7 Sampling native vs. non-native speakers 47
had lived in the United States and spoken American English since adolescence.
Adolescence (1012 years of age) was chosen as the cut-off point, because when
acquisition of a language occurs after this age, foreign accents tend to appear
(one marker of non-native speaker status). In non-native varieties of English, the
level of uency among speakers will vary considerably, and as Schmied (1996:
187) observes, it can be difcult to determine where an interlanguage ends
and educated English starts. Nevertheless, in selecting speakers for inclusion
in a corpus of non-native speech or writing, it is important to dene specic
criteria for selection, such as how many years an individual has used English,
in what contexts they have used it, how much education in English they have
had, and so forth.
To determine whether an individuals speech or writing is appropriate for
inclusion in a corpus (and also to keep track of the sociolinguistic variables
described in section 2.8), one can have individuals contributing texts to a corpus
ll out a biographical form in which they supply the information necessary for
determining whether their native or non-native speaker status meets the criteria
for inclusion in the corpus. For the American component of ICE, individuals
are asked to supply their residence history: the places they have lived over their
lives and the dates they have lived there. If, for instance, an individual has spent
the rst four years of life living outside the United States, it can probably be
safely assumed that the individual is a native speaker of English and came to
the United States at an early enough age to be considered a native speaker of
English. On the other hand, if an individual spent the rst fteen years of life
outside the United States, he or she is probably a non-native speaker of English
and has learned English as a second language. It is generally best not to ask
individuals directly on the biographical form whether they are a native speaker
or not, since they may not understand (in the theoretical sense) precisely what a
native speaker of a language is. If the residence history on the biographical form
is unclear, it is also possible to interviewthe individual afterwards, provided that
he or she can be located; if the individual does not t the criteria for inclusion,
his or her text can be discarded.
Determining native or non-native speaker status from authors of published
writing can be considerably more difcult, since it is often not possible to
locate authors and have them ll out biographical forms. In addition, it can
be misleading to use an individuals name alone to determine native speaker
status, since someone with a non-English-sounding name may have immigrant
parents and nevertheless be a native speaker of English, and many individuals
with English-sounding names may not be native speakers: one of the written
samples of the American component of ICE had to be discarded when the au-
thor of one of the articles called on the telephone and explained that he was
not a native speaker of American English but Australian English. Schmied
(1996: 187) discovered that a major newspaper in Kenya had a chief editor
and training editor who were both Irish and who exerted considerable editorial
inuence over the nal form of articles written by native Kenyan reporters.
This discovery led Schmied (1996) to conclude that there are real questions
about the authenticity of the English used in African newspapers. When deal-
ing with written texts from earlier periods, additional problems can be en-
countered: English texts were often translations from Latin or French, and a
text published in the United States, for instance, might have been written by
a speaker of British English. Ultimately, however, when dealing with written
texts, one simply has to acknowledge that in compiling a corpus of American
English, for instance, there is a chance that the corpus will contain some
non-native speakers, despite ones best efforts to collect only native-speaker
speech.
In spoken dialogues, one may nd out that one or more of the speakers in a
conversation do not meet the criteria for inclusion because they are not native
speakers of the variety being collected. However, this does not necessarily mean
that the text has to be excluded from the corpus, since there is annotation that
can be included in a corpus indicating that certain sections of a sample are
extra-corpus material (cf. section 4.1), that is, material not considered part
of the corpus for purposes of word counts, generating KWIC (key word in
context) concordances, and so forth.
2.8 Controlling for sociolinguistic variables
There are a variety of sociolinguistic variables that will need to be
considered before selecting the speakers and writers whose texts are being con-
sidered for inclusion in a corpus. Some of these variables apply to the collection
of both spoken and written texts; others are more particular to spoken texts. In
general, when selecting individuals whose texts will be included in a corpus,
it is important to consider the implications that their gender, age, and level of
education will have on the ultimate composition of the corpus. For the spoken
parts of a corpus, a number of additional variables need to be considered: the
dialects the individuals speak, the contexts in which they speak, and the rela-
tionships they have with those they are speaking with. The potential inuences
on a corpus that these variables will have are summarized below.
2.8.1 Gender balance
It is relatively easy when collecting speech and writing to keep track
of the number of males and females from whom texts are being collected.
Information on gender can be requested on a biographical form, and in written
texts, one can usually tell the gender of an individual by his or her rst name.
Achieving gender balance in a corpus involves more than simply ensuring
that half the speakers and writers in a corpus are female and half male. In certain
written genres, such as scientic writing, it is often difcult to achieve gender
balance because writers in these genres are predominantly male an unfortunate
2.8 Controlling for sociolinguistic variables 49
reality of modern society. To attempt to collect an equal proportion of writing
from males and females might actually misrepresent the kind of writing found
in these genres. Likewise, in earlier periods, men were more likely to be literate
than women and thus to produce more writing than women. To introduce more
writing by females into a corpus of an earlier period distorts the linguistic reality
of the period. Afurther complication is that much writing, particularly scientic
writing, is co-written, and if males and females collaborate, it will be difcult
to determine precisely whose writing is actually represented in a sample. One
could collect only articles written by a single author, but this again might lead
to a misrepresentation of the type of writing typically found in a genre. Finally,
even though an article may be written by a female or a male, there is no way of
determining how much an editor has intervened in the writing of an article and
thus distorted the effect that the gender of the author has had on the language
used in the article.
In speech, other complications concerning gender arise. Research has shown
that gender plays a crucial role in language usage. For instance, women will
speak differently with other women than they will with men. Consequently, to
adequately reect gender differences in language usage, it is best to include
in a corpus a variety of different types of conversations involving males and
females: women speaking only with other women, men speaking only with
other men, two women speaking with a single man, two women and two men
speaking, and so forth.
To summarize, there is no one way to deal with all of the variables affecting
the gender balance of a corpus. The best the corpus compiler can do is be
aware of the variables, confront them head on, and deal with them as much as
is possible during the construction of a corpus.
2.8.2 Age
Although there are special-purpose corpora such as CHILDES
(cf. section 1.3.8) that contain the language of children, adolescents, and adults,
most corpora contain the speech and writing of adults. The notable exception
to this trend is the British National Corpus. In the spoken section of the cor-
pus containing demographically sampled speech, there is a balanced grouping
of texts representing various age groups, ranging from the youngest group-
ing (014 years) to the oldest grouping (60+) (Composition of the BNC:
http://info.ox.ac.uk/bnc/what/balance.html ). In addition, there were texts in-
cluded in the BNC taken from the Bergen Corpus of London Teenager English
(COLT), which contains the speech of adolescents up to age 16 (Aston and
Burnard 1998: 32; also cf. section 1.3.4). However, the written part of the
BNC contains a sparser selection of texts from younger age groups, largely
because younger individuals simply do not produce the kinds of written texts
(e.g. press reportage or technical reports) represented in the genres typically
included in corpora. This is one reason why corpora have tended to contain
mainly adult language. Another reason is that to collect the speech of children
and adolescents, one often has to obtain the permission not just of the individual
being recorded but of his or her parents as well, a complicating factor in an
already complicated endeavor.
If it is decided that a corpus is to contain only adult language, it then becomes
necessary to determine some kind of age cut-off for adult speech. The ICE
project has set the age of 18 as a cut-off point between adolescent and adult, and
thus has included in the main body of the corpus only the speech and writing of
individuals over the age of 18. However, even though one may decide to include
only adult speech in a corpus, to get a full range of adult speaking styles, it is
desirable to include adults conversing with adolescents or children as well. In
the transcribed corpus itself, markup can be included to set off the speech of
adolescents or children as extra corpus material.
In most corpora attempting to collect texts from individuals of varying ages,
the ages of those in the corpus are collapsed into various age groups. For
instance, the British component of ICE has ve groups: 1825, 2645, 4665,
66+, and a category of co-authored for written texts with multiple authors.
2.8.3 Level of education
It is also important to consider the level of education of those whose
speech or writing is to be included in the corpus. One of the earlier corpora,
the London Corpus, set as its goal the collection of speech and writing from
educated speakers of British English. This goal presupposes not only that one
can dene an educated speaker but that it is desirable to include only the lan-
guage of this group in a corpus. The ICE project dened an educated speaker
very precisely, restricting texts included in all of the components of ICE to
those with at least a high school education. Arguably, such a restriction is arbi-
trary and excludes from the ICE Corpus a signicant range of speakers whose
language is a part of what we consider Modern English. Moreover, restricting
a corpus to educated speech and writing is elitist and seems to imply that only
educated speakers are capable of producing legitimate instances of Modern
English.
However, even though these are all valid criticisms, it is methodologically
valid to restrict a corpus to the speech and writing of whatever level of educa-
tion one wishes, provided that no claims are made that such a corpus is truly
representative of all speakers of a language. Consequently, the more represen-
tative one wishes a corpus to be, the more diverse the educational levels one
will want to include in the corpus. In fact, a corpus compiler may decide to
place no restrictions on the education level of those to be included in a corpus.
However, if this approach is taken, it is important to keep careful records of the
educational levels of those included in the corpus, since research has shown
that language usage varies by educational level, and future users of the corpus
may wish to use it to investigate language usage by level of education.
2.8 Controlling for sociolinguistic variables 51
2.8.4 Dialect variation
The issue of the validity of including only educated speech in a cor-
pus raises a more general consideration, namely the extent to which a corpus
should contain the range of dialects, both social and regional, that exist in any
language.
In many respects, those creating historical corpora have been more success-
ful in representing regional variation than those creating modern-day corpora:
the regional dialect boundaries in Old and Middle English are fairly well estab-
lished, and in the written documents of these periods, variant spellings reecting
differences in pronunciation can be used to posit regional dialect boundaries.
For instance, Rissanen (1992: 1902) is able to describe regional variation in
the distribution of (n)aught (meaning anything, something, or nothing)
in the Helsinki Corpus because in Old and Early Middle English this word
had variant spellings reecting different pronunciations: a spelling with<a>in
West-Saxon (e.g. Old English (n)awuht) and a spelling with<o>in Anglian
or Mercian (e.g. Old English (no)whit). Social variation is more difcult to
document because, as Nevalainen (2000: 40) notes, Early modern peoples
ability to write was conned to the higher social ranks and professional men.
In addition, more detailed information on writers is really not available until the
fteenth century. Therefore, it is unavoidable that sociolinguistic information
in a historical corpus will either be unavailable or skewed towards a particular
social class.
Because writing is now quite standardized, it no longer contains traces of
regional pronunciations. However, even though the modern-day corpus linguist
has access to individuals speaking many different regional and social varieties of
English, it is an enormous undertaking to create a spoken corpus that is balanced
by region and social class. If one considers only American English, a number
of different regional dialects can be identied, and within these major dialect
regions, one can isolate numerous subdialects (e.g. Boston English within the
coastal New England dialect). If social dialects are added to the mix of regional
dialects, even more variation can be found, as a social dialect such as African-
American Vernacular English can be found in all major urban areas of the
United States. In short, there are numerous dialects in the United States, and
to attempt to include representative samplings of each of these dialects in the
spoken part of a corpus is nothing short of a methodological nightmare.
What does one do, then, to ensure that the spoken part of a corpus contains a
balance of different dialects? In selecting speakers for inclusion in the British
National Corpus, twelve dialect regions were identied in Great Britain, and
from these dialect regions, 124 adults of varying social classes were randomly
selected as those whose speech would be included in the corpus (Crowdy 1993:
25960). Unfortunately, the speech of only 124 individuals can hardly be ex-
pected to represent the diversity of social and regional variation in a country the
size of Great Britain. Part of the failure of modern-day corpora to adequately
represent regional and social variation is that creators of these corpora have had
unrealistic expectations. As Chafe, Du Bois, and Thompson (1991: 69) note,
the thought of a corpus of American English to some individuals conjures
up images of a body of data that would document the full sweep of the lan-
guage, encompassing dialectal diversity across regions, social classes, ethnic
groups . . .[enabling] massive correlations of social attributes with linguistic
features. But to enable studies of this magnitude, the corpus creator would
have to have access to resources far beyond those that are currently available
resources that wouldenable the speechof thousands of individuals tobe recorded
and then transcribed.
Because it is not logistically feasible in large countries such as the United
States or Great Britain to create corpora that are balanced by region and social
class, it is more protable for corpus linguists interested in studying social
and regional variation to devote their energies to the creation of corpora that
focus on smaller dialect regions. Tagliamonte (1998) and Tagliamonte and
Lawrence (2000: 3256), for instance, contain linguistic discussions based on
a 1.5-million-word corpus of York English that has been subjected to extensive
analysis and that has yielded valuable information on dialect patterns (both
social and regional) particular to this region of England. Kirk (1992) has created
the Northern Ireland Transcribed Corpus of Speech containing transcriptions
of interviews of speakers of Hiberno English. Once a number of more focused
corpora like these are created, they can be compared with one another and valid
comparisons of larger dialect areas can then be conducted.
2.8.5 Social contexts and social relationships
Speech takes place in many different social contexts and among speak-
ers between whom many different social relationships exist. When we work,
for instance, our conversations take place in a specic and very common social
context the workplace and among speakers of varying types: equals (e.g.
co-workers), between whom a balance of power exists, and disparates (e.g. an
employer and an employee), between whom an imbalance of power exists. Be-
cause the employer has more power, he or she is considered a superordinate
in contrast to the employee, who would be considered a subordinate. At home
(another social context), other social relationships exist: a mother and her child
are not simply disparates but intimates as well.
There is a vast amount of research that has documented that both the social
context in which speech occurs and the social relationships between speakers
have a tremendous inuence on the structure of speech. As Biber and Burges
(2000) note, to study the inuence of gender on speech, one needs to consider
not just the gender of the individual speaking but the gender of the individual(s)
to whom a person is speaking. Because spontaneous dialogues will constitute a
large section of any corpus containing speech, it is important to make provisions
for collecting these dialogues in as many different social contexts as possible.
2.9 Conclusions 53
The LondonLund Corpus contains an extensive collection of spoken texts rep-
resenting various social relationships between speakers: spontaneous conversa-
tions or discussions between equals and between disparates; radio discussions
and conversations between equals; interviews and conversations between dis-
parates; and telephone conversations between equals and between disparates
(Greenbaum and Svartvik 1990: 2040).
The American component of ICE contains spontaneous conversations taking
place in many different social contexts: there are recordings of family mem-
bers conversing over dinner, friends engaging in informal conversation as they
drive in a car, co-workers speaking at work about work-related matters, teach-
ers and their students discussing class work at a university, individuals talk-
ing over the phone, and so forth. The Michigan Corpus of Academic Spoken
English (MICASE) collected academic speech in just about every conceivable
academic context, from lectures given by professors to students conversing in
study groups, to ensure that the ultimate corpus created represented the broad
range of speech contexts in which academic speech occurs (Simpson, Lucka,
and Ovens 2000: 48). In short, the more diversity one adds to the social contexts
in which language takes place, the more assured one can be that the full range
of contexts will be covered.
2.9 Conclusions
To create a valid and representative corpus, it is important, as this
chapter has shown, to plan the construction carefully before the collection of
data even begins. This process is guided by the ultimate use of the corpus. If one
is planning to create a multi-purpose corpus, for instance, it will be important
to consider the types of genres to be included in the corpus; the length not
just of the corpus but of the samples to be included in it; the proportion of
speech vs. writing that will be included; the educational level, gender, and
dialect backgrounds of speakers and writers included in the corpus; and the
types of contexts from which samples will be taken. However, because it is
virtually impossible for the creators of corpora to anticipate what their corpora
will ultimately be used for, it is also the responsibility of the corpus user to
make sure that the corpus he or she plans to conduct a linguistic analysis of
is valid for the particular analysis being conducted. This shared responsibility
will ensure that corpora become the most effective tools possible for linguistic
research.
Study questions
1. Is a lengthier corpus necessarily better than a shorter corpus? What kinds
of linguistic features can be studied in a lengthier corpus (such as the Bank
of English Corpus) that cannot be studied in a shorter corpus (such as the
Brown Corpus)?
2. Would a corpus containing ten 2,000-word samples of writing from the
humanities be lengthy enough to provide a representative sampling of hu-
manistic writing?
3. Why is achieving gender balance in a corpus more than a matter of simply
selecting texts from an equal number of males and females?
4. Why can FLOB (the Freiburg LondonOsloBergen Corpus) and FROWN
(the FreiburgBrown Corpus) be considered synchronic as well as diachronic
corpora?
5. If the goal is to create a corpus to study regional or social variation, why does
it make sense to target a very specialized regional or social dialect rather than
a larger one, such as British English or American English?
3 Collecting and computerizing data
Once the basic outlines of a corpus are determined, it is time to begin
the actual creation of the corpus. This is a three-part process, involving the
collection, computerization, and annotation of data. This chapter will focus on
the rst two parts of this process how to collect and computerize data. The
next chapter will focus in detail on the last part of the process: the annotation
of a corpus once it has been encoded into computer-readable form.
Collecting data involves recording speech, gathering written texts, obtaining
permission from speakers and writers to use their texts, and keeping careful
records about the texts collected and the individuals from whom they were
obtained. How these collected data are computerized depends upon whether
the data are spoken or written. Recordings of speech need to be manually
transcribed using either a special cassette tape recorder that can automatically
replay segments of a recording, or software that can do the equivalent with
a sample of speech that has been converted into digital form. Written texts
that are not available in electronic form can be computerized with an optical
scanner and accompanying OCR (optical character recognition) software, or
(less desirably) they can be retyped manually.
Even though the process of collecting, computerizing, and annotating texts
will be discussed as separate stages in this and the next chapter, in many senses
the stages are closely connected: after a conversation is recorded, for instance,
it may prove more efcient to transcribe it immediately, since whoever made
the recording will be available to answer questions about it and to aid in its
transcription. If a text is collected, and saved for later computerization and an-
notation, the individual who collected it may not be around to answer questions,
and information about the text may consequently be lost. Of course, logistical
constraints may necessitate collecting texts at one stage and computerizing and
annotating them at a later stage, in which case it is crucial that as much infor-
mation about the text as possible is recorded initially so that, at a later stage,
those working with the text will be able to recover this information easily.
3.1 General considerations
As chapter 2 demonstrated, before the actual data for a corpus are
collected, it is important to carefully plan exactly what will be included in the
55
corpus: the kinds and amounts of speech and/or writing, for instance, as well the
range of individuals whose speech and writing will become part of the corpus.
Once these determinations are made, the corpus compiler can begin to collect
the actual speech and writing to be included in the corpus. As was stressed at
the start of chapter 2, however, it is important not to become too rigidly invested
in the initial corpus design, since numerous obstacles and complications will be
encountered while collecting data that may require changes in the initial corpus
design: it may not be possible, for instance, to obtain recordings for all of the
genres originally planned for inclusion in the corpus, or copyright restrictions
might make it difcult to obtain certain kinds of writing. In these cases, changes
are natural and inevitable and, if they are carefully made, the integrity of the
corpus will not be compromised.
The ICE Project provides a number of examples of logistical realities that
forced changes in the initial corpus design. For instance, after the project began,
it was discovered that not all of the regional groups involved in the project would
be able to collect all of the text categories originally planned for inclusion in
the corpus. However, rather than drop these regional groups from the overall
project (or radically change the types of texts to be collected), it was decided
that the groups should produce a shorter corpus and collect the categories that
they were able to collect. Consequently, while a given regional component
of the ICE Corpus may not contain telephone conversations, for instance, it
will contain the required amount of some other part of the corpus (such as
spoken dialogues) that can be compared with other regional components of the
ICE Corpus containing comparable texts. Although this change means that the
overall ICE Corpus will not be as complete as originally envisioned, it will
nevertheless allow valid comparisons of the particular genres included in it.
Because collecting speech and writing involves quite different methodolo-
gies, the collection of each will be considered separately.
3.2 Collecting samples of speech
Speech is the primary mode of human communication. As a result, all
kinds of speech have evolved: there are not only spontaneous multi-party dia-
logues but scripted and unscripted monologues, radio and television interviews,
telephone conversations, class lectures, and so forth. Given the wealth of speech
that exists, as well as the logistical difculties involved in recording it, collect-
ing data for the spoken part of a corpus is much more difcult and involved than
collecting written samples. Consequently, there are numerous methodological
considerations that must be addressed in the collection of speech.
In collecting any kind of speech, the central concern is obtaining speech that
is natural. This is a particularly important issue when gathering spontaneous
multi-party dialogues, such as informal talks between two or more individuals.
If multi-party dialogues are not carefully collected, the result can be a series of
3.2 Collecting samples of speech 57
recordings containing very stilted and unnatural speech. Collecting natural
multi-party dialogues involves more than simply turning on a tape recorder and
having people converse. As anyone who has collected language data knows,
as soon as you turn on a tape recorder and tell people to speak, you have to
deal with the common scientic problem known as the observers paradox:
by observing something, you change its natural behavior. And this is especially
true with speech because many individuals are quite insecure about the way
that they speak, and when they know that their speech is being monitored, they
may adjust their speaking style and attempt to speak what they perceive to be
correct English.
To avoid this problem, those creating earlier corpora, such as the London
Lund Corpus, recorded people surreptitiously, and only after recordings were
secretly made were individuals informed that they had been recorded. While
it may have been acceptable and legal back in the1950s and 1960s to record
individuals without their knowledge, now such recordings are illegal in some
locales and considered unethical by most linguists.
1
It is therefore imperative
that anyone recording speech not only informs individuals that they are being
recorded but obtains written permission fromthe individuals to use their speech
in the corpus. This can be accomplished by having individuals sign a release
form prior to being recorded.
2
Since it is not possible to include surreptitious speech in a corpus, does this
mean that non-surreptitiously gathered speech is not natural? This is an open
question, since it is not possible to answer it with any denite certainty: so little
speech has been collected and included in corpora that it is not really possible
to know empirically precisely what natural speech is really like. However, there
are ways to increase the probability that the speech included in a corpus will
be natural and realistic.
First, before individuals are recorded, they should be given in written form
a brief description of the project they are participating in. In this description,
the purpose of the project should be described, and it should be stressed that
speech is being collected for descriptive linguistic research, not to determine
whether those being recorded are speaking correct or incorrect English.
In a sense, these individuals need to be given a brief introduction to a central
tenet of modern linguistics: that no instance of speech is linguistically better
or worse than any other instance of speech, and that it is possible to study any
kind of speech objectively, whether it is considered standard or non-standard.
Asecond way to enhance naturalness is to record as lengthy a conversation as
possible so that when the conversation is transcribed, the transcriber can select
1
See Murray and Ross-Murray (1992 and 1996) for a discussion of the ethical and legal issues
surrounding surreptitious recordings.
2
With broadcast recordings, it is not usually necessary to obtain the permission of individuals to use
their recordings but rather to contact the broadcast station, which owns rights to the program, and
receive the stations written authorization to use the program. How to obtain copyright clearance
for texts is discussed in section 3.3.
the most natural segment of speech from a much lengthier speech sample. How
much speech needs to be recorded is determined by the type of speech being
recorded and the length of segments that will be included in the corpus. If
the goal is to create a corpus consisting of 2,000-word samples, for instance,
then between ten and twenty minutes of speech is necessary to supply a single
sample. The type of speech being collected will determine the amount of speech
that is necessary. Spontaneous dialogues tend to consist of shorter speaker turns
and many pauses and hesitations. Consequently, such conversations require a
longer span of speech to reach the 2,000-word length requirement. On the other
hand, monologues (especially scripted monologues) contain far fewer pauses
and hesitations. Thus, less speech is needed to reach 2,000 words.
Work with ICE-USA has shown that it is most desirable for recordings of
speech to be thirty minutes or longer. This length of conversation increases
the probability that a natural and coherent 2,000-word speech sample can be
extracted from this longer sample. The initial part of the conversation can be
discarded, since people are sometimes nervous and hesitant upon rst being
recorded but after a while become less self-conscious and start speaking more
naturally.
3
Moreover, a lengthier segment allows the corpus compiler to select a
more coherent and unied 2,000-word segment of speech to ultimately include
in the corpus.
When it comes time to actually make recordings, whoever is making the
recordings needs to follow a few basic principles to ensure the most natural
recordings possible. Probably the least desirable way to make a recording is to
have the research assistant sitting silently nearby with microphone in hand while
the people being recorded converse. This all but ensures that the individuals
being recorded will constantly be reminded that they are part of a linguistic
experiment. As far as possible, those making recordings should try to record
individuals in the actual environments in which they typically speak. Samples
included in ICE-USAinclude individuals speaking while eating a meal, working
in ofces, riding in a car, participating in class discussions, and so forth. Because
the goal is to record people in natural speaking environments, it is best for the
research assistant not to be present during recordings. He or she can simply set
up the recording equipment, turn on the recorder, and leave. Alternatively, the
people being recorded can be loaned the recording equipment and taught to set
it up and turn on the recorder themselves.
Because people are being recorded in their natural speaking environments,
very often the quality of recordings is not as high as one would desire. If people
are recorded while driving in a car, for instance, the recording will contain trafc
noise. However, this noise can sometimes be edited out; in a recording of this
type included in ICE-USA, the conversation itself was audible, even though
there was considerable noise in the background. If high-quality recordings
3
Asimilar conclusion was reached in recordings made for the British National Corpus. See Crowdy
(1993: 2612).
3.2 Collecting samples of speech 59
are desired, individuals can be recorded in a recording studio, as was done
in collecting speech samples for the Map Task Corpus, a corpus of conversa-
tions between individuals giving directions to various locations (see Thompson,
Anderson, and Bader 1995 for more details). However, in speech collected
this way, while high-quality recordings will be obtained, the naturalness of the
speech collected will be compromised. Therefore, it is probably wise to sacrice
recording quality in favor of natural speech.
One of the more elaborate and interesting ways of collecting speech was used
to obtain speech samples for the British National Corpus (BNC).
4
Individuals
participating in the project were given portable tape recorders and an adequate
supply of cassettes, and were instructed to record all of the conversations they
had for periods ranging from two to seven days (Crowdy 1993: 260). To keep
track of the conversations they had, participants lled out a log book, indicating
when and where the recordings occurred as well as who was recorded. This
method of collection habituates participants to the process of being recorded
and, additionally, ensures that a substantial amount of speech is collected.
While collecting natural speech is a key issue when recording multi-party
dialogues, it is less of an issue with other types of speech. For instance, those
participating in radio and television broadcasts will undoubtedly be conscious
of the way they are speaking, and therefore may heavily monitor what they say.
However, heavily monitored speech is natural speech in this context, so this is
precisely the kind of speech one wants to gather. Likewise, with other types of
speech, such as public speeches (especially if they are scripted), heavily edited
speech is the norm.
In recording any kind of spoken English, the type of tape recorder to use is
an important issue to consider. In the past, most corpus compilers used analog
recorders and cassette tapes, since such recorders were small and unobtrusive
and cassette tapes were inexpensive. However, with the rise of digital tech-
nology, there are now widely available a variety of tape recorders that can
make high-quality digital recordings. Either an analog or digital recorder will
yield satisfactory recordings of multi-party dialogues: in pilot studies for the
British National Corpus, for instance, Crowdy (1993: 261) reports that even
though digital recorders produce high-quality recordings under good record-
ing conditions they were not signicantly better than analogue recorders under
the recording conditions likely to be encountered in this project [the British
National Corpus].
Despite this fact, however, there are distinct advantages to using digital rather
than analog recorders. Analog cassettes age over time, and even if they are stored
properly, they will eventually deteriorate in quality, rendering the recordings
that were made unusable. Digital tapes, on the other hand, do not degrade as
quickly. Although cassette tapes come in various lengths, they cannot record
4
This method of collection was also used to collect samples for the Bergen Corpus of London
Teenage English (COLT). See Haselrud and Stenstr om (1995) for details of this corpus.
more than sixty minutes of continuous speech on a single side (unless a special
recorder is purchased that automatically begins recording on the second side of
a cassette). If a fairly lengthy recording is being made, and a conversation has
to be stopped to turn over the cassette, the naturalness of the recording can be
seriously undermined. Digital cassettes, in contrast, can record up to four hours
of continuous speech.
When working with spoken recordings, it is often desirable to work with
digital recordings on a computer, for instance, to edit out unwanted background
noise or to prepare the recording for inclusion on a CDfor distribution as part of
the corpus. Although the speech on an analog cassette can be digitized, a digital
cassette can be transferred to disk directly with a minimal loss of recording
quality. Presently, digital tape recorders and cassettes are more expensive than
analog recorders and cassettes. However, current technology is going digital,
and as it does, the cost of digital recorders and tapes will continue to decline.
Therefore, unless cost is an important consideration, there is little reason for
corpus compilers to use analog recorders anymore.
In addition to considering the type of tape recorder to use, it is equally im-
portant to consider the quality and type of microphone to make recordings
with. An inexpensive microphone will produce recordings that are tinny even
if a good tape recorder is used. It is therefore advisable to invest resources
in good microphones, and to obtain microphones that are appropriate for the
kinds of recordings being made. To record a single individual, it is quite ac-
ceptable to use a traditional uni-directional microphone, one that records an
individual speaking directly into it. For larger groups, however, it is better to
use omni-directional microphones that can record individuals sitting at various
angles from the microphone. Lavaliere microphones, which are worn around
the neck, are useful for recording individuals who might be moving around
when they speak, as people lecturing to classes or giving speeches often do.
Wireless microphones are appropriate in recording situations of this type too,
and avoid the problem of the speaker being constrained by the length of the
cord attaching the microphone and recorder. There are also extra-sensitive mi-
crophones that can be used for recording individuals who are not close to the
microphone, as is the case in a class discussion, where those being recorded
are spread out all over a room and might not be close enough to a traditional
microphone for an audible recording to be made. For recording telephone con-
versations, special adapters can be purchased that record directly off the tele-
phone. High-quality microphones can be fairly expensive, but they are worth
the investment.
Even with the best recording equipment, however, assembling a large num-
ber of recordings suitable for inclusion in a corpus is a time-consuming and
sometimes frustrating process: for every ten recordings made, it may turn out
that only four or ve are usable. For instance, one recording made for ICE-USA
contains two individuals who converse quite naturally for periods of ten minutes
after which one of the individuals on more than one occasion remarks, Is the
3.3 Collecting samples of writing 61
recorder still working? Remarks such as this illustrate that no matter how hard
one tries, it is impossible to make many individuals forget that their speech is
being monitored and recorded. Other problems include excessive background
noise that makes all or part of a conversation inaudible, or people who are given
a recorder to record their speech, and then operate it improperly, in some cases
not recording any of the conversation they set out to record. Those compiling
spoken corpora should therefore expect to gather much more speech than they
will actually use to compensate for all the recordings they make that contain
imperfections preventing their use in the ultimate corpus being created.
While most of the recordings for a corpus will be obtained by recording
individuals with microphones, many corpora will contain sections of broadcast
speech. This type of speech is best recorded not with a microphone but directly
from a radio or television by running a cord from the audio output plug on the
radio or television to the audio input plug on the tape recorder. Since many
televisions do not have audio output plugs, if the television is connected to a
video cassette recorder or DVD player, a recording can be made by running a
line from the audio output plug on the recorder or player.
In sending individuals out to make recordings, it is important to realize that
making good recordings takes practice, and that as more recordings are made,
the quality will increase as mastering the art of recording improves.
3.3 Collecting samples of writing
Although collecting samples of writing is considerably less compli-
cated than collecting samples of speech, one additional and sometimes quite
frustrating complication is encountered when collecting writing: copyright re-
strictions. Under the fair use provisions of current US copyright laws, it is
acceptable to copy either all or part of a copyrighted text for purposes of non-
prot academic research. However, fair use of copyrighted material is exceeded
if a 2,000-word excerpt of a copyrighted article is computerized, for instance,
and then distributed as part of a corpus, even if it will be used for non-prot
academic research only. Therefore, the corpus compiler will need to get the
written permission of any author whose writing is to be included in a corpus.
This includes not just printed texts in magazines, journals, and books, which
will always be marked as copyrighted, but letters, online articles, and other un-
published works as well, which are automatically assigned copyright. In fact,
the only written materials in the United States that are not copyrighted are any
US government publications and works that are beyond seventy years from the
authors date of death. Although many countries have signed international copy-
right agreements, there is variation in practice. Therefore, the corpus compiler
needs to consider not just the laws of the country in which a corpus is being
assembled but the laws of the country in which the written documents were
created and copyrighted.
When making recordings of speech, one obtains permission from speakers
before they are recorded. Therefore, there is no time wasted making recordings
and then nding out that the speakers will not give permission for their record-
ings to be used in the corpus being created. However, with written texts, this
luxury is not available: a written text may be gathered for possible inclusion in
a corpus, and its author (or publisher) may not give permission for its use, or
(as was a common experience gathering written texts for inclusion in the
American component or ICE) a letter might be mailed requesting permission
but no reply to the letter is ever received. In cases like this, current copyright
law allows for the use of such materials, provided that it can be established
that a reasonable effort was made to obtain permission. But a cautious corpus
compiler might wish to avoid any future legal problems from authors of texts
from whom no written permission was obtained. For this reason, no written
text was included in ICE-USA without the written permission of its author (or
authors).
Because of the difculties in obtaining permission to use copyrighted mate-
rials, most corpus compilers have found themselves collecting far more written
material thantheyare able toobtainpermissiontouse: for ICE-USA, permission
has been obtained to use only about 25 percent of the written texts initially con-
sidered for inclusion in the corpus. Moreover, some authors and publishers will
ask for money to use their material: one publisher requested half an American
dollar per word for a 2,000-word sample of ction considered for inclusion in
ICE-USA! Therefore, if a corpus is to be used for non-prot academic research
only, this should be clearly stated in any letter of inquiry requesting permission
to use copyrighted material and many publishers will waive fees. However, if
there will be any commercial use of the corpus, special arrangements will have
to be made both with publishers supplying copyrighted texts and those making
use of the corpus.
Insome cases, obtainingpermissiontouse copyrightedmaterial has provenso
onerous that some corpora can only be used on-site where they were created. For
instance, because permission could not be obtained to distribute the English
Norwegian Parallel Corpus to any interested researcher, it is only available to
researchers able to travel to the University of Oslo, where it was created and
is now housed. Ask anyone who has created a corpus containing copyrighted
material, and they will state that their greatest frustration in compiling the
corpus was dealing with copyright complications.
In gathering written texts for inclusion in a corpus, the corpus compiler will
undoubtedly have a predetermined number of texts to collect within a range of
given genres: twenty 2,000-word samples of ction, for instance, or ten 2,000-
word samples of learned humanistic writing. However, because there is so much
writing available, it is sometimes difcult to determine precisely where to be-
gin to locate texts. Since most corpora are restricted to a certain time-frame,
this frame will of course narrow the range of texts, but even within this time-
frame, there is an entire universe of writing. Because whatever writing collected
3.3 Collecting samples of writing 63
will ultimately need to be computerized, a good place to start is journals and
magazines published on the World Wide Web: an increasing number of period-
icals are now available online, and once permission is obtained to use an article
in a periodical, the article can be easily downloaded in computerized form.
Of course, including texts of this type in a corpus raises a methodological
issue: are electronic texts essentially the same as traditionally published written
texts? With newspapers, this is less of an issue, since many newspapers pub-
lish online versions of articles that are identical to those included in printed
editions.
5
But many written texts exist exclusively online, and if a corpus con-
tains, for instance, ve electronic texts in the genre of scientic writing and ve
traditionally published texts in the same genre, do we have one genre here or
two subgenres: electronically published scientic writing and traditionally pub-
lished scientic writing? And since anyone can publish an article on the World
Wide Web, is an electronic text published on an individuals personal home page
any different from a text that has gone through the editorial process and been
published in a journal, magazine, or book? These are important questions, and
in answering them, some might conclude that electronically published texts are
different enough from traditionally published works that the two types of texts
should not be mixed together in a single genre within a corpus. But there is
no denying that the trend towards electronic publishing is increasing, and even
though traditional periodicals such as journals and magazines will probably
never disappear, including at least some electronically published writing in a
synchronic corpus accurately reects the current state of writing. A further ad-
vantage of electronically published writing is that it allows the corpus compiler
to avoid the tedious job of computerizing printed texts, either by typing them in
by hand or by scanning them with optical scanners, which ease the process of
computerizing written texts but which still produce an electronic copy of the text
albeit with a number of scanning errors which have to be manually corrected.
If a corpus compiler decides toinclude electronic texts ina corpus, a goodway
tondthemis touse anindexedsearchengine suchas Yahoo(http://www.yahoo.
com), that is, a search engine that organizes information on the World Wide
Web by putting it into categories. The opening page of Yahoo is divided into
a series of subject areas, such as news media, science (i.e. natural science),
social science, and arts and humanities all genres commonly found in cor-
pora. Selecting science yields a number of choices, including journals,
which in turn provides links to numerous online journals in areas such as zo-
ology (http://dir.yahoo.com/Science/Biology/Zoology/Journals/) and biology
(http://dir.yahoo.com/Science/Biology/Journals/). Each of the journals offers a
number of potential texts for inclusion in a corpus.
Alternatively, a search engine such as AltaVista (http://www.altavista.com/)
that searches for strings on the World Wide Web can be used. A search on
5
Even though most newspapers publish both printed and online versions of each issue, it is worth
checking the original printed edition with the online edition just to make sure that the two versions
match.
AltaVista of the key words ction and magazine produced thousands of
sites, and among these sites were many ction magazines containing full-text
pieces of ction. So many online texts exist that it is now possible to create an
entirely online corpus; that is, a corpus consisting exclusively of links to texts
that could be subjected to the same kinds of linguistic analyses now possible
only with corpora made available on disk (cf. Davies 2001 for an example).
Even though a corpus compiler may not want to include purely electronic
publications in a corpus, there are a number of traditionally printed texts that are
also available in electronic form. For instance, some newspapers have released
CD-ROMs containing previously published articles. It is also increasingly com-
mon, as was noted above, to nd newspapers with websites on which they pub-
lish all or some of the articles in their daily editions. If the paper does not have
its own website, very often its articles will be in the LEXIS/NEXUS database,
which many academic libraries subscribe to and fromwhich the articles in many
newspapers (and other written texts too) can be downloaded.
There are various ways to obtain printed texts. For instance, learned writing
in the areas of the humanities, natural sciences, social sciences, and technol-
ogy can be collected by randomly selecting texts from the current periodicals
sections of libraries. Other types of texts can be obtained in this manner as
well. Various periodical indexes and online catalogues can also be consulted
for references to books and articles. When gathering printed texts, it is impor-
tant that as good a photocopy as possible of each text is made so that if the text
is eventually included in the corpus, it can be scanned as accurately as possible:
poor photocopies tend to produce more numerous scanning errors.
3.4 Keeping records of texts gathered
As written texts are collected and spoken texts are recorded, it is imper-
ative that accurate records are kept about the texts and the writers and speakers
that created them. For ICE-USA, research assistants lled out a written check-
list supplying specic information for each spoken and written text that was
collected.
First of all, each text was assigned a number that designated a specic genre
in the corpus in which the sample might be included. For instance, a text num-
bered S1A-001 would be the rst sample considered for inclusion in the genre of
direct conversations; a text numbered S1B-001, on the other hand, would be
the rst sample considered for inclusion in the genre of classroom lectures.
A numbering system of this type (described in detail in Greenbaum 1996b:
60114) allows the corpus compiler to keep easy record of where a text belongs
in a corpus and how many samples have been collected for that part of the
corpus. After a text was numbered, it was given a short name providing de-
scriptive information about the sample. In ICE-Great Britain, sample S1A-001
(a spontaneous dialogue) was named Instructor and dance student, Middlesex
3.4 Keeping records of texts gathered 65
Polytechnic and sample S1B-001 (a class lecture) was entitled Jewish and
Hebrew Studies, 3rd year, UCL. The name given to a text sample is short and
mnemonic and gives the corpus compiler (and future users of the corpus) an
idea of the type of text that the sample contains.
The remaining information recorded about texts depended very much on
the type of text that was being collected. For each spoken text, a record was
kept of when the text was recorded, where the recording took place, who was
recorded, who did the recording, and how long the recording was. For each
person recorded, a short excerpt of something they said near the start of the
recording was written down so that whoever transcribed the conversation would
be able to match the speech being transcribed with the speaker. It can be very
difcult for a transcriber to do this if he or she has only the tape recording to
work with and has to gure out who is speaking. For speech samples recorded
fromtelevision or radio, additional information was written down, such as what
station the recording was made from, where the station is located, and who
should be contacted to obtain written permission to use the sample. For written
texts, a complete bibliographical citation for the text was recorded along with
the address of the publisher or editorial ofce from which permission to use the
written text could be obtained.
In addition to keeping records of the texts that are recorded, it is equally
important to obtain ethnographic information from individuals contributing
either a sample of speech or a written text. The particular information col-
