Doing Corpus Linguistics

Doing Corpus Linguistics
Doing Corpus Linguistics offers a practical step-by-step introduction to

corpus linguistics, making use of widely available corpora and of a regis
ter analysis-based theoretical framework to provide students in applied
linguistics and TESOL with the understanding and skills necessary to
meaningfully analyze corpora and carry out successful corpus-based
research. Divided into three parts—Introduction to Doing Corpus
Linguistics and Register Analysis; Searches in Available Corpora; and
Building Your Own Corpus, Analyzing Your Quantitative Results,
and Making Sense of Data—the book emphasizes hands-on experience
with performing language analysis research and interpreting findings in
a meaningful and engaging way. Readers are given multiple opportuni-
ties to analyze language data by completing smaller tasks and corpus
projects using publicly available corpora. The book also takes readers
through the process of building a specialized corpus designed to answer
a specific research question and offers detailed information on complet-
ing a final research project that includes both a written paper and an
oral presentation of the reader’s specific research projects. Doing Corpus
Linguistics provides students in applied linguistics and TESOL with the
opportunity to gain proficiency in the technical and interpretive aspects
of corpus research and to encourage them to participate in the growing
field of corpus linguistics.
William J. Crawford is an Associate Professor in the Department of Eng-

lish at Northern Arizona University.
Eniko Csomay is a Professor in the Department of Linguistics and Asian/

Middle Eastern Languages and Associate Dean in the College of Arts and
Letters at San Diego State University.
This page intentionally left blank
William J. Crawford and

Eniko Csomay
First published 2016
by Routledge
711 Third Avenue, New York, NY 10017
and by Routledge
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
Routledge is an imprint of the Taylor & Francis Group, an informa business
© 2016 Taylor & Francis
The right of William J. Crawford and Eniko Csomay to be identified as
authors of this work has been asserted by them in accordance with
sections 77 and 78 of the Copyright, Designs and Patents Act 1988.
All rights reserved. No part of this book may be reprinted or reproduced
or utilised in any form or by any electronic, mechanical, or other means,
now known or hereafter invented, including photocopying and recording,
or in any information storage or retrieval system, without permission in
writing from the publishers.
Trademark notice: Product or corporate names may be trademarks or
registered trademarks, and are used only for identification and explanation
without intent to infringe.
Library of Congress Cataloging-in-Publication Data
Crawford, William J.
Doing corpus linguistics / William J. Crawford and Eniko Csomay.
pages cm
Includes bibliographical references and index.
1. Corpora (Linguistics) 2. Linguistic analysis (Linguistics) I. Csomay,
Eniko. II. Title.
P128.C68C74 2016
420.1’88—dc23 2015013824
ISBN: 978-1-138-02460-1 (hbk)
ISBN: 978-1-138-02461-8 (pbk)
ISBN: 978-1-315-77564-7 (ebk)
Typeset in Sabon
by Apex CoVantage, LLC
Contents
List of Tables vii

List of Figures ix
Preface xi
Acknowledgments xiii
PART 1
Introduction to Doing Corpus Linguistics
and Register Analysis 1
1 Linguistics, Corpus Linguistics, and Language Variation 3

1.1 Language and Rules/Systems 3
1.2 What Is Corpus Linguistics? 5
1.3 Outline of the Book 11
2 Register Analysis 13
2.1 Why Register? 14
2.2 What Is Register Analysis? 16
2.3 Describing Situational Characteristics and
Identifying Variables 17
2.4 Providing a Functional Interpretation 24
2.5 Units of Analysis and Register Studies 28
2.6 End of Chapter Exercises 29
PART 2
Searches in Available Corpora 33
3 Searching a Corpus 35
3.1 KWIC 35
3.2 Collocates 40
vi Contents
3.3 N-Grams 41
3.4 POS Tags 53
4 Projects Using Publically Available Corpora 58

4.1 Word- and Phrase-Based Projects 60
4.2 Grammar-Based Projects 64
PART 3
Building Your Own Corpus, Analyzing Your Quantitative
Results, and Making Sense of Data 73
5 Building Your Own Corpus 75

5.1 Do-It-Yourself Corpora 75
5.2 Deciding on a Corpus Project 76
5.3 Building a Corpus 79
5.4 Software Programs and Your Corpus 85
6 Basic Statistics 94
6.1 Why Do Statistical Analyses? 94
6.2 Basic Terms, Concepts, and Assumptions 94
6.3 How to Go About Getting the Statistical Results 109
7 Statistical Tests (ANOVAs, Chi-square, Pearson Correlation) 117

7.1 Difference Tests 117
7.2 Relationship Tests 131
7.3 How to Go About Getting the Statistical Results 140
8 Doing Corpus Linguistics 142

8.1 Doing Register Analysis of Your Project 142
8.2 Situational Analysis 143
8.3 Linguistic Analysis 145
8.4 Functional Interpretation 151
8.5 Reporting on Your Project 152
9 A Way Forward 156
Index 161
Tables
1.1 Common right collocates of equal and identical in the

Corpus of Contemporary American English (COCA) 7
1.2 Text and corpus comparison (based on
Tognini-Bonelli, 2001) 10
2.1 Key situational differences between an email to a friend
and an email to a superior (boss) (Biber and Conrad
2009: 65) 21
2.2 Situational differences between news writing and
news talk 22
2.3 Key situational differences between student presentations
at a symposium and teacher presentations in the
classroom 23
2.4 Texts for news writing and a news talk 25
2.5 Texts from classroom presentation and symposium
presentation 26
2.6 Texts from email to a friend and email to a boss 30
3.1 Distribution of part of speech categories following the
word ‘say’ 38
3.2 Distribution of part of speech categories following the
word ‘say’ in spoken discourse and in academic prose 39
3.3 Raw and normed frequency counts for ‘said’ and ‘stated’
in three registers 39
3.4 Distribution of ‘said’ and ‘stated’ in two registers 40
3.5 Distribution of the sentence position of ‘on the other
hand’ in spoken and written discourse 52
4.1 Information on corpora available at byu.edu 59
4.2 Example of a regular and irregular verb in English 65
4.3 Passive voice 67
5.1 Examples of corpus projects 81
5.2 Coding scheme 83
6.1 Frequency of complement clauses 98
6.2 Complement clauses 98
viii Tables
6.3 Descriptive statistics for ‘hedges’ in three disciplines

through SPSS 115
7.1 First person pronoun data in SPSS data view format 119
7.2 Calculating sum of squares within groups 122
7.3 Calculating sum of squares between (across) groups 123
7.4 Calculating sum of squares total 123
7.5 Descriptive statistics for first person pronoun 125
7.6 One-way ANOVA results for first person pronouns 125
7.7 Post-hoc test (Scheffe) results 126
7.8 Distribution of cases in the dataset 129
7.9 Two-way ANOVA results for first person pronouns 130
7.10 Two-way ANOVA results with significant interaction 131
7.11 One-way Chi-square table 132
7.12 Two-way Chi-square table 133
7.13 Revised Chi-square table 134
7.14 Chi-Square calculations (row and column totals) 134
7.15 Calculating expected values 134
7.16 Calculating Chi-Square value 135
7.17 Correlations calculated by SPSS 138
8.1 A corpus of problem solution essays written by Thai
speakers of English 143
8.2 Situational characteristics of registers 144
8.3 Situational analysis of problem solution essay corpus 145
8.4 The five most frequent words in a corpus of problem
solution essays written by Thai speakers of English
(both raw and normed counts to 1000 words) 149
8.5 The five most frequent four-grams in a corpus of
problem solution essays written by Thai speakers of
English (both raw and normed counts to 1000 words) 150
8.6 Rubric to evaluate student presentations 154
Figures
2.1 Situational variables 18

3.1 Distributional patterns of the word ‘say’ in COCA 36
3.2 Concordance lines for the word ‘say’ in COCA 37
3.3 Concordance lines for the word ‘way’ in COCA 43
3.4 Text analysis based on vocabulary frequency in Word and
Phrase 44
3.5 Academic words highlighted in a text in Word and Phrase 45
3.6 Frequency patterns of individual words in Word and
Phrase 46
3.7 Text analysis through Word and Phrase 46
3.8 Part of speech search in COCA 48
3.9 Distributional patterns of the four-gram ‘on the other
hand’ in COCA 50
3.10 Concordance lines for the four-gram ‘on the other hand’
in COCA 51
3.11 Frequency search in Word and Phrase 54
5.1 Example of a text file with header information 84
5.2 Embedding header tags in AntConc 85
5.3 AntConc using three-word lists for vocabulary frequency
comparison 86
5.4 Searching your own corpus Concordance lines in
AntConc for the word ‘and’ 88
5.5 Sorting in AntConc 89
5.6 File view in AntConc 89
5.7 Keyword distribution in full texts 90
5.8 Keyword ‘and’ and its collocates 90
5.9 Collocates in concordance lines 91
5.10 Running your own n-grams in AntConc 92
5.11 Running a word list in your corpus in AntConc 92
6.1 Boxplot of the use of nouns by teachers and students 100
6.2 Variable view in SPSS 111
6.3 Important tabs in Variable view in SPSS 111
x Figures
6.4 Value labels in SPSS 112