lected will very much depend on the kind of corpus being created and the
variables future users of the corpus will want to investigate. Because ICE-USA
is a multi-purpose corpus, only fairly general ethnographic information was
obtained from contributors: their age, gender, occupation, and a listing of the
places where they had lived over the course of their lives. Other corpora have
kept different information on individuals, relevant to the particular corpus being
created. The Michigan Corpus of Academic Spoken English (MICASE) col-
lected samples of spoken language in an academic context. Therefore, not just
the age and gender of speakers in the corpus were recorded but their academic
discipline (e.g. humanities and arts, biological and health sciences), academic
level (e.g. junior undergraduates, senior faculty), native-speaker status, and rst
language (Powell and Simpson 2001: 35).
Ethnographic information is important because those using the ultimate cor-
pus that is created might wish to investigate whether gender, for instance, affects
conversational style, or whether younger individuals speakdifferentlythanolder
individuals. It is important for these researchers to be able to associate variables
such as these with specic instances of speech. While it is relatively easy to ob-
tain ethnographic information fromindividuals being recorded (they can simply
ll out the form when they sign the permission form), tracking down writers
and speakers on radio or television shows can be very difcult. Therefore, it is
unrealistic to expect that ethnographic information will be available for every
writer and speaker in a corpus. And indeed, many current corpora, such as the
British National Corpus and ICE, do not contain information on many speakers
and writers.
After all of the above information is collected, it can be very useful to enter it
into a database, which will allow the progress of the corpus to be tracked. This
database can contain not just information taken fromthe forms described above
but other information as well, such as whether the text has been computerized
yet, whether it has been proofread, and so forth. Creating a corpus is a huge
undertaking, and after texts are collected, it is very easy to le them away and
forget about them. It is therefore crucial to the success of any corpus undertaking
that accurate information is kept about each text to be considered for inclusion
in the corpus.
3.5 Computerizing data
Computerizing spoken and written texts for inclusion in a corpus is
a very labor-intensive part of creating a corpus. Transcribing speech is an ex-
tremely lengthy process, requiring the transcriber to listen to the same segments
of speech over and over again until an accurate transcription is achieved. Even
though printed texts can be scanned into a computerized format, there will
always be scanning errors so many errors, in fact, that some corpus compil-
ers have abandoned scanners altogether and have found that retyping texts is
a quicker and more efcient way to computerize them, particularly if it is not
possible to obtain a clear printed copy of the text to be scanned (which is es-
sential for accurate scanning). For both spoken and written texts, proofreading
a large corpus of texts after they have been computerized is painstaking work,
and no matter how carefully it is done, errors will remain.
There are a number of general considerations to bear in mind when beginning
the process of computerizing both spoken and written texts. First, because texts
to be included in a corpus will be edited with a word-processing program, it
may be tempting to save computerized texts in a le format used by a word-
processing program (such as les with the extension .doc in Microsoft Word).
However, it is best from the onset to save texts in ASCII (or text) format, since
this is the standard format used for texts included in corpora, and to use a simple
text editor to work with texts rather than a standard word-processing program:
even though such programs claim to create pure ASCII les, very often they
insert formatting commands that can be problematic when a given corpus is
analyzed by a tagging or concordancing program.
The ASCII le format has both advantages and disadvantages. The main
advantage is that ASCII is a universally recognized text format, one that can be
used with any word-processing program and the numerous software programs
designed to work on corpora, such as taggers and parsers (cf. sections 4.3
and 4.5) and concordancers (cf. section 5.3.2). The disadvantage is that be-
cause ASCII has a fairly limited set of characters, many characters and symbols
3.5 Computerizing data 67
cannot be represented in it. The creators of the Helsinki Corpus had to create
a series of special symbols to represent characters in earlier periods of English
that are not part of the ASCII character set: the OldEnglishwordt (that), for
instance, is encoded in the corpus as +t+at with the symbol +t correspond-
ing to the Old English thorn and the symbol +a to Old English ash
(Kyt o 1996). This system of symbols is specic to the Helsinki Corpus. There
exists a successor to ASCII, Unicode, containing an expanded character set able
to represent all of the characters found in the languages of the world. Unfortu-
nately, this standard has not been around long enough to have replaced ASCII in
the corpus linguistics community. As a result, most corpus compilers use some
kind of systemof annotation to represent non-ASCII characters (cf. section 4.1),
though with the emergence of the Text Encoding Initiative (TEI), this can be
done more systematically so that each corpus project does not have to create a
unique system of annotation (different from somebody elses system) to repre-
sent non-ASCII characters. The TEI has developed a formalism, called a writ-
ing systemdeclaration, for identifying the specic character set used in a given
electronic document (see Overall Structure of Writing System Declaration
in Sperberg-McQueen and Burnard 1994a, http://etext.lib.virginia.edu/bin/tei-
tocs-p3?div=DIV2&id=WDOV). Whichever system is used, because informa-
tion will be lost when a text is encoded in ASCII format, it is important that
a printed copy of every written text is kept on le, so that when the text
is annotated, the appropriate markup can be added to it to mark non-ASCII
characters.
When creating a corpus, it is easiest to save individual texts in separate les
stored in directories that reect the hierarchical structure of the corpus. This
does not commit one to distributing a corpus in this format: the ICAME CD-
ROM(2nd edn.) allows users to work with an entire corpus saved in a single le.
But organizing a corpus into a series of directories and subdirectories makes
working with the corpus much easier, and allows the corpus compiler to keep
track of the progress being made on corpus as it is being created. Figure 3.1
contains a partial directory structure for ICE-USA.
ICE-USA consists of two main directories one containing all the spoken
texts included in the corpus, the other all the written texts. These two directories,
in turn, are divided into a series of subdirectories containing the main types of
speech and writing that were collected: the spoken part into monologues and
dialogues, the written part into printed and non-printed material.
The remainder of gure 3.1 contains additional subdirectories for the specic
types of dialogues that were gathered (for reasons of space, the other subdirec-
tories are not included): business transactions, classroom discussions, political
debates, and so forth. And within each of these directories would be the indi-
vidual texts collected to ll the category. Since an individual text goes through
many stages of analysis, various versions of a text are kept in separate directo-
ries. Texts that are in the process of being transcribed or scanned are kept in the
draft directory. Once a draft version of a text has been fully annotated with
Spoken Written
Dialogues Monologues Printed Non-Printed
Business Classroom Political Spontaneous Broadcast Broadcast Legal Phone
transactions discussions
Draft
S1B-071d
S1B-072d
etc.
Lexical version
S1B-071L
S1B-072L
etc.
Proofread version (I)
S1B-071p1
S1B-072p1
etc.
Proofread version (II)
S1B-071p2
S1B-072p2
etc.
discussions debates conversations interviews cr. ex. conversations
Figure 3.1 Partial directory structure for American component of ICE
structural markup (cf. section 4.1), it is moved into the lexical directory
to indicate that the text is ready for inclusion in a lexical version of ICE-USA
(i.e. a version of ICE-USA containing the text and structural annotation). The
text then undergoes the rst round of proofreading after which it is placed in the
directory Proofread version (I). A second round of proofreading is done after
completion of the entire corpus so that any editorial changes made during the
creation of the corpus can be incorporated into the nal version of the corpus,
which is placed in the directory Proofread version (II).
While a text is being worked on at a particular stage of analysis, it receives
an additional le extension to indicate that work on the text is in progress. For
instance, while a draft version of a text in the category of business transactions
is being created, the text is saved as S1B-071di, the i indicating that work
on the text is incomplete. As a particular text is being worked on, a log is
maintained that notes what work was done and what work needs to be done. At
each stage of analysis, to avoid duplication of work, it is most efcient to have
a single person work on a text; at the proofreading stage, it is best to have the
text proofread by someone not involved with any prior version of the text, since
he or she will bring a fresh perspective to the text.
Finally, although inserting structural markup into a text is separate from
the process of actually computerizing the text, there are many instances where
markup can be inserted while texts are being computerized. For instance, in
3.6 Computerizing speech 69
transcribing spontaneous conversations, the transcriber will encounter numer-
ous instances of overlapping speech individuals speaking at the same time.
The segments of speech that overlap need to be marked so that the eventual
user of the corpus knows which parts overlap in the event that he or she wishes
to study overlapping speech. If annotating overlapping segments of speech is
done separately fromthe actual transcription of the text, the individual doing the
annotation will have to go through the tape over and over again to reconstruct
the overlaps a process that could be done more efciently by the person doing
the transcription. Likewise, speaker identication tags tags indicating who is
speaking are more efciently inserted during the transcription of texts. With
written texts, if two-line breaks are inserted between paragraphs while a text is
being computerized, then paragraph tags can be inserted automatically at a later
stage. Of course, some markup is probably better inserted after a text sample
is computerized. But because computerizing and annotating a text is such an
integrated process, it is best to combine the processes whenever this is possible.
3.6 Computerizing speech
Traditionally, speech has been transcribed using a special transcription
machine which has a foot pedal that stops and starts a tape and also automat-
ically rewinds the tape to replay a previous segment. As anyone who has ever
transcribed speech knows, the ow of speech is much faster than the ability of
the transcriber to type. Therefore, it is extremely important to be able to replay
segments automatically, and anyone attempting to transcribe a tape using a tra-
ditional tape recorder is wasting their time and adding unnecessary hours to
their transcription efforts.
Because of recent advances in computer technology, it is now possible to use
software programs designed specically to transcribe speech that has been dig-
itized. VoiceWalker 2.0 was developed to aid in the transcription of texts incl-
udedwithinthe Santa Barbara Corpus of SpokenAmericanEnglish(http://www.
linguistics.ucsb.edu/resources/computing/download/download.htm). Sound-
Scriber is a similar program used to transcribe texts that are part of the
Michigan Corpus of Academic Spoken English (MICASE) (http://www.lsa.
umich.edu/eli/micase/soundscriber.html). Both of these programs are available
at the above URLs as freeware, and work very much like cassette transcription
machines: the transcriber opens a word-processing program in one window
and the transcription program in another. After a sample of digitized speech
is loaded into the program, short segments of the sample can be automatically
replayed until an accurate transcription is achieved. The advantage of software
like this is that it has made obsolete the need for cassette transcription machines,
which because of all of their moving parts tend to break down after frequent use.
The disadvantage of working with digitized speech is that high-quality digital
recordings can require very large amounts of disk space for storage. But this
problem can be minimized if recordings are saved at settings resulting in lower
quality (but still highly audible) recordings. ICE-USA saves digital recordings
as .wav les in Windows PCMformat, mono, 16 bit, 16000 Hz. Atypical 2,000-
word sample of speech saved with these specications will require 2030 MB
of hard disk space. If the same le were saved in stereo at CD quality, it would
require 80 MB (or more) of hard disk space. Since the ICE Project began, a
new le format, .mp3, has gained popularity, primarily because it creates au-
dio les much smaller than .wav les: a 26 MB ICE .wav le, for instance,
was converted into a 2.5 MB .mp3 le, and there was very little loss of sound
quality.
To digitize analog cassettes, all that is needed is a sound board and a program
that can digitize speech, such as Syntrilliums Cool Edit (available as share-
ware at http://www.syntrillium.com/cooledit/index.html). To use a program
such as Cool Edit to digitize speech, one simply connects the audio out-
put of the cassette recorder to the audio input of the sound board and has the
program digitize the speech sample as it is playing on the cassette player. After
the sample is digitized, it can be saved on either a hard drive or some other
media, such as a CD.
While transcription machines and software can ease the process of transcrib-
ing speech, there is no getting around the fact that speech has to be manually
transcribed. Although there are no immediate prospects that this process will
be automated, speech-recognition programs have improved considerably in re-
cent years, and in the near future, offer the hope that they can at least partially
automate the process of transcribing speech.
Early transcription programs, such as Dragon Dictate (http://www.voice-
recognition.com/1998/products/dragon/dictate.html), offeredlittle hope totran-
scribers, since they required words in a recording to be carefully pronounced and
followed by a pause. With a little training, such programs produced reasonable
speech recognition but only of a very articial type of speech. In the early 1990s,
a major technological breakthrough occurred: the ability of speech-recognition
programs to recognize continuous speech. There now exist a number of pro-
grams, suchas DragonNaturallySpeaking (http://voicerecognition.com/1998/
products/dragon/vrstandard.html), that have large lexicons and, with training,
can quite accurately recognize carefully articulated monologues. Current pro-
grams under development have progressed to the point where they are beginning
to work with certain kinds of dialogues. Nguyen et al. (1999) describe a system
that recognizes the kinds of monologues and dialogues spoken on broadcast
news programs. Built into the system is the ability to adjust to the speech of
males and females and to handle the various dysuencies that occur in spon-
taneous speech: repetitions, pauses, partially articulated words, and vocalized
pauses such as uh and uhm. Although speech-recognition programs may
never be able to take a multi-party dialogue, for instance, and in a matter of
seconds produce an accurate transcription of the dialogue, it is likely that at
some time in the future the technology may help ease the time-consuming task
of transcribing speech manually. And even if such systems are not completely
accurate, they could at least provide a draft transcription of a conversation that
could then be edited.
In terms of transcription time, speech runs a continuum, with multi-party dia-
logues with numerous overlaps taking the most time to transcribe, and scripted
monologues taking the least time to transcribe. For ICE-USA, a 2,000-word
multi-party dialogue takes on average fteen to twenty hours to transcribe, an-
notate with markup, and proofread. A scripted monologue, on the other hand,
takes much less time (approximately ve to eight hours). The remaining kinds
of speech, such as broadcast dialogues or telephone conversations, require tran-
scription times somewhere between these two extremes.
Transcribing speech is in essence a highly articial process, since an exclu-
sivelyoral formof language is representedinwrittenform. Consequently, before
any transcription is undertaken, it is important to decide just how much that
exists in a spoken text one wishes to include in a transcription of it. Compilers
of corpora have varied considerably in how much detail they have included in
their transcriptions of speech. In creating the Corpus of Spoken Professional En-
glish, Barlow(CPSADescription: http://www.athel.com/corpdes.html) made
a number of compromises. This is not a corpus of speech that Barlow recorded
and transcribed. Instead, it consists entirely of transcriptions of academic meet-
ings and White House press conferences that are in the public domain and that
were created by a third party. Consequently, while the transcriptions, accord-
ing to Barlow, appear relatively unedited and thus include hesitations, false
starts, and so on, ultimately there is no way to determine how faithful they
are to the original speech. Given the cost and effort of creating a corpus of
speech, it is understandable why corpora of this type exist, and undoubtedly,
they will yield much valuable information about speech. But there will always
be an element of doubt concerning results taken from a corpus created in this
manner.
At the other extreme are corpora of speech that attempt to replicate in a tran-
scription as much information as possible about the particular text being tran-
scribed. The Santa Barbara Corpus of Spoken American English, for instance,
contains not only an exact transcription of the text of a spoken conversation
(including hesitations, repetitions, partially uttered words, and so forth) but an-
notation marking various features of intonation in the text, such as tone unit
boundaries, pauses, and pitch contours. This kind of detail is included because
creators of this corpus attached a high value to the importance of intonation in
speech. The main drawback of this kind of detailed transcription is the amount of
time it takes toannotate a text withinformationabout intonation(cf. section4.1).
The advantage is that a tremendous amount of information about a spoken text
is provided to the user, thus ensuring that a broad range of studies can be con-
ducted on the corpus without any doubt about the authenticity of the data.
Whether one does a minimal or detailed transcription of speech, it is impor-
tant to realize that it is not possible to record all of the subtleties of speech in a
written transcription. As Cook (1995: 37) notes, a spoken text is made mean-
ingful by more than the words one nds in a transcription: how a conversation
is interpreted depends crucially upon such contextual features as paralanguage
(e.g. gestures and facial expressions), the knowledge the participants have about
the cultural context in which the conversation takes place, their attitudes towards
one another, and so forth. All of this extra-linguistic information is very dif-
cult to encode in a written transcription without the corpus compiler developing
an elaborate system of annotation to mark this information and the transcriber
spending hours both interpreting what is going on in a conversation and in-
serting the relevant markup. It is therefore advisable when transcribing speech
to nd a middle ground: to provide an accurate transcription of what people
actually said in a conversation, and then, if resources permit, to add extra infor-
mation (e.g. annotation marking features of intonation). In reaching this middle
ground, it is useful to follow Chafes (1995) principles governing the transcrip-
tion of speech. A transcription system, Chafe (1995: 55) argues, should (1)
adopt as far as possible standard conventions of orthography and capitalize on
habits [of literacy] already formed by corpus users; (2) strive to be as iconic
as possible; and (3) be compatible with current computer technology. The fol-
lowing sections contain a brief survey of some of the issues that the corpus
creator must address in creating a satisfactory system for transcribing speech.
The survey is not exhaustive but illustrative, since sketching out a complete
system of transcription is beyond the scope of the discussion.
3.6.1 Representing speech in standard orthographic form
Many of the expressions found in speech can easily be represented
within the constraints of standard orthography, particularly if annotation is
added to a transcription to mark the particular feature being transcribed.
3.6.1.1 Vocalized pauses and other lexicalized expressions
Speech contains a group of one- or two-syllable utterances sometimes
referred to as vocalized pauses, verbal expressions that allow the speaker to
pause and plan what is to be said next. These expressions form a closed class
and include pronunciations such as [
e
] or [a:], [
e
m], and [
e
h
e
]. Although there is
no universally agreed spelling for these expressions, in the ICE corpora, these
pronunciations are transcribed, respectively, as uh, uhm, and uhuh.
Speech also contains a group of expressions that can be used to answer yes/no
questions or to express various kinds of emotions. To represent these expres-
sions in a transcription, the Michigan Corpus of Academic Spoken English
(MICASE) has developed a number of orthographic representations. For in-
stance, for the various expressions that can be used to answer yes/no questions
in English, MICASE has created the following orthographic representations:
yes: mhm, mm, okey-doke, okey-dokey, uhuh, yeah, yep, yuhuh
no: uhuh, huhuh, mm, huhuh (MICASE Transcription and Spelling Conventions
http://www.lsa.umich.edu/eli/micase/transcription.html)
To represent various kinds of verbal expressions that people use, MICASE has
representations such as ach, ah, ahah, gee, and jeez.
3.6.1.2 Linked expressions
In speech, there are examples of expressions that are spelled as two
separate words, but pronounced as one word. In ctional dialogue, they are
sometimes spelled as gotta (for got to), hafta (for have to), and gonna (for
going to). If an expression such as gotta is viewed merely as a phonetic merging
of got and to, then all instances of this pronunciation can simply be transcribed
as got to. However, if gotta is felt to be a single lexical item, then it should be
transcribed differently than got to. How these words are transcribed depends
crucially upon which of these two analyses one accepts. If, for instance, all
instances of gotta and got to are transcribed as got to, then whoever uses the
corpus will not be able to recover the distinct forms of this expression. For this
reason, the ICE project includes gotta, hafta, and gonna in its transcriptions
of spoken texts. Other corpus projects, such as MICASE, include additional
examples, such as kinda, sorta, and lotsa (MICASETranscription and Spelling
Conventions: http://www.lsa.umich.edu/eli/micase/transcription.html).
3.6.1.3 Partially uttered words and repetitions
Speech (especially unscripted speech) contains a number of false starts
and hesitations resulting in words that are sometimes not completely uttered.
In the example below, the speaker begins uttering the preposition in but only
pronounces the vowel beginning the word.
<$D> There are more this year than <.> i </.> in in your year werent there (ICE-GB
s1a-040-050)
Inthe ICEproject, suchincomplete utterances are givenanorthographic spelling
that best reects the pronunciation of the incompletely uttered word, and then the
incomplete utterance is enclosed in markup, <.> and </.>, that labels the expres-
sion as an instance of an incomplete word. Other corpora, such as the London
Lund Corpus, sometimes provide a phonetic transcription of such utterances.
Repetitions can be handled in a similar manner. When speaking, an individual
will often repeat a word more than once as he or she is planning what to say
next. In the example below, the speaker repeats the noun phrase the police boat
twice before she completes the utterance:
<$B> <} ><- >the police boat<-/> <= >the police boat<=/> <}/ >we know but the war-
dens boat we dont you knowhe could just be a in a little rowboat shing and (ICE-USA)
To accurately transcribe the above utterance, the transcriber will want to in-
clude both instances of the noun phrase. However, this will have the unfortunate
consequence of skewing a lexical analysis of the corpus in which this utterance
occurs, since all instances of the, police, and boat will be counted twice. To
prevent this, the ICE project has special markup that encloses the entire se-
quence of repetitions (<} ><}/ >) and then places special markup (<= >the
police boat<=/>) around the last instance of the repetition, the only instance
counted in analyses done by ICECUP (cf. section 5.3.2), the text analysis pro-
gram used in the ICE Project.
3.6.1.4 Unintelligible speech
Very often when people speak, their speech is not intelligible. If two
people speak simultaneously, for instance, they may drown out each others
words and the speech of both speakers will become unintelligible. Anyone
doing any transcription of spoken dialogues will therefore encounter instances
where speech cannot be transcribed because it is not understandable. In the
ICE project, stretches of speech of this type are annotated as an uncertain
transcription or as an unclear word or syllable. In the example below, the
tags <?>and </?>surround the word thembecause the transcriber was uncertain
whether this was actually the word the speaker uttered.
<$C> <#9:1:C> <sent> What was Moses doing going off in <?> them</?> jeans
3.6.1.5 Punctuation
Because speech is segmented by intonation rather than by punctuation,
if speech is transcribed without any punctuation, it can be very difcult to read,
as sentences and utterances are run together and difcult to separate:
<$A> you can sh off our dock interesting not this summer last summer I was sitting in
the boat listening to the radio and a man walked to the end of the peer put his tackle box
down put some I think articial bait on he started casting and he caught two bass two
peeper bass within while I was sitting there watching him for half an hour it depends on
on the condition you dont have to go a long ways in the evening they would be in the
shallow water <unintelligible> (ICE-USA)
To make speech easier to read, it is tempting to add standard punctuation to
a spoken transcription. However, by punctuating speech, the corpus compiler
is in a sense interpreting the spoken text for future users of the corpus and
therefore making decisions that the users really ought to make themselves as
they analyze a spoken text and (if possible) listen to a recording of it. For this
reason, the ICE project includes only a very limited number of punctuation
marks: apostrophes for contractions and possessives, capitalization of proper
nouns and words beginning text units, and hyphens for hyphenated words. If
the corpus compiler wishes to segment spoken texts in transcriptions, it is far
better to include prosodic transcription in a corpus; that is, annotation that marks
such features of intonation as pauses, tone unit boundaries, and pitch changes
(cf. section 4.1).
3.6.1.6 Background noise
As speakers converse, they very frequently laugh, cough, sneeze, and
make other non-linguistic sounds. In many transcription systems, these noises
are simply ignored and not transcribed. In the ICE project, such sounds are
enclosed in markup: <O>cough<O/> or <O>sneeze<O/>. In other corpora,
comparable conventions are used.
3.6.1.7 Changing the names of individuals referred
to in spoken texts
In any conversation, speakers will address each other by name, and
they will talk about third-party individuals, sometimes in unattering ways
one spoken sample fromthe American component of ICE contains two brothers
talking quite disparagingly about their parents.
In transcribing a recording taken from a public broadcast, such as a radio
talk show, it is of little concern whether the actual names of individuals are
included in a transcription, since such a conversation was intended for public
distribution. Intranscribingprivate conversations betweenindividuals, however,
it is crucial that names are changed in transcriptions to protect the privacy of
the individuals conversing and the people they are conversing about. And if the
recordings accompanying the transcriptions are to be made available as well,
any references to peoples names will have to be edited out of the recordings
something that can be done quite easily with software that can be used to edit
digitizedsamples of speech. Inchangingnames intranscriptions, one cansimply
substitute new names appropriate to the gender of the individual being referred
to or, as was done in the LondonLund Corpus, substitute ctitious names
that are prosodically equivalent to the originals (Greenbaum and Svartvik
1990: 19).
3.6.2 Iconicity and speech transcription
Because writing is linear, it is not difcult to preserve the look of a
printed text that is converted into an electronic document and transferred from
computer to computer in ASCII format: although font changes are lost, markup
can be inserted to mark these changes; double-spaces can be inserted to sep-
arate paragraphs; and standard punctuation (e.g. periods and commas) can be
preserved. However, as one listens to the ow of a conversation, it becomes
quite obvious that speech is not linear: speakers very often talk simultaneously,
and while someone is taking their turn in a conversation, another party may ll
in brief gaps in the turn with backchannels, expressions such as yeah and right
that tell the speaker that his or her speech is being actively listened to and sup-
ported. Attempting to transcribe speech of this nature in a purely linear manner
is not only difcult but potentially misleading to future users of the corpus,
especially if they have access only to the transcription of the conversation, and
not the recording.
To explore the many options that have evolved for making transcriptions
more iconic, it is instructive, rst of all, to view how conversations were tran-
scribed by early conversational analysts, whose transcriptions occurred mainly
in printed articles, not in computerized corpora, and how using this early con-
vention of transcription in computerized corpora has certain disadvantages. The
conversational excerpt below contains a system of transcription typical of early
systems.
A: its gure three that we have to edit now
[ ]
B: no its gure
four we already did gure three
[ ]
A: oh yeah I remember we did it before
(Blachman, Meyer, and Morris 1996:61)
In the excerpt above, the brackets placed above and below segments of speech
in successive speaker turns indicate overlapping segments of speech. However,
instead of representing these overlaps with markup (as an SGML-conformant
system would do; cf. section 4.1), this system attempts to indicate them icon-
ically by vertically aligning the parts of the conversation that overlap to give
the reader a sense of how the ow of the conversation took place. While such
visual representations of speech are appealing, the implementation of such a
system, as Blachman, Meyer, and Morris (1996) note, is very problematical.
To vertically align segments of speech, it is necessary to insert tabs into a text.
However, as a text gets transferred from computer to computer, font sizes can
change and throwthe vertical alignment off, giving the corpus user an erroneous
representation of the segments of speech that overlap.
To overcome problems like this, Blachman Meyer, and Morris (1996) advo-
cate that segments of a conversation containing overlapping speech should be
represented in tabular form, a manner of presentation that provides both an ac-
curate and iconic representation of a conversation (cf. also Meyer, Blachman,
and Morris 1994). Below is how the above conversational excerpt would be
represented in a system of this type.
A B
its gure three that we have to
edit now no its g
ure four
oh yeah I remember we already did gure three
we did it before
(Blachman, Meyer, and Morris 1996: 62)
In the excerpt above, segments of speech in adjoining cells of the table overlap:
edit now in speaker As turn, for instance, overlaps with no its g in speaker Bs
turn.
An alternative way to represent iconically not just overlapping speech but the
ow of conversation in general is to lay it out as though it were a musical score.
In the HIAT system (Ehlich 1993), a speakers contribution to a conversation
is represented on a single horizontal line. When the speaker is not conversing,
his or her line contains blank space. If speakers overlap, their overlaps occur
when they occupy the same horizontal space. In the example below, T begins
speaking and, midway through his utterance of the word then, H overlaps her
speech with Ts. Speakers S1, S2, and Sys lines are blank because they are not
contributing to the conversation at this stage.
T: at once, then the same res/ No, leave it! Wouldve been (immediately) the same
result. (instantly)
H: Shall I (wipe it out)?
S1:
S2:
Sy:
Ehlich (1993: 134)
Other attempts at inconicity in the above excerpt include the use of the slash
in Ts turn to indicate that Hs overlap is an interruption. Slashes, according to
Ehlich (1993: 128), help convey a sense of jerkiness in speech. Periods and
commas mark various lengths of pauses, and the parentheses indicate uncertain
transcriptions.
While iconicity is a worthy goal to strive for in transcriptions of speech, it is
not essential. As long as transcriptions contain clearly identied speakers and
speaker turns, and appropriate annotation to mark the various idiosyncrasies
of speech, a transcription will be perfectly usable. Moreover, many users of
corpora containing speech will be interested not in howthe speech is laid out but
in automatically extracting information from it. Bibers (1988) study of speech
and writing, for instance, was based on tagged versions of the LondonLund and
LOB corpora; and through the implementation of a series of algorithms, Biber
was able to extract much valuable information from these corpora, without
having to examine them manually and make sense out of them with all the
markup that they contained. Nevertheless, one of the great challenges that
corpus linguists must face is the development of software with user interfaces
that permit users to browse spoken corpora easily and effortlessly.
3.7 Computerizing written texts
Because written texts are primarily linear in structure, they can eas-
ily be encoded in an ASCII text format: most features of standard orthogra-
phy, such as punctuation, can be maintained, and those features that cannot,
such as font changes or special characters, can be annotated with an SGML-
conformant tag. Although computerizing a written text takes time, particularly
if it has to be scanned, this process is far less time-consuming than comput-
erizing a spoken text. However, if one is creating a historical corpus and thus
working with texts from earlier periods of English, converting a text into elec-
tronic form can be a formidable task and, in addition, raise methodological
concerns that the corpus linguist working with modern texts does not need to
consider.
Because written texts fromearlier periods may exist only in manuscript form,
they cannot be optically scanned but have to be typed in manually. Moreover,
manuscripts can be illegible in sections, requiring the corpus creator to recon-
struct what the writer might have written. Describing a manuscript extract of the
Middle English religious work Hali Meidenhad, Markus (1997: 21112) details
numerous complications that the corpus creator would encounter in attempting
to create a computerized version of this manuscript: spelling inconsistencies
(e.g. <v> and <u> are used interchangeably), accent marks over certain vowels,
and colons and dots marking prosodic groupings rather than syntactic construc-
tions.
Although only two versions of Hali Meidenhad have survived, thus reducing
the level of difference between various versions of this text that the corpus
compiler would have to consider, other texts, such as Ancrene Riwle, can be
found in numerous manuscript editions: eleven versions in English, four in
Latin, and two in French (Markus 1997: 212). Because the corpus compiler is
concerned with absolute delity to the original, having more than one version
of a single text raises obvious methodological concerns which can be dealt with
in a number of different ways.
In theory, the corpus compiler could create a corpus containing every manu-
script version of a text that exists, and then let the user decide which version(s)
3.7 Computerizing written texts 79
to analyze, or provide some kind of interface allowing the user to compare
the various versions of a given manuscript. The Canterbury Project gives users
access to all eighty-eight versions of Chaucers Canterbury Tales and allows
various kinds of comparisons between the differing versions (Robinson 1998).
Because of the enormous amount of work involved in computerizing all versions
of a particular text, it is much more practical to computerize only one version.
One possibility is to computerize an edited version of a manuscript, that is,
a version created by someone who has gone through the various manuscript
versions, made decisions concerning howvariations in the differing manuscript
versions ought to be reconciled, and produced, in a sense, a version that never
existed.
6
A variation on this second alternative (and one advocated by Markus
1997) is for the corpus compiler to become editor and normalize the text as he or
she computerizes it, making decisions about which variant spellings to consider,
which diacritics to include, and so forth. From a modern perspective, none of
these alternatives is ideal, but when working with texts from earlier periods, the
corpus compiler has to make compromises given the kinds of texts that exist.
Fortunately with modern-day texts, none of these complications exists, and
the process of computerizing a written text involves mainly getting the text into
electronic form. As was noted in section 3.3, if a written text is not available in
electronic form, a printed copy of the text can be converted into an electronic
format with an optical scanner. There are two variables that will affect the
success that one has with an optical scanner. The rst variable is the quality
of the original printed text. If the text is blurred or does not contain a distinct
typeface, when scanned, it will contain numerous typographical errors that may
take longer to correct than it would to simply retype the entire text by hand.
The second variable is the quality of scanner and the OCR (optical character
recognition) software used with it.
There are two types of scanners: form-feed scanners and atbed scanners.
Experience with ICE-USA has shown that atbed scanners are slightly more
accurate than form-feed scanners but that satisfactory results can be obtained
with the use of a relatively inexpensive form-feed scanner if the printed copy
to be scanned is of high quality and no attempt is made to scan pages with
columns, a common format in newspapers and magazines and a format that
cannot be handled by the OCR software commonly bundled with inexpensive
form-feed scanners. However, this problemcan be solved if columns are copied,
clipped, and scanned separately.
Because so many kinds of written texts are now available on the World Wide
Web, it makes more sense to obtain as many written texts as possible from
relevant websites. However, even though such texts can be easily downloaded
in electronic form, in many cases they will contain as much HTML coding
6
Because edited versions of older texts are always copyrighted, the corpus compiler will need to
obtain copyright clearance for use of the version. This can greatly complicate the creation of
a historical corpus. ARCHER, for instance, has not been publicly released because copyright
clearance could not be obtained for many of the texts included in the corpus.
as text. To attempt to manually delete this coding will require a considerable
amount of time and effort. Fortunately, there exists software (e.g. HTMASC
http://www.bitenbyte.com/) that can automatically strip HTML coding from
texts and produce in seconds an ASCII text le with no coding.
3.8 Conclusions
The process of collecting and computerizing texts is, as this chapter
has demonstrated, a labor-intensive effort. Collecting speech requires the corpus
creator to become, in a sense, a eld linguist and go out and record individuals
in the various locations where speech takes place: homes, ofces, schools,
and so forth. And once speech is recorded, considerable time and effort must
be expended transcribing it. Written texts are somewhat easier to collect and
encode in electronic format, but obtaining copyright clearance for copyrighted
texts can be a very complicated undertaking; if texts from earlier periods are
being collected and computerized, decisions will need to be made about what
exactly goes into the version of the text that is to be included in a corpus if,
for instance, more than one manuscript version of the text exists. Once the
process of collecting and computerizing texts is completed, texts to be included
in a corpus are ready for the nal stage of preparation annotation a topic
discussed in the next chapter.
Study questions
1. What makes collecting natural speech so difcult?
2. To record speech, those creating the British National Corpus gave individuals
tape recorders and had them record conversations they had with others. Why
is this potentially a better way of collecting speech than sending out a research
assistant with a tape recorder and having him or her make the recordings?
3. What are some of the advantages and disadvantages of including in a corpus
written texts obtained from the World Wide Web?
4. Why, when collecting texts for a corpus, is it important to keep records
documenting ethnographic information for writers and speakers whose texts
will be included in the corpus?
5. What problems do texts taken from earlier periods of English pose when it
comes time to computerize the texts?
4 Annotating a corpus
For a corpus to be fully useful to potential users, it needs to be an-
notated. There are three types of annotation, or markup, that can be inserted
in a corpus: structural markup, part-of-speech markup, and grammatical
markup.
1
Structural markup provides descriptive information about the texts. For in-
stance, general information about a text can be included in a le header, which
is placed at the start of a text, and can contain such information as a complete
bibliographic citation for a written text, or ethnographic information about the
participants (e.g. their age and gender) in a spoken dialogue. Within the ac-
tual spoken and written texts themselves, additional structural markup can be
included to indicate, for instance, paragraph boundaries in written texts or over-
lapping segments of speech in spoken texts. Part-of-speech markup is inserted
by a software program called a tagger that automatically assigns a part-of-
speech designation (e.g. noun, verb) to every word in a corpus. Grammatical
markup is inserted by a software program called a parser that assigns labels
to grammatical structures beyond the level of the word (e.g. phrases, clauses).
This chapter focuses on the process of annotating texts with these three kinds
of markup. The rst section discusses why it is necessary for corpora to be
annotated with structural markup and then provides an overview of the various
systems of structural markup that have evolved over the years as well as the
various tools that have been developed for inserting markup in corpora. The next
two sections are concerned with the process of tagging and parsing a corpus.
These sections will describe some of the more common taggers and parsers that
are available, provide descriptions of what tagged and parsed texts look like,
detail how the various taggers and parsers available reect different views of
English grammar, and discuss some of the constructions that lead to tagging
and parsing errors. In addition, other types of tagging, such as semantic and
discourse tagging, are briey described. Even though much of the underlying
work on taggers and parsers has been done by computational linguists doing
research in the area of natural language processing, this chapter will be primarily
concerned with how tagged and parsed corpora can be of benet to corpus
linguists interested in creating and analyzing corpora for purposes of descriptive
1
There is also annotation, such as semantic annotation, that can be used to mark higher-level
structures larger than the word, clause, or sentence. Cf. section 4.4.3 for details.
81
linguistic research; theoretical issues of computation will be broached only as
they clarify issues related to linguistic theory and description.
4.1 Structural markup
A document created by any word-processing program can be format-
ted in numerous ways: font types and sizes can be varied, sections of the text
can be italicized or placed in boldface, footnote indicators can be displayed in
superscript form, and so forth. However, as soon as this document is converted
into ASCII (or text) format, all of this formatting is lost, and only pure text can
be displayed (cf. section 3.5 for a discussion of ASCII text format). Therefore,
for texts passed from computer to computer in ASCII format, some system of
markup needs to exist so that anyone receiving a text can reconstruct how
it originally looked. And in language corpora, this issue becomes especially
important: not only are there features of written corpora that will need to be pre-
served, but there are characteristics of speech (such as overlapping speech) that
need to be marked so that they can be easily identied by the user of the corpus.
In earlier corpora, a fairly minimal and non-standardized set of markup was
used to annotate various features of speech and writing. For instance, in the
LondonLund Corpus, speakers were identied with alphabetic letters (A, B,
C, etc.) followed by colons, and overlapping segments of speech were enclosed
with asterisks:
A: yes that I think you told me *I*
B: *and* none of them have been what you might call very successful in this world
(LLC S.1.13 98100)
More recently, as electronic documents have proliferated, a standard for the
markup of electronic documents has developed. This standard, known as
Standard Generalized Markup Language (SGML), is not a series of predeter-
mined symbols, or tags, as in the example above, but instead a metalanguage
that provides a mechanism for describing the structure of electronic docu-
ments. For corpus compilers, the main advantage of placing SGML-conformant
markup in their corpora is that information about a corpus can be consistently
and unambiguously described and maintained as the corpus is transferred from
computer to computer. Although the idea behind SGML is simple and straight-
forward, the actual implementation of an SGML-based systemof markup can be
quite complex. For this reason, this section will provide only a general overview
of SGML.
SGML markup is purely descriptive, that is, it does not tell the computer to
perform any operations on a text but instead consists of markup codes which
simply provide names to categorize parts of a document (Sperberg-McQueen
andBurnard1994b: http://etext.lib.virginia.edu/bin/tei-tocs-p3?div=DIV1&id=
SG). The example below shows how the previous excerpt would be anno-
tated with the SGML-conformant structural tags used in the ICE project
4.1 Structural markup 83
(cf. Nelson 1996 for a complete discussion of the structural markup used to
annotate ICE texts).
<$A><#:1> yes that I think you told me <{><[>I<[/>
<$B><#:2> <[>and<[/><{/> none of themhave been what you might call very success-
ful in this world
Some of the markup provides a description of the structure of the conversation.
Two of the tags, <$A> and <$B>, indicate who is speaking at a given time in
the conversation; two other tags, <#:1> and ><#:2>, set off and number con-
secutively what are called text units, that is, segments of a spoken or written
text that are either grammatical sentences or, in speech, utterances: stretches of
speech that may be grammatically incomplete but that form a coherent gram-
matical unit. For instance, the expression Yeah is not a grammatical sentence but
an utterance: when used in a conversation as a response to a yes/no question, it
makes perfect sense and is semantically coherent. Other tags in the excerpt de-
scribe individual features of the conversation. Two tags, <{ > and <{/>, mark the
beginning and end of a speech segment containing overlapping speech; within
this lengthier segment are two sets of additional tags, <[ > and <[/>, that mark
the two individual segments of speech in each speaker turn that overlap. The
ICE Project has developed markup for written texts as well to annotate such
features of writing as italics, boldface, small caps, superscript and subscript
symbols, and foreign letters not represented in ASCII format.
Typically, SGML-conformant tags contain a start tag and an end tag, as was
the case in the above excerpt with the tags used to mark overlapping speech.
However, sometimes only one tag is necessary. For instance, a speaker identi-
cation tag simultaneously marks the beginning of a new speakers turn and the
end of a previous speakers turn. Likewise, a text unit tag marks the beginning
of a new text unit and the end of a previous one.
In addition to containing markup placed directly in a text, ICEtexts also begin
with le headers, that is, a series of statements enclosed in ICEtags that provide,
for instance, bibliographic information about written texts, or ethnographic
information about individuals in spoken texts. Below is the beginning section
of a le header for a written text included in the British component of ICE:
<text.info>
<le.description>
<textcode>ICE-GB-W2A-008</textcode>
<number.of.subtexts>1</number.of.subtexts>
<category>printed;informational:popular:humanities</category>
<wordcount>2101</wordcount>
<version>tagged by TOSCA tagger using the ICE tagset</version>
<free.comments> </free.comments>
</le.description>
.
.
.
<text.info> (Nelson 1996: 51)
The tags in the above header provide important descriptive information about
the text, such as the genre in which the text has been classied, the number of
words that the text contains, and a statement that the TOSCA tagger assigned
part-of-speech designations taken from the ICE tagset (cf. section 4.2). It must
be emphasized that the tags used in the ICE Corpus are not SMGL tags per
se, but rather SGML-conformant tags: the names and symbols used in the tags
above were developed by the ICE Project within the constraints stipulated by
SGML.
The tags for the ICE Project were developed in the late 1980s and therefore
do not reect recent work done by the Text Encoding Initiative (TEI) to develop
a comprehensive SGML-conformant markup system for electronic documents,
particularly those used in the humanities and in language corpora. As Burnard
(1995) notes, an electronic document annotated with TEI-conformant markup
would contain three types of tagsets: core tagsets (a series of tags, such as
those associated with le headers or paragraph divisions, available for insertion
in any document); base tagsets (a series of tags associated with particular
kinds of texts, such as verse, drama, or transcribed speech); and additional
tagsets (any tags the user wishes to add to the core and base tagsets already
in a document). The advantage of the TEI system is that it provides both a
standardized set of tags for insertion in a document and the exibility for the
insertion of tags designed by the corpus compiler. Since it appears that TEI is
the emerging standard for the markup of corpora (the entire British National
Corpus was annotated with TEI markup), anyone contemplating the creation of
a corpus ought to give serious consideration to using the TEI standards for the
markup of texts. The more standardized creators of corpora can be in the markup
they insert, the easier it will be for users to analyze corpora. A full description
of the TEI system can be found in Sperberg-McQueen and Burnard (1994a).
More recently, there has been interest in the corpus linguistics community
in a newly emerging markup system: Extensible Markup Language (XML).
XML is a restricted version of SGML that has been designed mainly for use in
web documents. The American National Corpus (a corpus of American English
comparable to the British National Corpus in length and coverage) plans to use
XML to annotate texts included in the American National Corpus (American
National Corpus: http://www.cs.vassar.edu/ide/anc/). The EAGLES Project
has an extensive working document to develop a Corpus Encoding Standard
that describes howto use XML to annotate a corpus (XCES: Corpus Encoding
Standard for XML: http://www.cs.vassar.edu/XCES/). And the TEI Project has
been working to incorporate XML within its standards. At this point, it is not
clear how widely used XML will become in the corpus linguistics community.
But a clear advantage of XML is its ability to be used on web pages, and as
the web develops, it will become an increasingly important medium for the
distribution of corpora.
There is one additional type of annotation that can be incorporated within an
SGML-conformant systemof markup but that has typically not been: annotation
4.1 Structural markup 85
that is used to mark features of intonation in spoken texts, such as changes in
pitch or tone unit boundaries. The example below illustrates the system of
annotation used in the Santa Barbara Corpus of Spoken American English
(for additional details, cf. Chafe, Du Bois, and Thompson 1991: 75; Du Bois,
Schuetze-Coburn, Cumming, and Paolino 1993):
G: For

most people, its

celebration,\
for

me =,
its .. its a%

ti=me,
to=
(H) to get in

be=d,\/ (Chafe, Du Bois, and Thompson 1991: 77)
In the above example, all of the annotation indicates particular characteristics
of intonation, such as falling pitch (\), primary accent (
), lengthened syllables
(=), and a short pause (..). While prosodic annotation provides a more accu-
rate rendering of the ow of speech than punctuation does (cf. section 3.6.1.5),
inserting the necessary prosodic markup can increase transcription time sig-
nicantly in some cases, doubling or tripling the time of transcription. To
minimize transcription time, the ICE project decided to mark two lengths of
pause only short and long and to dispense with any other prosodic markup.
Although SGML-conformant systems of annotation provide important de-
scriptive information about a text, they can pose numerous difculties for both
corpus compilers and corpus users. First of all, because much markup has to
be manually inserted, if a corpus compiler wants to exploit the full possibilities
of the TEI system, for instance, he or she will need considerable resources to
hire assistants to insert the markup. There are tools that can assist in the in-
sertion of markup. To assist in the annotation of ICE texts, the ICE Markup
Assistant was developed (Quinn and Porter 1996: 657). This program uses
a series of WordPerfect macros to insert ICE markup in texts. These macros
are inserted either manually or automatically, depending upon how easy it is
to predict where a given tag should be inserted. For instance, because over-
lapping segments of speech occur randomly throughout a conversation, their
occurrence cannot be predicted and they therefore have to be inserted manu-
ally. Conversely, in a written text, text unit tags can be inserted automatically at
sentence boundaries, and then those that are erroneously marked can be post-
edited manually. Arange of SGMLresources can be found on the W3Cwebsite:
http://www.w3.org/MarkUp/SGML/.
Another way to minimize the amount of time it takes to annotate texts is
to use a reduced system of annotation. The ICE Project reduced the amount
of structural markup required in ICE texts to the most essential markup for
those ICE teams lacking the resources to insert all of the ICE markup that
had been developed (Meyer 1997). Likewise, there exists a version of TEI
called TEI Lite, which contains a minimal set of TEI conformant markup
(Burnard and Sperberg-McQueen 1995). In determining how much markup
should be included in a text, it is useful to adopt Burnards (1998) Chicago
pizza metaphor. One can view markup as toppings on a pizza: the particular
toppings (i.e. markup) that are added depend upon both what the corpus
compiler sees as important to annotate and what resources are available to
insert this annotation. Although some object to this view Cook (1995), for
instance, views annotation used in spoken texts not as an embellishment, an
extra topping, but crucial to the interpretation of a speech event realistically
the corpus compiler has to draw a line in terms of how detailed texts should be
annotated.
While texts annotated with structural markup greatly facilitate the automatic
analysis of a corpus, the user wishing to browse through a text (particularly a
spoken text) will nd it virtually unreadable: the text will be lost among the
markup. The British component of ICE (ICE-GB) works around this problem
by enabling the user to select how much markup he or she wishes to see in a
text: all, none, or only some. Likewise, some concordancing programs, such
as MonoConc Pro 2.0 (cf. section 5.3.2), can display markup or turn it off.
One of the future challenges in corpus linguistics is the development of tools
that provide an easy interface to corpora containing signicant amounts of
annotation. The introduction of XML markup into corpora offers potential
help in this area, since as web browsers are designed to read XML annotated
documents, they will be able to convert a corpus into the kinds of documents
displayed by browsers on websites.
2
The discussion in this section might lead one to think that there has been
an endless proliferation of markup systems in corpus linguistics to annotate
corpora with structural markup. However, it is important to realize that all of
these systems have one important similarity: they are all SGML-conformant.
The ICE system was developed to annotate ICE documents, the TEI system a
wider range of spoken and written corpora as well as various kinds of documents
commonly found in the humanities, and XML various corpora for possible
distribution on the World Wide Web. Thus, all of these markup conventions are
simply instantiations of SGML put to differing uses.
4.2 Tagging a corpus
In discussing the process of tagging a corpus, it is important, rst of
all, to distinguish a tagset a group of symbols representing various parts of
speech from a tagger a software program that inserts the particular tags
making up a tagset. This distinction is important because tagsets differ in the
number and types of tags that they contain, and some taggers can insert more
than one type of tagset. In the example below, the Brill tagger (described in
Brill 1992), adapted for use in the AMALGAM Tagging Project, has been used
2
Cf. Edwards (1993) for further information on how markup can be made more readable, partic-
ularly markup used to annotate spoken texts.
4.2 Tagging a corpus 87
to assign part-of-speech designations to each word in the sentence Im doing
the work:
3
I/PRON(pers,sing)
m/V(cop,pres,encl)
doing/V(montr,ingp)
the/ART(def )
work/N(com,sing)
./PUNC(per)
In this example, which contains tags from the ICE tagset (Greenbaum and
Ni 1996), each word is assigned to a major word class: I to the class of pro-
nouns, m and doing to the class of verbs, the to the class of articles; work to
the class of nouns; and the end stop to the class of punctuation. In parentheses
following each major word class are designations providing more specic infor-
mation about the word: I is a personal pronoun that is singular; mis the enclictic
(i.e. contracted) present-tense formof the copula be;
4
doing is an -ing participle
that is monotransitive (i.e. takes a single object); the is a denite article; work
is a common noun; and the specic punctuation marked that is used is a pe-
riod. The manner in which the ICE tagset has been developed follows Leechs
(1997: 256) suggestion that those creating tagsets strive for Conciseness,
Perspicuity (making tag labels as readable as possible), and Analysability
(ensuring that tags can be decomposable into their logical parts, with some
tags, such as noun, occurring hierarchically above more specic tags, such as
singular or present tense).
Over the years, a number of different tagging programs have been developed
to insert a variety of different tagsets. The rst tagging programwas designed in
the early 1970s by Greene and Rubin (1971) to assign part-of-speech labels to
the Brown Corpus. Out of this programarose the various versions of the CLAWS
program, developed at the University of Lancaster initially to tag the LOB
Corpus (Leech, Garside, and Atwell 1983) and subsequently to tag the British
National Corpus (Garside, Leech, and Sampson 1987; Garside and Smith 1997).
The TOSCA team at the University of Nijmegen has developed a tagger that
can insert two tagsets: the TOSCA tagset (used to tag the Nijmegen Corpus)
and the ICE tagset (Aarts, van Halteren, and Oostdijk 1996). The AUTASYS
Tagger can also be used to insert the ICE tagset as well as the LOB tagset
(Fang 1996). The Brill Tagger is a multi-purpose tagger that can be trained
3
The example provided here is in a vertical format generated by the AMALGAM tagging project.
This project accepts examples to be tagged by e-mail, and will tag the examples with most of the
major tagsets, suchas those createdtotagthe LOB, Brown, andICEcorpora. For more information
on this project and on howthe Brill Tagger was trained to insert the various tagsets, see Atwell et al.
(2000) as well as the project website: http://agora.leeds.ac.uk/amalgam/amalgam/amalghome.
htm. The UCREL research centre at the University of Lancaster also provides online tagging
with CLAWS4 at: http://www.comp.lancs.ac.uk/computing/research/ucrel/claws/trial.html.
4
In the current ICE tagset, m would be tagged as aux (prog, pres, encl), that is, as a present-tense
progressive enclictic auxiliary.
to insert whatever tagset the user is working with into texts in English or any
other language. These taggers do not exhaust the number of taggers that have
been created, but provide an overview of the common taggers that have been
developed to annotate the types of corpora that corpus linguists typically work
with.
Taggers are of two types: rule-based or probablistic. In a rule-based tagger,
tags are inserted on the basis of rules of grammar written into the tagger. One of
the earlier rule-based taggers was the TAGGIT program, designed by Greene
and Rubin (1971) to tag the Brown Corpus and described in detail in Francis
(1979: 198206). The rst step in the tagging process, Francis (1979) notes, is
to look up a given word in the programs lexicon, and if the word is found, it
is assigned however many tags associated with the word in the lexicon. If after
this search the word is not found, an attempt is made to match the ending of the
word with a list of sufxes and the word-class tags associated with the sufxes.
If this search also fails to nd a match, the word is arbitrarily assigned three
tags: singular or mass noun, verb (base form), or adjective three form class
designations that the majority of words in English will fall into.
Of the words reaching this stage of analysis, 61 percent will have one tag,
and 51 percent of the remaining words will have sufxes associated with one
tag. The remaining words will have more than one tag and thus are candidates
for disambiguation. Initially, this is done automatically by a series of context
frame rules that look at the context in which the word occurs. For instance,
in the sentence The ships are sailing, the word ships will have two tags: plural
noun and third-person-singular verb. The context frame rules will note that ships
occurs following an article, and will therefore remove the verb tag and assign the
plural noun tag to this word. Although the context frame rules can disambiguate
a number of tags, the process is complicated, as Francis (1979: 202) observes,
by the fact that many of the words surrounding a word with multiple tags
will have multiple tags themselves, making the process of disambiguation quite
complicated. Consequently, 23 percent of the remaining tags had to be manually
disambiguated: analysts had to look at each example and decide which tag was
most appropriate, a process that itself was subject to error and inconsistency
and that led to further post-editing.
Although rule-based taggers have been largely superseded by probability-
based taggers, there is one current rule-based tagger, EngCG-2 (cf. Samuelsson
and Voutilainen 1997), that has been designed to overcome some of the prob-
lems inherent in early rule-based taggers like TAGGIT. In particular, rules in
EngCG-2 have wider application than in TAGGIT and are able to refer up to
sentence boundaries (rather thanthe local context alone) (Voutilainen1999: 18).
This capability has greatly improved the accuracy of EngCG-2 over TAGGIT.
Because rule-based taggers rely on rules written into the tagger, most taggers
developed since TAGGIT have been probabilistic in nature: they assign tags
based on the statistical likelihood that a given tag will occur in a given context.
Garside and Smith (1997: 104) give the example of the construction the run
4.2 Tagging a corpus 89
beginning a sentence. Because run starts the sentence and is preceded by the
determiner the, there is high probability that run is a noun rather than a verb.
The advantage of probabilistic taggers is that they can be trained on corpora
and over time develop very high and accurate probabilities, making them quite
accurate. However, even though probabilitistic and some rule-based taggers
such EngCG-2 can achieve accuracy rates exceeding 95 percent, the remaining
inaccuracies can be more extensive than one might think.
Kennedy (1996: 253) notes that such rates of accuracy involve averaging:
computing the success rate of tagging by combining gures for constructions
(such as the denite article the) that have high frequencies and that can be tagged
accurately almost all the time with other constructions (such as the word once,
which can be an adverbial or a subordinating conjunction) that can be tagged
accurately only 8085 percent of the time. If a corpus is not post-edited after
tagging is done (as was the case with the initial release of the British National
Corpus), the rate of inaccurate tagging for certain words can be quite high.
For instance, after the Wellington Corpus of Written New Zealand English was
tagged with the CLAWS tagger, Kennedy (1996: 255) found that more than
20 percent of the instances of once were wrongly tagged. In the sentence Once
implemented, it will be with us, once was tagged erroneously as an adverbial
rather than a subordinating conjunction. In other instances, adverbial uses of
once were tagged as conjunctions. Multiplied over a large corpus, error rates
as high as 20 percent can lead to considerable inaccuracy, a point that users of
corpora need to be aware of, especially if they intend to do an automatic analysis
of the corpus without actually looking carefully at the data being analyzed.
Because of the time and effort it takes to post-edit a tagged corpus, in the
future we are more likely to see corpora that are released for use without any
post-editing.
Because taggers cannot sanitize the data they have to work with, it is easy to
understand why they cannot tag with complete accuracy the words of any human
language. Human language is full of unusual and idiosyncratic phenomena
(Smith 1997: 147) that taggers have to account for. For example, consider how
the AMALGAM tagger analyzed the sentence Whats he want to prove?
what/PRON(nom)
s/V(cop,pres,encl)
he/PRON(pers,sing)
want/V(montr,pres)
to/PRTCL(to)
prove/V(montr,inn)
?/PUNC(qm)
The full uncontracted form of this sentence would be What does he want to
prove? However, because s is usually a contracted form of is, the tagger has
wrongly tagged s as the present-tense enclitic form of the copula be. To deal
with problematic cases of this type, those designing the CLAWS tagger have
developed a series of corpus patches (Smith 1997: 1457): context-sensitive
rules that take the output of data tagged with CLAWS and attempt to correct
persistently problematic cases that have been identied by previous tagging.
However, even with these rules, there remains a 2 percent rate of error in the
tagging.
The tagsets used to annotate corpora are as varied as the taggers used to
insert them. For instance, the Brown Corpus tagset contains seventy-seven tags
(Francis and Ku cera 1982), the ICE tagset 262 tags (Greenbaum and Ni 1996:
93).
5
A listing of tags for the auxiliary verb do illustrates the manner in which
these two tagsets differ. The Brown tagset contains only three tags for do that
specify the main surface structure forms that this verb takes:
Tag Form
DO do
DOD did
DOZ does
The ICE tagset, on the other hand, contains eight tags for do that provide a more
exhaustive listing of the various forms that this auxiliary has (Greenbaum and
Ni 1996: 101):
Tag Form
Aux (do, inn) Do sit down
Aux (do, inn, neg) Dont be silly
Aux (do, past) Did you know that?
Aux (do, past, neg) You didnt lock the door
Aux (do, present) I do like you
Aux (do, pres, encl) Whats he want to prove?
Aux (do, pres, neg) You just dont understand
Aux (do, pres, procl) Dyou like ice-cream?
The differences in the tagsets reect differing conceptions of English grammar.
The Brown tagset with its three forms of do is based on a more traditional view
of the forms that this auxiliary takes, a view that is quite viable because this
tagset was designed to apply to a corpus of edited written English. The ICE
tagset, in contrast, is based on the view of grammar articulated in Quirk et al.
(1985). Not only is this grammar very comprehensive, taking into consideration
constructions with low frequencies in English, but it is one of the few reference
grammars of English to be based on speech as well as writing. Consequently,
the ICE tagset is not just detailed but intended to account for forms of do that
would be found in speech, such as the proenclictic formDyou. The more recent
Longman Grammar of Spoken and Written English is also based on a corpus,
the Longman Spoken and Written English Corpus, that was tagged by a tagger
whose tagset was sensitive to distinctions made in speech and writing (cf. Biber
5
The number of tags for the ICEtagset is approximate. As GreenbaumandNi (1996: 93) emphasize,
newconstructions are always being encountered, causing additional tags to be added to the tagset.
4.3 Parsing a corpus 91
et al. 1999: 358). Although it might seem that a larger tagset would lead to
more frequent tagging errors, research has shown that just the opposite is true:
the larger the tagset, the greater the accuracy of tagging (Smith 1997: 1401;
Le on and Serrano 1997: 1547).
4.3 Parsing a corpus
Tagging has become a very common practice in corpus linguistics,
largely because taggers have evolved to the point where they are highly accurate:
many taggers can automatically tag a corpus (with no human intervention) at
accuracy rates exceeding 95 percent. Parsing programs, on the other hand,
have much lower accuracy rates (7080 percent at best, depending upon how
correctness of parsing is dened [Leech and Eyes 1997: 35 and 51, note 3]),
and they require varying levels of human intervention.
Because tagging and parsing are such closely integrated processes, many
parsers have taggers built into them. For instance, the Functional Dependency
Grammar of English (the EngFDG parser) contains components that assign not
only syntactic functions to constituents but part-of-speech tags to individual
words (Voutilainen and Silvonen 1996). The TOSCA Parser has similar capa-
bilities (Aarts, van Halteren, and Oostdijk 1996). Like taggers, parsers can be
either probabilistic or rule-based, and the grammars that underlie them reect
particular conceptions of grammar, even specic grammatical theories, result-
ing in a variety of different parsing schemes, that is, different systems of gram-
matical annotation that vary both in detail and in the types of grammatical con-
structions that are marked (cf. the AMALGAM MultiTreebank for examples
of differing parsing schemes: http://www.scs.leeds.ac.uk/amalgam/amalgam/
multi-tagged.html).
Although there is an ongoing debate among linguists in the eld of natural
language processing concerning the desirability of probabilistic vs. rule-based
parsers, both kinds of parsers have been widely used to parse corpora. Propo-
nents of probabilistic parsers have seen them as advantageous because they are
able to parse rare or aberrant kinds of language, as well as more regular, run-of-
the-mill types of sentence structures (LeechandEyes 1997: 35). This capability
is largely the result of the creation of treebanks, which aid in the training of
parsers. Treebanks, such as the Lancaster Parsed Corpus and the Penn Tree-
bank, are corpora containing sentences that have been either wholly or partially
parsed, anda parser canmake use of the alreadyparsedstructures ina treebankto
parse newly encountered structures and improve the accuracy of the parser. The
example below contains a parsed sentence from the Lancaster Parsed Corpus:
A01 2
[S[N a AT move NN [Ti[Vi to TO stop VB Vi][N \0Mr NPT Gaitskell NP
N][P from IN [Tg[Vg nominating VBG Vg][N any DTI more AP labour NN
life NN peers NNS N]Tg]P]Ti]N][V is BEZ V][Ti[Vi to TO be BE
made VBN Vi][P at IN [N a AT meeting NN [Po of INO [N labour NN
\0MPs NPTS N]Po]N]P][N tomorrow NR N]Ti] . . S]
The rst line of the example indicates that this is the second sentence from
sample A01 (the press reportage genre) of the LOBCorpus, sections of which
(mainly shorter sentences) are included in the treebank. Open and closed brack-
ets mark the boundaries of constituents: [S marks the opening of the sentence,
S] the closing; the [N preceding a move marks the beginning of a noun
phrase, N] following to stop its ending. Other constituent boundaries marked
in the sentence include Ti (to-innitive clause to stop Gaitskell from. . . ),
Vi (non-nite innitive clause to stop), and Vg (non-nite ing participle
clause nominating). Within each of these constituents, every word is assigned
a part-of-speech tag: a, for instance, is tagged AT, indicating it is an article;
move is tagged NN, indicating it is a singular common noun; and so forth.
6
Although many treebanks have been released and are available for linguistic
analysis, their primary purpose is to train parsers to increase their accuracy.
To create grammatically analyzed corpora intended more specically for
linguistic analysis, the TOSCA Group at Nijmegen University developed the
TOSCAParser, a rule-based parser that was used to parse the Nijmegen Corpus
and sections of the British component of ICE (ICE-GB). As the parse tree taken
from ICE-GB in gure 4.1 illustrates, the grammar underlying the TOSCA
Parser (described in detail in Oostdijk 1991) recognizes three levels of descrip-
tion: functions, categories, and features.
Figure 4.1 Parse tree from ICE-GB
6
A full listing of the labels in the above example can be found in Garside, Leech, and V aradi
(1992).
The rst level of description functions is specied both within the clause and
the phrase. For instance, Heavy rainfall is functioning as subject (SU) in the
main clause in which it occurs; in the noun phrase itself, Heavy is functioning as
an adjective premodier (AJP). Categories are represented at both the phrase
and word level: Heavy rainfall is a noun phrase (NP), rainfall a noun
(N) that is also head of the noun phrase (NPHD). Features describe various
characteristics of functions or categories. For instance, the noun rainfall is a
common noun (com) that is singular (sing).
7
The TOSCAParser provides such detailed grammatical information because
it was designed primarily not to create a treebank to be used to increase the accu-
racy of the parser but to produce a grammatically annotated corpus that yields
databases containing detailed information that may be used by linguists . . . and
[that] allows for the testing of linguistic hypotheses that have been formulated
in terms of the formal grammar (Oostdijk 1991: 64). And, indeed, the two
corpora parsed by the TOSCA Parser, the Nijmegen Corpus and ICE-GB, can
be used with software programs that extract information from the corpora: the
Linguistic Database (LDB) Program for the Nijmegen Corpus (van Halteren
and van den Heuvel 1990) and the ICE Corpus Utility Program (ICECUP) for
ICE-GB (cf. section 5.1.3).
While the development of taggers has ourished in recent years (Fang
1996: 110), parsers are still in a state of development. Taggers such as the Brill
Tagger or CLAWS4 are readily available, and can be easily run on a corpus,
once the corpus is in a form the tagger can accept. The output of the tagger
will have to be post-edited, a process that can take some time. Nevertheless,
tagging a corpus is currently a fairly common practice in corpus linguistics.
Parsing, in contrast, is a much more involved undertaking, largely because of
the complexity of structures that a parser is required to analyze, especially if
spoken as well as written data is being parsed, and the subsequent increase in
the amount of post-editing of the parsed output that needs to be done.
A parser has a much greater range of structures to analyze than a tagger: not
just individual words but phrases and clauses as well. And because phrases and
clauses are considerablycomplexinstructure, there are numerous constructions,
suchas those that are coordinated, that are notoriouslydifcult toparse correctly.
Figure 4.2 contains the parsed output of the sentence The child broke his armand
his wrist and his mother called a doctor after it was submitted to the EngFDG
Parser.
8
7
A listing of function, category, and feature labels used in the parsing of ICE-GB can be
found in the Quick Guide to the ICE-GB Grammar (http://www.ucl.ac.uk/english-usage/ice-
gb/grammar.htm).
8
Cf. Tapanainen and J arvinen (1997) for a more detailed discussion of the EngFDG Parser and
the grammatical annotation it inserts into parsed texts such as those in gure 4.2. This is a rule-
based parser inuenced by the view of grammar articulated in dependency grammars, that
is, grammars such as those described in Mel cuk (1987) that instead of grouping constituents
hierarchically (as parsers based on phrase structure grammars do), break each word down into
its constituent parts, noting dependency relationships between constituents. The parsed output in
0
1 The the det:>2 @DN> DET SG/PL
2 child child subj:>3 @SUBJ N NOM SG
3 broke break main:>0 @+FMAINV V PAST
4 his he attr:>5 @A> PRON PERS MASC GEN SG3
5 arm arm obj:>3 @OBJ N NOM SG
6 and and @CC CC
8 wrist wrist @SUBJ N NOM SG @OBJ N NOM SG @PCOMPL-S N NOM SG
@A>N NOM SG
9 and and cc:>3 @CC CC
11 mother mother subj:>12 @SUBJ N NOM SG
12 called call cc:>3 @+FMAINV V PAST
13 a a det:>14 @DN> DET SG
14 doctor doctor obj:>12 @OBJ N NOM SG
Figure 4.2 Parsed output from the EngFDG Parser
The example in gure 4.2 contains the coordinator and conjoining two noun
phrases, his armand his wrist, as well as two main clauses. Because coordinators
such as and can be used to conjoin phrases as well as clauses, it can be difcult
for parsers to distinguish phrasal from clausal coordination. In the example in
gure 4.2, the parser is unable to determine whether his wrist is coordinated
with his arm or his mother. Therefore, his wrist is assigned two function labels:
subject (@SUBJ) to indicate that his wrist and his mother are coordinated
noun phrases functioning as subject of the second clause; and object (@OBJ),
which is actually the correct parse, to indicate that his arm and his wrist are
coordinated noun phrases functioning as object of the verb broke in the rst
clause. The difculty that parsers have with coordinated constructions is further
compounded by the fact that coordination is so common (the one-million-word
ICE-GB contains over 20,000 instances of and), increasing the likelihood of
multiple parses for coordinated constructions.
An additional level of complexity is introduced if a parser is used to parse
spoken as well as written data. In this case, the parser has to deal with the
sheer ungrammaticality of speech, which poses major challenges for parsers,
especiallythose that are rule-based. For instance, the utterance below, takenfrom
gure 4.2 was obtained by using the demonstration version of the EngFDG parser available on
the Conexor website at: http://www.conexor./testing.html#1.
ICE-USA, is typical of the ungrammatical nature of speech: it has numerous
false starts, leading to a series of partially formed sentences and clauses:
but anyway you were saying Peggy that well I was asking you the general question what
um how when youve taken courses in the linguistics program have you done mainly
just textbook stuff or have you actually this is actual hardcore linguistic scholarship
(cited in Greenbaum 1992: 174)
In a rule-based parser, it is impossible to write rules for utterances containing
repetitions, since such utterances have no consistent, predictable structure and
thus conform to no linguistic rule. In addition, while a written text will contain
clearly delineated sentences marked by punctuation, speech does not. Therefore,
with examples such as the above, it is difcult to determine exactly what should
be submitted to the parser as a parse unit: the stretch of language the parser
is supposed to analyze (Gerald Nelson, personal communication).
The increased complexity of structures that a parser has to analyze leads
to multiple parses in many cases, making the process of tag disambiguation
more complex than it is with the output of a tagger. With tagging, disambigua-
tion involves selecting one tag from a series of tags assigned to an individual
word. With parsing, in contrast, disambiguation involves not single words but
higher-level constituents; that is, instead of being faced with a situation where
it has to be decided whether a word is a noun or a verb or adjective, the an-
alyst has to select the correct parse tree from any number of different parse
trees generated by the parser. In using the TOSCA parser, for instance, Willis
(1996: 238) reports that after running the parser on a sentence, if a single cor-
rect parse tree was not generated, he had to select from a dozen to (on some
occasions) 120different parse trees. There is software tohelpinthis kindof post-
editing, such as ICETree (described at: http://www.ucl.ac.uk/english-usage/ice-
gb/icetree/download.htm), a general editing program that was designed to aid
in the post-editing of corpora tagged and parsed for the ICE project, and the
Nijmegen TOSCA Tree editor (described at: http://lands.let.kun.nl/TSpublic/
tosca/page4.html), which can also be used to post-edit the output of the TOSCA
parser. But still, any time there is manual intervention in the process of cre-
ating a corpus the amount of time required to complete the corpus multiplies
exponentially.
To increase the accuracy of parsing and therefore reduce the amount of post-
editing that is required, parsers take differing approaches. The TOSCA Parser
requires a certain amount of manual pre-processing of a text before it is sub-
mitted to the parser: syntactic markers are inserted around certain problematic
constructions (cf. Oostdijk 1991: 2637; Quinn and Porter 1996: 714), and
spoken texts are normalized. To ensure that coordinated structures are correctly
parsed, their beginnings and ends are marked. In the sentence The child broke
his arm and his wrist and his mother called a doctor (the example cited above),
markers would be placed around his arm and his wrist to indicate that these
two noun phrases are coordinated, and around the two main clauses that are
coordinated to indicate that the coordinator and following wrist conjoins two
clauses. Other problematic constructions that had to be marked prior to pars-
ing were noun phrase postmodiers, adverbial noun phrases, appositive noun
phrases, and vocatives.
To successfully parse instances of dysuency in speech, the parser requires
that, in essence, an ungrammatical utterance is made grammatical: markup has
to be manually inserted around certain parts of an utterance that both tells the
parser to ignore this part of the utterance and creates a grammatically well-
formed structure to parse. In the example below, the speaker begins the utter-
ance with two repetitions of the expression can I before nally completing the
construction with can we.
can I can I can we take that again (ICE-GB S1A-001016)
To normalize this construction, markup is inserted around the two instances
of can I (<-> and </->) that tells the parser to ignore these two expres-
sions and parse only can we take that again, which is a grammatically well
formed:
<sent> <}> <-> Can I can I </-> <=> can we </=> </}> take that again <$?>
Arguably, normalization compromises the integrity of a text, especially if the
individual doing the normalization has to make decisions about how the text
is to be edited. And because normalization must be done manually, it can be
time-consuming as well. But normalization has advantages too. In constructions
such as the above, if the repetitions are not annotated, they will be included in
any lexical analysis of the corpus, leading to inaccurate word counts, since can,
for instance, will be counted three times, even though it is really being used
only once. In addition, even though repetitions and other features of speech are
excluded from analysis, if they are properly annotated, they can be recovered
if, for instance, the analyst wishes to study false starts.
Even though parsing a corpus is a formidable task, there do exist a num-
ber of parsed corpora available for linguistic research, though many of them
are relatively short by modern standards. In addition to the parsed corpora de-
scribed above the Nijmegen Corpus (130,000 words), the Lancaster Parsed
Corpus (140,000 words), and ICE-GB (one million words) there are also
the Lancaster/IBM Spoken English Corpus (52,000 words), the Polytechnic
of Wales Corpus (65,000 words), the PennHelsinki Parsed Corpus of Middle
English (1.3 million words), and the Susanne Corpus (128,000 words). The Penn
Treebank is larger (3,300,000 words) than all of these corpora, but aside from
the texts taken from the Brown Corpus, this corpus is not balanced, consisting
primarily of reportage from Dow Jones news reports. The ENGCG Parser (an
earlier version of the EngFDG Parser) has been used to parse sections of the
Bank of English Corpus (over 100 million words as of 1994), with the goal of
parsing the entire corpus (cf. J arvinen 1994 for details).
4.4 Other types of tagging and parsing 97
4.4 Other types of tagging and parsing
The previous sections described the tagging of words and the parsing of
constituents into syntactic units. While these processes are very well established
in the eld of corpus linguistics, there are other less common types of tagging
and parsing as well.
4.4.1 Semantic tagging
Semantic tagging involves annotating a corpus with markup that spec-
ies various features of meaning in the corpus. Wilson and Thomas (1997)
describe a number of systems of semantic tagging that have been employed. In
each of these systems, words in corpora are annotated with various schemes
denoting their meanings. For instance, in one scheme, each word is assigned a
semantic eld tag (Wilson and Thomas 1997: 61): a word such as cheeks is
given the tag Body and Body Parts, a word such as lovely the tag Aesthetic
Sentiments, and so forth. Schemes of this type can be useful for the creation
of dictionaries, and for developing systems that do content analysis, that is,
search documents for particular topics so that users interested in the topics can
have automatic analysis of a large database of documents.
4.4.2 Discourse tagging
In addition to annotating the meaning of words in corpora, semantic
systems of tagging have also looked at such semantic phenomena as anaphora,
the chaining together of co-referential links in a text. This is a type of discourse
tagging, whereby features of a text are annotated so that analysts can recover the
discourse structure of the text. Rocha (1997) describes a system he developed
to study topics in conversations. In this system, Rocha (1997) developed an
annotation scheme that allowed him to classify each anaphor he encountered in
a text and mark various characteristics of it. For instance, each anaphor receives
annotation specifying the type of anaphor it is: personal pronouns are marked
as being either subject pronouns (SP) or object pronouns (OP); demonstratives
as (De); possessives as (PoP); andsoforth(Rocha 1997: 269). Other information
is also supplied, such as whether the antecendent is explicit (ex ) or implicit
(im ). Rocha (1997) used this scheme to compare anaphor resolution in English
and Portuguese, that is, what chain of links are necessary to ultimately re-
solve the reference of any anaphor in a text. Rochas (1997: 277) preliminary
results show that most anaphors are personal pronouns, and their reference is
predominantly explicit.
4.4.3 Problem-oriented tagging
De Haan (1984) coined the term problem-oriented tagging to de-
scribe a method of tagging that requires the analyst to dene the tags to be used
and to assign them manually to the constructions to be analyzed. For instance,
Meyer (1992) used this method to study appositions in the Brown, London
Lund, and Survey of English Usage Corpora. Each apposition that Meyer (1992)
identied was assigned a series of tags. An apposition such as one of my closest
friends, John in the sentence I called one of my closest friends, John would be
assigned various tag values providing such information as the syntactic form
of the apposition, its syntactic function, whether the two units of the apposition
were juxtaposed or not, and so forth. Within each of these general categories
were a range of choices that were assigned numerical values. For instance,
Meyer (1992: 1367) found that the appositions in the corpora he analyzed had
seventy-eight different syntactic forms. The above apposition had the form of
an indenite noun phrase followed by a proper noun, a form that was assigned
the numerical value of (6). By assigning numerical values to tags, the tags could
be subjected to statistical analysis.
Because problem-oriented tagging has to be done manually, it can be very
time-consuming. However, there is a software program, PC Tagger, that can be
used to expedite this type of tagging (Meyer and Tenney 1993). The advantage
of problem-oriented tagging is that the analyst can dene the tags to be used
and is not be constrained by someone elses tagset. In addition, the tagging can
be quite detailed and permit grammatical studies that could not be carried out
on a corpus that has only been lexically tagged.
The termproblem-orientedtaggingsuggests that this type of analysis involves
mainly tagging. But it is really a process of parsing as well: constituents larger
than the word are assigned syntactic labels.
4.5 Conclusions
In recent years, corpus linguists of all persuasions have been ac-
tively involved in developing systems of annotation for corpora. The Linguistic
Data Consortium (LDC) at the University of Pennsylvania, for instance, has
a whole web page of links to various projects involved in either annotat-
ing corpora or developing tools to annotate them (Linguistic Annotation:
http://www.ldc.upenn.edu/annotation/). This is a signicant and important de-
velopment in corpus linguistics, since an annotated corpus provides the corpus
user with a wealth of important information.
But while it is important that systems of annotation are developed, it is equally
important that corpus linguists develop tools that help reduce the amount of
time it takes to annotate corpora and that help users understand the complex
systems underlying many annotation systems. The Text Encoding Initiative
(TEI) has developed a comprehensive system of structural markup. However, it
is very time-consuming to insert the markup, and many users may have difculty
understanding how the TEI system actually works. Likewise, parsing a corpus