6.5 Measures in SPSS 113
6.6 Adding variables in SPSS 113
6.7 Data view in SPSS 114
6.8 Descriptive statistics through ‘Explore’ in SPSS 115
7.1 Correlation between ‘ok’ and ‘I mean’ 137
7.2 Overlap between ‘I mean’ and ‘ok’ 138
8.1 Word list in a corpus of problem solution essays written
by Thai speakers of English 146
8.2 Four-grams in a corpus of problem solution essays
written by Thai speakers of English 147
Preface
In our experiences teaching introductory corpus linguistics classes, we

have found that undergraduate and graduate students gain both confi-
dence and ability doing corpus analysis when they are given the opportu-
nity to work with corpora and are exposed to hands-on experience with
corpus tools and corpus analysis. While an understanding of principles,
approaches, and advantages of using corpora provides the necessary
foundational knowledge of this approach to language analysis, there is a
technical side to corpus linguistics that is best acquired through practice
and experience with corpora. We have found that students are sympa-
thetic to the benefits and advantages of using language corpora, but the
real challenge is teaching them how to work with corpora. When “doing”
corpus linguistics, students need to gain experience searching a corpus
and interpreting the results of their corpus searches so that they can use
this information to explain why their analysis and findings are important
or relevant. In this book, we offer multiple opportunities to work on cor-
pus projects by first including an entire chapter dedicated to smaller cor-
pus projects (Chapter 4) and then providing students with information on
how to build and analyze their own corpora (Part 3). We have found that
offering students the opportunity to build and analyze their own corpora
gives them valuable experience in corpus building and sometimes even
encourages them to build other corpora for projects outside of the class.
In order to allow the students to gain experience in “doing” corpus
linguistics, we have intentionally limited the corpora and software used in
the corpus projects. There are many different corpora available with dif-
ferent levels of public availability and numerous software programs that
can be used for analysis (some free and some not). Each corpus and soft-
ware program has its own idiosyncrasies and we have found that these
different corpora and software programs are sometimes confusing to stu-
dents who do not have access to the corpora and/or struggle to learn one
program or search interface in a corpus and then have to learn another.
To address this issue, all of the projects in this book use the suite of cor-
pora created by Mark Davies at Brigham Young University (http://corpus.
xii Preface
byu.edu/). In this way, students can use the different corpora available on
this site so they have multiple opportunities working with a single search
interface. For the larger corpus project, we have focused on one free-ware
program, AntConc, developed and supported by Laurence Anthony at
Waseda University (http://www.laurenceanthony.net/software/antconc/).
We acknowledge that the corpora and software used in this book are not
the only ones, but we feel that they provide a strong foundation for cor-
pus analysis and allow students to start “doing” corpus linguistics with
these readily available and user-friendly tools.
Because a good deal of corpus work involves quantitative data analysis,
we also included some elementary statistical information (Chapter 5) and
tests (Chapter 6). Keeping with one of the guiding principles of this book,
we see this introductory information as a way to have students learn the
basics of analysis with the hope that they may apply this knowledge in
other projects or learn more about more advanced statistical techniques
that they can use in the future.
There are many different descriptive and theoretical frameworks that
are used in corpus linguistics. We have selected one particular frame-
work to guide the students in their interpretation of their corpus findings.
Register analysis has strongly influenced our work and we believe that
this approach to understanding language use is broad enough to encom-
pass the various types of projects that students choose to do. In Chap-
ter 2, we outline the basics of a register analysis and then ask students
to refer to this same framework when building their corpus and doing
their corpus study. We recognize the relevance of other descriptive and
theoretical agendas but feel that focusing on a single approach provides
students with more extended experience interpreting their corpus results
and motivating the significance of their findings. Without knowledge and
practice using a specific framework, we have found that students can be
quite enamored with the “button-pushing” aspects of corpus linguistics
at the expense of interpreting the results of their searches.
We also recognize the importance of reporting on research in a cohe-
sive way. To this end, we have included material dedicated to the specifics
of writing a research paper and presenting research (Chapter 8). Our goal
in this chapter is to provide both students and teachers with some guide-
lines for how to demonstrate and present their specific research projects.
In the final chapter (Chapter 9), we ask students to consider more
advanced types of corpus research with the hope that this book will
serve as an introduction to the field and encourage students to pursue
these ideas at a more advanced level and perhaps even impact the field in
significant ways.
Acknowledgments
We would like to thank the following people who have influenced this
book. We first acknowledge the work of Douglas Biber at Northern Ari-
zona University and Susan Conrad at Portland State University. Their
work in corpus linguistics and register analysis has strongly influenced
this book. Also, Mark Davies at Brigham Young University and Laurence
Anthony at Waseda University have provided free online corpora and
tools that are vital to the structure of this book. Finally, we would like
to thank our students in the previous corpus linguistics classes we have
taught. It is through these experiences that we have seen the need for such
a book.
Part 1
Introduction to Doing
Corpus Linguistics and
Register Analysis
Chapter 1
Linguistics, Corpus
Linguistics, and Language
Variation
1.1 Language and Rules/Systems

1.2 What Is Corpus Linguistics?
1.3 Outline of the Book
1.1 Language and Rules/Systems

While all humans use language to communicate, the ability to describe
language is not nearly as advanced as our ability to actually use language.
One defining component of the scientific study of language (i.e., linguis-
tics) includes a description of how language works. Native speakers of
English are able to produce plural nouns that end in different sounds—we
say batS and bagZ, not batZ and bagS. These same speakers can also
produce plurals of nonsense words that we have never heard before—we
would say bligZ and not bligS. Native speakers of English also know that
We worked out the problem and We worked the problem out are both
acceptable sentences but We worked it out and We worked out it may not
be equally acceptable (the latter is likely to sound strange to many native
speakers of English). The fact that we can agree on these aspects of Eng-
lish related to the pronunciation of plurals and word order point to the
fact that language, in many respects, is predictable (i.e., systematic). Such
aspects are not only related to sounds and the order of words, but they
are also related to how we might use language in different contexts and
for different purposes. For example, we would not be likely to ask a pro-
fessor for an extension on an assignment by saying: “Hey, man. Gimme
an extension.” Instead, we are more likely to make such a request by say-
ing: “Would you please allow me to hand in that assignment tomorrow?
I have been so busy that I completely forgot about it.”
While it may be difficult to explain these particular aspects of the Eng-
lish language, native speakers apply these “rules” of language flawlessly.
In other words, one important component of linguistic description is to
make implicit “rules” or patterns of language (knowledge we use) explicit
(knowledge we can describe). It is safe to say that language users follow
4 Introduction
rules (and sometimes choose not to follow rules) for specific reasons even
though they may not be able to explain the rules themselves. An impor-
tant part of linguistic study focuses on analyzing language and explaining
what may seem on the surface to be a confusing circumstance of facts that
may not make much sense.
When many people think of language rules, they may think of the
grammar and spelling rules that they learned in school. Rules such as
“don’t end a sentence with a preposition” or “don’t start a sentence with
the word and” are rules that many people remember learning in school.
Very often people have strong opinions about these types of rules. For
example, consider the excerpt below taken from a grammar website on
whether or not to follow the grammar rule of “don’t split an infinitive.”
Even if you buy the sales pitch for language being descriptive rather
than prescriptive, splitting infinitives is at the very best inelegant and
most certainly hound-dog lazy. It is so incredibly simple to avoid
doing it with a second or two of thought that one wonders why it is
so common. There are two simple solutions.
(1) “The President decided to not attend the caucus” can be fixed
as easily as moving the infinitive: “The President decided not to
attend the caucus.” I’d argue that works fine, and not using that
simple fix is about as hound-dog lazy as a writer can get, but split
infinitives can be avoided entirely with just a bit more thought.
How about:
(2) “The President has decided he will not attend the caucus.” What
the heck is wrong with that?
It’s hound-dog lazy, I say. Where has the sense of pride gone in writ-
ers? (https://gerryellenson.wordpress.com/2012/01/02/to-not-split-
infinitives/)
Examples such as these are not uncommon. One would only have to look
at letters to the editor in newspapers or at blog posts to find many more
instances of people who have very strong opinions about the importance
of following particular grammar rules.
So far, we have looked at “rules” as doing two different things: 1)
describing implicit, naturally occurring language patterns and 2) prescrib-
ing specific, socially accepted forms of language. Although both descrip-
tive and prescriptive perspectives make reference to language rules,
prescriptive rules attempt to dictate language use while descriptive rules
provide judgment-free statements about language patterns. Both prescrip-
tive and descriptive aspects of language are useful. When writing an aca-
demic paper or formal letter, certain language conventions are expected.
Corpus Linguistics and Language Variation 5
A prescriptive rule can provide useful guidelines for effective commu-

nication. However, descriptive approaches can be useful in uncovering
patterns of language that are implicit (as in the examples described
above). Descriptive approaches can also be used to see how prescriptive
rules are followed by language users.
The concept of language rules raises another interesting question: Why
are these rules sometimes followed and sometimes “violated”? Con-
sider the prescriptive infinitive rule described above. Is it accurate to say
that those who write to not attend are not following a rule? In some
respects, this may be the case, but there is another—perhaps somewhat
misunderstood—issue related to language that deserves some attention
and serves as a basis for this book: the role of language variation. It is
an incontrovertible fact that language varies and changes. The type of
English used in the British Isles is quite different from the type of English
used in the United States. The type of English used in the British Isles or
the United States also varies from region to region or among people from
different socio-economic classes. The type of English used 150 years ago
in the United States is quite different from the type of English used in the
United States today. Language even changes and varies in a single person.
The study of language variation seeks to gain an understanding of how
language changes and varies for different reasons and in different con-
texts. There are different perspectives on how to investigate and under-
stand language variation. The perspective that we will take is, as you can
tell from the title of the book, related to an area of language study called
corpus linguistics.
1.2 What Is Corpus Linguistics?