is a very involved process that requires the corpus creator not just to spend
Study questions 99
considerable time post-editing the output of the parser but to have a conceptual
understanding of the grammar underlying the parser. Because the process of
annotating corpora involves the manual intervention of the corpus creator at
various stages of annotation, in the foreseeable future, annotating a corpus will
continue to be one of the more labor-intensive parts of creating a corpus.
Study questions
1. Why is it necessary for a corpus in ASCII text format to contain structural
markup? What kinds of features would users of such a corpus not be able to
recover if the samples within the corpus did not contain structural markup?
2. How would a probabilistic tagger determine that in the sentence The child
likes to play with his friends, the word play is a verb rather than a noun?
3. Make up a short sentence and submit it to the online taggers below:
CLAWS part-of-speech tagger:
http://www.comp.lancs.ac.uk/computing/research/ucrel/claws/trial.html
EngCG-2:
http://www.conexor./testing.html#1
Do the two taggers assign similar or different part-of-speech labels to your
sample sentence? Briey summarize the similarities and differences.
4. Why is parsing a much more difcult and involved process than tagging?
5 Analyzing a corpus
The process of analyzing a completed corpus is in many respects sim-
ilar to the process of creating a corpus. Like the corpus compiler, the corpus
analyst needs to consider such factors as whether the corpus to be analyzed is
lengthy enough for the particular linguistic study being undertaken and whether
the samples in the corpus are balanced and representative. The major difference
between creating and analyzing a corpus, however, is that while the creator of a
corpus has the option of adjusting what is included in the corpus to compensate
for any complications that arise during the creation of the corpus, the corpus
analyst is confronted with a xed corpus, and has to decide whether to continue
with an analysis if the corpus is not entirely suitable for analysis, or nd a new
corpus altogether.
This chapter describes the process of analyzing a completed corpus. It begins
with a discussion of how to frame a research question so that from the start,
the analyst has a clear hypothesis to test out in a corpus and avoids the
common complaint that many voice about corpus-based analyses: that many
such analyses do little more than simply count linguistic features in a corpus,
payinglittle attentiontothe signicance of the counts. The next sections describe
the process of doing a corpus analysis: howto determine whether a given corpus
is appropriate for a particular linguistic analysis, how to extract grammatical
information relevant to the analysis, how to create data les for recording the
grammatical information taken from the corpus, and how to determine the
appropriate statistical tests for analyzing the information in the data les that
have been created.
To illustrate the process of analyzing a corpus, this chapter will focus on a spe-
cic corpus analysis: the study of a particular grammatical construction, termed
the pseudo-title (Bell 1988), in various components of the International Cor-
pus of English (ICE). Pseudo-titles are constructions such as linguist in linguist
Noam Chomsky that occur in positions in which we normally nd titles like
Professor or Doctor. However, unlike titles, pseudo-titles do not function as
honorics (markers of deference and respect) but instead provide descriptive
information about the proper nouns that they precede (e.g. that NoamChomsky
is a linguist). The case study approach is taken in this chapter to avoid us-
ing disparate and unrelated examples to illustrate the process of doing a corpus
analysis and to demonstrate that each stage of corpus analysis is related.
1
1
For additional case studies, see Biber, Conrad, and Reppen (1998).
100
5.1 The pseudo-title in English: framing a research question 101
5.1 The pseudo-title in English: framing a
research question
To begin to frame a research question about pseudo-titles, it is rst
of all necessary to understand the history of the construction and how it has
evolved in English.
Pseudo-titles are thought to have originated in American English, specically
in journalistic writing. Quirk et al. (1985: 276, note) refer to pseudo-titles as
characteristic of Timestyle, a style of journalistic writing popularized by
Time magazine. Although pseudo-titles may have originated in American press
writing, they can currently be found in the press writing of both British and
New Zealand English. However, as Ryd en (1975), Bell (1988), and Meyer
(1992) have documented, while pseudo-titles are not stylistically marked in
American English they occur widely in all types of American newspapers and
magazines they are stigmatized in British English: their use is avoided in more
formal newspapers such as The Times and the Guardian, and they are found
mainly in newspapers associated with what is nowtermed tabloid journalism.
The stigma against pseudo-titles in British English is well documented in
style guides for British newspapers. For instance, in The Independent Style
Manual, it is remarked that pseudo-titles should be avoided because peoples
jobs are not titles (p. 6). In newspapers prohibiting the use of pseudo-titles, the
information contained in the pseudo-title is instead expressed in an equivalent
appositional construction (e.g. the linguist Noam Chomsky or Noam Chomsky,
a linguist). But while style guides may prohibit certain usages such as the
pseudo-title, it is an open question whether practice follows prescription, and
a well-designed corpus study can be used to examine the relationship between
prescription and practice. Bell (1988), for instance, studied 3,500 instances of
pseudo-titles taken from a large corpus of British, American, and New Zealand
newspapers. Using this corpus, Bell (1988) was able to conrm that while
pseudo-titles were quite common in American press writing, they were absent
from British newspapers that prohibited their usage. Additionally, Bell (1998)
was able to document that New Zealand press writing had moved away from
the British norm for pseudo-titles to the American norm, that is, that over the
years, pseudo-titles had gained in prominence and acceptability in NewZealand
press writing. This change in usage norms, Bell (1988: 3267) argues, reects
the fact that the colonial inuence of British culture on New Zealand culture
had been weakening in recent years, and that New Zealand English was
refocusing towards the United States . . . , a change that has led New Zealand
English closer to the American normfor pseudo-titles. The size of Bells (1988)
corpus also enabled him to study the linguistic structure of pseudo-titles, and to
document, for instance, the extent that newspapers favored pseudo-titles over
equivalent appositional structures, and to showthat there were certain linguistic
structures that favored or disfavored the use of a pseudo-title. For instance, he
found that pseudo-titles favored minimal post-modication. Consequently, a
construction such as lawyer Frederick Smith was more common than ?lawyer
for Hale, Brown, and Jones Frederick Smith.
To evaluate whether Bells (1988) study of pseudo-titles uses a corpus to an-
swer valid research questions, and to determine whether his study warrants fur-
ther corpus studies of pseudo-titles, it is important to understand the kinds of lin-
guistic evidence that corpora are best able to provide. Since the Chomskyan rev-
olution of the 1950s, linguistics has always regarded itself as a science in which
empirical evidence is used to advance linguistic theory. For the Chomskyan
linguist, this evidence was often obtained through introspection: the gathering
of data based on the linguists own intuitions. Because many corpus linguists
have found this kind of evidence limiting (cf. section 1.1), they have turned to the
linguistic corpus as a better source of empirical evidence for real rather than
contrived examples of linguistic constructions and for statistical information
on how frequently a given linguistic construction occurs in a corpus.
In turning to the corpus for evidence, however, many corpus linguists have
regarded the gathering of evidence as a primary goal: in some corpus studies,
it is not uncommon to see page after page of tables with statistical information
on the frequency of grammatical constructions, with little attention paid to the
signicance of the frequencies. Corpus-based research of this nature, Aarts
(2001: 7) notes,
invariably elicits a so what response: so what if we know that there are 435 instances
of the conjunction because in a particular category of written language, whereas there
are only 21 instances in conversations? So what if we are told that subordination is much
more common in womens speech than in mens speech?
To move beyond simply counting features in a corpus, it is imperative before
undertaking a corpus analysis to have a particular research question in mind,
and to regard the analysis of a corpus as both qualitative and quantitative
research research that uses statistical counts or linguistic examples to test a
clearly dened linguistic hypothesis.
In using a corpus to study pseudo-titles, Bell (1988) has obviously done more
thansimplycount the number of pseudo-titles inhis corpus. He has useda corpus
to, for instance, describe the particular linguistic structures that pseudo-titles
favor, and he has used frequency counts to document the increased prevalence
of pseudo-titles in NewZealand press writing. Because Bells (1988) study sug-
gests that pseudo-titles may be spreading beyond British and American English
to other regional varieties of English, his ndings raise additional questions
worthy of further investigation of corpora:
(a) To what extent have pseudo-titles spread to other varieties of English, and
in these varieties, is the American norm followed or the British norm?
(b) Do pseudo-titles have the same structure in these varieties that they have in
British, American, and New Zealand English?
5.2 Determining whether a corpus is suitable for answering a particular research question 103
(c) To what extent do newspapers in these varieties prefer pseudo-titles over
equivalent appositional structures?
The rst step in addressing these questions is to determine whether suitable
corpora exist for answering them.
5.2 Determining whether a corpus is suitable for
answering a particular research question
Bells (1988) study of pseudo-titles was based on a large corpus of
British, American, and NewZealand press writing that Bell assembled himself.
Although this is the ideal way to ensure that the corpus being used is appropriate
for the analysis being conducted, this approach greatly increases the work that
the corpus analyst must do, since creating ones own corpus prior to analyzing
it is obviously a large undertaking. It is therefore most desirable to work with
a corpus already available not just to decrease the work time but to add to
the growing body of work that has been based on a given corpus. The Brown
Corpus, for instance, has been around for close to forty years and has become
a type of benchmark for corpus work: a corpus on which a large amount of
research has been conducted over the years research that has generated results
that can be compared and replicated on other corpora, and that has added greatly
to our knowledge of the types of written English included in the Brown Corpus.
To study the spread of pseudo-titles to various regional varieties of English,
it is necessary not only to nd corpora of differing varieties of English but to
make sure that the corpora selected are as comparable as possible, that is, that
they contain texts that have been collected during a similar time period, that
represent similar genres, that are taken from a variety of different sources, and
that are of the appropriate length for studying pseudo-title usage. In other words,
the goal is to select corpora that are as similar as possible so that extraneous
variables (such as texts covering different time periods) are not introduced into
the study that may skew the results.
Since pseudo-titles have been extensively studied in British, American, and
New Zealand English, it might appear that corpora of these varieties need not
be sought, since analyzing these varieties again will simply repeat what others
have already done. However, there are important reasons why corpora of British,
American, and New Zealand English ought to be included in the study. First,
it is always desirable to have independent conrmation of results obtained in
previous studies. Such conrmation ensures that the results have validity beyond
the initial study from which they were obtained. Second, the previous studies of
pseudo-titles were conducted on very different kinds of corpora. Bells (1988)
study was conducted on a privately created corpus; Meyers (1992) study was
based on the Brown Corpus and the London Corpus, corpora consisting of
samples of written British and American English collected as far back as the late
1950s. Comparing very different corpora increases the potential that extraneous
variables will inuence the results. For instance, because Bells (1988) study
suggests that pseudo-title usage seems to be increasing over time, it is important
to study corpora collected during a similar time period so that the variable of
time does not distort the results.
There are several well-establishedcorpora that potentiallyqualifyas resources
for studying pseudo-title usage: the Brown and LOB corpora of American
and British English, respectively (and their updated counterparts: the Freiburg
Brown Corpus [FROWN] and Freiburg LOB Corpus [FLOB]); the Wellington
Corpus of New Zealand English; and the Kolhapur Corpus of Indian English.
Because each of these corpora has a similar design (e.g. comparable written
genres divided into 2,000-word samples), they are very suitable for compar-
ison. However, the Wellington Corpus contains texts collected at a later date
than the other corpora: 1986-90 as opposed to 1961. FLOB and FROWN con-
tain texts from 1991. These corpora thus introduce a time variable that is less
than desirable. If these corpora are eliminated from consideration, only three
corpora remain (Brown, LOB, and Kolhapur), and only one of these corpora
the Kolhapur Corpus of Indian English represents a variety of English in
which pseudo-titles have never before been studied. It is therefore obvious that
a different set of corpora are necessary if any information is to be obtained
on the spread of pseudo-titles to other varieties of English than American and
British English.
A better choice of corpora to consider are those included in the International
Corpus of English (ICE). The regional components of the ICE contain not just
identical genres but samples of speech and writing collected during a similar
time-frame (1990present). In addition, ICE corpora offer numerous regional
varieties not just British, American, and New Zealand English but Philippine,
Singaporean, Jamaican, and East African English.
2
The ICE corpora therefore
offer a potentially useful dataset for studying pseudo-titles. But to determine the
full potential of the ICEcorpora, two further points must be considered: whether
the components of ICE selected for analysis provide a representative sampling
of the kind of writing that pseudo-titles occur in, and if the components do,
whether the samples selected from the corpora are of the appropriate length to
nd a sufcient number of pseudo-titles and corresponding appositives.
Inhis studyof pseudo-titles, Bell (1988: 327) restrictedhis analysis topseudo-
titles occurring in hard news reporting and excluded fromconsideration other
types of press writing, such as editorials or feature writing. Because each ICE
corpus contains twenty samples of press reportage as well as twenty samples of
broadcast news reports, it will be quite possible to base an analysis of pseudo-
titles on the kinds of press reportage that Bell (1988) analyzed in his study.
The next question, however, is whether the analysis should be based on all
2
The ICE Project contains other varieties of English than these, but these were the only varieties
available at the time this study was conducted.
5.2 Determining whether a corpus is suitable for answering a particular research question 105
forty texts from each ICE corpus, or some subset of these texts. There are a
number of considerations that will affect choices of this nature. First of all, if
the construction to be analyzed has to be extracted by manual analysis, then
it becomes imperative to use as short a corpus as possible, since extensive
manual analysis will add to the length of time a given study will take. As will be
demonstrated in section 5.3, while pseudo-titles can be automatically extracted
from the British component of ICE because it is fully tagged and parsed, they
must be retrieved manually fromthe other components, since unlike, say, regular
verbs in English, which have verb sufxes such as -ing or -ed, pseudo-titles have
no unique marker that can be searched for. Therefore, to minimize the analysis
time, it is advisable to examine only a subset of the eligible texts in the various
ICE components.
One possibility is to analyze ten samples of press reporting and ten samples
of broadcast news from each ICE corpus. Another possibility is to analyze only
the twenty samples of press reportage from each corpus, or only the twenty
samples of broadcast news. Ultimately, it was decided to restrict the sample to
the twenty samples of press reportage, primarily because broadcast news is a
mixed category, containing not just scripted monologues but interviews as well.
Since interviews are not strictly hard news reporting, they are less likely to
contain pseudo-titles.
3
Moreover, since the goal of the study is to determine how
widespread pseudo-titles are in press reportage, it is quite sufcient to attempt
to investigate as many different newspapers as possible from the various ICE
corpora, to determine how common pseudo-title usage is.
Having restricted the number of texts to be examined, it is next necessary
to examine the individual samples from each component of ICE to ensure that
the samples represent many different newspapers rather than a fewnewspapers.
In other words, if a given component of ICE contains twenty samples taken
from only ve different newspapers, the samples to be investigated are less
than representative, since they provide information only on how ve different
newspapers use pseudo-titles. With a sample this small, it will be difcult to
make strong generalizations about the usage of pseudo-titles in general within
the press writing of a given variety of English. Table 5.1 lists the number of
different newspapers represented in the samples selected for study from the
various components of ICE.
As table 5.1 illustrates, the components vary considerably in terms of the
number of different newspapers that are represented. While ICE-USA contains
3
Robert Sigley (personal communication) notes that pseudo-titles would be perfectly appropriate
in a broadcast interview as a way of introducing an interviewee: We will be speaking with
Boston Globe Editor Robert Storrin. In ICE-GB, the only pseudo-title found in a broadcast
interview (designer in designer Anthony Baker) occurred in an interviewees description of a
play he recently saw:
The staging by Francisco Negrin and designer Anthony Baker wittily plays off the opera house
and its conventions nally allowing us to glimpse a three tier gilt and scarlet plush auditorium
into which the cast make their escape (ICE-GB:S1B-044 #2:1:A; emphasis added)
Table 5.1 Number of different newspapers included
within the various ICE components
Component Total
Great Britain 15
United States 20
New Zealand 12
Philippines 10
Jamaica 3
East Africa 3
Singapore 2
twenty different newspapers, ICE-East Africa contains only three. The relatively
few newspapers represented in some of the components is in part a reection
of the fact that in an area such as East Africa, in which English is a second
language, there exist only a few English language newspapers. Therefore, three
newspapers fromthis region is a representative sample. Still, the results obtained
from such a small sample will have to be viewed with some caution.
Because all ICE corpora contain 2,000-word samples, it is also important
to consider whether text fragments, rather than entire texts, will yield enough
examples of pseudo-titles and, additionally, whether pseudo-titles serve some
kind of discourse function, necessitating that a whole newspaper article be
analyzed as opposed to simply a short segment of it.
4
One way to answer these
questions is to do a pilot study of a few randomly selected samples taken
from each ICE component. Table 5.2 lists the number of pseudo-titles and
corresponding appositive structures occurring in nine samples from various
randomly selected ICE corpora.
Table 5.2 demonstrates that even though the frequency of the various struc-
tures being investigated differs in the nine newspapers, the structures do occur
with some regularity in the 2,000-words samples: as many as sixteen times in
one sample from ICE-New Zealand to as few as ve times in one sample from
ICE-Philippines. Of course, ultimately statistical tests will have to be applied
to determine the true signicance of these numbers. But at this stage, it seems
evident that pseudo-titles are such a common grammatical construction that
adequate numbers of them can be found in short 2,000-word excerpts.
One drawback of analyzing text fragments is that certain discourse features
cannot be studied in them because only partial texts (rather than whole texts)
are available for analysis. In particular, if one is investigating a construction
that contributes to the structure of a given text, then a text fragment will prove
insufcient for carrying out such a study. Because pseudo-titles serve mainly
4
Many ICE samples are actually lengthier than 2,000 words. However, the text of a sample beyond
2,000 words is marked as extra-corpus text and for this study was excluded from analysis.
5.3 Extracting information from a corpus 107
Table 5.2 The number of pseudo-titles and corresponding appositives in
selected newspapers from various ICE corpora
Pseudo-title Equivalent appositive Total
Independent (ICE-GB) 9 9
Guardian (ICE-GB) 7 7
Press (ICE-NZ) 13 13
Daily Nation (ICE-EA) 4 9 13
NY Times (ICE-US) 2 6 8
Jamaica Herald (ICE-JA) 2 4 6
Phil. Star (ICE-PH) 4 1 5
Manila Standard (ICE-PH) 11 0 11
Dominion (ICE-NZ) 16 0 16
to provide descriptive information about individuals, they play no role in the
overall structure or organization of a text. Therefore, it is perfectly justiable to
study pseudo-title usage in the 2,000-word excerpts occurring in samples from
ICE corpora, particularly if the goal is to study the structure and frequency
of pseudo-titles and corresponding appositives in various national varieties of
English.
5.3 Extracting information from a corpus
After a research question has been framed and a corpus selected to
analyze, it is next necessary to plan out exactly what kinds of grammatical
informationwill be extractedfromthe corpus, todetermine howthis information
will be coded and recorded, and to select the appropriate software that can most
efciently assist in nding the construction being investigated in the corpus
being studied.
5.3.1 Dening the parameters of a corpus analysis
The rst step in planning a corpus analysis is the development of a
clear working denition of the grammatical construction(s) to be studied. If a
corpus analysis is begun without an adequate denition of terms, the analyst
runs the risk of introducing too much inconsistency into his/her analysis. In
the case of pseudo-titles, this becomes an especially crucial issue because re-
search has shown that there is not always a clear boundary between a full title
and a pseudo-title. If some kind of working denition of each kind of title is
not determined prior to analysis, it is highly likely that as the analysis is con-
ducted what is counted as a pseudo-title will vary from instance to instance,
and as a consequence, considerable inconsistency will result that will seriously
compromise the integrity of the ndings. Fortunately, Bell (1988) has specied
quite precisely the kinds of semantic categories into which full titles can be
categorized, and his categories can be used to distinguish what counts as a title
from what counts as a pseudo-title:
Professional (Doctor, Professor)
Political (President, Chancellor, Senator)
Religious (Bishop, Cardinal, Mother)
Honors (Dame, Earl, Countess)
Military (General, Corporal )
Police (Commissioner, Constable, Detective-Sergeant)
Foreign (Monsieur, Senorita) (Bell 1988: 329)
In the present study, then, a pseudo-title will be dened as any construction
containing an initial unit (such as city employee in city employee Mark Smith)
that provides descriptive information about an individual and that cannot be
classied into any of the above semantic classes.
Because pseudo-titles will be compared to corresponding appositives, it is
alsonecessarytohave a workingdenitionof the types of appositives that will be
considered equivalent to a pseudo-title. Bell (1988) claims that pseudo-titles
and equivalent appositives are related through a process he terms determiner
deletion. Thus, for Bell (1988: 328), a pseudo-title such as fugitive nancier
Robert Vesco would be derived from an appositive containing the determiner
the in the rst unit: the fugitive nancier Robert Vesco. However, if ones goal
is to study the relationship between the use of pseudo-titles and all possible
equivalent appositives, then a wider range of appositive structures needs to
be investigated, largely because there exist corresponding structures where the
unit of the appositive equivalent to the pseudo-title occurs after, not before, the
noun phrase it is related to, and in such constructions, a determiner is optional:
Robert Vesco, [the] fugitive nancier. In the current study, then, a corresponding
appositive will include any appositive that is related to a pseudo-title through
the process systematic correspondence (Quirk et al. 1985: 57f.), a process that
species that two constructions are equivalent if they contain roughly the same
lexical content and have the same meaning. Although constructions such as
fugitive nancier Robert Vesco and Robert Vesco, fugitive nancier certainly
differ in emphasis and focus, they are close enough in form and meaning that
they can be considered equivalent structures.
Of course, a working denition is just that: a preliminary denition of a
grammatical category subject to change as new data are encountered. And as
anyone who has conducted a corpus study knows, once real data are exam-
ined, considerable complications can develop. As the study of pseudo-titles was
conducted, examples were found that required modication of the initial deni-
tion (cf. section 5.3.3 for a discussion of the actual methods that were employed
to locate pseudo-titles). For instance, the construction former Vice President
Dan Quayle (ICE-USA) contains an initial unit, former Vice President, that
is semantically a Political designation and therefore qualies as a title; more-
over, Vice President is capitalized, a common orthographic characteristic of
titles. But the inclusion of the adjective former before Vice President suggests
that the focus here is less on honoring Dan Quayle and more on his pre-
vious occupation as Vice President of the United States. Consequently, this
construction (and others like it) was counted as a pseudo-title, even though
Vice President has semantic characteristics normally associated with titles.
Decisions of this nature are common in any corpus analysis that involves
the close examination of data, and there are basically two routes that the cor-
pus analyst can take to decide how to classify problematic data. As was done
above, if enough evidence exists to place a construction into one category
rather than another, then the construction can be classied as x rather than
y, even though the construction may have characteristics of both x and
y. Alternatively, the analyst can create an ad hoc category in which construc-
tions that cannot be neatly classied are placed. Either approach is acceptable,
provided that whichever decision is made about the classication of a partic-
ular grammatical construction is consistently applied throughout a given cor-
pus analysis, and that the decisions that are made are explicitly discussed in
the article or book in which the results of the study are reported. Moreover,
if an ad hoc category is created during a corpus analysis, once the analysis
is completed, all examples in the category should be examined to determine
whether they have any characteristics in common that might lead to additional
generalizations.
5.3.2 Coding and recording grammatical information
Once key terms and concepts have been dened, it is next necessary
to decide what grammatical information needs to be extracted from the corpus
being examined to best answer the research questions being investigated, and
to determine how this information can best be coded and recorded. To nd
precise answers to the various linguistic questions posed about pseudo-titles in
section 5.1, the following grammatical information will need to be obtained:
Information A: To determine howwidespread pseudo-title usage has become,
and to reveal which newspapers permit the use of pseudo-titles and which prefer
using only equivalent appositive constructions, it will be necessary to obtain
frequency counts of the number of pseudo-titles and equivalent appositives
occurring in the various regional varieties of ICE being investigated.
Information B: To determine how common pseudo-titles are in newspapers
that allowtheir use, it will be important to examine the role that style plays in the
use of pseudo-titles and equivalent appositives, and to ascertain whether given
the choice between one construction and the other, a particular newspaper will
choose a pseudo-title over a corresponding appositive construction. To study
stylistic choices, three types of correspondence relationships will be coded:
those in which an appositive can be directly converted into a pseudo-title:
(1) Durk Jager, executive vice president (ICE-USA)
executive vice president Durk Jager
those in which only a determiner needs to be deleted for the appositive to be
converted into a pseudo-title:
(2) the Organising Secretary, Mr Stephen Kalonzo Musyoka (ICE-East Africa)
Organising Secretary Mr Stephen Kalonzo Musyoka
and those in which only part of the appositive can be converted into a pseudo-
title:
(3) Peter Houliston, Counsellor andProgramme Director of Development at the Canadian
High Commission (ICE-Jamaica)
Counsellor and Programme Director Peter Houliston
Information C: To investigate why only part of the appositive can form a
pseudo-title in constructions such as (3), it is necessary to study the linguistic
considerations that make conversion of the entire unit of an appositive to a
pseudo-title stylistically awkward. Bell (1988: 336) found that an appositive
construction was favored (4a and 5a) over an equivalent pseudo-title (4b and
5b) if the unit that could potentially become a pseudo-title contained either a
genitive noun phrase (e.g. bureaus in 4a) or postmodication of considerable
complexity (e.g. . . . of the CIA station in Rome in the 1970s in 5a):
(4) a. the bureaus litigation and prosecution division chief Osias Baldivino (ICE-
Philippines)
b. ?*bureaus litigation and prosecution division chief Osias Baldivino
(5) a. Ted Shackley, deputy chief of the CIA station in Rome in the 1970s (ICE-
GB:W2C-010 #62:1)
b. ?deputy chief of the CIA station in Rome in the 1970s Ted Shackley
Information D: It was decided to record the length of a pseudo-title after a
casual inspection of the data seemed to suggest that in the Philippine component
there was greater tolerance for lengthy pseudo-titles (6) than for what seem-
ed to be the norm in the other varieties: pseudo-titles of only a few words in
length (7).
(6) Salamat and Presidential Adviser on Flagship Projects in Mindanao Robert Aven-
tajado (ICE-Philippines)
(7) Technology editor Kenneth James (ICE Singapore)
Often after a corpus analysis has begun, newinformation will be discovered that
is of interest to the study being conducted and that really needs to be recorded.
Usually such discoveries can be easily integrated into the study, provided that
the analyst has kept accurate record of the constructions that have already been
recorded, and has coded the data in a manner that allows for the introduction of
new grammatical information into the coding scheme.
Data can be coded and recorded in numerous ways. Because the goal of
the current study is to record specic information about each construction be-
ing studied, it was decided to use a type of manual coding (or tagging) called
problem-oriented tagging (De Haan 1984; cf. also section 4.4.3). This method
of tagging allows the analyst to record detailed information about each gram-
matical construction under investigation. Table 5.3 outlines the coding system
used for recording the grammatical information described in (A)(D) above.
The coding scheme in table 5.3 allows for every pseudo-title or equivalent
appositive construction to be described with a six-sequence series of numbers
that not only label the construction as being a pseudo-title or appositive but
describe the regional variety in which the construction was found; the partic-
ular sample from which it was taken; the type of correspondence relationship
existing between an appositive and potential pseudo-title; the form of the con-
struction; and the length of the pseudo-title or the unit of the apposition that
could potentially be a pseudo-title. Coding the data this way will ultimately
allow for the results to be viewed from a variety of different perspectives. For
instance, because each construction is given a number identifying the regional
variety in which it occurred, it will be possible to compare pseudo-title usage
in, say, British English and Singapore English. Likewise, each sample repre-
sents a different newspaper. Therefore, by recording the sample from which a
construction was taken, it will be possible to know which newspapers permit
pseudo-title usage, and which do not, and the extent to which newspapers use
pseudo-titles and equivalent appositives similarly or differently.
According to the scheme in table 5.3, a construction such as nancial
adviser David Innes (ICE-GB:W2C-009 #41:2) would be coded in the fol-
lowing manner:
Country: Great Britain 6
Sample: W2C-009 9
Type: pseudo-title 1
Correspondence relationship 4
Form: simple NP 1
Length: two words 2
The advantage of using a coding scheme such as this is that by assigning a series
of numerical values to each construction being studied, the results can be easily
exported into any statistical program(cf. section 5.4 for a discussion of the kinds
of statistical programs into which quantitative data such as this can be exported).
The disadvantage of using numerical sequences of this nature is that they
increase the likelihood that errors will occur during the coding process. If, for
instance, a pseudo-title is mistakenly tagged as an apposition, it is difcult for
the analyst to recognize this mistake when each of these categories is given the
arbitrary number 1 and 2, respectively. Moreover, if each variable is coded
with an identical sequence of numbers, it will be quite difcult to determine
T
a
b
l
e
5
.
3
C
o
d
i
n
g
s
c
h
e
m
e
f
o
r
s
t
u
d
y
o
f
p
s
e
u
d
o
-
t
i
t
l
e
s
a
n
d
e
q
u
i
v
a
l
e
n
t
a
p
p
o
s
i
t
i
v
e
s
C
o
r
r
e
s
p
o
n
d
e
n
c
e
C
o
u
n
t
r
y
S
a
m
p
l
e
n
u
m
b
e
r
T
y
p
e
r
e
l
a
t
i
o
n
s
h
i
p
F
o
r
m
L
e
n
g
t
h
U
S
(
1
)
W
2
C
0
0
1
(
1
)
P
s
e
u
d
o
-
t
i
t
l
e
(
1
)
T
o
t
a
l
e
q
u
i
v
a
l
e
n
c
e
(
1
)
S
i
m
p
l
e
N
P
(
1
)
O
n
e
w
o
r
d
(
1
)
P
h
i
l
i
p
p
i
n
e
s
(
2
)
W
2
C
0
0
2
(
2
)
A
p
p
o
s
i
t
i
v
e
(
2
)
D
e
t
e
r
m
i
n
e
r
d
e
l
e
t
i
o
n
(
2
)
G
e
n
i
t
i
v
e
N
P
(
2
)
T
w
o
w
o
r
d
s
(
2
)
E
a
s
t
A
f
r
i
c
a
(
3
)
W
2
C
0
0
3
(
3
)
P
a
r
t
i
a
l
e
q
u
i
v
a
l
e
n
c
e
(
3
)
M
u
l
t
i
p
l
e
p
o
s
t
-
m
o
d
i
c
a
t
i
o
n
(
3
)
T
h
r
e
e
w
o
r
d
s
(
3
)
J
a
m
a
i
c
a
(
4
)
W
2
C
0
0
4
(
4
)
N
/
A
(
4
)
F
o
u
r
w
o
r
d
s
(
4
)
N
e
w
Z
e
a
l
a
n
d
(
5
)
W
2
C
0
0
5
(
5
)
F
i
v
e
w
o
r
d
s
(
5
)
G
r
e
a
t
B
r
i
t
a
i
n
(
6
)
e
t
c
.
S
i
x
o
r
m
o
r
e
w
o
r
d
s
(
6
)
S
i
n
g
a
p
o
r
e
(
7
)
Figure 5.1 PC-Tagger pop-up menu
whether a 2 assigned for Variable 1, for instance, is an error when the same nu-
merical value is used for each of the other variables being studied. To reduce the
incidence of errors, pseudo-titles and equivalent appositives were coded with a
software program called PC Tagger (described in detail in Meyer and Tenney
1993). This programdisplays both the actual variables being studied in a partic-
ular corpus analysis, and the values associated with each variable (cf. gure 5.1).
Figure 5.1 contains a pseudo-title selected fromICE-Philippines, MILF legal
counsel Ferdausi Abbas, and a pop-up window containing in the left column
the Tag Names (i.e. variables) and in the right column Tag Values. Because
the length variable is selected, only the values for this variable appear. Since
the pseudo-title is three words in length, the value three is selected. The
programproduces as output a separate le that converts all values to the numbers
in table 5.3 associated with the values, and that can be exported to a statistical
program. Because the analyst is working with the actual names of tags and their
values, the results can be easily checked and corrected.
An alternative to using a program such as PC Tagger is to use a system con-
taining codes that are mnemonic. For instance, the CHAT system (Codes for
the Human Analysis of Transcripts), developed within the CHILDES Project
(Child Language Data Exchange System; cf. section 1.3.8), contains numerous
mnemonic symbols for coding differing kinds of linguistic data. To describe
the morphological characteristics of the phrase our family, the codes 1P,
POSS, and PRO are used to describe our as a rst-person-plural possessive
pronoun; the codes COLL and N are used to describe family as a collective
noun (CHAT Manual: http://childes.psy.cmu.edu/pdf/chat.pdf ). For the current
study, a pseudo-title, for instance, could be coded as PT and a corresponding
apposition as AP. And even though a system such as this lacks numerical val-