One way to understand linguistic analysis and language is through cor-
pus linguistics, which looks at how language is actually used in certain
contexts and how it can vary from context to context. While understand-
ing variation and contextual differences is a goal shared by researchers
in other areas of linguistic research, corpus linguistics describes language
variation and use by looking at large amounts of texts that have been
produced in similar circumstances. The concept of a “circumstance”
or “context” or “situation” depends on how each researcher defines it.
Corpus linguistic studies have frequently noted the general distinction
between two different modes of language production—written language
and spoken language. From a written perspective, one may be interested
in contexts such as news writing, text messaging or academic writing.
From an oral perspective, one may be interested in language such as news
reporting, face-to-face conversation or academic lectures. Although text
messaging and academic writing are both written, the purpose of text
messaging is quite different from the purpose of academic writing and we
6 Introduction
would expect, therefore, some degree of language variation in these differ-

ent written contexts. The same may be said with face-to-face conversation
and academic lectures; both are spoken but they have different purposes
and consequently have different linguistic characteristics. More generally,
we might also expect that spoken language (in all of its various purposes
and contexts) would likely differ from written forms of language. Spoken
language does not generally have the same type of planning and opportu-
nities for revision that we find in many types of written language. We will
consider how different circumstances (or situational variables) can affect
language use in the following chapter. But, before we do, we would like to
briefly describe what we mean by a corpus.
A corpus is a representative collection of language that can be used to
make statements about language use. Corpus linguistics is concerned with
understanding how people use language in various contexts. A corpus is a
collection of a fairly large number of examples (or, in corpus terms, texts)
that share similar contextual or situational characteristics. These texts are
then analyzed collectively in order to understand how language is used in
these different contexts. The result of this analysis is a collection of lan-
guage patterns that are recurrent in the corpus and either provide an expla-
nation of language use or serve as the basis for further language analysis.
One common method used in corpus research is to look at the environment
of a particular word or phrase to see what other words are found (i.e.,
“collocate”) with the reference word. As an example, we will use the Cor-
pus of Contemporary American English (available at http://corpus.byu.
edu/coca/), a publically available collection of over 450 million words of
American English, to investigate the use of two words: equal and identical.
In many respects, equal and identical can mean the same thing (two
things that are similar to each other), and they are often taken as synonyms
of one another. For example, we can use both of these words in a sentence
such as: These two students are equal/identical in their performance on the
exam with the same general meaning. If we were asked to define the word
equal we may use the word identical in our definition (and vice versa).
However, if we use a corpus and look at how these words are actually
used, a different picture emerges. The Corpus of Contemporary American
English (COCA) shows us that, although they may sometimes be synonyms,
these two words behave very differently. We are more likely to use expres-
sions such as equal opportunity, equal rights, and equal protection rather
than identical opportunity, identical rights, or identical protection. We are
not likely to talk about equal twins or equal copies but instead use the
phrase identical twins and identical copies. A consideration of the words
that follow equal and identical suggest that equal is more likely to modify
abstract concepts such as opportunities, rights, and protection while iden-
tical is more likely to modify concrete nouns such as twins, items, and
houses. Without reference to large amounts of texts, we would likely not
be able to make such an observation. This is one example of how corpus

linguistics can provide information about language use that can help lin-
guists understand how language is actually used in authentic contexts.
Additionally, the corpus can tell us about frequency differences between
equal and identical (see Table 1.1). The top five collocates of equal occur
between 950 and 405 times in the COCA corpus and the top five col-
locates of identical occur between 417 and 20 times in the corpus. In
other words, we can see that the word equal is more frequent than the
word identical because the frequency of collocates shows a large differ-
ence between the two words. In fact, the word equal occurs 22,480 times
in the corpus, and the word identical occurs 8,080 times.
In addition to information on collocation and frequency, a corpus will
also allow us to examine the extent to which certain types of prescriptive
rules are followed. Let us look at what a corpus might tell us about split-
ting infinitives. Earlier in this chapter, we saw that this rule can raise the
ire of some people—to the point of associating some serious character
flaws in those writers who do not follow it. The Corpus of Contempo-
rary American English shows that we have examples such as to better
understand (874 times in the corpus) compared with to understand better
(94 times) and to really get (349 times) compared with to get really (151
times). Additionally, the sequence of words to better understand is most
commonly found in academic writing while the sequence to get really is
most commonly found in spoken language contexts. This evidence sug-
gests that a strong prescriptive statement such as “don’t ever split an
infinitive” runs into serious problems when looking at actual language
use. Some examples of split infinitives are actually more frequent in for-
mal academic writing; others are more frequent in spoken language. In
other words, even though there are strong views on certain usage rules
of English grammar, many of these rules may run counter to authentic
examples of how language is actually used. That is to say, the “rule”
against splitting an infinitive is not always followed (i.e., there is variation
in the application of the rule).
Table 1.1 Common right collocates of equal and identical in the Corpus of Contempo-
rary American English (COCA)
First word right Number First word Number

right
Equal OPPORTUNITY 950 Identical TWINS 417

RIGHTS 881 TWIN 247
PROTECTION 733 COPIES 29
ACCESS 485 ITEMS 27
EMPLOYMENT 405 HOUSES 20
8 Introduction
We have already seen a few examples of what corpus information can

tell us. Now we will consider the defining characteristics of corpus lin-
guistics as they will be used in this book. In a general sense, a corpus
can refer to any collection of texts that serve as the basis for analysis.
A person might, for example, collect examples of news editorials that are
on a particular topic and refer to this collection as a “corpus.” However,
would we say that this person is “doing” corpus linguistics? In their 1998
book, Corpus Linguistics: Investigating Language, Structure and Use,
Biber, Conrad, and Reppen define corpus research as having the follow-
ing characteristics:
• it is empirical, analyzing the actual patterns of use in natural

language texts
• it utilizes a large and principled collection of natural texts, known as
a “corpus,” as the basis for analysis
• it makes extensive use of computers for analysis, using both auto-
matic and interactive techniques
• it depends on both quantitative and qualitative analytical techniques
(Biber, Conrad, & Reppen, 1998: 4)
In our discussion above, we have already seen examples of the first

two characteristics. Corpus linguistics is concerned with language that
is produced for reasons other than linguistic investigation. Instead of
using judgments or intuitions about language (or making up exam-
ples of language that illustrate a specific point), corpus linguistics uses
language that represents actual language use. In this sense, the lan-
guage in a corpus is “natural” or “authentic” because it represents
how language is actually used in a given context. For the second char-
acteristic, we have already seen the benefit of referring to very large
samples of language; there are certain patterns of language that are
found only by referring to large amounts of language. However, it is
also vital to include only those texts that represent how language is
used in a given context. For example, if one were interested in looking
at characteristics of face-to-face conversation, they would not want
to include examples of spoken language found in academic lectures
or political speeches. Even though conversation and lectures share a
common feature of being spoken (as opposed to written), they do not
share other important characteristics (such as purpose of communica-
tion and amount of interaction between participants). Thus, a specific
corpus must be large but also must contain texts of the same type (i.e.,
share certain characteristics). The concept of a text in corpus linguistics
may be different from how you generally view a text. A text refers to
any sample of language that is used for some authentic purpose. From
this perspective, texts include language that has been written (such as
newspaper articles, letters, and fiction writing) but also written rep-
resentations of spoken language (such as face-to-face conversations,
sitcoms, or academic lectures).
The third and fourth characteristics of corpus linguistics make reference
to the importance of computers in the analysis of language as well as dif-
ferent analytical approaches. It would be hard to imagine how one might
use a 450-million-word corpus such as COCA without using a computer
to help identify certain language features. Despite the large number of
texts and the relative ease of obtaining numerous examples, a corpus
analysis does not only involve counting things (quantitative analysis); it
also depends on finding reasons or explanations for the quantitative find-
ings. In Chapters 3 and 4, we will cover some of the specific directions
we can explore in the corpus through software programs that allow for
different types of analysis. It is important to remember that corpus meth-
ods do not just involve using computers to find relevant examples; these
methods also focus on analyzing and characterizing the examples for a
qualitative interpretation.
In addition to these four characteristics of a corpus, Elena Tognini-
Bonelli, in her book Corpus Linguistics at Work (2001), also provides
some useful ideas in defining corpus linguistics. She describes the differ-
ences between reading a single text and using corpus linguistic tools to
investigate a collection of texts (i.e., a corpus). To illustrate this differ-
ence, we will make reference to one specific type of text: a newspaper
article. A single newspaper article is generally read from start to finish
so that the reader can understand the content of the text and relate it
to other points of view on a given topic. In English, the text is read
horizontally from left to right, and all texts from this perspective can
be viewed as a single communicative event or act (see Table 1.2). If one
were to compile a collection of 300 editorials and use a corpus approach
to analyze these texts, the ways these texts are used are quite different.
The texts in a corpus are not read from start to finish in a horizontal
manner as with a single news editorial; instead, the texts are a collection
of different (but related) events and are investigated not as whole but as
fragments, in the sense that many examples of a single feature are seen
in relation to each other. In this sense, the corpus is not read horizontally
but is instead read vertically—many examples of a particular language
feature are examined at one time.
A final point to consider when looking at corpus research relates to
various views that researchers may have about corpus linguistics. Elena
Tognini-Bonelli (2001) has made a distinction between “corpus-based”
research and “corpus-driven” research. In a “corpus-based” approach,
corpus linguistic researchers are guided by previous corpus findings or by
specific issues concerning language use. That is, researchers have a very
specific idea before searching the corpus as to what linguistic item they are
Table 1.2 Text and corpus comparison (based on Tognini-Bonelli, 2001)
(Part of ) a newspaper text A corpus of newspaper texts
The site appeared to be working again late

Thursday. A spokesman for the Motion
Picture Association of America said in an
emailed statement that the group’s site also
had been hacked, but it too appeared to be
working later in the evening. “The motion
picture and television industry has always
been a strong supporter of free speech,”
the spokesman said. “We strongly condemn
any attempts to silence any groups or
individuals.” Matthew Barakat reported from
McLean, Virginia.
looking for in a corpus. We have already seen an example of this with the
split infinitive discussion above. In this case, the perspective on the perils
of using a split infinitive was outlined in the article and this perspective
served as the basis for our corpus investigation on how split infinitives are
actually used. Given this prescriptive rule against splitting an infinitive,
we decided to see what a corpus, i.e., text samples of a naturally occur-
ring discourse, would tell us about how this rule is applied by language
users. Other examples include any linguistic feature that we know that we
want to search for, such as individual words or individual grammatical
items. In all instances, however, we already had an idea a priori (before
the fact) as to what we were going to search for in the corpus.
The other approach to finding out about language use that Tognini-
Bonelli has identified is through a “corpus-driven” method. In contrast
to the method described above, researchers following a “corpus-driven”
approach do not attempt to do corpus analysis with any predetermined
or fixed set of search criteria; instead, they use specific methods to see
what types of language patterns surface from the corpus. They extract
those patterns from the texts and document them, after which they
describe them and interpret them. Examples of research following this
approach are collocation and lexical bundle studies. Lexical bundles are
most frequently occurring four-word sequences in a register. We cannot
know ahead of time what the most frequently occurring sequences are.
Therefore, we rely on special computer programs that can extract those
patterns from the corpus for us. Researchers then analyze them grammat-
ically (Biber et al. 1999) as well as functionally (Biber et al. 2004). Not
only lexical items can be searched this way; grammatical features can be,
as well. Typically, researchers look for co-occurring grammatical patterns
in texts to characterize registers that way. We will describe this method
briefly in the last chapter as a way forward to doing corpus linguistics.
In these types of studies, however, we do not have one specific language
feature in mind a priori.
1.3 Outline of the Book

This book is divided into three main parts. In Part 1, we introduce the
concept of a corpus and locate corpus linguistics as an approach to lan-
guage study that is concerned with the analysis of authentic language,
and a focus on language variation, using large amounts of texts (Chap-
ter 1). We then provide a register analytical framework for interpreting
corpus findings (Chapter 2). In Part 2 of this book we focus on how
to use existing corpora. We introduce a set of search tools and a set of
language units that could serve as the basis for the analysis of already
existing corpora (Chapter 3). We then provide twelve different projects
that use existing online corpora in order to introduce the basics of how
12 Introduction
to work with corpora, interpret data, and present findings (Chapter 4).
Once these basics of corpus analysis and an analytical framework have
been addressed, readers will be ready to build their own corpora and
conduct their own research study. Part 3 of the book takes you through
the steps of building a corpus (Chapter 5) and then covers some statistical
procedures that can be used when interpreting data (Chapters 6 and 7).
Chapter 8 then provides a step-by-step process for writing up and pre-
senting your research. Because this introductory book contains some of
the basics of how to conduct a corpus research project, we do not cover
many of the relevant issues that corpus linguistics is presently address-
ing in its research. In Chapter 9, we discuss some of these issues with
the hope that this book has taught you enough about corpus research to
pursue a more advanced study of the field.
References
Biber, D., S. Conrad & R. Reppen (1998) Corpus Linguistics: Investigating Lan-
guage, Structure and Use. Cambridge: Cambridge University Press
Biber, D., S. Conrad & V. Cortes (2004) ‘If you look at . . . : Lexical Bundles in
university teaching and textbooks’, Applied Linguistics 25/3: 371–405
Biber, D., S. Johansson, G. Leech, S. Conrad & E. Finegan (1999) Longman
Grammar of Spoken and Written English, New York: Longman
Tognini-Bonelli, E. (2001) Corpus Linguistics at Work, Amsterdam: John
Benjamins
Chapter 2
Register Analysis
2.1 Why Register?

2.2 What Is Register Analysis?
2.3 Describing Situational Characteristics and Identifying Variables
2.4 Providing a Functional Interpretation
2.5 Units of Analysis and Register Studies
2.6 End of Chapter Exercises
As we have discussed in Chapter 1, language variation is one prevalent

characteristic of human language. There are, however, many different
ways of understanding variation. We could choose to look at how lan-
guage varies in different places of the world (for example, the differences
between British and American English). We could also investigate how
language varies by ethnicity or social class. Another way of looking at
variation would be to consider the differences among individual writ-
ers or speakers. We could, for example, study the speeches of Martin
Luther King Jr. in order to understand how his “style” might differ from
speeches given by other people such as, Winston Churchill. We could also
take a much broader viewpoint and look at language variation by exam-
ining the relationship between language use and language context. Such a
perspective is considered broad not only because it considers general con-
cepts such as “language use” and “context” but also because this view
of language variation also requires a representative sample of language
from a specific context. This perspective involves looking at fairly large
amounts of data, determining which linguistic features are frequent in a
given context, and then proposing possible reasons for the relationship
between the context and the language features. This approach to under-
standing language variation is known as register analysis. In this chapter,
we will take a closer look at the concept of a register and introduce the
basic components of register analysis. The chapter will end by providing
some examples of register analysis and provide the reader with an oppor-
tunity to see how register analysis works.
14 Introduction
2.1 Why Register?