ues, the mnemonic tags can be easily converted into numbers (e.g. all instances
of PT can be searched for and replaced with the value 1).
5.3.3 Locating relevant constructions for a particular
corpus analysis
The greatest amount of work in any corpus study will be devoted to lo-
cating the particular construction(s) being studied, and then assigning to these
constructions the particular linguistic values being investigated in the study.
In the past, when corpora were not available in computer-readable form, the
analyst had to painstakingly extract grammatical information by hand, a very
tedious process that involved reading through printed texts and manually record-
ing grammatical information and copying examples by hand. Now that corpora
exist in computer-readable form, it is possible to reduce the time it takes to
conduct a corpus analysis by using software programs that can automate (to
varying degrees) the extraction of grammatical information. In conducting a
grammatical analysis of a corpus, the analyst can either learn a programming
language such as Perl and then write scripts to extract the relevant gram-
matical information, or use a general-purpose software application, such as a
concordancing program (described in detail below), that has been developed
for use on any corpus and that can perform many common tasks (e.g. word
searches) associated with any corpus analysis. Both approaches to corpus anal-
ysis have advantages and disadvantages, and neither approach will guarantee
that the analyst retrieves precisely what is being sought: just about any corpus
analysis will involve sorting through the data manually to eliminate unwanted
constructions and to organize the data in a manner suitable to the analysis.
There exist a number of programming languages, such as Python, Visual
Basic, or Perl, that can be powerful tools for analyzing corpora. Fernquest
(2000) has written a number of Perl scripts that can, for instance, search a corpus
and extract from it lines containing a specic phrase; that can calculate word
frequencies, sorting words either alphabetically or by frequency of occurrence;
and that can count bigrams in a text (i.e. two-word sequences) and organize
the bigrams by frequency. Perl scripts are quite commonly available (cf., for
instance, Melamuds 1996 le of Perl scripts), and can be modied to suit the
specic grammatical analysis being conducted. Of course, much of what can
be done with many Perl scripts can be more easily accomplished with, say, a
good concordancing program. But as Sampson (1998) quite correctly observes,
if those analyzing corpora have programming capabilities, they do not have to
rely on software produced by others which may not meet their needs.
If one is using someone elses scripts, using the scripts on a corpus is not that
difcult. But writing new scripts requires programming skills that are probably
beyond the capabilities of the average corpus linguist. For these individuals,
there exist a number of very useful software programs that require no program-
ming skills and that can extract much useful information from corpora: the
Linguistic Data Base (LDB) for analyzing the Nijmegen Corpus (cf. section
4.3); ICECUP for analyzing the British component of ICE (described below);
Sara to analyze the British National Corpus (cf. Aston and Burnard 1998);
and a variety of different concordancing programs available for use on PCs,
Macintoshes, Unix work stations, even the World Wide Web (cf. appendix 2 for
a listing of concordancing programs and also Hockey 2000: 4965 for a survey
of some of the more common programs). Of course, as is the case with using any
new software program, the user will need to experiment with the program both
to learn how to use it and to determine whether it can do what the user needs to
accomplish in a given corpus analysis. Thus, before any corpus analysis is begun
in earnest, the corpus analyst will want to experiment with more than one con-
cordancing program to ensure that the best program is used for a given corpus
analysis.
The ability of such programs to retrieve grammatical information depends
crucially upon how easy it is to construct a search for a given grammatical con-
struction. In a lexical corpus (a corpus containing raw text and no grammatical
annotation), it is easiest to construct searches for particular strings of char-
acters: a particular word or group of words, for instance, or classes of words
containing specic prexes or sufxes (e.g. the derivational morpheme -un, or
a verb inection such as -ing). In corpora that have been tagged or parsed, it is
possible to search for particular tags and therefore retrieve the particular gram-
matical constructions (common nouns, adverbs, etc.) associated with the tags.
However, even in tagged and parsed corpora, it may be difcult to automatically
extract precisely what the analyst is studying. Therefore, just about any corpus
study will at some stage involve the manual analysis of data. Pseudo-titles pro-
vide a good illustration of the strengths and limitations of the various tools that
have been developed for corpus analysis.
A pseudo-title such as linguist Noam Chomsky consists of a common noun
followed by a proper noun. Because nouns are open-class items, they are a
very broad category. Consequently, it is highly likely that every pseudo-title
discovered in a given corpus will contain a unique series of common and proper
nouns. This characteristic of pseudo-titles makes it very difcult to use the most
common corpus tool the concordancing program to automatically retrieve
pseudo-titles.
The most common capability of any concordancing program is to conduct
simple searches for words, groups of words, or words containing prexes and
sufxes; to display the results of a given search in KWIC (key word in context)
format; and to calculate the frequency of the item being searched in the corpus
in which it occurs. Figure 5.2 contains an example concordancing window
displaying the results of a search for all instances of the conjunction but in
the fully tagged and parsed British component of ICE (ICE-GB). This corpus
Figure 5.2 Concordancing window for the conjunction but in ICE-GB
can be searched by a text retrieval program, ICECUP, that has concordancing
capabilities.
In gure 5.2, the conjunction but is displayed in KWIC format: all instances
of but are vertically aligned for easy viewing, and displayed in the text unit
(cf. section 4.1) in which they occur. In ICECUP (and most other concordancing
programs) the context can be expanded to include, for instance, not just a single
text unit but groups of text units occurring before and after the search item. At
the bottom of the screen is a gure (5,909) specifying how many instances of
but were found in the corpus.
While all concordancing programs can perform simple word searches, some
have more sophisticated search capabilities. For instance, some programs can
search not just for strings of characters but for lemmas (e.g. all the forms
of the verb be) if a corpus has been lemmatized, that is, if, for example,
all instances of be (am, was, are, etc.) have been linked through the process of
lemmatization. In a corpus that has been lemmatized, if one wants information
on the frequency of all forms of be in the corpus, it is necessary to simply search
for be. In an unlemmatized corpus, one would have to search for every form of
be separately.
5
Many concordancing programs can also perform wild card searches. That
is, a search can be constructed that nds not just strings of characters occurring
5
Yasumasa Someya has produced a le of lemmatized verb forms for English that can be integrated
into WordSmith and save the analyst the time of creating such a le manually. The le can be
downloaded at http://www.liv.ac.uk/ms2928/wordsmith/index.htm.
in an exact sequence but strings that may have other words intervening between
them. For instance, a search of the correlative coordinator not just . . . but, as
in The movie was not just full of excessive violence but replete with foul lan-
guage, will have to be constructed so that it ignores the words that occur be-
tween just and but. To conduct such a search in a program such as Mono Conc
Pro 2.0 (http://www.athel.com/mono.html#monopro), the number of words be-
tween the two parts of the search expression can be set, specifying that the
search should nd all instances of not just . . . but where, say, up to twenty words
can occur between just and but. Wild card searches can additionally nd parts
of words: inWordSmith(http://www.liv.ac.uk/ms2928/wordsmith/index.htm),
searching for *ing will yield all instances of words in a corpus ending in -ing.
And wild card searches (as well as searches in general) can be extended to
search for words tagged in a specic way in a given corpus, or a particular word
tagged in a specic manner (e.g. all instances of like tagged as a conjunction).
Because pseudo-titles (and many other grammatical constructions) do not
have unique lexical content, a simple lexical search will not retrieve any in-
stances of pseudo-titles in a corpus. However, if pseudo-titles are studied in
corpora that have been tagged or parsed, it is at least possible to narrow the
range of structures generated in a search.
In a tagged or parsed corpus, it is possible to search not just for strings of
characters but for actual grammatical categories, such as nouns, verbs, noun
phrases, verb phrases, and so forth. Consequently, a tagged or parsed corpus
will allow for a greater range of structures to be recovered in a given search.
For instance, a concordancing program can be used to nd possessive nouns
by searching for the strings s or s. Although a search for these strings will
certainly recover numerous possessive nouns, it will also retrieve many un-
wanted constructions, such as contractions (e.g. Johns leaving). If possessives
are tagged, however, it will be possible to search for the possessive tag, in a
spoken or written text, and to recover only possessives, not extraneous construc-
tions such as contractions. In ICE-GB, possessives are assigned the tag genm
(genitive marker), and a search for all constructions containing this tag turned
up all the instances of possessives in ICE-GBin a couple of seconds. Other pro-
grams, such as WordSmith or Mono Conc Pro 2.0, can also be used to search
for tags.
Because in a tagged corpus only individual words are annotated, it can be dif-
cult to construct searches for grammatical constructions having different part-
of-speech congurations. Examples (8)(10) belowcontain tagged instances of
three pseudo-titles in ICE-GB:
(8) Community <N(com,sing):1/3> liaison <N(com,sing):2/3> ofcer <N(com,sing):
3/3> John <N(prop,sing):1/2> Hambleton <N(prop,sing):2/2>
(ICE-GB:W2C-009 #109:7)
(9) Liberal <N(com,sing):1/3> Democrats <N(com,sing):2/3> leader <N(com,sing):
3/3> Cllr <N(prop,sing):1/3> John <N(prop,sing):2/3> Hammond <N(prop,sing):
3/3> (ICE-GB:W2C-009 #59:3)
Figure 5.3 Parse tree for a sample pseudo-title in ICE-GB
(10) 59-year-old<ADJ(ge)>caretaker <N(com,sing)>Rupert <N(prop,sing):1/2>Jones
<N(prop,sing):2/2> : <PUNC(col)> (ICE-GB:W2C-011 #67:2)
While the pseudo-titles in (8)(10) consist of common and proper nouns, there
are some notable differences: (8) and (10) contain two proper nouns, (9) three
proper nouns; (10) begins with an adjective, (8) and (9) do not. There is thus no
sequence of common nouns and proper nouns to be searched for that uniquely
denes a pseudo-title. One could search for part of the pseudo-title that each
of the constructions has in common (e.g. a proper noun preceding a common
noun). But such a search is likely to turn up many structures other than pseudo-
titles. When such a search was conducted on ICE-GB, a noun phrase followed
by a vocative (11) was retrieved, as was an adverbial noun phrase followed by
a proper noun (12):
(11) Thats rubbish Aled (ICE-GB:S1A-068 #319:1:C)
(12) This time Hillier does nd his man (ICE-GB:S2A-018 #120:1:A)
To retrieve particular grammatical constructions, it is therefore more desirable
to use a parsed corpus rather than a tagged corpus because a parsed corpus will
contain annotation describing higher-level grammatical constructions.
Of the various ICE corpora being used to study pseudo-titles, only ICE-GB
has been parsed. To search ICE-GB for instances of pseudo-titles and corre-
sponding appositives, it is rst necessary to determine how such constructions
have been parsed. Figure 5.3 contains a parse tree fromICE-GBfor the pseudo-
title general manager Graham Sunderland.
The pseudo-title in gure 5.3 has been parsed into two noun phrases, the rst
containing two common nouns and the second two proper nouns. The second
noun phrase has the feature appos, indicating that in ICE-GB pseudo-titles
are considered a type of apposition.
Figure 5.4contains a parse tree for what is consideredanequivalent appositive
structure, Allen Chase, head of strategic exposure for NatWest Bank.
Figure 5.4 Parse tree for a sample appositive in ICE-GB
5.4 Subjecting the results of a corpus study to statistical analysis 119
Figure 5.5 FTF for appositive proper nouns
Like the pseudo-title in gure 5.3, the construction in gure 5.4 has the form
of two noun phrases containing various other structures. But unlike the pseudo-
title in gure 5.3, the appositive in gure 5.4 contains the feature appos for
both noun phrases, and additionally contains a parse marker for the mark of
punctuation, the comma, that separates both noun phrases.
There are two ways to search for pseudo-titles and appositives in ICE-GB.
Two separate searches can be constructed: one searching for two noun phrases
with the feature appos, and a second searching for two noun phrases with
only the second one containing the feature appos. Punctuation is irrelevant
to both searches, as the search tool for ICE-GB, ICECUP, can be instructed to
ignore punctuation. A more economical way to search for both pseudo-titles
and equivalent appositives is to construct a search containing one structure that
both constructions have in common: a single noun phrase containing proper
nouns that have the feature appos.
To nd all instances of proper nouns annotated with the feature appos,
ICECUP requires that a fuzzy tree fragment (FTF) is constructed, that is, a
partial tree structure that can serve as the basis for a search that will retrieve all
structures in ICE-GB that t the description of the FTF. Figure 5.5 contains an
FTF that will nd all instances of proper nouns with the feature appos.
This search was restricted to the press reportage section of ICE-GB, and in a
matter of seconds, all of the relevant constructions were retrieved and displayed
in KWIC format. Because the other components of ICE used in the study were
not tagged or parsed, pseudo-titles in these components had to be identied
manually, a process that took several days.
5.4 Subjecting the results of a corpus study to
statistical analysis
After information is extracted from a corpus (or series of corpora),
through either manual analysis or the use of a text retrieval program, it is
necessary to subject this information to some kind of statistical analysis. This
analysis can be quite simple and involve no more than providing frequency
counts of the particular constructions being investigated. However, if the goal
of ones analysis is to determine whether similarities or differences exist in a
corpus, it then becomes more necessary to apply specic statistical tests (such as
chi-square) to determine whether the differences or similarities are statistically
signicant or not, that is, whether they are real or merely the result of chance.
Because many modern-day corpus linguists have been trained as linguists,
not statisticians, it is not surprising that they have been reluctant to use statistics
in their studies. Many corpus linguists come from a tradition that has provided
themwith ample background in linguistic theory and the techniques of linguistic
description, but little experience in statistics. And as they begin doing analyses
of corpora, they nd themselves practicing their linguistic tradition within the
realm of numbers, the discipline of statistics, which many corpus linguists
nd foreign and intimidating. As a consequence, many corpus linguists have
chosen not to do any statistical analysis of the studies they conduct, and to work
instead mainly with frequency counts. For instance, the study of coordination
ellipsis described in section 1.2 was based solely on frequency differences in
the samples of speech and writing that were analyzed, and no statistical tests
were applied to determine whether these differences were statistically different.
Recently, however, many corpus linguists have begun to take the role of
statistics in corpus analysis much more seriously. A number of books have
been written, such as Woods, Fletcher, and Hughes (1986), Kretzschmar and
Schneider (1996), and Oakes (1998), that discuss statistical analysis of language
data in considerable detail. Some text analysis programs, such as WordSmith,
have statistical capabilities built into them so that as corpus linguists study
particular linguistic constructions (e.g. collocations) they can performstatistical
analyses to see whether the results they obtain are signicant. And many of the
commonly available statistical programs, such as SAS, SPSS, SYSTAT, or G,
can now be run on PCs or Macintoshes and have very friendly user interfaces
(cf. Kretzschmar 2000 for a review of the student version of SPSS, and a listing
of other statistical packages that can be run on a PC or Macintosh).
If the corpus linguist nds any of these resources intimidating, he or she
can always consult a specialist in the area of statistics for advice. However, if
this route is taken, it is important to realize that statistics is a vast and complex
eld of inquiry, and different academic disciplines will approach the use of
statistical tests from differing perspectives: mathematicians, for instance, use
statistics for different purposes than social scientists. Since corpus linguists are
interested in studying linguistic variation within a corpus, it is most appropriate
for them to follow the methodology of statistical analysis used in the social
sciences.
However one obtains information on the use of statistics in corpus analysis, it
is important to realize that by conducting a more rigorous statistical evaluation
of their results, corpus linguists can not only be more condent about the results
5.4 Subjecting the results of a corpus study to statistical analysis 121
they obtain but may even gain new insights into the linguistic issues they are
investigating.
The statistical analysis of data is essentially a three-part process that in-
volves:
1. evaluating the corpus from which the results are to be obtained to determine
its suitability for statistical analysis;
2. running the appropriate statistical tests;
3. interpreting the results, and nding linguistic motivations for them.
Whether the latter part of step 3 nding linguistic motivations for the results
is done rst in a series of hypothesis tested out in a corpus or last depends
upon the type of study one is doing, a point that will be discussed in greater
detail in section 5.4.1. But following all of these steps is the best way to insure
the validity and soundness of ones statistical analysis of a corpus.
5.4.1 Judging the suitability of a corpus for statistical
analysis and determining the appropriate statistical
tests to apply
Prior to conducting the study of pseudo-titles described in section 5.1,
great care was taken in selecting a corpus that would yield valid results. Because
pseudo-title usage has changed over time, texts were selected from a similar
time-frame to ensure that the results reected current usage trends. Even though
only text fragments were used, a pilot study was conducted to ensure that suf-
cient numbers of pseudo-titles could be found in 2,000-word samples. Had
texts from different time periods been mixed, or had the samples not turned
up enough examples of pseudo-titles, it would have been difcult to trust the
results that had been obtained, and considerable time would have been wasted
conducting a study that for all practical purposes was invalid. The more the
analyst is able to keep extraneous variables out of a corpus study, the more he
or she will be able to trust the validity of the results obtained.
Many times, however, the analyst does not have the advantage of basing an
analysis on a corpus that is methodologically pure. This is particularly the
case when using corpora developed in earlier eras, when those creating corpora
did not have a pilot corpus to guide their designs (Biber 1993: 256) and the
benet of the years of experience that now exists in the design and creation of
corpora. Also, the analyst may wish to do an analysis in an area where no good
corpora are available. However, even if all or part of a corpus is not designed
as ideally as the analyst would like, it is still possible to analyze the corpus and
make generalizations based on the results that are obtained. Woods, Fletcher,
and Hughes (1986: 55-6) advise that analysts accept the results of each study,
in the rst place, as though any sampling had been carried out in a theoretically
correct fashion . . . [and] then look at the possibility that they may have been
distorted by the way the sample was, in fact, obtained. If there are problems
with the sample, it is advisable to attempt to foresee some of the objections that
might be made about the quality of that material and either attempt to forestall
criticism or admit openly to any serious defects.
Ideally, it would be most desirable to analyze only fully representative corpora
so that it is not necessary to make the concessions that Woods, Fletcher, and
Hughes (1986) advocate. But as long as the analyst is clear about the kind of
corpus that was examined, open to the variables that might have affected the
results, and sensitive to criticisms that may be leveled against the results, it is
perfectly acceptable to base a linguistic analysis on a less than ideal corpus.
After determining whether the corpus being studied is suitable to the analysis
being undertaken, the analyst needs to next consider what statistical tests are ap-
propriate, and after applying the tests, whether any signicant results obtained
have some linguistic motivation: statistical signicance is meaningless if it has
no linguistic motivation. In deciding which statistical tests to apply, and howthe
results of such tests are best evaluated, it is useful to make the distinction that
Biber (1988) makes between macroscopic and microscopic analyses. In
macroscopic analyses, Biber (1988: 61) observes, the goal is to determine
the overall dimensions of variation in a language, that is, to investigate spoken
vs. written English, or a variety of genres of English, and to determine which
linguistic constructions dene and distinguish these kinds of English. Macro-
scopic analyses are in a sense inductive. In his analysis of linguistic variation
in speech and writing, although Biber (1988: 6575) pre-selected the corpora
his analysis would be based on and the particular linguistic constructions he
investigated, he did not initially go through the descriptive statistics that he
generated, based on the occurrence of the constructions in the corpora, and at-
tempt to determine which results looked meaningful and which did not before
applying further statistical tests. Instead, he simply did a factor analysis of the
results to determine which constructions tended to co-occur in specic genres
and which did not. And it was not until he found that passives and nominal-
izations, for instance, tended to co-occur that he attempted to nd a functional
motivationfor the results (Biber 1988: 80): that passives andnominalizations oc-
curredtogether ingenres (suchas academic prose) inwhichabstract information
predominated (Biber 1988: 119).
In microscopic analyses, on the other hand, the purpose is to undertake
a detailed description of the communicative functions of particular linguistic
features: the analyst takes a particular grammatical construction, such as the
relative clause, and attempts to study its function in various genres or linguistic
contexts. Microscopic analyses are typically deductive rather than inductive:
the analyst begins with a series of hypotheses and then proceeds to conrm or
disconrm them in the corpus being investigated. Of course, there will always
be results that are unanticipated, but these results can be handled by proposing
newhypotheses or reformulating old ones. Since the majority of corpus analyses
involve microscopic rather than macroscopic analyses, the remainder of this
chapter will focus on the various kinds of statistical tests that were applied to
the results of the study of pseudo-titles detailed in section 5.3.
5.5 The statistical analysis of pseudo-titles 123
5.5 The statistical analysis of pseudo-titles
A good corpus study, as was argued in section 5.1, combines quali-
tative and quantitative research methods. In any corpus analysis, the balance
between these methods will vary. This section explores this balance with re-
spect to the various hypotheses proposed in section 5.2 concerning the use and
structure of pseudo-titles, and demonstrates how these hypotheses can be con-
rmed or disconrmed through simple exploration of the various samples
of ICE included in the study as well as the application of various statistical
tests.
5.5.1 Exploring a corpus
Previous research on pseudo-titles has documented their existence in
American, British, and New Zealand press reportage, and demonstrated that
because their usage was stigmatized, certain newspapers (particularly in the
British press) prohibited their usage. To determine whether pseudo-titles have
spread to the other varieties of English represented in ICE and whether their
usage is stigmatized in these varieties, it is only necessary to examine examples
from the various samples to see whether a given newspaper allows pseudo-title
usage or not. And in simply examining examples in a corpus, one is likely to
encounter interesting data. In the case of pseudo-titles, for instance, it was
found that certain newspapers that prohibited pseudo-title usage did in fact
contain pseudo-titles, a nding that reects the well-known fact that practice
does not always follow prescription.
Table 5.4 lists the number of newspapers containing or not containing pseudo-
titles in the various regional components of ICE investigated. As table 5.4
illustrates, pseudo-titles have spread to all of the regional varieties of English
investigated, and it is only in Great Britain that there were many newspapers
prohibiting the usage of pseudo-titles.
Table 5.4 Number of newspapers containing pseudo-titles in various
ICE components
Newspapers without Newspapers with
Country pseudo-titles pseudo-titles Total
Great Britain 7 8 15
United States 1 19 20
New Zealand 1 11 12
Philippines 0 10 10
Jamaica 0 3 3
East Africa 0 3 3
Singapore 0 2 2
Totals 9 (14%) 56 (86%) 65 (100%)
Although the results displayed in table 5.4 reveal specic trends in usage,
there were some interestingexceptions tothese trends. InICE-USA, after further
investigation, it was found that the one US newspaper that did not contain any
pseudo-titles, the Cornell Chronicle, actually did allow pseudo-titles. It just
so happened that the 2,000-word sample included in ICE did not have any
pseudo-titles. This nding reveals an important limitation of corpora: that the
samples included within them do not always contain the full range of usages
existing in the language, and that it is often necessary to look further than the
corpus itself for additional data. In the case of the Cornell Chronicle, this meant
looking at additional samples of press reportage from the newspaper. In other
cases, it may be necessary to supplement corpus ndings with elicitation tests
(cf. Greenbaum 1984; de M onnink 1997): tests that ask individuals to identify
constructions as acceptable or not, or that seek information from individuals
about language usage. For pseudo-titles, one might give newspaper editors and
writers a questionnaire that elicits their attitudes towards pseudo-titles, and asks
them whether they consider certain borderline cases of pseudo-titles (such as
former President George Bush; section cf. 5.3.1) to actually be pseudo-titles,
and so forth.
Two newspapers whose style manuals prohibited pseudo-title usage actually
contained pseudo-titles. The New York Times and one British newspaper, the
Guardian, contained pseudo-titles in sports reportage. This suggests that at
least in these two newspapers the prohibition against pseudo-titles sometimes
does not extend to less formal types of writing. In addition, the New York
Times contained in its news reportage instances of so-called borderline cases of
pseudo-titles (see above). Some of the British-inuenced varieties of English
contained a mixture of British and American norms for pseudo-title usage.
Example (13) (taken from ICE-East Africa) begins with a pseudo-title, Lawyer
Paul Muite, but two sentences later contains a corresponding apposition a
lawyer, Ms. Martha Njoka that contains features of British English: a title, Ms.,
before the name in the second part of the appositive, and no punctuation marking
the title as abbreviated. In American press writing, typically an individuals full
name would be given without any title, and if a title were used, it would end in
a period (Ms.) marking it as an abbreviation.
(13) Lawyer Paul Muite and his co-defendants in the LSK contempt suit wound up their
case yesterday and accused the Government of manipulating courts through proxies
to silence its critics . . . Later in the afternoon, there was a brief drama in court when
a lawyer, Ms. Martha Njoka, was ordered out after she deed the judges directive
to stop talking while another lawyer was addressing the court. (ICE-East Africa)
Exploring a corpus qualitatively allows the analyst to provide descriptive in-
formation about the results that cannot be presented strictly quantitatively. But
because this kind of discussion is subjective and impressionistic, it is better to
devote the bulk of a corpus study to supporting qualitative judgements about a
corpus with quantitative information.
5.5.2 Using quantitative information to support qualitative
statements
In conducting a microscopic analysis of data, it is important not to
become overwhelmed by the vast amount of statistical information that such a
study will be able to generate, but to focus instead on using statistical analysis
to conrm or disconrm the particular hypotheses one has set out to test. In
the process of doing this, it is very likely that new and unanticipated ndings
will be discovered: a preliminary study of pseudo-titles, for instance, led to the
discovery that the length of pseudo-titles varied by national variety, a discovery
that will be described in detail below.
One of the most common ways to begin testing hypotheses is to use the
cross tabulation capability found in any statistical package. This capabil-
ity allows the analyst to arrange the data in particular ways to discover as-
sociations between two or more of the variables being focused on in a par-
ticular study. In the study of pseudo-titles, each construction was assigned a
series of tags associated with six variables, such as the regional variety the
construction was found in, and whether the construction was a pseudo-title
or a corresponding apposition (cf. section 5.3.2). To begin investigating how
pseudo-titles and corresponding appositives were used in the regional vari-
eties of ICE being studied, a cross tabulation of the variables country and
type was generated. This cross tabulation yielded the results displayed in
table 5.5. Because so few examples of pseudo-titles and equivalent apposi-
tives were found in ICE-East Africa, ICE-Singapore, and ICE-Jamaica, the
cross tabulations in table 5.5 (and elsewhere in this section) were restricted
to the four ICE varieties (ICE-USA, ICE-Philippines, ICE-New Zealand, and
ICE-Great Britain) from which a sufcient number of examples could be
taken.
The results of the cross tabulation in table 5.5 yield rawnumbers and percent-
ages which suggest various trends. In ICE-USA, Phil, and NZ, more pseudo-
titles than corresponding appositives were used, though ICE-Phil and NZ have
a greater percentage of pseudo-titles than does ICE-USA. In ICE-GB, just the
opposite occurs: more corresponding appositives than pseudo-titles were used,
Table 5.5 The frequency of pseudo-titles and corresponding appositives in the
national varieties of ICE
Country Pseudo-title Appositive Total
USA 59 (54%) 51 (46%) 110 (100%)
Phil 83 (69%) 38 (31%) 121 (100%)
NZ 82 (73%) 31 (27%) 113 (100%)
GB 23 (23%) 78 (77%) 101 (100%)
Total 247 198 445
ndings reecting the fact that there is a greater stigma against the pseudo-titles
in British press reportage than in the reportage of the other varieties.
When comparing results from different corpora, in this case differing com-
ponents of ICE, it is very important to compare corpora of similar length. If
different corpora of varying length are compared and the results are not nor-
malized, then the comparisons will be distorted and misleading. For instance,
if one were to count and then compare the number of pseudo-titles in one corpus
of 40,000 words and another of 50,000 words, the results would be invalid, since
a 50,000-word corpus is likely to contain more pseudo-titles than a 40,000-word
corpus, simply because it is longer. This may seem like a fairly obvious point,
but in conducting comparisons of the many different corpora that nowexist, the
analyst is likely to encounter corpora of varying length: corpora such as Brown
or LOB are one million words in length and contain 2,000-word samples; the
LondonLund Corpus is approximately 500,000 words in length and contains
5,000-word samples; and the British National Corpus is 100 million words long
and contains samples of varying length. Moreover, often the analyst will wish to
compare his or her results with the results of someone elses study, a comparison
that is likely to be based on corpora of differing lengths.
To enable comparisons of corpora that differ in length, Biber, Conrad, and
Reppen (1998: 2634) provide a convenient formula for normalizing frequen-
cies. Using this formula, to calculate the number of pseudo-titles occurring
per 1,000 words in the four varieties of ICE in table 5.5, one simply divides
the number of pseudo-titles (247) by the length of the corpus in which they
occurred (80,000 words) and multiplies this number by 1,000:
(247/80,000) 1,000 = 3.0875
The choice of norming to 1,000 words is arbitrary, but as larger numbers and
corpora are analyzed, it becomes more advisable to normto a higher gure (e.g.
occurrences per 10,000 words).
Although the percentages in table 5.5 suggest various differences in how
pseudo-titles and corresponding appositives are used, without applying any
statistical tests there is no way to know whether the differences are real or due
to chance. Therefore, in addition to considering percentage differences in the
data, it is important to apply statistical tests to the results to ensure that any
claims made have validity. The most common statistical test for determining
whether differences are signicant or not is the t-test, or analysis of variance.
However, because linguistic data do not typically have normal distributions, it
is more desirable to apply what are termed non-parametric statistical tests:
tests that make no assumptions about whether the data on which they are being
applied have a normal or non-normal distribution.
Data that are normally distributed will yield a bell curve: most cases will
be close to the mean, and the remaining cases will fall off quickly in fre-
quency on either side of the curve. To understand why linguistic data are not
normally distributed, it is instructive to examine the occurrence of pseudo-titles
Table 5.6 Frequency of occurrence of pseudo-titles in the samples from
ICE components
1 2 3 4 5 6 7 8 9 10 Total
USA 2 3 0 10 16 2 3 15 1 7 59
Phil 15 11 9 6 4 8 6 15 5 4 83
NZ 24 13 4 10 6 5 4 6 10 0 82
GB 0 0 0 0 8 3 3 2 0 7 23
Standard
Minimum Maximum Average deviation Kurtosis Skewness
0 24 6.175 5.514026 2.97711 1.147916
in the forty texts that were examined (cf. table 5.6), and the various statistical
measurements that calculate whether a distribution is normal or not.
As the gures in table 5.6 indicate, the distribution of pseudo-titles across
the forty samples was quite varied. Many samples contained no pseudo-titles;
one sample contained 24. The average number of pseudo-titles per sample was
around six. The standard deviation indicates that 68 percent of the pseudo-titles
occurring in the samples clustered within about 5.5 points either belowor above
the average, that is, that 68 percent of the samples contained between one and
11 pseudo-titles. But the real signs that the data are not normally distributed
are the gures for kurtosis and skewness.
If the data were normally distributed, the gures for kurtosis and skewness
would be 0 (or at least close to 0). Kurtosis measures the extent to which
a distribution deviates from the normal bell curve: whether the distribution is
clustered around a certain point in the middle (positive kurtosis), or whether the
distribution is clustered more around the ends of the curve (negative kurtosis).
Skewness measures how asymmetrical a distribution is: the extent to which
more scores are above or below the mean. Both of the scores for kurtosis and
skewness are veryhigh: a negative kurtosis of 2.97711indicates that scores are
clustering very far from the mean (the curve is relatively at), and the gure
of 1.147916 for skewness indicates that more scores are above the mean than
below. Figure 5.6 illustrates the at and skewed nature of the curve. The
horizontal axis plots the various number of pseudo-titles found in each of the
forty texts, and the vertical axis the number of texts having a given frequency.
The resultant curve is clearly not a bell curve.
Because most linguistic data behave the way that the data in table 5.5 do, it is
more desirable to apply non-parametric statistical tests to the results, and one of
the more commonly applied tests of this nature in linguistics is the chi-square.
The chi-square statistic is very well suited to the two-way cross tabulation in
table 5.5: the dependent variable (i.e. the variable that is constant, in this case the
country) is typically put in the left-hand column, and the independent variable
8
7
6
5
4
3
2
1
0
N
u
m
b
e
r