As noted above, linguists have taken different approaches to investi-
gate language variation. When traveling to different parts of the United
States, one may notice that there are different words for sandwiches
(subs, heroes, grinders) or soft drinks (soda, pop, cola). Not only do
people have different words for things, but also the way that people
say certain words can sound very different from region to region. In
some parts of the United States, the words pin and pen both sound like
pen. In some parts of the United States, people say caught and cot with
different vowel sounds; in other parts of the United States, the words
sound the same. The role that geographic location plays in lexical and
phonological variation in these examples is generally covered in a field
of linguistics known as sociolinguistics. Researchers in this field seek to
understand how language variation is related to factors such as, geo-
graphic region, gender, ethnicity, age, and socio-economic status. The
traditional sociolinguistic approach frequently considers variation to be
present when the same concept is expressed in different ways. From this
standpoint, what counts as a variable must be a concept that is similar
in meaning but different in the words used to describe it or in the pho-
nological form of the word, or, in some cases, a different grammatical
structure that describes the same concept (for example, the comparative
structure more better).
Traditional sociolinguistic researchers will likely acknowledge that lan-
guage doesn’t just vary in the lexical, phonological, or syntactic form
of similar concepts. Linguistic differences between spoken and written
languages are not necessarily due to region or age differences but instead
are the result of differences in mode or context. Even within a specific
mode or context we find variation in specific written and spoken forms
of language. Academic writing, for example, occurs in a different context
than newspaper writing or letter writing, and register analysis in a sense
“predicts” that these contextual differences will result in the variation
of linguistic features. Similarly, face-to-face conversation occurs in dif-
ferent contexts than academic lectures do, and these contexts also result
in linguistic variation. Seen from this perspective, traditional sociolin-
guistic approaches to variation would show a different type of variation
than what is illustrated by register variation. Register variation considers
variation from the perspective of context and takes a broader view of dif-
ferences, restricting them not only to cases where similar concepts are at
play but also in seeking to describe variation that occurs in these different
contexts. In basic terms, a register is a variety of language that is charac-
terized by both a specific context and the language used in the context.
Variables in register analysis are not restricted to linguistic characteristics
that are not meaning changing; register analysis considers the context as
Register Analysis 15
a variable and looks at the different linguistic features that are found in
specific situations.
In some respects, a register perspective is similar to traditional sociolin-
guistic approaches. Both sociolinguistic variation and register variation
studies are interested in how social or situational characteristics relate to
language use; however, register analysis considers a wider range of factors
that are not only due to what are traditionally viewed as “social” factors
(e.g., age, gender, socio-economic status). For example, when looking
at potential differences between speaking and writing, the communica-
tive purpose and topic are likely not as socially conditioned as are other
components accounted for in register variation such as, the relationship
between participants. Seen from this perspective, register analysis takes
into consideration a broader range of factors that may include social
factors but may also include other factors, for example, topic, purpose
of communication, and mode of communication. Another difference
between sociolinguistics and register analysis relates to the linguistic fea-
tures under investigation. Sociolinguistic studies are generally focused
on a small number of language features that vary for purely social rea-
sons. This approach allows us to understand why some people use the
word grinder and others using the word hoagie. Register variation takes
a different view of language variation by using corpora to investigate
and identify different linguistic features. A register approach also uses a
different type of analysis to investigate how different linguistic features
co-occur in given situations of use. From this perspective, the focus can
either be on specific linguistic features or on the co-occurrence of multiple
features found in particular situations of language use. Because register
variation considers how linguistic features co-occur in a given context,
a corpus linguistic approach is well suited to register analysis because
corpora provide large amounts of authentic texts for analysis. In fact, it
would be hard to see how a register analysis could be achieved without
the use of corpora. Looking at a smaller number of texts would likely not
provide a representative sample of language use to allow for a character-
ization of a given register.
The relevance of the register perspective in corpus linguistics relates
closely to the definition of corpus linguistics discussed in Chapter 1.
Recall that corpus linguistics includes both quantitative and qualitative
analysis. While the quantitative information can sometimes be fairly easy
to obtain (after all, many times all one has to do is push a few buttons
to obtain results!), proposing reasons for the quantitative information
can be more challenging. One way to interpret the quantitative results
(but not by any means the only way) is through register analysis. Because
register variation serves as the basis of much of the corpus analysis in the
remainder of the book, we will take a closer look at the steps in a register
analysis in the next section.
16 Introduction
2.2 What Is Register Analysis?

If we see register as a variety of language, then we can describe register
analysis as a framework used to understand language variation. Register
analysis is most readily associated with the work of Douglas Biber and
colleagues. According to Biber and Conrad (2009), a register analysis has
three main components: 1) an analysis of the context in which a text is
produced; 2) an analysis of the linguistic features that are found in the
texts; 3) a functional interpretation of the relationship between the con-
text and the language produced in a given context. Register analysis illus-
trates systematic aspects of variation by showing how different contexts
(called situational variables) relate to different forms of language (called
linguistic variables).
As we touched on in Chapter 1 and will explore further in Chapter 9,
scholars are typically interested in taking one of two approaches to study
variation in different registers. On the one hand, they focus on varia-
tion in the use of one individual lexical unit (e.g., words, collocations,
n-grams, or lexical bundles) or in the use of an individual grammatical
unit (e.g., subordinate clauses, modals, pronouns). They use a corpus to
find out how these individual features vary in different registers. On the
other hand, scholars such as Biber and his colleagues are interested in
describing language variation in registers from another point of view.
Instead of focusing on individual linguistic features, they are interested in
characterizing texts from a comprehensive linguistic perspective. In order
to do this, they search multiple linguistic variables at the same time in
a corpus. These studies report on how these multiple linguistic features
work together (i.e., how they co-occur) in texts, and then examine how
their co-occurrence patterns vary in the different registers. This approach
provides a comprehensive linguistic picture of registers and the varia-
tion within them. With either approach scholars take, however, a register
analytical framework can then help understand why the features vary by
proposing functional reasons for language variation.
Following the three components of register analysis described above,
we focus on describing situational variables in this chapter. In Chapter 3,
we will show you how to search an existing corpus for the individual
linguistic variables you may be interested in (whether lexical or gram-
matical), and in Chapter 4, we will take you through a number of projects
that will help you learn how to do studies of this kind.
Before we move on to describing situational variables further, we need
to clarify what the distinguishing element is between register, genre,
and style, as these three terms are often used interchangeably. Biber &
Conrad (2009) compare the perspective of register with those of genre
and style. From a genre perspective, the focus is on linguistic features that
are conventionalized in specific text types but the features themselves are
not necessarily common. When one reads (or hears) “Once upon a time
in a land far, far away” we understand that a fairy tale is likely to follow
because this phrase is used to identify a particular genre. From a style
perspective, the focus is more on the language employed by individual
writers or speakers or by historical factors reflected of language from a
particular time period. When Shakespeare wrote “What light through
yonder window breaks” in Romeo and Juliet, his choice of words not
only came from his much-celebrated ability to use language but was also
reflective of the type of language used in late 16th-century England. In
this soliloquy, Romeo does not say something like “Hey, a light came on
in the window!” We can think of a register as a variety of language, and
register analysis as a method of understanding the relationship between
language context and language use in many ways, and regardless of any
individual “style.”
2.3 Describing Situational Characteristics

and Identifying Variables
Prior to any linguistic analyses, register studies examine multiple aspects
of the communicative context (often referred to as the “speech situa-
tion”) that the sample texts are taken from. During the past few decades,
a number of scholars (e.g., Hymes 1974, Halliday 1978, Duranti 1985)
have studied and discussed the nature of a communicative context, tak-
ing different perspectives on what constitutes the components of a speech
situation. Mode is one aspect, for example, through which language
used in these contexts could be described. Accordingly, spoken and writ-
ten modes are considered very different and register studies have illus-
trated the ways that the language used in these two modes is drastically
different.
Biber and Conrad (2009: 40–47) have identified the following basic
components through which the situational characteristics of registers
could be described:
1) Participants
2) Relations among participants
3) Channel
4) Production circumstances
5) Setting
6) Communicative purposes
7) Topic
Figure 2.1 provides an overview of the situational framework and pro-

vides some examples of each of the seven components. Although the spe-
cific examples of each component are not exhaustive, the figure and the
18 Introduction
Parcipants
•Speaker/writer
•Hearer/listener
Communicave Relaonships among
•Onlookers
purposes parcipants
• Narrate •Interacon
• Inform •Social roles
• Entertain •Personal
relaonships
•Shared knowledge
Situaonal Channel
Seng
Variables •Speech,wring,
gestures
•Shared space and
me •Permanence
•Private vs. public (recorded,
•Present vs. past wrien)
•Transience
(spoken)
Producon
circumstances
•Planned vs. Topic
unplanned
•Edited
Figure 2.1 Situational variables
following short description of each component should help you to under-

stand the range of factors that are considered in a situational analysis.
The first two factors that we will discuss are those of participants and
the relationships between the participants. Participant information from
a situational perspective includes who is doing the writing/speaking (the
addressor) and who is doing the reading/listening (addressee). There are
many possible configurations of participants: in an academic lecture, it
is generally the case that there is a single person lecturing to a larger
group of people; in a panel discussion, there may be a group of people
that are addressing another group of people; in face-to-face conversation
there may be a single addressor and addressee or, in the case of a lively
family discussion at the dinner table, there many be multiple addressors
and addressees at the same time. Additionally, there are also times where
an interaction may include people who can hear what is being said even
though they are not actively part of the conversation—for example, those
sitting on an airplane and listening to the conversation of others around
them. In addition to the participants themselves, another relevant situ-

ational factor includes the relationships between the participants. Does
one person have more knowledge about a topic than another? Does
one person have a higher or a lower level of status or power than the
other person? Are the people strangers, co-workers, or family members?
Answers to these questions will likely have an effect on the type of lan-
guage that is used.
In addition to participant information, a situational analysis also
requires a description of the environment and conditions in a given con-
text. Relevant in this aspect are the channel and the production circum-
stances. Channel refers to both the mode and medium of the language.
Mode refers to the way the language is transmitted: speaking and writ-
ing are generally the two main modes of using language, but there are
also gestural systems such as signing that can convey meaning. Medium
refers to the relative permanence of the language. We may compare many
forms of written language as being more permanent than many forms of
spoken language. Written forms of language can be preserved for mul-
tiple readers or multiple opportunities for reference while spoken lan-
guage is generally more short-lived. This is not to say that all written
forms of language are permanent or all spoken forms of language are
transient. We need to differentiate between mode and medium to dis-
tinguish the differences between a written grocery list and a recorded
speech. The grocery list may be written out, but it is not as permanent
as a recorded speech. In addition to channel, a situational analysis also
characterizes the conditions in which the language has been constructed.
We may also want to use mode and medium when referring to different
types of written language found on the internet or cellular phones (such
as social media posts, tweets, and text messages), which vary in grades
of permanence depending on the topic and potential effect they have. We
likely can think of examples when a writer has either deleted or regretted
a particular tweet, social media post, or text messages. The production
circumstances of language may also relate to the process of planning,
drafting, and revising. Some types of written or spoken language require
extensive planning, drafting, or revising while other forms do not. In the
written mode, an academic research article has gone through an extensive
planning, drafting, and revising process; in a social media post or text
message, generally this is not the case. Even in the spoken mode, we can
acknowledge the differences in planning between an academic lecture or
business presentation and face-to-face conversation. We can see the effect
of production circumstances in spoken language where it has been shown
that planned discourse contains fewer filled pauses (uh, um) and hesita-
tions than unplanned discourse (Biber et al. 1999).
Next, we will take a closer look at the variables of setting and communi-
cative purpose. Setting describes the time and place of the communicative
20 Introduction
events. A face-to-face conversation involves a shared physical space but

may take place in a private or public setting. A telephone conversation
may also be in a private or public setting but generally does not take place
in a shared physical space. Another relevant variable related to setting
includes whether the language has been produced in the present or in
the past. For example, newspaper articles written in the 21st century are
very different from those written in the nineteenth century. In addition to
setting, variation also can be the result of the communicative purpose. In
some contexts, the purpose is to inform, persuade, or tell a story while
in other contexts the purpose may be to just interact and share thoughts,
ideas, or feelings. There are also cases where we would expect that the
language would be more or less factual. We would hope that a newspa-
per article or academic textbook would contain factual information. We
generally do not expect facts in a fairy tale or work of fiction. Commu-
nicative purpose also includes the extent to which the speaker or writer
uses language that expresses their attitude about the topic (something we
might not expect in news reporting but might expect in a news editorial).
A final aspect of the situational analysis relates to topic. This is a very
broad situational variable that has not been investigated in much detail.
A conversation about where to find a suitable place to eat will likely have
very different linguistic characteristics than a conversation about how to
fix a broken refrigerator. In a similar way, an introductory psychology
textbook will likely have very different linguistic characteristics than an
introductory music theory textbook. However, the situational variable of
communicative purpose sometimes is also relevant in relation to topic.
One might argue that the communicative purpose of a conversation on
finding a place to eat and fixing a refrigerator are quite different, but
the two different textbook types share the same communicative purpose.
Thus, topic and communicative purpose sometimes “overlap” or share
relevance, but other times they do not.
Although we have covered seven different situational variables, there
are cases in which all seven variables are not involved in a situational
analysis. In a register analysis of potential differences between different
types of news writing, many situational variables associated with par-
ticipants, mode, channel, and production circumstances may not differ
although the communicative purpose may. Editorial news writing often
seeks to persuade readers to adopt or at least consider a specific view-
point; news reporting does not share this same communicative purpose
but instead seeks to inform readers.
In Chapter 3, you will have the opportunity to do situational analy-
ses in some of the projects, and in Chapter 5 you will do a situational
analysis of your own corpus. In order to understand how to apply the
situational variables to different text types, we provide three examples of
situational analyses below.
2.3.1 Email to a Friend and Email to a Boss

There are at least two situational differences between a letter to a friend
and a letter to one’s boss: certain aspects of the relationship between par-
ticipants and certain aspects of the production circumstances (see Table
2.1). Unlike the boss-employee relationship in which the boss has more
authoritative power over an employee (i.e., there is a power differential),
the relationship between friends does not have this same type of power dif-
ferential (at least we would hope it does not!). At the same time, for other
situational characteristics such as production circumstances or mode, they
will remain relatively the same because they are both in writing. Yet, it
could be that the text will be more carefully planned, revised, and edited
when an email is composed to a boss. As we compare these two registers
situationally, the “linguistic features will mark particular components of
the situation” (Biber 1988: 28), and so the differences in language use
could be attributed to the differences in the situational parameters.
Table 2.1 Key situational differences between an email to a friend and an email to a
superior (boss) (Biber and Conrad 2009: 65)
Email to a friend Email to a boss
Participants One author addressing One author addressing

one reader one reader
Relations among No interaction No interaction
participants Author and addressee Author and addressee
have the same have the same
knowledge knowledge
Author and addressee Author and addressee
know each other well know each other
Equal relationship, Unequal relationship,
personal professional
Channel Written Written
Production Text planned, may Text carefully planned,
circumstances be edited, may be edited, and revised
revised
Setting Participants not in same Participants not in same
physical space physical space
Communicative Convey information Convey information
purpose
Topic Varied Varied
2.3.2 News Writing and News Talk

In this section, we will look at two related contexts that also differ in
some situational respects: news writing and news talk show language.
While both of these texts share the situational characteristic of conveying
22 Introduction
Table 2.2 Situational differences between news writing and news talk
News writing News talk show
Participants One addressor, multiple One addressor, multiple

addressees addressees (although the
discussion is in the studio,
the television audience is an
“unnamed” addressee)
Relations among Interaction between Interaction between participants
participants addressor and addressee possible
not possible
Channel Written Spoken
Production Text has been edited Text may be edited on the spot
circumstances
Setting Addressor and addressees Addressor and addressees are
are not physically in the physically in the same room
same room
Communicative Convey information to the Convey information to the
purposes audience audience
Provide opinion; express
suitability for a position
Topic Varied Varied
(potentially new) information to an audience (under Communicative Pur-

pose), there are many differences in other areas of the situational analysis.
Even the communicative purpose could be argued to be different in these
situations in that the news talk show involves not only providing infor-
mation but also presenting an opinion.
2.3.3 Classroom Versus Symposium Presentations

Csomay (2015) explored the language of presentations in academic con-
texts. She started with the situational analysis of two contexts: classroom
presentations and symposium presentations. While the channel (spoken), the
setting (participants sharing the same physical space), and the overarching
communicative purpose (inform or expose to information) is similar in both
contexts, there are important differences in other aspects of the situations.
More specifically, the main differences between the two situations (see
Table 2.2) are related to a) participants; b) relations among participants;
and c) production circumstances. First, the social characteristics of the
participants differ in the two contexts. In the classroom, the presenter is
an expert professional, while at the student symposium, the presenter is
a novice, or an emerging professional. Secondly, the most pertinent dif-
ferences are apparent as we examine the relationship among participants.
For example, while classroom settings most likely allow questions at any
point in time during the teacher’s presentation, presentation settings have

set routines and questions can only be asked after the presentation has
been delivered. In terms of the presenter’s familiarity with the participants,
it is clear that the teacher knows (or is at least familiar with) most of the
participants in the classroom because the same set of people would meet
weekly for twelve or more weeks at a time in the same physical space. In
contrast, the presenter at the student symposium may or may not know
the audience. Yet another difference in this area is the presenter’s social/
academic status in the group they are presenting. The teacher (addressor)
in the classroom setting is superior to the audience (addressee) and has a
high social/academic status in the community. In contrast, the presenter at
the student symposium plays a subordinate role in the academic commu-
nity (especially if teachers are sitting in the audience) and has a relatively
low social status in the community. If the student is an undergraduate,
s/he has an even lower status than that of graduate students. If the student
is a master’s student, s/he will have a lower status than a doctoral student,
and so on. Finally, the teacher holds the power in the classroom while
the audience holds the power at this particular symposium because the
presenters are being judged for their performance.
Thirdly, the production circumstances are also different in the two set-
tings. Primarily, this difference is attributed to the ability of instantaneous
revisions of the text. In the classroom, there is plenty of room to negotiate
and revise the text “online” since questions can be asked any time and clari-
fications can be requested. At the symposium presentation, there is a lack of
spontaneous interaction, or “online” requests for clarification; there is no
room for immediate negotiation of the text. Clarification questions can be
made after the presentation has been delivered. Table 2.3 summarizes the
aspects we have discussed so far.
Table 2.3 Key situational differences between student presentations at a symposium and
teacher presentations in the classroom (Csomay 20015:4)
Classroom presentation Symposium presentation

(Instructor) (Student)
Participants One addressor, multiple One addressor, multiple

addressees addressees
Social characteristics: Social characteristics: novice
expert professional professional
Relations among Interaction is possible Interaction is possible but only
participants during presentation after presentation
Addressor has more Addressor has more knowledge
knowledge than audience than audience
(Continued)
24 Introduction
Table 2.3 (Continued)
Classroom presentation Symposium presentation