o
f

t
e
x
t
s

w
i
t
h

t
h
i
s

f
r
e
q
u
e
n
c
y
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Frequency per text
Figure 5.6 Pseudo-title frequency across samples
(i.e. the variable that changes, in this case the type: whether the construction
is a pseudo-title or corresponding apposition) is in the top-most row. Table 5.7
presents the results of a chi-square analysis of the data in table 5.5.
In essence, the chi-square test calculates the extent to which the distribution
in a given dataset either conrms or disconrms the null hypothesis: in this
case, whether or not there are differences in the distribution of pseudo-titles and
equivalent appositives in the four regional varieties of ICE being compared. To
perform this comparison, the chi-square test compares observed frequencies
in a given dataset with expected frequencies (i.e. the frequencies one would
expect to nd if there were no differences in the distribution of the data). The
higher the chi-square value, the more signicant the differences are.
The application of the chi-square test to the frequencies in table 5.5 yielded
a value of 65.686. To interpret this number accurately, one rst of all needs to
knowthe degrees of freedom in a given dataset (i.e. the number of data points
Table 5.7 Chi-square results for differences in the distribution of pseudo-titles
and corresponding appositives in the samples from ICE components
Statistical test Value Degrees of freedom Signicance level
Chi-square 65.686 3 p .000
that may vary). Since table 5.5 contains four rows and two columns, the degrees
of freedom can be calculated using the formula below:
(4 1) (2 1) = 3 degrees of freedom
With three degrees of freedom, the chi-square value of 65.686 is signicant at
less than the .000 level.
While it is generally accepted that any level below .05 indicates statistical
signicance, it is quite common for more stringent signicance levels to be
employed (e.g. p
<
=
.001). Because the signicance level for the data in table 5.7
is considerably beloweither of these levels, it can be safely and reliably assumed
that there are highly signicant differences in the distributions of pseudo-titles
and appositives across the four varieties of English represented in the table.
The chi-square test applied to the data in table 5.5 simply suggests that
there are differences in the use of pseudo-titles in the four national varieties
of English being investigated. The chi-square test says nothing about differ-
ences between the individual varieties (e.g. whether ICE-USA differs from
ICE-NZ). To be more precise about how the individual varieties differ from
one another, it is necessary to compare the individual varieties themselves in
a series of 22 chi-square tables. However, in examining a single dataset as
exhaustively as this, it is important to adjust the level that must be reached for
statistical signicance because, as Sigley (1997: 231) observes, If . . . many
tests are performed on the same data, there is a risk of obtaining spuriously
signicant results. This adjustment can be made using the Bonferroni correc-
tion, which determines the appropriate signicance level by dividing the level
of signicance used in a given study by the number of different statistical tests
applied to the dataset. The Bonferroni-corrected critical value for the ICE data
being examined is given below and is based on the fact that to compare all
four ICE components individually, six different chi-square tests will have to be
performed:
.05 / 6 = .0083
Signicance level Number of tests performed Corrected value
Table 5.8 contains the results of the comparison, from most signicant differ-
ences down to least signicant differences.
The results in table 5.8 illustrate some notable differences in the use of
pseudo-titles and equivalent appositives in the various national varieties. First
of all, the use of these constructions in ICE-GB is very different from their
use in the other varieties: the levels of signicance are very high and reect
the deeply ingrained stigma against the use of pseudo-titles in British press re-
portage, a stigma that does not exist in the other varieties. Second, even though
Table 5.8 Comparison of the distribution of pseudo-titles and corresponding
appositives in individual ICE components
Countries Statistical test Degrees of freedom Value
6
Signicance level
NZ and GB Chi-square 1 50.938 p << 0.0001
Phil and GB Chi-square 1 44.511 p << 0.0001
US and GB Chi-square 1 19.832 p << 0.0001
US and NZ Chi-square 1 7.796 p = .005
US and Phil Chi-square 1 4.830 p = .028 (non-sig.)
Phil and NZ Chi-square 1 .273 p = .601 (non-sig.)
pseudo-titles may have originated in American press reportage, their use is
more widespread in ICE-NZ and ICE-Phil, though with the Bonferroni correc-
tion the values are just below the level of signicance to indicate a difference
between ICE-USA and ICE-Phil. Finally, there were no signicant differences
between ICE-NZ and ICE-Phil. These results indicate that pseudo-title usage
is widespread, even in British-inuenced varieties such as New Zealand En-
glish, and that there is a tendency for pseudo-titles to be used more widely than
equivalent appositives in those varieties other than British English into which
they have been transplanted from American English.
While the chi-square statistic is very useful for evaluating corpus data, it
does have its limitations. If the analyst is dealing with fairly small numbers
resulting in either empty cells or cells with low frequencies, then the reliability
of chi-square is reduced. Table 5.9 lists the correspondence relationships for
appositives in the four varieties of English examined.
Three of the cells in the category of total equivalence contain fewer than
ve occurrences, making the chi-square statistic invalid for the data in table
5.9. One way around this problem is to combine variables in a principled
manner to increase the frequency for a given cell and thus make the results
Table 5.9 Correspondence relationships for appositives in the samples from
ICE components
Country Total equivalence Determiner deletion Partial equivalence Total
USA 1 (2%) 14 (28%) 36 (71%) 51 (101%)
Phil 1 (2.6%) 8 (21%) 29 (76%) 38 (100%)
NZ 0 (0%) 13 (42%) 18 (58%) 31 (100%)
GB 8 (10%) 22 (28%) 48 (62%) 78 (100%)
Total 10 (5%) 57 (29%) 131 (66%) 198 (100%)
6
In a 2 2 chi-square table, as Sigley (1997: 226) observes, the distribution is binomial rather
than continuous. It is therefore customary in a 2 2 table to use Yates correction rather the
normal Pearson chi-square value.
Table 5.10 Correspondence relationships for appositives in the samples from
ICE components (with combined cells)
Total equivalence/
Country determiner deletion Partial equivalence Total
USA 15 (29%) 36 (71%) 51 (100%)
Phil 9 (24%) 29 (76%) 38 (100%)
NZ 13 (42%) 18 (58%) 31 (100%)
GB 30 (39%) 48 (62%) 78 (101%)
Total 67 (34%) 131 (66%) 198 (100%)
Statistical Degrees of Signicance
test Value freedom level
Chi-square 3.849 3 p = .278
of the chi-square statistic more valid. As was noted in section 5.1, one rea-
son for recording the particular correspondence relationship for an apposi-
tive was to study the stylistic relationship between pseudo-titles and various
types of equivalent appositives: to determine, for instance, whether a newspa-
per prohibiting pseudo-titles relied more heavily than those newspapers allow-
ing pseudo-titles on appositives related to pseudo-titles by either determiner
deletion (e.g. the acting director, Georgette Smith acting director Geor-
gette Smith) or total equivalence (Georgette Smith, acting director acting
director Georgette Smith). Because these two correspondence relationships in-
dicate similar stylistic choices, it is justiable to combine the results for both
choices to increase the frequencies and make the chi-square test for the data
more valid.
Table 5.10 contains the combined results for the categories of total equiv-
alence and determiner deletion. This results in cells with high enough fre-
quencies to make the chi-square test valid. The results indicate, however, that
there was really no difference between the four varieties in terms of the cor-
respondence relationships that they exhibited: the chi-square value (3.849) is
relatively low and as a result the signicance level (.278) is above the level
necessary for statistical signicance.
It was expected that ICE-GBwould contain more instances of appositives ex-
hibiting either total equivalence or determiner deletion, since in general British
newspapers do not favor pseudo-titles and would therefore favor alternative ap-
positive constructions. And indeed the newspapers in ICE-GBdid contain more
instances. But the increased frequencies are merely a consequence of the fact
that, in general, the newspapers in ICE-GB contained more appositives than
the other varieties. Each variety followed a similar trend and contained fewer
appositives related by total equivalence or determiner deletion and more related
by partial equivalence. These ndings call into question Bells (1988) notion
Table 5.11 The length of pseudo-titles in the various components of ICE
Country 14 words 5 or more words Total
USA 57 (97%) 2 (3%) 59 (100%)
Phil 71 (86%) 12 (15%) 83 (101%)
NZ 66 (81%) 16 (20%) 82 (101%)
GB 23 (100%) 0 (0%) 23 (100%)
Total 217 (88%) 30 (12%) 247 (100%)
Chi-square 12.005 3 p = .007
Likelihood ratio 15.688 3 p = .001
of determiner deletion, since overall such constructions were not that com-
mon and whether a newspaper allowed or disallowed pseudo-titles had little
effect on the occurrence of appositives related by determiner deletion. Having
a greater effect on the particular correspondence relation that was found was
whether the appositive contained a genitive noun phrase or some kind of post-
modication, structures that led to a partial correspondence with a pseudo-title
and that occurred very commonly in all varieties.
While combining values for variables can increase cell values, often such
a strategy does not succeed simply because so few constructions occur in a
particular category. In such cases, it is necessary to select a different statistical
test to evaluate the results. To record the length of a pseudo-title or appositive,
the original coding system had six values: one word in length, two words, three
words, four words, ve words, and six or more words. It turned out that this
coding scheme was far too delicate and made distinctions that simply did not
exist in the data: many cells had frequencies that were too low to apply the chi-
square test. And combining categories, as is done in table 5.11, still resulted
in two cells with frequencies lower than ve, making the chi-square results for
this dataset invalid.
In cases like this, it is necessary to apply a different statistical test: the log-
likelihood (or G
2
) test. Dunning (1993: 656) has argued that, in general, this
test is better than the chi-square test because it can be applied to very much
smaller volumes of text . . . [and enable] comparisons to be made between the
signicance of the occurrences of both rare and common phenomena. Dunning
(1993: 623) notes that the chi-square test was designed to work with larger
datasets that have items that are more evenly distributed, not with corpora
containing what he terms rare events (e.g. two instances in ICE-USA of
pseudo-titles lengthier than ve words). Applied to the data in table 5.11, the
log-likelihood test (termed the likelihood ratio in SPSS parlance) conrmed
that the length of pseudo-titles varied by variety.
The results of the log-likelihood test point to a clear trend in table 5.11:
that lengthier pseudo-titles occur more frequently in ICE-Phil and NZ than in
ICE-USA and GB. In fact, ICE-GB had no pseudo-titles lengthier than ve
Table 5.12 The length of appositives in the various components of ICE
Country 14 words 5 or more words Total
USA 22 (43%) 29 (57%) 51 (100%)
Phil 14 (37%) 24 (63%) 38 (100%)
NZ 14 (45%) 17 (55%) 31 (100%)
GB 32 (41%) 46 (59%) 78 (100%)
Total 82 (41%) 116 (59%) 198 (100%)
Chi-square .574 3 p = .902
words, and ICE-USA had only two instances. These ndings are reected in
the examples in (14) and (15), which contain pseudo-titles lengthier than ve
words that occurred predominantly in newspapers in ICE-Phil and ICE-NZ.
(14) a. Salamat and Presidential Adviser on Flagship Projects in Mindanao Robert
Aventajado (ICE-Philippines)
b. Time Magazine Asia bureau chief Sandra Burton (ICE-Philippines)
c. Marikina Metropolitan Trial Court judge Alex Ruiz (ICE-Philippines)
d. MILF Vice Chairman for Political Affairs Jadji Murad (ICE-Philippines)
e. Autonomous Region of Muslim Mindanao police chief Damming Unga (ICE-
Philippines)
(15) a. Oil and Gas planning and development manager Roger OBrien (ICE-NZ)
b. NewPlymouth Fire Services deputy chief re ofcer Graeme Moody (ICE-NZ)
c. corporate planning and public affairs executive director Graeme Wilson (ICE-
NZ)
d. Federated Gisborne-Wairoa provincial president Richard Harris (ICE-NZ)
e. Wesley and former New Zealand coach Chris Grinter (ICE-NZ)
The pseudo-title is a relatively new and evolving structure in English. There-
fore, it is to be expected that its usage will show variation, in this case in the
length of pseudo-titles in the various components of ICE under investigation.
The appositive, on the other hand, is a well-established construction in English,
and if the length of appositives is considered, there were no differences between
the varieties, as is illustrated in table 5.12. Table 5.12 demonstrates that it is
more normal for appositives to be lengthier, and that while ICE-GB has more
appositives than the other varieties, the proportion of appositives of varying
lengths is similar to the other varieties.
One reason for the general difference in length of appositives and pseudo-
titles is that there is a complex interaction between the form of a given pseudo-
title or appositive and its length. In other words, three variables are interacting:
type (pseudo-title or appositive), form (simple noun phrase, genitive noun
phrase, noun phrase with post-modication), and length (one to four words
or ve words or more). Table 5.13 provides a cross tabulation of all of these
variables.
Table 5.13 The form and length of pseudo-titles and corresponding
appositives
Type Form 1-4 words 5 or more words Total
PT Simple NP 216 (90%) 23 (10%) 239 (100%)
Gen. NP 0 (0%) 0 (0%) 0 (0%)
Post. Mod. 1 (13%) 7 (87%) 8 (100%)
Total 217 (88%) 30 (12%) 247 (100%)
Appos Simple NP 52 (84%) 10 (16%) 62 (100%)
Gen. NP 18 (67%) 9 (33%) 27 (100%)
Post. Mod. 12 (11%) 97 (89%) 109 (100%)
Total 82 (41%) 116 (59%) 198 (100%)
A chi-square analysis of the trends in table 5.13 would be invalid not only
because some of the cells have values lower than ve but because the chi-square
test cannot pinpoint specically which variables are interacting. To determine
what the interactions are, it is more appropriate to conduct a loglinear analysis
of the results.
A loglinear analysis considers interactions between variables: whether, for
instance, there is an interaction between type, form, and length; between
type and form; between form and length; and so forth. In setting up
a loglinear analysis, one can either investigate a predetermined set of associa-
tions (i.e. only those associations that the analyst thinks exist in the data), or
base the analysis on a saturated model: a model that considers every possible
interaction the variables would allow. The drawback of a saturated model, as
Oakes (1998: 38) notes, is that because it includes all the variables and inter-
actions required to account for the original data, there is a danger that we will
select a model that is too good. . . [and that nds] spurious relationships. That
is, when all interactions are considered, it is likely that signicant interactions
between some interactions will be coincidental. Thus, it is important to nd
linguistic motivations for any signicant associations that are found.
Because only three variables were being compared, it was decided to use a
saturated model to investigate associations. This model generated the following
potential associations:
(16) a. type*form*length
b. type*form
c. type*length
d. form*length
e. form
f. type
g. length
Table 5.14 Associations between various variables
K Degrees of freedom Likelihood ratio Probability Probability
3 2 .155 .9254 .9300
2 7 488.010 .0000 .0000
1 11 825.593 .0000 .0000
Likelihood ratio and chi-square tests were conducted to determine whether there
was a signicant association between all three variables (16a), and between all
possible combinations of two-way interactions (16bd). In addition, the vari-
ables were analyzed individually to determine the extent to which they affected
the three- and two-way associations in 16ad. The results are presented in table
5.14.
The rst line in table 5.14 demonstrates that there were no associations
between the three variables: the likelihood ratio score had probability where
p > .05. On the other hand, there were signicant associations between the
two-way and one-way variables.
To determine which of these associations were strongest, a procedure called
backward elimination was applied to the results. This procedure works in a
step-by-step manner, at each stage removing from the analysis an association
that is least strong and then testing the remaining associations to see which is
strongest. This procedure produced the two associations in table 5.15 as being
the strongest of all the associations tested. Interpreted in conjunction with the
frequency distributions in table 5.13, the results in table 5.14 suggest that
while appositives are quite diverse in their linguistic form, pseudo-titles are
not. Even though a pseudo-title and corresponding appositive have roughly the
same meaning, a pseudo-title is mainly restricted to being a simple noun phrase
that is, in turn, relatively short in length. In contrast, the unit of an appositive
corresponding to a pseudo-title can be not just a simple noun phrase but a
genitive noun phrase or a noun phrase with post-modication as well.
These linguistic differences are largely a consequence of the fact that the
structure of a pseudo-title is subject to the principle of end-weight (Quirk
et al. 1985: 13612). This principle stipulates that heavier constituents are best
placed at the end of a structure, rather than at the beginning of it. A pseudo-
title will always come at the start of the noun phrase in which it occurs. The
lengthier and more complex the pseudo-title, the more unbalanced the noun
Table 5.15 Strongest associations between variables
Degrees of freedom Likelihood ratio Probability
Type*form 2 246.965 .0000
Length*form 2 239.067 .0000
phrase will become. Therefore, pseudo-titles typically have forms (e.g. simple
noun phrases) that that are short and non-complex structures, though as table
5.12 illustrated, usage does vary by national variety. In contrast, an appositive
consists of two units, one of which corresponds to a pseudo-title. Because this
unit is independent of the proper noun to which it is related in speech it
occupies a separate tone unit, in writing it is separated by a comma from the
proper noun to which it is in apposition it is not subject to the end-weight
principle. Consequently, the unit of an appositive corresponding to a pseudo-
title has more forms of varying lengths.
The loglinear analysis applied to the data in table 5.12 is very similar to the
logistic regression models used in the Varbrul programs: IVARB(for MS-DOS)
and GoldVarb (for the Macintosh) (cf. Sankoff 1987; Sigley 1997: 23852).
These programs have been widely used in sociolinguistics to test the interaction
of linguistic variables. For instance, Tagliamonte and Lawrence (2000) used
GoldVarb to examine which of seven linguistic variables favored the use of
three linguistic forms to express the habitual past: a simple past-tense verb,
used to, or would. Tagliamonte and Lawrence (2000: 336) found, for instance,
that the type of subject used in a clause signicantly affected the choice of verb
form: the simple past was used if the subject was a second-person pronoun,
used to was used if the subject was a rst-person pronoun, and would was
used if the subject was a noun phrase with a noun or third-person pronoun as
head.
Although the Varbrul programs have been used primarily in sociolinguistics
to study the application of variable rules, Sigley (1997) demonstrates the value
of the programs for corpus analysis as well in his study of relative clauses in
the Wellington Corpus of New Zealand English. The advantage of the Varbul
programs is that they were designed specically for use in linguistic analy-
sis and are thus easier to use than generic statistical packages, such as SPSS
or SAS. But it is important to realize that these statistical packages, as the
loglinear analysis in this section demonstrated, can replicate the kinds of statis-
tical analyses done by the Varbrul programs. Moreover, as Oakes (1998: 151)
demonstrates, these packages can perform a range of additional statistical anal-
yses quite relevant to the concerns of corpus linguists, from those, such as the
Pearson-Product Moment, that test correlations between variables to Regres-
sion tests, which test the effects that independent variables have on dependent
variables.
5.6 Conclusions
To conduct a corpus analysis effectively, the analyst needs to plan out
the analysis carefully. It is important, rst of all, to begin the process with a very
clear research question in mind, so that the analysis involves more than simply
counting linguistic features. It is next necessary to select the appropriate
5.6 Conclusions 137
corpus for analysis: to make sure, for instance, that it contains the right kinds
of texts for the analysis and that the corpus samples to be examined are lengthy
enough. And if more than one corpus is to be compared, the corpora must
be comparable, or the analysis will not be valid. After these preparations are
made, the analyst must nd the appropriate software tools to conduct the study,
code the results, and then subject them to the appropriate statistical tests. If
all of these steps are followed, the analyst can rest assured that the results
obtained are valid and the generalizations that are made have a solid linguistic
basis.
Study questions
1. What is the danger of beginning a corpus analysis without a clearly thought-
out research question in mind?
2. How does the analyst determine whether a given corpus is appropriate for
the corpus analysis to be conducted?
3. What kinds of analyses are most efciently conducted with a concordancing
program?
4. What kinds of information can be found in a tagged or parsed corpus that
cannot be found in a lexical corpus?
5. The data in the table below are adapted from a similar table in Meyer (1996:
38) and contain frequencies for the distribution of phrasal (e.g. John and
Mary) and clausal (e.g. We went to the store and our friends bought some
wine) coordination in various samples of speech and writing from the Inter-
national Corpus of English. Go to the web page given below at Georgetown
University and use the Web Chi-square Calculator on the page to deter-
mine whether there is a difference between speech and writing in the distribu-
tion of phrasal and clausal coordination: http://www.georgetown.edu/cball/
webtools/web chi.html.
Syntactic structures in speech and writing
Medium Phrases Clauses Total
Speech 168 (37%) 289 (63%) 457 (100%)
Writing 530 (77%) 154 (23%) 684 (100%)
6 Future prospects in corpus linguistics
In describing the complexity of creating a corpus, Leech (1998: xvii)
remarks that a great deal of spadework has to be done before the research
results [of a corpus analysis] can be harvested. Creating a corpus, he comments,
always takes twice as much time, and sometimes ten times as much effort
because of all the work that is involved in designing a corpus, collecting texts,
and annotating them. And then, after a given period of time, Leech (1998: xviii)
continues, the corpus becomes out of date, requiring the corpus creator to
discard the concept of a static corpus of a given length, and to continue to collect
and store corpus data indenitely into the future . . . The process of analyzing a
corpus may be easier than the description Leech (1998) gives above of creating
a corpus, but still, many analyses have to be done manually, simply because we
do not have the technology that can extract complex linguistic structures from
corpora, no matter how extensively they are annotated. The challenge in corpus
linguistics, then, is to make it easier both to create and analyze a corpus. What
is the likelihood that this will happen?
Planning a corpus. As more and more corpora have been created, we have
gained considerable knowledge of howto construct a corpus that is balanced and
representative and that will yield reliable grammatical information. We know,
for instance, that what we plan to do with a corpus greatly determines how it is
constructed: vocabulary studies necessitate larger corpora, grammatical studies
(at least of relatively frequently occurring grammatical constructions) shorter
corpora. The British National Corpus is the culmination of all the knowledge
we have gained since the 1960s about what makes a good corpus.
But while it is of prime importance to descriptive corpus linguistics to create
valid and representative corpora, in the eld of natural language processing this
is an issue of less concern. Obviously, the two elds have different interests: it
does not require a balanced and representative corpus to train a parser or speech-
recognition system. But it would greatly benet the eld of corpus linguistics if
descriptive corpus linguists and more computationally oriented linguists and en-
gineers worked together to create corpora. The British National Corpus is a good
example of the kind of corpus that can be created when descriptive linguists,
computational linguists, and the publishing industry cooperate. The TalkBank
Project at Carnegie Mellon University and the University of Pennsylvania is a
multi-disciplinary effort designed to organize varying interest groups engaged
in the computational study of human and animal communication. One of the
138
Future prospects in corpus linguistics 139
interest groups, Linguistic Exploration, deals withthe creationandannotationof
corpora for purposes of linguistic research. The Michigan Corpus of Academic
Spoken English (MICASE) is the result of a collaborative effort involving both
linguists and digital librarians at the University of Michigan. Cross-disciplinary
efforts such as these integrate the linguistic and computational expertise that
exists among the various individuals creating corpora, they help increase the
kinds and types of corpora that are created, and they make best use of the limited
resources available for the creation of corpora.
Collecting and computerizing written texts. Because so many written texts
are now available in computerized formats in easily accessible media, such as
the World Wide Web, the collection and computerization of written texts has
become much easier than in the past. It is no longer necessary for every written
text to be typed in by hand or scanned with an optical scanner and then the
scanning errors corrected. If texts are gathered from the World Wide Web, it
is still necessary to strip them of html formatting codes. But this process can
be automated with software that removes such markup from texts. Creating
a corpus of written texts is now an easy and straightforward enterprise. The
situation is more complicated for those creating corpora containing texts from
earlier periods of English: early printed editions of books are difcult to scan
optically with any degree of accuracy; manuscripts that are handwritten need
to be manually retyped and the corpus creator must sometimes travel to the
library where the manuscript is housed. This situation might be eased in coming
years. There is an increased interest both among historical linguists and literary
scholars in computerizing texts from earlier periods of English. As projects
such as the Canterbury Project are completed, we may soon see an increased
number of computerized texts from earlier periods.
Collecting and computerizing spoken texts. While it is now easier to prepare
written texts for inclusion in a corpus, there is little hope for making the col-
lection and transcription of spoken texts easier. For the foreseeable future, it
will remain an arduous task to nd people who are willing to be recorded, to
make recordings, and to have the recordings transcribed. There are advantages
to digitizing spoken samples and using specialized software to transcribe them,
but still the transcriber has to listen to segments of speech over and over again
to achieve an accurate transcription. Advances in speech recognition might au-
tomate the transcription of certain kinds of speech (e.g. speeches and perhaps
broadcast news reports), but no software will be able to cope with the dysu-
ency of a casual conversation. Easing the creation of spoken corpora remains
one of the great challenges in corpus linguistics, a challenge that will be with
us for some time in the future.
Copyright restrictions. Obtaining the rights to use copyrighted material has
been a perennial problem in corpus linguistics. The rst release of the British
National Corpus could not be obtained by anyone outside the European Union
because of restrictions placed by copyright holders on the distribution of certain
written texts. The BNC Sampler and second release of the entire corpus have
140 Future prospects in corpus linguistics
no distribution restrictions but only because the problematic texts are not in-
cluded in these releases. The distribution of ARCHERhas been on hold because
it has not been possible to obtain copyright permission for many of the texts
included in the corpus. As a result, access to the corpus is restricted to those
who participated in the actual creation of the corpus a method of distribution
that does not violate copyright law. It is unlikely that this situation will ease
in the future. While texts are more widely available in electronic form, partic-
ularly on the World Wide Web, getting permission to use these texts involves
the same process as getting permission for printed texts, and current trends in
the electronic world suggest that access to texts will be more restrictive in the
future, not less. Therefore, the problem of copyright restrictions will continue
to haunt corpus linguists for the foreseeable future.
Annotating texts with structural markup. The development of SGML-based
annotation systems has been one of the great advances in corpus linguistics,
standardizing the annotation of many features of corpora so that they can be
unambiguously transferred from computer to computer. The Text Encoding
System (TEI) has provided a system of corpus annotation that is both detailed
and exible, and the introduction of XML (the successor to HTML) to the eld
of corpus linguistics will eventually result in corpora that can be made available
on the World Wide Web. There exist tools to help in the insertion of SGML-
conformant markup to corpora, and it is likely that such tools will be improved
in the future. Nevertheless, much of this annotation has to be inserted manually,
requiring hours of work on the part of the corpus creator. We will have much
better annotated corpora in the future, but it will still be a major effort to insert
this annotation into texts.
Tagging and parsing. Tagging is now a standard part of corpus creation, and
taggers are becoming increasingly accurate and easy to use. There will always be
constructions that will be difcult to tag automatically and will require human
intervention to correct, but tagged corpora should become more widespread
in the future we may even see the day when every corpus released has been
tagged.
Parsing is improving too, but has a much lower accuracy than tagging. There-
fore, much human intervention is required to correct a parsed text, and we have
not yet reached the point that the team not responsible for designing the parser
can use it effortlessly. One reason the British component of ICE (ICE-GB) took
nearly ten years to complete was that considerable effort had to be expended
correcting the output of the parsing of the corpus, particularly the spoken part.
At some time in the future, parsing may be as routine as tagging, but because a
parser has a much more difcult job than a tagger, we have some time to wait
before parsed corpora will be widely available.
Text analysis. The most common text analysis program for corpora, the con-
cordancer, has become an established xture for the analysis of corpora. There
are many such programs available for use on PCs, Macintoshes, and even the
World Wide Web. Such programs are best for retrieving sequences of strings
Future prospects in corpus linguistics 141
(such as words), but many can now search for particular tags in a corpus, and if
a corpus contains le header information, some concordancing programs can
sort les so that the analyst can specify what he or she wishes to analyze in a
given corpus: journalistic texts, for instance, but not other kinds of texts.
More sophisticated text analysis programs, such as ICECUP, are rare, and
it is not likely that we will see as many programs of this nature in the future
as concordancers. And a major problem with programs such as ICECUP and
many concordancers is that they were designed to work on a specic corpus
computerized in a specic format. Consequently, ICECUP works only on the
British component of the International Corpus of English, and Sara on the
BNC (though there are plans to extend the use of Sara to other corpora as well).
The challenge is to systemize the design of corpora and concordancers so that
any concordancer can work on any corpus. Of course, it is highly likely that
the next generation of corpus linguists will have a much better background in
programming. Thus, these corpus linguists will be able to use their knowledge
of languages such as Perl or Visual Basic to write specic scripts to analyze
texts, and as these scripts proliferate, they can be passed from person to person
and perhaps make obsolete the need for specic text analysis programs to be
designed.
Corpus linguistics has been one of the more exciting methodological devel-
opments in linguistics since the Chomskyan revolution of the 1950s. It reects
changing attitudes among many linguists as to what constitutes an adequate
empirical study of language, and it has drawn upon recent developments in
technology to make feasible the kinds of empirical analyses of language that cor-
pus linguists wish to undertake. Of course, doing a corpus analysis will always
involve work more work than sitting in ones ofce or study and making up the
data for a particular analysis but doing a corpus analysis properly will always
have its rewards and will help us advance the study of human language, an area
of study that linguists of all persuasions would agree we still know relatively
little about.
Appendix 1
Corpus resources
Cross references to resources listed in this table are indicated in boldface. The
various resources are alphabetized by acronym or full name, depending upon which
usage is most common.
The publisher has used its best endeavors to ensure that the URLs for external websites
referred to in this book are correct and active at the time of going to press. However, the
publisher has no responsibility for the websites and can make no guarantee that a site
will remain live or that the content is or will remain appropriate.
Resource Description Availability
American National Corpus Currently in progress; is
intended to contain spoken
and written texts that
model as closely as
possible the texts in the
British National Corpus
Project website:
http://www.cs.vassar.edu/
ide/anc/
American Publishing
House for the Blind Corpus
25 million words of edited
written American English
Originally created by IBM;
described in Fillmore
(1992)
ARCHER (A
Representative Corpus of
English Historical
Registers)
1.7 million words
consisting of various
genres of British and
American English covering
the period 16501990
In-house corpus (due to
copyright restrictions); an
expanded version,
ARCHER II, is underway
Bank of English Corpus 415 million words of
speech and writing (as of
October 2000); texts are
continually added
Collins-Cobuild:
http://titania.cobuild.collins.
co.uk/boe info.html
Bergen Corpus of London
Teenage English (COLT)
500,000-word corpus of
the speech of London
teenagers from various
boroughs; available online
or as part of the British
National Corpus
Project website:
http://www.hd.uib.no/colt/
Birmingham Corpus 20 million words of written
English
Evolved into Bank of
English Corpus
142
Corpus resources 143
(BNC)
100 million words of
samples of varying length
containing spoken (10
million words) and written
(90 million words) British
English
BNC website:
http://info.ox.ac.uk/bnc/
index.html
(BNC) Sampler
2 million words of speech
and writing representing
184 samples taken from the
BNC website:
http://info.ox.ac.uk/bnc/
getting/sampler.html
Brown Corpus One million words of
edited written American
English; created in 1961;
divided into 2,000-word
samples from various
genres (e.g. press
reportage, ction,
government documents)
See: ICAME CD-ROM
Cambridge International
Corpus
100 million words of
varying amounts of spoken
and written British and
American English, with
additional texts being
added continuously
CUP website:
http://uk.cambridge.org/elt/
reference/cic.htm
Cambridge Learners
Corpus
10 million words of student
essay exams, with
additional texts being
added continuously
CUP website:
http://uk.cambridge.org/elt/
reference/clc.htm
Canterbury Project Ultimate goal of the project
is to make available in
electronic form all versions
of the Canterbury Tales
and to provide an interface
to enable, for instance,
comparisons of the various
versions
Project website:
http://www.cta.dmu.ac.uk/
projects/ctp/
Chemnitz Corpus A parallel corpus of
English and German
translations
Project website:
http://www.tu-chemnitz.de/
phil/english/real/
transcorpus/index.htm
144 Appendix 1
Child Language Data
Exchange System
(CHILDES) Corpus
Large multi-lingual
database of spoken
language from children and
adults engaged in rst or
second language
acquisition
http://childes.psy.cmu.edu/
See also: MacWhinney
(2000)
Corpora Discussion List
(made available through
the Norwegian Computing
Centre for the Humanities)
Internet discussion list for
issues related to corpus
creation, analysis, tagging,
parsing, etc.
http://www.hit.uib.no/
corpora/welcome.txt
Corpus of Early
English Correspondence
Two versions of English
correspondence: the full
version (2.7 million words)
and a sampler version
(450,000 words)
Project website:
http://www.eng.helsinki./
doe/projects/ceec/
corpus.htm
Sampler version available
on ICAME CD-ROM
Corpus of Middle English
Prose and Verse
A large collection of
Middle English texts
available in electronic
format
Project website:
http://www.hti.umich.edu/c/
cme/about.html
Corpus of Spoken
Professional English
Approximately 2 million
words taken from spoken
transcripts of academic
meetings and White House
press conferences
Athelstan website:
http://www.athel.com/
cpsa.html
The Electronic Beowulf A digital version of the Old
English poem Beowulf that
can be searched
Project website:
http://www.uky.edu/
kiernan/eBeowulf/
guide.htm
EnglishNorwegian
Parallel Corpus
A parallel corpus of
English and Norwegian
translations: 30 samples of
ction and 20 samples of
non-ction in the original
and in translation
Project website:
http://www.hf.uio.no/iba/
prosjekt/
The Expert Advisory
Group on Language
Engineering Standards
(EAGLES)
Has developed A Corpus
Encoding Standard
containing guidelines for
the creation of corpora
Project website:
http://www.cs.vassar.edu/
CES/
FLOB (Freiburg
LancasterOslo
Bergen) Corpus
One million words of
edited written British
English published in 1991;
samples in varying genres
intended to replicate the
LOB Corpus
See: ICAME CD-ROM
FROWN (FreiburgBrown)
Corpus
edited written American
English published in 1991;
samples in varying genres
intended to replicate the
Brown Corpus
See: ICAME CD-ROM
Helsinki Corpus Approximately 1.5 million
words of Old, Middle, and
Early Modern English
divided into samples of
varying length
See: ICAME CD-ROM
Helsinki Corpus of Older
Scots
Approximately 400,000
words of transcribed
speech (recorded in the
1970s) from four rural
dialects in England and
Ireland
See: ICAME CD-ROM
Hong Kong University of
Science and Technology
Learner Corpus
25 million words of learner
English written by
rst-year university
students whose native
language is Chinese
Contact:
John Milton, Project
Director, [email protected]
ICAME Bibliography Extensive bibliography of
corpus-based research
created by Bengt Altenberg
(Lund University, Sweden)
ICAME website:
1989:
http://www.hd.uib.no/icame
/icame-bib2.txt
19908:
http://www.hd.uib.no/icame
/icame-bib3.htm
ICAME CD-ROM 20 different corpora (e.g.
Brown, LOB, Helsinki) in
various computerized
formats (DOS, Windows,
Macintosh and Unix)
ICAME website:
http://www.hit.uib.no/icame
/cd/
146 Appendix 1
International Corpus of
English (ICE)
A variety of million-word
corpora (600,000 words of
speech, 400,000 words of
writing) representing the
various national varieties
of English (e.g. American,
British, Irish, Indian, etc.)
Three components now
complete:
Great Britain:
http://www.ucl.ac.uk/
english-usage/ice-
gb/index.htm
East Africa: http://www.tu-
chemnitz.de/phil/english/
real/eafrica/corpus.htm
New Zealand:
http://www.vuw.ac.nz/lals/
corpora.htm#The New
Zealand component of the
International
ICECUP (ICE Corpus
Utility Program)
Text retrieval software for
use with ICE-GB
Survey of English Usage
website:
http://www.ucl.ac.
uk/english-usage/ice-gb
/icecup.htm
ICE-GB (British
component of the
English)
spoken and written British
English fully tagged and
parsed
See: International Corpus
of English
Learner English (ICLE)
words of written English
composed by non-native
speakers of English from
14 different linguistic
backgrounds
Project website:
http://www.tr.ucl.ac.be/FL
TR/GERM/ETAN/CECL/
introduction.html
Lampeter Corpus Approximately 1.1 million
words of Early Modern
English tracts and
pamphlets taken from
various genres (e.g.
religion, politics) from the
period 16401740;
contains complete texts,
not text samples
Project website:
http://www.tu-
chemnitz.de/phil/english/
real/lampeter/
lamphome.htm
Lancaster Corpus A precursor to the
million-word
LancasterOsloBergen
(LOB) Corpus of edited
written British English
See: LancasterOslo
Bergen (LOB) Corpus
Lancaster/IBM Spoken
English Corpus
53,000 words of spoken
British English (primarily
radio broadcasts); available
in various formats,
including a parsed version
See: ICAME CD-ROM
Lancaster Parsed Corpus A parsed corpus containing
approximately 140,000
words from various genres
in the LancasterOslo
Bergen (LOB) Corpus
See: ICAME CD-ROM
LancasterOsloBergen
(LOB) Corpus
edited written British
English published in 1961
and divided into 2,000-
word samples; modeled
after the Brown Corpus
See: ICAME CD-ROM
Linguistic Data
Consortium (LDC)
For an annual fee, makes
available to members a
variety of spoken and
written corpora of English
and many other languages
LDC website:
http://www.ldc.upenn.edu/
London Corpus The original corpus of
English rst created in the
1960s by Randolph Quirk
at the Survey of English
Usage, University College
London; sections of the
spoken part are included in
the LondonLund Corpus
Can be used on-site at the
Survey of English Usage:
http://www.ucl.ac.uk/
english-usage/home.htm
LondonLund Corpus Approximately 500,000
words of spoken British
English from various
genres (e.g. spontaneous
dialogues, radio
broadcasts) that has been
prosodically transcribed
See: ICAME CD-ROM
148 Appendix 1
LongmanLancaster
Corpus
A corpus available in
orthographic form that
contains approximately 30
million words of written
English taken from various
varieties of English
world-wide
Longman website:
http://www.longman-
elt.com/dictionaries/corpus/
lclonlan.html
Longman Learners Corpus 10 million words of writing
by individuals from around
the world learning English
as a second or foreign
language
Longman website:
http://www.longman-
elt.com/dictionaries/corpus/
lclearn.html
The Longman Spoken and
Written English Corpus
(LSWE)
words of samples of
and American English
Described in Biber et al.
(1999)
Map Task Corpus Digitized transcriptions of
individuals engaged in
task-oriented dialogues
in which one speaker helps
another speaker replicate a
route on a map
Project website:
http://www.hcrc.ed.ac.uk/
maptask/
Michigan Corpus of
Academic Spoken English
(MICASE)
Various types of spoken
American English recorded
in academic contexts: class
lectures and discussions,
tutorials, dissertation
defenses
Searchable on the Web:
http://www.hti.umich.edu/
m/micase/
Nijmegen Corpus A 130,000-word parsed
corpus of written English
TOSCA Research website:
http://lands.let.kun.nl/
TSpublic/tosca/
research.html
The Northern Ireland
Transcribed Corpus of
Speech
400,000 words of
interviews with individuals
speaking Hiberno-English
from various regions of
Northern Ireland
See Kirk (1992)
PennHelsinki Parsed
Corpus of Middle English
1.3 million words of parsed
Middle English taken from
55 samples found in the
Helsinki Corpus
Project website:
http://www.ling.upenn.edu/
mideng/
Penn Treebank (Releases I
and II)
A heterogeneous collection
of speech and writing
totaling approximately 4.9
million words; sections
have been tagged and
parsed
Linguistic Data
Consortium (LDC)
website:
Catalog/LDC95T7.html
Polytechnic of Wales
Corpus
A 65,000-word parsed
corpus of the speech of
children ages 612
conversing in playgroups
of three and in interviews
with adults
See: ICAME CD-ROM
Santa Barbara Corpus of
Spoken American English
Large corpus containing
samples of varying length
of different kinds of spoken
American English:
spontaneous dialogues,
monologues, speeches,
radio broadcasts, etc.
Project website:
http://www.linguistics.ucsb.
edu/research/sbcorpus/
default.htm
First release of corpus can
be purchased from the
Linguistic Data
Consortium (LDC):
Catalog/LDC2000S85.html
Susanne Corpus 130,000 words of written
English based on various
genres in the Brown
Corpus that have been
parsed and marked up
based on an annotation
scheme developed for the
project
Project website:
http://www.cogs.susx.ac.uk/
users/geoffs/RSue.html
Switchboard Corpus 2,400 telephone
conversations between two
speakers from various
dialect regions in the
United States; topics of
conversations were
suggested beforehand
Linguistic Data
Consortium (LDC)
website:
TalkBank Project Cross-disciplinary effort to
use computational tools to
study human and animal
communication
Project website:
http://www.talkbank.org/
150 Appendix 1
Tampere Corpus A corpus proposed to
consist of various kinds of
scientic writing for
specialized and
non-specialized audiences
Described in Norri and
Kyt o (1996)
Text Encoding Initiative
(TEI)
Has developed standards
for the annotation of
electronic documents
Project website:
http://www.tei-c.org/
TIMIT Acoustic-Phonetic
Continuous Speech Corpus
Various speakers from
differing dialects of
American English reading
ten sentences containing
phonetically varied sounds
Linguistic Data
Consortium (LDC)
website:
TIPSTER Corpus Collection of various kinds
of written English, such as
Wall Street Journal and
Associated Press news
stories; intended for
research in information
retrieval
Linguistic Data
Consortium (LDC)
website:
Catalog/LDC93T3A.html
Wellington Corpus One million words of
written New Zealand
English divided into genres
that parallel the Brown and
LOB corpora but that were
collected between 1986
and 1990
See: ICAME CD-ROM
York Corpus 1.5 million words taken
from sociolinguistic
interviews with speakers of
York English
See Tagliamonte (1998)
Appendix 2
Concordancing programs
PC/Macintosh-based programs
Conc (for the Macintosh)
John Thomson
available from Summer Institute of Linguistics
http://www.indiana.edu/letrs/help-services/QuickGuides/about-conc.html
Concordancer for Windows
Zdenek Martinek in collaboration with Les Siegrist
http://www.ifs.tu-darmstadt.de/sprachlit/wconcord.htm
Corpus Presenter
Raymond Hickey
http://www.uni-essen.de/lan300/corpus presenter.htm
Corpus Wizard
Kobe Phoenix Lab, Japan
http://www2d.biglobe.ne.jp/htakashi/software/CWNE.HTM
Lexa
Raymond Hickey
Available from Norwegian Computing Centre for the Humanities
http://www.hd.uib.no/lexainf.html
MonoConc Pro 2.0
Athelstan
http://www.athel.com/mono.html#monopro
ParaConc (for multilingual corpora)
Michael Barlow
http://www.athel.com/
Sara
http://info.ox.ac.uk/bnc/sara/client.html
Tact
Centre for Computing in the Humanities, University of Toronto
http://www.chass.utoronto.ca:8080/cch/tact.html
WordCruncher (now called Document Explorer)
Hamilton-Locke, Inc. (an older version is also available on the ICAME
CD-ROM, 2nd edn.)
http://hamilton-locke.com/DocExplorer/Index.html
151
152 Appendix 2
WordSmith
Mike Scott
Oxford University Press
http://www.oup.com/elt/global/catalogue/multimedia/wordsmithtools3/
Web-based programs
CobuildDirect
http://titania.cobuild.collins.co.uk/direct info.html
KWiCFinder
http://miniappolis.com/KWiCFinder/KWiCFinderHome.html
The Michigan Corpus of Academic Spoken English (MICASE)
http://www.hti.umich.edu/micase/
Sara
Online version of the British National Corpus
http://sara.natcorp.ox.ac.uk
TACTWeb
http://kh.hd.uib.no/tactweb/homeorg.htm
References
Aarts, Bas (1992) Small Clauses in English: The Nonverbal Types. Berlin and NewYork:
Mouton de Gruyter.
(2001) Corpus Linguistics, Chomsky, and Fuzzy Tree Fragments. In Mair and Hundt
(2001). 513.
Aarts, Bas and Charles F. Meyer (eds.) (1995) The Verb in Contemporary English.
Cambridge University Press.
Aarts, Jan and Willem Meijs (eds.) (1984) Corpus Linguistics: Recent Developments
in the Use of Computer Corpora. Amsterdam: Rodopi.
Aarts, Jan, Pieter de Haan, and Nelleke Oostdijk (eds.) (1993) English Language Cor-
pora: Design, Analysis, and Exploitation. Amsterdam: Rodopi.
Aarts, Jan, Hans van Halteren, and Nelleke Oostdijk (1996) The TOSCA Analysis
System. In Koster and Oltmans (1996). 18191.
Aijmer, Karin and Bengt Altenberg (eds.) (1991) English Corpus Linguistics. London:
Longman.
Altenberg, Bengt and Marie Tapper (1998) The Use of Adverbial Connectors in
Advanced Swedish Learners Written English. In Granger (1998). 8093.
Ammon, U., N. Dittmar, and K. J. Mattheier (eds.) (1987) Sociolinguistics: An Interna-
tional Handbook of the Science of Language and Society, vol. 2. Berlin: de Gruyter.
Aston, Guy and Lou Burnard (1998) The BNCHandbook: Exploring the British National
Corpus with SARA. Edinburgh University Press.
Atwell, E., G. Demetriou, J. Hughes, A. Schiffrin, C. Souter, and S. Wilcock (2000)
A Comparative Evaluation of Modern English Corpus Grammatical Annotation
Schemes. ICAME Journal 24. 723.
Barlow, Michael (1999) MonoConc 1.5 and ParaConc. International Journal of Corpus
Linguistics 4 (1). 31927.
Bell, Alan (1988) The British Base and the American Connection in NewZealand Media
English. American Speech 63. 32644.
Biber, Douglas (1988) Variation Across Speech and Writing. New York: Cambridge
University Press.
(1990) Methodological Issues Regarding Corpus-based Analyses of Linguistic
Variation. Literary and Linguistic Computing 5. 25769.
(1993) Representativeness in Corpus Design. Literary and Linguistic Computing 8.
24157.
(1995) Dimensions of Register Variation: ACross-Linguistic Comparison. Cambridge
University Press.
Biber, Douglas, Edward Finegan, and Dwight Atkinson (1994) ARCHER and its
Challenges: Compiling and Exploring a Representative Corpus of English
Historical Registers. In Fries, Tottie, and Schneider (1994). 113.
153
154 References
Biber, Douglas, Susan Conrad, and Randi Reppen (1998) Corpus Linguistics:
Investigating Language Structure and Language Use. Cambridge University
Press.
Biber, Douglas, Stig Johansson, Geoffrey Leech, Susan Conrad, and Edward Finegan
(1999) The Longman Grammar of Spoken and Written English. London: Longman.
Biber, Douglas and Jen a Burges (2000) Historical Change in the Language Use of
Women and Men: Gender Differences in Dramatic Dialogue. Journal of English
Blachman, Edward, Charles F. Meyer, and Robert A. Morris (1996) The UMBIntelligent
ICE Markup Assistant. In Greenbaum (1996a). 5464.
Brill, Eric (1992) A Simple Rule-Based Part-of-Speech Tagger. Proceedings of the 3rd
Conference on Applied Natural Language Processing. Trento: Italy.
Burnard, Lou (1995) The Text Encoding Initiative: An Overview. In Leech, Myers, and
Thomas (1995). 6981.
(1998) The Pizza Chef: A TEI Tag Set Selector. http://www.hcu.ox.ac.uk/TEI/pizza.
html.
Burnard, Lou and C. M. Sperberg-McQueen (1995) TEI Lite: An Introduction to Text
Encoding for Interchange. http://www.tei-c.org/Lite/index.html.
Burnard, Lou and Tony McEnery (eds.) (2000) Rethinking Language Pedagogy from a
Corpus Perspective. Frankfurt: Peter Lang.
Chafe, Wallace (1994) Discourse, Consciousness, and Time. Chicago: University of
Chicago Press.
(1995) Adequacy, User-Friendliness, and Practicality in Transcribing. In Leech,
Myers, and Thomas (1995). 5461.
Chafe, Wallace, John Du Bois, and Sandra Thompson (1991) Towards a New Corpus of
American English. In Aijmer and Altenberg (1991). 6482.
Chomsky, Noam (1995) The Minimalist Program. Cambridge, MA: MIT Press.
Coates, Jennifer (1983) The Semantics of the Modal Auxiliaries. London: Croom
Helm.
Collins, Peter (1991a) The Modals of Obligation and Necessity in Australian English.
In Aijmer and Altenberg (1991). 14565.
(1991b) Cleft and Pseudo-Cleft Constructions in English. Andover: Routledge.
Composition of the BNC. http://info.ox.ac.uk/bnc/what/balance.html.
Cook, Guy (1995) Theoretical Issues: Transcribing the Untranscribable. In Leech,
Corpus Encoding Standard (2000) http://www.cs.vassar.edu/CES/122.
Crowdy, Steve (1993) Spoken Corpus Design. Literary and Linguistic Computing 8.
25965.
Curme, G. (1947) English Grammar. New York: Harper and Row.
Davies, Mark (2001) Creating and Using Multi-million Word Corpora from Web-based
Newspapers. In Simpson and Swales (2001). 5875.
de Haan, Pieter (1984) Problem-Oriented Tagging of English Corpus Data. In Aarts and
Meijs (1984). 12339.
DuBois, John, StephanSchuetze-Coburn, Susanna Cumming, andDanae Paolino(1993)
Outline of Discourse Transcription. In Edwards and Lampert (1993). 4589.
Dunning, Ted (1993) Accurate Methods for the Statistics of Surprise and Coincidence.
Computational Linguistics 19 (1). 6174.
References 155
Eckman, Fred (ed.) (1977) Current Themes in Linguistics. Washington, DC: John Wiley.
Edwards, Jane (1993) Principles and Contrasting Systems of Discourse Transcription.
In Edwards and Lampert (1993). 331.
Edwards, Jane and Martin Lampert (eds.) (1993) Talking Data. Hillside, NJ: Lawrence
Erlbaum.
Ehlich, Konrad (1993) HIAT: A Transcription System for Discourse Data. In Edwards
and Lampert (1993). 12348.
Elsness, J. (1997) The Perfect and the Preterite in Contemporary and Earlier English.
Berlin and New York: Mouton de Gruyter.
Fang, Alex (1996) AUTASYS: Automatic Tagging and Cross-Tagset Mapping. In
Greenbaum (1996a). 11024.
Fernquest, Jon (2000) Corpus Mining: Perl Scripts and Code Snippets. http://www.
codearchive.com/home/jon/program.html.
Fillmore, Charles (1992) Corpus Linguistics or Computer-Aided Armchair Linguistics.
In Svartvik (1992). 3560.
Finegan, Edward and Douglas Biber (1995) That and Zero Complementisers in Late
ModernEnglish: ExploringARCHERfrom16501990. InAarts andMeyer (1995).
24157.
Francis, W. Nelson (1979) A Tagged Corpus Problems and Prospects. In Greenbaum,
Leech, and Svartvik (1979). 192209.
(1992) Language Corpora B.C. In Svartvik (1992). 1732.
Francis, W. Nelson and H. Ku cera (1982) Frequency Analysis of English Usage: Lexicon
and Grammar. Boston: Houghton Mifin.
Fries, Udo, Gunnel Tottie, and Peter Schneider (eds.) (1994) Creating and Using English
Language Corpora. Amsterdam: Rodopi.
Garside, Roger, Geoffrey Leech, and Geoffrey Sampson (1987) The Computational
Analysis of English. London: Longman.
Garside, Roger, Geoffrey Leech, and Tam as V aradi (1992) Lancaster Parsed Corpus.
Manual to accompany the Lancaster Parsed Corpus. http://khnt.hit.uib.no/
icame/manuals/index.htm.
Garside, Roger, Geoffrey Leech, and Anthony McEnery (eds.) (1997) Corpus Annota-
tion. London: Longman.
Garside, Roger and Nicholas Smith (1997) A Hybrid Grammatical Tagger: CLAWS 4.
In Garside, Leech, and McEnery (1997). 102121.
Gavioli, Laura (1997) Exploring Texts through the Concordancer: Guiding the Learner.
In Anne Wichmann, Steven Fligelstone, Tony McEnery, and Gerry Knowles (eds.)
(1997) Teaching and Language Corpora. London: Longman. 8399.
Gillard, Patrick and Adam Gadsby (1998) Using a Learners Corpus in Compiling ELT
Dictionaries. In Granger (1998). 15971.
Granger, Sylvianne (1993) International Corpus of Learner English. In Aarts, de Haan,
and Oostdijk (1993). 5771.
(1998) Learner English on Computer. London: Longman.
Greenbaum, Sidney (1973) Informant Elicitation of Data on Syntactic Variation. Lingua
31. 20112.
(1975) Syntactic Frequency and Acceptability. Lingua 40. 99113.
(1984) Corpus Analysis and Elicitation Tests. In Aarts and Meijs (1984). 195201.
(1992) A New Corpus of English: ICE. In Svartvik (1992). 17179.
156 References
(ed.) (1996a) Comparing English Worldwide: The International Corpus of English.
Oxford: Clarendon Press.
(1996b) The Oxford English Grammar. Oxford: Oxford University Press.
Greenbaum, Sidney, Geoffrey Leech, and Jan Svartvik (eds.) (1979) Studies in English
Linguistics. London: Longman.
Greenbaum, S. and Meyer, C. (1982) Ellipsis and Coordination: Norms and Preferences.
Language and Communication 2:13749.
Greenbaum, Sidney and Jan Svartvik (1990) The LondonLund Corpus of Spoken
English. In Svartvik (1990). 1145.
Greenbaum, Sidney and Ni Yibin (1996) About the ICE Tagset. In Greenbaum (1996a).
92109.
Greenbaum, Sidney, Gerald Nelson, and Michael Weizman (1996) Complement Clauses
in English. In Thomas and Short (1996). 7691.
Greene, B. B. and G. M. Rubin (1971) Automatic Grammatical Tagging. Technical
Report. Department of Linguistics: Brown University.
Hadley, Gregory (1997) Sensing the Winds of Change: An Introduction to Data-Driven
Learning. http://web.bham.ac.uk/johnstf/winds.htm.
Haegeman, Lilliane (1987) Register Variation in English: Some Theoretical Observa-
tions. Journal of English Linguistics 20 (2). 23048.
(1991) Introduction to Government and Binding Theory. Oxford: Blackwell.
Halteren, Hans van and Theo van den Heuvel (1990) Linguistic Exploitation of Syntactic
Databases. The Use of the Nijmegen Linguistic DataBase Program. Amsterdam:
Rodopi.
Haselrud, V. and Anna-Brita Stenstr om (1995) Colt: Mark-up and Trends. Hermes 13.
5570.
Hasselg ard, Hilde (1997) Sentence Openings inEnglishandNorwegian. InLjung(1997).
320.
Hickey, Raymond, Merja Kyt o, Ian Lancashire, and Matti Rissanen (eds.) (1997) Tracing
the Trail of Time. Proceedings from the Second Diachronic Corpora Workshop.
Amsterdam: Rodopi.
Hockey, Susan (2000) Electronic Texts in the Humanities. Oxford: Oxford University
Press.
J arvinen, Timo (1994) Annotating 200 Million Words: The Bank of English Project.
Proceedings of COLING 94, Kyoto, Japan. http://www.lingsoft./doc/engcg/
Bank-of-English.html.
Jespersen, Otto (190949) A Modern English Grammar on Historical Principles.
Copenhagen: Munksgaard.
Johansson, Stig and Knut Hoand (1994) Towards an EnglishNorwegian Parallel
Corpus. In Fries, Tottie, and Schneider (1994). 2537.
Johansson, Stig and Jarle Ebeling (1996) Exploring the EnglishNorwegian Parallel
Corpus. In Percy, Meyer, and Lancashire (1996).
Johns, Tim F. (1994) From Printout to Handout: Grammar and Vocabulary Teaching in
the Context of Data-driven Learning. In Odlin (1994). 293313.
Kalton, Graham (1983) Introduction to Survey Sampling. Beverly Hills, CA: Sage.
Kennedy, Graeme (1996) Over Once Lightly. In Percy, Meyer, and Lancashire (1996).
25362.
Kettemann, Bernhard (1995) On the Use of Concordancing in ELT. TELL&CALL 4.
415.
References 157
Kirk, John (1992) The Northern Ireland Transcribed Corpus of Speech. In Leitner