(Instructor) (Student)
All participants have some All participants have some

specialist knowledge specialist knowledge
Addressor gets to know Addressor does not
most or all participants know most or all participants
Social/academic status of Social/academic status of
addressor is superior to addressor is subordinate to
addressee (high status) most of the addressees in the
audience (low status)
Power is held in Power is held in addressee’s
addressor’s hand hand (judges evaluate
performance)
Channel Spoken Spoken
Production Text has been planned and Text has been planned and may
circumstances may have been revised have been revised or edited
or edited prior to prior to production
production
Text can be negotiated, and Text cannot be negotiated,
revised on the spot revised, or edited on the spot
Text can be read out Text can be read out
Setting Addressor and addresses Addressor and addresses are
are physically in the same physically in the same room
room
Communicative Convey information Convey information potentially
purposes potentially new to the new to the audience
audience
Explain concepts and Report on the procedures
methods and results of a project
Comprehension by Comprehension by addressees is
addressees is not assumed
assumed
Convey personal
attitudes
Direct students what
to do
In sum, there are enough situational differences in some of the aspects

of this context to predict that language would be used differently.
2.4 Providing a Functional Interpretation

The third step in a register analysis requires the researcher to provide
a functional reason for the linguistic features in a given text. Because
register analysis seeks to describe the relationship between situational
variables and linguistic variables, the occurrence of linguistic features
requires a description of the context. In fact, from this perspective, lan-

guage features occur because they are fitting to a specific context. In other
words, the situational variables in a sense “lead” language users to adopt
particular linguistic features. As we discussed in the previous section,
these linguistic features can be single features, but they can also be a set
of co-occurring features.
2.4.1 News Writing and News Talk

In this example, we will show a single feature analysis focusing on first
person pronouns in subject position.
In Table 2.4, the text length is fairly similar (the news writing text con-
tains 134 words and the news talk text contains 128 words); however,
they differ in the number of first person pronouns, with two in the news
writing text and six in the news talk. What situational characteristics
might account for this difference? One possibility can be found in the
communicative purpose. Because the participants in news talk shows are
often asked to present their opinion (and in this particular case, make
a case for why one political candidate is preferable over another), we
would expect that a personal opinion (stance) would be expressed by
Table 2.4 Texts for news writing and a news talk (COCA)
News writing News talk show
A playwright hikes into the woods with A: Except that Obama is likely to give
his laptop, works until the battery runs it to them. I mean, that is the case.
down and then hikes back. A theater They’re expecting Barack Obama
company sets up huge pieces of to inject the enthusiasm into the
scaffolding that double as musical Republican base.
instruments. An author sits in a cabin, B: One of the great phrases that has
working on a new play, wearing nothing been used in defense of venture
at all. capitalism and Bain Capital is
“When I’m in New York, there are Schumpeter’s creative destruction.
neighbors across the way, “ says Whenever I hear Republicans say
Cassandra Medley, the playwright who that, I want to say, you know what
takes au naturel to heart. “I just like to America has been looking for five
shed and think and feel.” years at a lot of destruction, creative
The opportunity to return to a “primeval, and non-creative. They’re not going
primordial” state is one reason that to like that defense. They’re going to
Ms. Medley—and other authors—love like a defense that says, guess what,
attending the growing network of I can create jobs, I have a plan, we
summer retreats where writers and can move this thing forward, we can
others who work in the theater get save our country. Treatises on the
away from the urban grind and try to essential nature of capitalism, I think,
reconnect with their muses. won’t do it for Mr. Romney.
26 Introduction
referring to oneself as well as by including verbs such as, think, and that
suitability for candidacy would be expressed through verbs such as, can
and have. Another relevant contextual variable is related to the mode of
production. In the spoken mode, people often refer to themselves (and
others) in the conversation. Further support for this is found in the news
writing excerpt where the two instances of a first person pronoun occur
in a direct quotation. This short analysis does not mean that first person
pronouns do not occur in written contexts, but we would not expect
them to occur at the same frequency nor for the same reasons.
2.4.2 Classroom Versus Symposium Presentation

In Table 2.5 we show you a multi-feature analysis in two presentation
types (analyzed above for the situational characteristics).
Table 2.5 Texts from classroom presentation and symposium presentation
Classroom presentation by teacher Symposium presentation by student
So what I’m suggesting to you And we found that an immature cynofields resides
then, is, is that this second in the kidney that’s where we found the most
dynamic, which accounts for the cells with those characteristics and I interpreted
popularity, the contemporary that we found also oh . . . oh . . . a relationship
popularity of civilian review, has for those cynofields but those were more
to do with money, and civil mature. We can say that because . . . The
liability, and the ways in which electro- microscopy results with that we can
the behavior of law enforcement see the morphology and chronology and this is a
institutions can, render, human cynofield with a transmission electronic
municipalities liable for millions microscopy of the human cynofield and we did
and millions and millions of with a zebrafish we found very similar morphology
dollars, uh, in, uh, civil liability that granules are round as same as the human
lawsuits. Not only that, usual ones and the nucleus is big at this stage so we
contingency, um, uh, rules, are found the cell that looks like cynofields so now
waived in these kinds of lawsuits. we want to study its function we study the
All right? What that means is function by migration of recommendation to the
that usually, when you pursue infection and then we see they change their
a civil claim, against somebody, morphology. So we know that cycles-sum in
you ask for a hundred thousand human cynofields includes information response
bucks, OK? And, you get it, and and we inject the fish with the cycles-sum we
your lawyer takes a third. All let them live for 6 hours in order to provide an
right? What happens if you sue order response and then to (syll) we sacrifice the
a municipality and they say yeah single cell suspension and within the facts analysis
we think you’re right but [short of photometry and those are our results. We
laugh] the situation was so much found we use a control also and we can see in
more complicated, we award, the control the populations of cynofields are in not
one dollar, OK? Is your lawyer increase as dramatically with the one that we
gonna take thirty three cents? or injected we cycle-sum and it was 20% more of
in these kinds of lawsuits, right? population of those cell that we found in this gate.
In these segments, we would like you to see differences in terms of the

use of nouns and grammatical features specific to conversation such as,
non-clausal units, tags, and so on (Biber et al. 1999). We have italicized
the nouns in each segment, and bolded and italicized the grammatical fea-
tures specific to conversation. Just by looking at the patterns of the itali-
cized features, we can see a difference. If we do the calculations, it turns
out that the classroom text segment has 17.19 nouns and the symposium
presentation text segment has 19.50, and the classroom text has 6.37
“conversational features” while the symposium presentation has 2.07.1
Another type of analysis could tell us about the lexical density of the
two texts. This is typically measured by the text’s type-token ratio.2 The
higher the ratio, the fewer repetitions occur, generally indicating a more
lexically dense text. In the classroom text, we have 157 words (tokens)
and 102 types of words, so the type-token ratio for the classroom text
is 157/102 = 1.54. For the symposium text, we have 241 words (tokens)
and 116 types, so the type-token ratio is 241/116 = 2.07. This number
simply means that in one text there seems to be more repetition than in
the other text.
The question is: Why are these two texts so different in their noun and
conversational features and in their lexical density? In other words, what
may account for the differences? If we look at the situational differences
between the two texts, there are many. However, perhaps most pertinent
to these two segments is the production circumstances and the commu-
nicative purpose. In the classroom, there is no pre-set script to follow;
that is, there is always room for asking questions, and the potential for
interaction is always present. In contrast, at a symposium presentation,
the rules are strict, and questions may be asked only at the end of the talk.
Therefore, the presenter is expected to talk continuously for a period of
time, after which the questions from the audience may be asked.
In terms of communicative purpose, there are two major areas that may
account for the differences. On the one hand, in the classroom, the pur-
pose is to explain concepts and methods; at the symposium, the purpose
is to report on the processes and results of a research project. In addition,
in the classroom, the addressor (the teacher) periodically checks for com-
prehension to see whether the students are understanding the material.
In contrast, at the symposium presentation, the addressor (the presenter)
assumes comprehension and expects questions to be asked afterwards.
For these reasons, there seems to be quite a difference in the frequency of
the conversational features between the two texts. However, the same is
not true for the use of nouns. Because the communicative purpose in both
contexts is to convey information that is potentially new to the audience,
the actual information seems to be packaged in similar ways; that is, the
information is delivered through nouns either embedded in noun-noun
sequences or in a series of prepositional phrases. In terms of how the
information is conveyed, we see differences in the type-token ratio. The
28 Introduction
teacher uses more repetitions (hence the lower ratio number) while the
presenter is conveying the information without that many repetitions.
2.5 Units of Analysis and Register Studies