(1992). 6573.
Koster, C. and E. Oltmans (eds.) (1996) Proceedings of the rst AGFL Workshop.
Nijmegen: CSI.
Kretzschmar, William A., Jr. (2000) Review of SPSS Student Version 9.0 for Windows.
Journal of English Linguistics 28 (3). 31113.
Kretzschmar, WilliamA., Jr. and E. Schneider (1996) Introduction to Quantitative Anal-
ysis of Linguistic Survey Data. Los Angeles: Sage.
Kretzschmar, William A., Jr., Charles F. Meyer, and Dominique Ingegneri (1997) Uses
of Inferential Statistics in Corpus Linguistics. In Ljung (1997). 16777.
Kyt o, M. (1991) Variation and Diachrony, with Early American English in Focus. Stud-
ies on can/may and shall/will. University of Bamberg Studies in English
Linguistics 28. Frankfurt am Main: Peter Lang.
(1996) Manual to the Diachronic Part of the Helsinki Corpus of English Texts: Coding
Conventions and Lists of Source Texts. 3rd edn. Department of English: University
of Helsinki.
Labov, W. (1972) The Transformation of Experience in Narrative Syntax. In Language
in the Inner City. Philadelphia: University of Pennsylvania Press. 35496.
Landau, Sidney (1984) Dictionaries: The Art and Craft of Lexicography. New York:
Charles Scribner.
Leech, Geoffrey (1992) Corpora and Theories of Linguistic Performance. In Svartvik
(1992). 10522.
(1997) Grammatical Tagging. In Garside, Leech, and McEnery (1997). 1933.
(1998) Preface. In Granger (1998). xivxx.
Leech, Geoffrey, Roger Garside, and Eric Atwell (1983) The Automatic Grammatical
Tagging of the LOB Corpus. ICAME Journal 7. 1333.
Leech, Geoffrey, Greg Myers, and Jenny Thomas (eds.) (1995) Spoken English on Com-
puter. Harlow, Essex: Longman.
Leech, Geoffrey and Elizabeth Eyes (1997) Syntactic Annotation: Treebanks. In Garside,
Leech, and McEnery (1997). 3452.
Leitner, Gerhard (ed.) (1992) New Directions in English Language Corpora. Berlin:
Mouton de Gruyter.
Le on, Fernando S anchez and Amalio F. Nieto Serrano (1997) Retargeting a Tagger. In
Garside, Leech, and McEnery (eds.). 15165.
Ljung, Magnus (ed.) (1997) Corpus-based Studies in English. Amsterdam: Rodopi.
MacWhinney, Brian (1996) The CHILDES System. American Journal of Speech-
Language Pathology 5. 514.
(2000) The CHILDES Project: Tools for Analyzing Talk. 3rd edn., vol. 1: Transcription
Format and Programs, vol 2: The Database. Mahwah, NJ: Erlbaum.
Mair, Christian (1990) Innitival Complement Clauses in English. Cambridge Uni-
versity Press.
(1995) Changing Patterns of Complementation, and Concomitant Grammaticalisa-
tion, of the Verb Help in Present-Day British English. In Aarts and Meyer (1995).
25872.
Mair, Christian and Marianne Hundt (eds.) (2001) Corpus Linguistics and Linguistic
Theory. Amsterdam: Rodopi.
Maniez, Fran cois (2000) Corpus of English Proverbs and Set Phrases. Message posted
on the Corpora List, 24 January. http://www.hit.uib.no/corpora/20001/0057.html.
158 References
Marcus, M., B. Santorini, and M. Marcinkiewicz (1993) Building a Large Annotated
Corpus of English: The Penn Treebank. Computational Linguistics 19. 31430.
Markus, Manfred (1997) Normalization of Middle English Prose in Practice. In Ljung
(1997). 21126.
Mel cuk, Igor A. (1987) Dependency Syntax: Theory and Practice. Albany: State Uni-
versity of New York Press.
Melamed, Dan (1996) 170 General Text Processing Tools (Mostly in PERL5).
http://www.cis.upenn.edu/melamed/genproc.html.
Meurman-Solin, Anneli (1995) ANewTool: The Helsinki Corpus of Older Scots (1450
1700). ICAME Journal 19. 4962.
Meyer, Charles F. (1992) Apposition in Contemporary English. Cambridge University
Press.
(1995) Coordination Ellipsis in Spoken and Written American English. Language
Sciences 17. 24169.
(1996) Coordinate Structures in the British and American Components of the Inter-
national Corpus of English. World Englishes 15. 2941.
(1997) Minimal Markup for ICE Texts. ICE NEWSLETTER 25. http://www.cs.umb.
edu/meyer/icenews2.html.
(1998) Studying Usage on the World Wide Web. http://www.cs.umb.edu/
meyer/usage.html.
Meyer, Charles F. and Richard Tenney (1993) Tagger: An Interactive Tagging Program.
In Souter and Atwell (1993). 30212.
Meyer, Charles F., Edward Blachman, and Robert A. Morris (1994) Can You See Whose
Speech Is Overlapping? Visible Language 28 (2). 11033.
Milton, John and Robert Freeman (1996) Lexical Variation in the Writing of Chinese
Learners of English. In Percy, Meyer, and Lancashire (1996). 12131.
Mindt, Dieter (1995) An Empirical Grammar of the English Verb. Berlin: Cornelsen.
M onnink, Inga de (1997) Using Corpus and Experimental Data: A Multimethod
Approach. In Ljung (1997). 22744.
Murray, Thomas E. and Carmen Ross-Murray (1992) On the Legality and Ethics of
Surreptitious Recording. Publication of the American Dialect Society 76. 1575.
(1996) Under Cover of Law: More on the Legality of Surreptitious Recording.
Publication of the American Dialect Society 79. 182.
Nelson, Gerald (1996) Markup Systems. In Greenbaum (1996a). 3653.
Nevalainen, Terttu (2000) Gender Differences in the Evolution of Standard English:
Evidence from the Corpus of Early English Correspondence. Journal of English
Nevalainen, Terttu, and Helena Raumolin-Brunberg (eds.) (1996) Sociolinguistics and
Language History: Studies Based on the Corpus of Early English Correspondence.
Amsterdam: Rodopi.
Newmeyer, Frederick (1998) Language Form and Language Function. Cambridge,
MA: MIT Press.
Nguyen, Long, Spyros Matsoukas, Jason Devenport, Daben Liu, Jay Billa, Francis
Kubala, and John Makhoul (1999) Further Advances in Transcription of Broadcast
News. Proceedings of the 6th European Conference on Speech Communication and
Technology, vol. 2. Edited by G. Olaszy, G. Nemeth, K. Erdohegy (aka EuroSpeech
99). European Speech Communication Association (ESCA). 66770.
References 159
Norri, Juhani and Merja Kyt o (1996) A Corpus of English for Specic Purposes: Work
in Progress at the University of Tampere. In Percy, Meyer, and Lancashire (1996).
15969.
Oakes, Michael P. (1998) Statistics for Corpus Linguistics. Edinburgh University Press.
Odlin, Terrence (ed.) (1994) Perspectives on Pedagogical Grammar. New York:
Cambridge University Press.
Ooi, Vincent (1998) Computer Corpus Lexicography. Edinburgh University Press.
Oostdijk, Nelleke (1991) Corpus Linguistics and the Automatic Analysis of English.
Amsterdam: Rodopi.
Pahta, P aivi and Saara Nevanlinna (1997) Re-phrasing in Early English. Expository
Apposition with an Explicit Marker from 1350 to 1710. In Rissanen, Kyt o, and
Heikkonen (1997). 12183.
Percy, Carol, Charles F. Meyer, and Ian Lancashire (eds.) (1996) Synchronic Corpus
Linguistics. Amsterdam: Rodopi.
Porter, Nick and Akiva Quinn (1996) Developing the ICE Corpus Utility Program. In
Greenbaum (1996a). 7991.
Powell, Christina and Rita Simpson (2001) Collaboration between Corpus Linguists and
Digital Librarians for the MICASE Web Search Interface. In Simpson and Swales
(2001). 3247.
Prescott, Andrew (1997) The Electronic Beowulf and Digital Restoration. Literary and
Linguistic Computing 12. 18595.
Quinn, Akiva and Nick Porter (1996) Annotation Tools. In Greenbaum (1996a). 6578.
Quirk, Randolph (1992) On Corpus Principles and Design. In Svartvik (1992). 45769.
Quirk, Randolph, Sidney Greenbaum, Geoffrey Leech, and Jan Svartvik (1972) AGram-
mar of Contemporary English. London: Longman.
(1985) A Comprehensive Grammar of the English Language. London: Longman.
Renouf, Antoinette (1987) Corpus Development. In Sinclair (1987). 140.
Rissanen, Matti (1992) The Diachronic Corpus as a Window to the History of English.
In Svartvik (1992). 185205.
(2000) The World of English Historical Corpora: fromCdmon to the Computer Age.
Journal of English Linguistics 28 (1). 720.
Rissanen, Matti, Merja Kyt o, and Kirsi Heikkonen (eds.) (1997) English in Transition:
Corpus-based Studies in English Linguistics and Genre Styles. Topics in English
Linguistics 23. Berlin and New York: Mouton de Gruyter.
Robinson, Peter (1998) New Methods of Editing, Exploring, and Reading The Canter-
bury Tales. http://www.cta.dmu.ac.uk/projects/ctp/desc2.html.
Rocha, Marco (1997) A Probabilistic Approach to Anaphora Resolution in Dialogues
in English. In Ljung (1997). 26179.
Ryd en, Mats (1975) Noun-Name Collocations in British Newspaper Language. Studia
Neophilologica 67. 1439.
Sampson, Geoffrey (1998) Corpus Linguistics User Needs. Message posted to the
Corpora List, 29 July. http://www.hd.uib.no/corpora/19983/0030.html.
Samuelsson, Christer and Atro Voutilainen (1997) Comparing a Linguistic and a
Stochastic Tagger. Proceedings of the 35th Annual Meeting of the Association for
Computational Linguistics and the 8th Conference of the European Chapter of the
Association for Computational Linguistics. Madrid: Association for Computational
Linguistics. 24653.
160 References
S anchez, Aquilino and Pascual Cantos (1997) Predictability of Word Forms (Types)
and Lemmas in Linguistic Corpora. A Case Study Based on the Analysis of the
CUMBRE Corpus: An 8-Million-Word Corpus of Contemporary Spanish. Inter-
national Journal of Corpus Linguistics 2. 25980.
Sanders, Gerald (1977) A Functional Typology of Elliptical Coordinations. In Eckman
(1977). 24170.
Sankoff, David (1987). Variable rules. In Ammon, Dittmar, and Mattheier (1987).
98497.
Schmied, Josef (1996) Second-Language Corpora. In Greenbaum (1996a). 18296.
Schmied, Josef and Hildegard Sch afer (1996) Approaching Translationese Through
Parallel and Translation Corpora. In Percy, Meyer and Lancashire (1996).
4155.
Schmied, Josef, and Claudia Claridge (1997) Classifying Text- or Genre-Variation in
the Lampeter Corpus of Early Modern English Texts. In Hickey, Kyt o, Lancashire,
and Rissanen (1997). 11935.
Sigley, Robert J. (1997) Choosing Your Relatives: Relative Clauses in New Zealand
English. Unpublished PhD thesis. Wellington: Department of Linguistics, Victoria
University of Wellington.
Simpson, Rita, Bret Lucka, and Janine Ovens (2000) Methodological Challenges of
Planning a Spoken Corpus with Pedagogical Outcomes. In Burnard and McEnery
(2000). 439.
Simpson, Rita and John Swales (eds.) (2001) Corpus Linguistics in North America. Ann
Arbor: University of Michigan Press.
Sinclair, John (ed.) (1987) Looking Up: An Account of the COBUILD Project. London:
Collins.
(1991) Corpus, Concordance, Collocation. Oxford University Press.
(1992) Introduction. BBC English Dictionary. London: HarperCollins. xxiii.
Smith, Nicholas (1997) Improving a Tagger. In Garside, Leech and McEnery (1997).
13750.
Souter, Clive and Eric Atwell (eds.) (1993) Corpus-Based Computational Linguistics.
Amsterdam: Rodopi.
Sperberg-McQueen, C. M. and Lou Burnard (eds.) (1994a) Guidelines for Electronic
Text Encoding and Interchange ( TEI P3). http://etext.virginia.edu/TEI.html.
(1994b) A Gentle Introduction to SGML. In Guidelines for Electronic Text En-
coding and Interchange ( TEI P3). http://etext.lib.virginia.edu/bin/tei-tocs?div=
DIV1&id=SG.
Stenstr om, Anna-Brita and Gisle Andersen (1996) More Trends in Teenage Talk: A
Corpus-Based Investigation of the Discourse Items cos and innit. In Percy, Meyer,
and Lancashire (1996). 189203.
Svartvik, J. (ed.) (1990) The LondonLund Corpus of Spoken English. Lund University
Press.
(1992) Directions in Corpus Linguistics. Berlin: Mouton de Gruyter.
Svartvik, Jan and Randolph Quirk (eds.) (1980) ACorpus of English Conversation. Lund
University Press.
Tagliamonte, Sali (1998) Was/Were Variation across the Generations: View from the
City of York. Language Variation and Change 10 (2). 15391.
References 161
Tagliamonte, Sali and Helen Lawrence (2000) I Used to Dance, but I Dont Dance
Now: The Habitual Past in English. Journal of English Linguistics 28 (4). 324
53.
Tannen, D. (1989) Talking Voices: Repetition, Dialogue, and Imagery in Conversational
Discourse. Cambridge University Press.
Tapanainen, Pasi and Timo J arvinen (1997) A Non-projective Dependency Parser.
http://www.conexor./anlp97/anlp97.html [also published in Procs. ANLP-97.
ACL. Washington, DC].
The Independent Style Manual, 2nd edn. (1988) London: The Independent.
The Spoken Component of the BNC. http://info.ox.ac.uk/bnc/what/spok design.html.
The Written Component of the BNC. http://info.ox.ac.uk/bnc/what/writ design.html.
Thomas, Jenny and Mick Short (eds.) (1996) Using Corpora for Language Research.
London: Longman.
Thompson, Henry S., Anne H. Anderson, and Miles Bader (1995) Publishing a Spoken
and Written Corpus on CD-ROM: The HCRC Map Task Experience. In Leech,
Tottie, G. (1991) Negation in English Speech and Writing. A Study in Variation. San
Diego: Academic Press.
van Halteren, Hans and Theo van den Heuvel (1990) Linguistic Exploitation of Syntactic
Databases. Amsterdam: Rodopi.
Voutilainen, Atro (1999) A Short History of Tagging. In Hans van Halteren (ed.)
Syntactic Wordclass Tagging (1999). Dordrecht: Kluwer.
Voutilainen, Atro and Mikko Silvonen (1996) A Short Introduction to ENGCG.
http://www.lingsoft./doc/engcg/intro.
Wheatley, B., G. Doddington, C. Hemphill, J. Godfrey, E.C. Holliman, J. McDaniel, and
D. Fisher (1992) Robust Automatic Time Alignment of Orthographic Transcriptions
with Unconstrained Speech. Proceedings of ICASSP-92, vol. 1. 5336.
Willis, Tim (1996) Analysing the Lancaster/IBM Spoken English Corpus (SEC) Using
the TOSCA Analysis System (for ICE): Some Impressions from a User. In Percy,
Meyer, and Lancashire (1996). 23751.
Wilson, A. and Tony McEnery (1994) Teaching and Language Corpora. Technical
Report. Department of Modern English Language and Linguistics, University of
Lancaster.
Wilson, Andrewand Jenny Thomas (1997) Semantic Annotation. In Garside, Leech and
McEnery (1997). 5365.
Woods, A, P. Fletcher, and A. Hughes (1986) Statistics in Language Studies. Cambridge
University Press.
Index
Aarts, Bas, 4, 102
adequacy, 23, 1011
age, 4950
Altenberg, Bengt, 267
AMALGAM Tagging Project, 867, 89
American National Corpus, 24, 84, 142
American Publishing House for the Blind
Corpus, 17, 142
analyzing a corpus, 100
determining suitability, 1037, 107t
exploring a corpus, 1234
extracting information: dening parameters,
1079; coding and recording, 10914,
112t; locating relevant constructions,
11419, 116f, 118f
framing research question, 1013
future prospects, 1401
see also pseudo-titles (corpus analysis
case study); statistical analysis
anaphors, 97
annotation, 989
grammatical markup, 81
parsing, 916, 98, 140
part-of-speech markup, 81
structural markup, 689, 816
tagging, 8691, 978, 111, 11718, 140
types, 81
appositions, 42, 98
case study)
ARCHER (A Representative Corpus of
English Historical Registers), 21,
22, 79 n6, 140, 142
Aston, Guy, 19
AUTASYS Tagger, 87
Bank of English Corpus, 15, 96, 142
BBC English Dictionary, 15, 1617
Bell, Alan, 100, 1013, 104, 108, 110, 131
Bergen Corpus of London Teenage English
see COLT Corpus
Biber, Douglas, 10, 1920, 22, 32, 33,
36, 3940, 41, 42, 52, 78, 121,
122, 126
Biber, Douglas, et al. (1999) 14
Birmingham Corpus, 15, 142
Blachman, Edward, 767
BNC see British National Corpus
Brill, Eric, 86
Brill Tagger, 868
British National Corpus (BNC), 143
annotation, 84
composition, 18, 31t, 34, 36, 38, 401, 49
copyright, 13940
planning, 302, 33, 43, 51, 138
record keeping, 66
research using, 15, 36
speech samples, 59
tagging, 87
time-frame, 45
British National Corpus (BNC) Sampler,
13940, 143
Brown Corpus, xii, 1, 143
genre variation, 18
length, 32
research using, 6, 910, 12, 42, 98, 103
sampling methodology, 44
tagging, 87, 90
time-frame, 45
see also FROWN (FreiburgBrown)
Corpus
Burges, Jen, 52
Burnard, Lou, 19, 82, 84, 856
Cambridge International Corpus, 15, 143
Cambridge Learners Corpus, 15, 143
Canterbury Project, 79, 143
Cantos, Pascual, 33 n2
Chafe, Wallace, 3, 32, 52, 72, 85
CHAT system (Codes for the Human Analysis
of Transcripts), 26, 11314
Chemnitz Corpus, 23, 143
CHILDES (Child Language Data Exchange
System) Corpus, xiii, 26, 113, 144
Chomsky, Noam, 2, 3
CIA (contrastive interlanguage analysis), 26
CLAN software programs, 26
CLAWS tagger, 25, 87, 8990
Coates, Jennifer, 12, 13
162
Index 163
collecting data
general considerations, 556
record keeping, 646
speech samples, 56; broadcasts, 61;
future prospects, 139; microphones, 60;
natural speech, 568, 59; permission,
57; problems, 601; recording, 589;
sample length, 578; tape recorders, 5960
writing samples: copyright, 38, 612, 79 n6,
13940; electronic texts, 634; future
prospects, 139; sources, 624
see also sampling methodology
Collins, Peter, xiixiii
Collins COBUILD English Dictionary, 15
Collins COBUILD Project, 14, 15
COLT Corpus (Bergen Corpus of London
Teenage English), xiiixiv, 18, 49, 142
competence vs. performance, 4
computerizing data
directory structure, 67, 68f
le format, 667
markup, 67, 689 see also annotation
speech, see speech, computerizing
written texts, 7880, 139
concordancing programs
KWIC format, 11516, 116f
for language learning, 278
lemma searches, 116
programs, 115, 117, 1501
with tagged or parsed corpus, 11718
uses, 16, 86, 114
wild card searches, 11617
Conrad, Susan, 126
contrastive analysis, 224
contrastive interlanguage analysis (CIA), 26
Cook, Guy, 72, 86
copyright, 38, 44, 57, 612, 79 n6, 13940
Corpora Discussion List, 144
corpus (corpora)
balanced, xii
construction see planning corpus
construction
denitions, xixii
diachronic, 46
historical, 202, 378, 46, 51, 789
learner, 267
monitor, 15
multi-purpose, 36
parallel, 224
parsed, 96
resources, 1429
special-purpose, 36
synchronic, 456
corpus linguistics, xi, xiiixiv, 12, 34
Corpus of Early English Correspondence, 22,
37, 144
Corpus of Middle English Prose and Verse, 144
Corpus of Spoken Professional English, 71,
144
corpus-based research, 11
contrastive analysis, 224
grammatical studies, 1113
historical linguistics, 202
language acquisition, 267
language pedagogy, 278
language variation, 1720
lexicography, 1417
limitations, 124
natural language processing (NLP), xiii,
246
reference grammars, 1314
translation theory, 224
Crowdy, Steve, 43, 59
Curme, G., 13
data-driven learning, 278
de Haan, Pieter, 978
descriptive adequacy, 2, 3
diachronic corpora, 46
dialect variation, 512
dictionaries, 1417
Du Bois, John, 32, 52, 85
Dunning, Ted, 132
EAGLES Project see Expert Advisory Group
on Language Engineering Standards, The
Ebeling, Jarle, 23
education, 50
Ehlich, Konrad, 77
Electronic Beowulf, The, 21, 144
electronic texts, 634
elliptical coordination
frequency, 7, 12
functional analysis, 611
genres, 6, 910
position, 67
repetition in speech, 9
serial position effect, 78, 8t
speech vs. writing, 89
suspense effect, 78, 8t
empty categories, 45
ENGCG Parser, 96
EngCG-2 tagger, 88
EngFDG parser, 91, 934, 934 n8, 96
EnglishNorwegian Parallel Corpus, 23,
62, 144
ethnographic information, 656
see also sociolinguistic variables
Expert Advisory Group on Language
Engineering Standards, The (EAGLES),
xi, 84, 144
explanatory adequacy, 2, 3, 1011
164 Index
Extensible Markup Language see XML
Eyes, Elizabeth, 91
Fernquest, Jon, 114
Fillmore, Charles, 4, 17
Finegan, Edward, 22
Fletcher, P., 1212
FLOB (FreiburgLancasterOsloBergen)
Corpus, 21, 45, 145
frame semantics, 17
Francis, W. Nelson, 1, 88
FROWN (FreiburgBrown) Corpus, 21, 145
FTF see fuzzy tree fragments
functional descriptions of language
elliptical coordination, 611, 8t, 12
repetition in speech, 9
voice, 56
fuzzy tree fragments (FTF), 119, 119f
Gadsby, Adam, 27
Garside, Roger, 889
Gavioli, Laura, 28
gender, 18, 22, 489
generative grammar, 1, 35
genre variation, 18, 1920, 31t, 348, 35t, 402
Gillard, Patrick, 27
government and binding theory, 45
grammar
generative, 1, 35
universal, 23
Grammar Safari, 28
grammars, reference, 1314
grammatical markup see parsers
grammatical studies, 1113
Granger, Sylvianne, 26
Greenbaum, Sidney, 7, 14, 22, 35t, 64, 75, 95
Greene, B. B., 87, 88
Haegeman, Lilliane, 23, 45, 6
Hasselg ard, Hilde, 23
Helsinki Corpus, 145
composition, 201, 38
planning, 46
research using, 22, 37, 51
symbols system, 67
Helsinki Corpus of Older Scots, 145
historical corpora, 202, 378, 46, 51, 789
see also ARCHER; Helsinki Corpus
Hoand, Knut, 23
Hong Kong University of Science and
Technology (HKUST) Learner Corpus,
26, 145
Hughes, A., 1212
ICAME Bibliography, 145
ICAME CD-ROM, 67, 145
ICE (International Corpus of English), 146
annotation, 823, 84, 85, 87, 90
composition, 34, 35t, 36, 38, 39, 402, 104
computerizing data, 72, 73
copyright, 38, 44
criteria, 50
record keeping, 66
regional components, 104, 1056, 106t,
110, 123, 124
research using, 6, 9 see also pseudo-titles
(corpus analysis case study)
sampling, 44, 56
time-frame, 45
see also ICECUP; ICE-GB; ICE-USA
ICE Markup Assistant, 85, 86
ICE Tree, 95
ICECUP (ICE Corpus Utility Program), 19,
116, 119, 146
ICE-East Africa, 106, 106t, 107t, 110,
123t, 124
ICE-GB, 146
annotation, 25, 834, 86, 923, 92f, 96,
11718, 118f, 140
composition, 106t
computerizing data, 73
criteria, 50
record keeping, 645
research using, 14, 19, 11516, 116f
case study)
ICE-Jamaica, 106t, 107t, 110, 123t
ICE-New Zealand, 106, 106t, 107t, 123t,
125, 1303
ICE-Philippines, 106, 106t, 107t, 110, 123t,
125, 1303
ICE-Singapore, 106t, 110, 123t
ICE-USA
composition, 53, 106t
computerizing data, 70, 71, 734, 79
copyright, 62
criteria, 467
directory structure, 678, 68f
length, 323
record keeping, 64, 65
research using see pseudo-titles (corpus
analysis case study)
sampling, 58, 601
ICLE see International Corpus of Learner
English
Ingegneri, Dominique, 423
International Corpus of English see ICE
International Corpus of Learner English
(ICLE), 26, 27, 146
Jespersen, Otto, xii, 13
Johansson, Stig, 23
Index 165
Kalton, Graham, 43
Kennedy, Graeme, 89
Kettemann, Bernhard, 278
Kirk, John, 52
Kolhapur Corpus of Indian English, 104
Kretzschmar, William A., Jr., 423
Kucera, Henry, 1
KWIC (key word in context), 11516, 116f
Kyt, M., 37
Kyt, Merja, 42
Labov, W., 9
Lampeter Corpus, 38, 146
Lancaster Corpus, 12, 147
see also LOB (LancasterOsloBergen)
Corpus
LancasterOsloBergen Corpus see LOB
(LancasterOsloBergen) Corpus
Lancaster Parsed Corpus, 912, 96, 147
Lancaster/IBM Spoken English Corpus, 96,
147
Landau, Sidney, 16
language acquisition, 267, 47
language pedagogy, 278
language variation, 3, 1720
dialect variation, 512
genre variation, 18, 1920, 31t, 348,
35t, 402
sociolinguistic variables, 1819, 22, 4853
style-shifting, 19
Lawrence, Helen, 52, 136
LDB see Linguistic Database Program
LDC see Linguistic Data Consortium
learner corpora, 267
Leech, Geoffrey, xi, 4, 87, 91, 138
lemmas, 16, 116, 116 n5
length
of corpus, 324, 126
of text samples, 3840
lexicography, 1417
Linguistic Data Consortium (LDC), 24,
98, 147
Linguistic Database (LDB) Program, 93, 115
linguistic theory
adequacy, 23, 1011
corpus linguistics, xi, xiiixiv, 12, 34
generative grammar, 1, 35
government and binding theory, 45
minimalist theory, 3
see also functional descriptions of language
LOB (LancasterOsloBergen) Corpus, 12,
1415, 19, 39, 45, 87, 147
see also FLOB
(FreiburgLancasterOsloBergen)
Corpus
London Corpus, 12, 13, 50, 103, 147
LondonLund Corpus, 147
annotation, 82
composition, 53
names in, 75
research using, 12, 19, 39, 42, 98
Longman Dictionary of American English, 15
Longman Dictionary of Contemporary
English, 15
Longman Essential Activator, 27
LongmanLancaster Corpus, 12, 148
Longman Learners Corpus, 26, 27, 148
Longman Spoken and Written English Corpus,
The (LSWE), 14, 90, 148
LSWE see Longman Spoken and Written
English Corpus, The
Mair, Christian, 45
Map Task Corpus, 59, 148
markup, 67, 689
see also annotation
Markus, Manfred, 78, 79
Melcuk, Igor A., 93 n8
Meyer, Charles F., 6, 7, 28, 423, 767,
98, 101, 103
Michigan Corpus of Academic Spoken
English (MICASE), 148, 151
composition, 36, 53
computerization, 69, 723
planning, 139
record keeping, 65
Mindt, Dieter, 1213
minimalist theory, 3
modal verbs, xiixiii, 1213
monitor corpora, 15
Morris, Robert A., 767
multi-purpose corpora, 36
Murray, James A. H., 16
native vs. non-native speakers, 468
natural language processing (NLP), xiii, 246
Nelson, Gerald, 22, 83
Nevalainen, Terttu, 22, 51
Nguyen, Long et al., 70
Nijmegen Corpus, 87, 92, 93, 96, 148
NLP see natural language processing
Norri, Juhani, 42
Northern Ireland Transcribed Corpus
of Speech, The, 52, 148
noun phrases, 56, 1314
null-subject parameter, 23
Oakes, Michael P., 134, 136
observational adequacy, 2
Ooi, Vincent, 15
Oostdijk, Nelleke, 93
Oxford English Dictionary (OED), 16
166 Index
ParaConc, 24, 150
parallel corpora, 224
parsed corpora, 96
parsers
probabilistic, 91
rule-based, 924, 934 n8, 956
parsing a corpus
accuracy, 91, 95
complexity, 935
disambiguation, 95
manual pre-processing, 956
normalization, 96
parsers, 914, 95
post-editing, 95
problem-oriented tagging, 978
speech, 945, 96
treebanks, 912
part-of-speech markup see taggers
PC Tagger, 98, 113, 113f
PennHelsinki Parsed Corpus of Middle
English, 96, 148
Penn Treebank, xii, xiii, 25, 37, 91, 96, 149
Perl programming language, 114
planning corpus construction, 30, 445, 53
British National Corpus, 302, 33, 43, 51,
138
genres, 31t, 348, 35t, 402
length of text samples, 3840
native vs. non-native speakers, 468
number of texts, 403
overall length, 324
range of speakers and writers, 402, 434
sociolinguistic variables, 1819, 4853
time-frame, 456
Polytechnic of Wales Corpus, 96, 149
pro-drop, 23
programming languages, 114
pseudo-titles (corpus analysis case study),
1012
determining suitability, 1037, 107t
extracting information: dening parameters,
1079; coding and recording, 10914,
112t, 113f; locating relevant
constructions, 115, 11719, 118f, 119f
framing research question, 1013
statistical analysis: exploring a corpus,
1234; using quantitative information,
12536, 125t, 127t, 128f, 128t, 130t,
131t, 132t, 133t, 134t, 135t
Quirk, Randolph et al., 1314, 101, 108, 135
record keeping, 646
reference grammars, 1314
Reppen, Randi, 126
research see analyzing a corpus; corpus-based
research
Rissanen, Matti, 20, 21, 378, 46, 51
Rocha, Marco, 97
Rubin, G. M., 87, 88
Ryd en, Mats, 101
sampling methodology
non-probability sampling, 44, 45
probability sampling, 434
sampling frames, 423
see also speech samples
Sampson, Geoffrey, 114
S anchez, Aquilino, 33 n2
Sanders, Gerald, 67
Santa Barbara Corpus of Spoken American
English, 32, 44, 69, 71, 85, 149
Sara concordancing program, 1819,
150, 151
scanners, 79
Sch afer, Hildegard, 234
Schmied, Josef, 234, 47
SEU see Survey of English Usage
SGML (Standard Generalized Markup
Language), 825, 86
Sigley, Robert J., 105 n3, 129, 136
Sinclair, John, 14, 15
small clauses, 4
Smith, Nicholas, 889, 90
social contexts and relationships, 523
sociolinguistic variables, 1819
age, 4950
dialect, 512
education, 50
gender, 18, 22, 489
social contexts and relationships, 523
see also ethnographic information
software programs, 1819, 24, 26, 115
Someya, Yasumasa, 114
special-purpose corpora, 36
speech, computerizing
background noise, 75
detail, 712
extra-linguistic information, 72
iconicity and speech transcription, 758
lexicalized expressions, 723
linked expressions, 73
names of individuals, 75
partially uttered words, 73
principles, 72
punctuation, 745
repetitions, 734
speech-recognition programs, 246, 701
transcription programs, 6970
Index 167
transcription time, 71
unintelligible speech, 74
vocalized pauses, 72
speech, repetition in, 9
speech samples, 56
broadcasts, 61
microphones, 60
natural speech, 568, 59
parsing, 945, 96
permission, 57
problems, 601
recording, 589
sample length, 578
tape recorders, 5960
see also planning corpus construction;
speech, computerizing
speech-recognition programs, 246, 701
Sperberg-McQueen, C. M., 82
Standard Generalized Markup Language see
SGML
statistical analysis, 11921
backward elimination, 135
Bonferroni correction, 129
chi-square test, 12732, 127t, 134, 135
cross tabulation, 125
degrees of freedom, 129
determining suitability, 1212
exploring a corpus, 1234
frequency counts, 120
frequency normalization, 126
kurtosis, 127, 127t
length of corpora, 126
linguistic motivation, 122
log-likelihood (G
2
) test, 132, 135
loglinear analysis, 134, 136
macroscopic analysis, 122
non-parametric tests, 126, 127
normal distribution, 1267
programs, 120, 136
pseudo-titles (case study), 12336, 125t,
127t, 128f, 128t, 130t, 131t, 132t,
133t, 134t, 135t
saturated models, 134
skewness, 127, 127t
using quantitative information, 12536
Varbrul programs, 136
structural markup, 81
detail, 856
display, 86
intonation, 85
SGML, 825, 86
TEI, 67, 84, 85, 86, 98, 149
timing, 689
XML, 84, 86
style-shifting, 19
Survey of English Usage (SEU), 42, 98
Susanne Corpus, 96, 149
Svartvik, Jan, 75
Switchboard Corpus, 25, 149
synchronic corpora, 456
taggers, 8690
accuracy, 8990
probabilistic, 889
rule-based, 88
tagging a corpus, 8691
accuracy, 8990, 91
disambiguation, 88
discourse tagging, 97
limitations, 11718
post-editing, 89
problem-oriented tagging, 978, 111
semantic tagging, 97
taggers, 8690
tagsets, 86, 87, 901
see also SGML; Text Encoding Initiative
(TEI)
TAGGIT, 88
Tagliamonte, Sali, 52, 136
tagsets, 86, 87, 901
TalkBank Project, 138, 149
Tampere Corpus, 42, 150
Tannen, D., 9
Tapper, Marie, 267
Text Encoding Initiative (TEI), 67, 84, 85, 86,
98, 150
text samples
length, 3840
number, 403
that, 22
Thomas, Jenny, 97
Thompson, Sandra, 32, 52, 85
time-frame, 456
TIMIT AcousticPhonetic Continuous Speech
Corpus, 245, 150
TIPSTER Corpus, 25, 150
TOSCA parser, 91, 923, 92f, 956
TOSCA tagset, 87
TOSCA Tree editor, 95
transcription programs, 6970
translation theory, 224
tree editors, 95
treebanks, 912
see also Penn Treebank
universal grammar, 23
Varbrul programs, 136
verb complements, 45
verb phrases, 13
168 Index
verbs, modal, xiixiii, 1213
voice, 56
Weizman, Michael, 22
Wellington Corpus, 89, 136, 150
Willis, Tim, 95
Wilson, Andrew, 97
Woods, A., 1212
World Wide Web, 28, 634, 7980,
140, 14251
written texts
collecting data, 614, 139
computerizing, 7880, 139
copyright, 38, 44, 612, 79 n6, 13940
electronic texts, 634
see also planning corpus construction
XML (Extensible Markup Language), 84, 86
York Corpus, 52, 150

(Charles F. Meyer) English Corpus Linguistics An

Uploaded by

Copyright:

Available Formats

(Charles F. Meyer) English Corpus Linguistics An

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(Charles F. Meyer) English Corpus Linguistics An

Uploaded by

Copyright:

Available Formats

This page intentionally left blank

English Corpus Linguistics

To Libby and Freddie

You might also like