As we mentioned at the beginning of the chapter, many corpus research-
ers choose to investigate a single linguistic feature to see its variation
in multiple registers. Among the lexico-grammatical studies is Fortanet’s
(2004), which, taking the Michigan Corpus of Academic Spoken English
(MICASE), looks at the use of pronouns in academic lectures. More spe-
cifically, Fortanet classifies ‘we’ into functional categories, which she then
tracks to see how more or less interactive classes use this particular pro-
noun and their associated functions differently. Another example is found
in Wulff et al. (2012), which examines the Michigan Corpus of Upper-level
Student Papers (MICUSP) to describe ‘attended’ versus ‘un-attended’ this
in student writing. Wulff et al. look at sentence initial uses of this and
classify them first based on their part of speech category (whether they are
pronouns taking the full noun phrase slot or they take the determiner slot
in the noun phrase). Then, they investigate the frequency patterns of these
two types of this, associate them with communicative functions, and look
at their functional distributions. Because these studies considered the dis-
tribution of a single feature in different contexts, would this be included
in register analysis? Although perhaps not everyone would agree, we tend
to see these as examples of register analyses because they 1) make refer-
ence to specific situational contexts; and 2) provide functional reasons
for the different uses of specific words. In fact, part of the value of these
studies is found not necessarily in frequency counts of a particular word
but in the frequency of different types of functions in single words. For
example, in Wulff et al. (2012), the lexical word this was associated with
two different functions, so it was the frequency of function that varied
across texts rather than the frequency of the word this.
Another example of single feature analysis in corpus studies is found
in studies focusing on two- or multi-word sequences (collocations, and
n-grams or lexical bundles, respectively). Lexical bundles are the most
frequently occurring word combinations in a register; that is, in situa-
tional language use. The most often investigated bundles are four-word
combinations. Biber, Conrad & Cortes (2004) investigate lexical bundles
in university language use based on the T2KSWAL corpus (Biber et al.
2002) and classify them based on their grammatical characteristics and
their discourse functions. Biber et al. (2004) specifically look at the dis-
tributional patterns of the identified bundles in two sub-registers with
the university setting—namely, teaching and textbooks. Csomay (2013)
takes this set of bundles and places them back into the discourse flow of
university lectures to see how frequently the different functions occur
within the structure of discourse.
These approaches to studying language use in registers provide detailed

analyses of these individual features and their individual patterns. There-
fore, we can learn a good deal about the use of that one feature. While
these studies are interesting and very informative for these features sepa-
rately, as Csomay indicates, “comprehensive descriptions of variation in
language use cannot be based on investigating a single linguistic form or
a single linguistic feature in isolation” (2015: 5). When describing the lin-
guistic characteristics of texts, relying on one feature at a time is difficult
mostly because a) an a priori selection of that one feature is hard to predict,
since we would not really know which feature will mark the difference
in the situations we are comparing, and because, as mentioned above, b)
language features are typically in relationship with each other and do not
occur in a vacuum. In order to characterize a register, we would need to
provide a comprehensive analysis. For a comprehensive analysis, we need
to look at all linguistic features in texts. In addition, we need to examine
their distributional patterns to gain a ‘full’ picture as to what linguistic fea-
tures registers may be made up of. While this is possible, we would still lack
the understanding of how various language features relate to one another.
However, as mentioned above, in this book, we will mainly focus
on individual linguistic features and their variation in registers for two
main reasons. On the one hand, these kinds of studies are relatively easy
to carry out without the necessary background knowledge for a more
sophisticated type of study. On the other hand, and tied to the previous
point, our book is for a relatively novice audience. It is written for some-
one doing corpus linguistics perhaps even for the first time and for some-
one who has limited or no basic statistical knowledge. In contrast, to
carry out comprehensive linguistic analyses from the start, the researcher
must have a solid background in computer programming and in multi-
variate statistical methods. While such methodology is not the focus of
this book, based on the results of previous comprehensive studies, we will
point to ways a cluster of already identified features could be analyzed
and discuss how they could be useful for our own analyses.
2.6 End of Chapter Exercises

1. In this chapter, we have discussed some benefits of investigating lan-
guage from a register perspective. In your opinion, what is the advan-
tage of register variation in corpus linguistics as opposed to other
possibilities, such as genre analysis or analysis of style? What types
of linguistic methods are suited to genre or style analysis?
2. Go to COCA and pick two subcategories. Line up the situational
characteristics for each, the way it is done in this chapter, to see what
the differences are.
3. Relationship between sociolinguistics and register variation: Given
what you know about traditional sociolinguistic variation, which of
30 Introduction
Table 2.6 Texts from email to a friend and email to a boss
Email to a friend Email to a boss
hey! Hello __,

guess what? As I may have mentioned to you before,
today leah announced she had a I’m currently taking a Ph. D seminar in
present for me and handed me speech perception, and Dr. ___ ____
a Marriott card that said “tad has been guiding all of us towards
cross” and i said “WTF?” and she conducting some sort of research
informed me that our colleague project this semester. The project I have
flora had met him at a conference, chosen relates to teaching pronunciation
one in which he played a leading skills in a reading classroom, and
role, and he’d mentioned my name. whether or not this (constituting a
this evening in the parking lot she brief and non-interfering classroom
told me, “he’s VERY handsome.” treatment) will improve students reading
flora’s 65 and likely thinks fluency and receptive vocabulary (as
everyone’s handsome, but i had to reflected in the week 1 assessment).
admit i’d had a huge crush on him If you’re willing, I’d like to meet with you at
when i was a wee one. haven’t seen your earliest convenience, so that I can
him since, but i’d love a reminder ask you a couple of questions regarding
on what he’s up to. i seem to this. I would like to work with level 3
remember something funny (uber students if possible, which I believe you
christian?), but also that lynn loves coordinate (for reading and writing
him. and i plan to attend that class).
conference next year, so eventually I have cc’d Dr. ___ on this email, as
i might find out for meself. per what I am told is the policy for
discussing potential research in classes.
Please let me know when and if you have
some free time to talk.
the seven situational factors described in Figure 2.1 would you attri-
bute to traditional sociolinguistic factors?
4. Look back on the situational analysis presented in 2.3.1 and think
about what linguistic features are worthy of closer investigation.
Before you focus on one or two features for comparison, look at the
text in Table 2.6 and identify features found in one text but not in the
other.
Notes
1 Since these text segments are uneven in length (one is 157 words long and the
other is 241 words long), we had to scale the raw frequency counts as if both
texts were 100 words long. To do this, we need to norm the feature count
with a simple statistic: (Raw feature count/actual text length)*desired text
length. We will discuss this technique more in subsequent chapters.
2 To calculate type-token ratios, we take the number of words in a text and
divide it by the number of word types. If a word is repeated, it counts as a new
token but not as a new type. For example, in the following two sentences,
there are 10 tokens (i.e., number of words) and 8 types (because ‘the’ and ‘cat’
are repeated): He saw the cat. The cat was in the garage.
References
Biber, D. (1988) Variation across Speech and Writing, Cambridge: Cambridge
University Press
Biber, D. & S. Conrad (2009) Register, Genre and Style, Cambridge: Cambridge
University Press
Biber, D., S. Conrad, R. Reppen, P. Byrd & M. Helt (2002) ‘Speaking and writ-
ing in the university: A multidimensional comparison’, TESOL Quarterly 36:
9–48
Cortes, V. & E. Csomay (2015) Corpus-based Research in Applied Linguistics.
Studies in Honor of Doug Biber. Amsterdam: Benjamins
Csomay, E. (ed) (2012) Contemporary Perspectives on Discourse and corpora.
Special issue of Corpus Linguistics and Linguistic Theory. 8/1.
Csomay, E. (2013) ‘Lexical bundles in discourse structure: A corpus-based study
of classroom discourse’ Applied Linguistics 34: 369–388
Csomay, E. (2015) ‘A corpus-based analysis of linguistic variation in teacher and
student presentations in university settings’, in Cortes, V. and E. Csomay (eds)
2015: 1–23
Duranti, A. (1985) ‘Sociocultural dimensions of discourse’, in Teun van Dijk (ed)
1985: 193–230
Fortanet, I. (2004). ‘The use of ‘we’ in university lectures: Reference and func-
tion’, English for Specific Purposes 23: 45–66
Halliday, Michael A.K. (1978) Language as Social Semiotic: The Social Interpre-
tation of Language and Meaning. London: Edward Arnold
Hymes, D. (1974) Foundations of Sociolinguistics: An Ethnographic Approach,
Philadelphia: University of Pennsylvania Press
Wulff, S., U. Römer & J. Swales (2012) ‘Attended/unattended this in academic
student writing: Quantitative and qualitative perspectives’, in Csomay, E. (ed)
2012: 129–157
Part 2
Searches in Available
Corpora
Chapter 3
Searching a Corpus
3.1 KWIC
3.2 Collocates
3.3 N-Grams
3.4 POS Tags
When researchers use corpora for their analyses, they are interested in
exploring the use of lexical items, or certain grammatical constructions.
They may also investigate lexical or grammatical patterns to see how
variation in those patterns may relate to different contexts. In register
studies, as we saw in the previous chapters, by contextual differences we
basically mean differences in particular aspects of the situational charac-
teristics in the construction and production of a given text.
In this chapter, we will use the Corpus of Contemporary American English
(COCA—corpus.byu.edu/coca/) to illustrate the most commonly identified
units of language that researchers use for their analyses: words, collocations,
n-grams/lexical bundles for lexical patterns, and (POS) tags for grammatical
patterns. We will illustrate how to identify these units of language by pro-
viding different tasks that will give you practice in searching and analyzing
these units of language. This chapter then is divided into four sections: 1)
KWIC (keyword in context); 2) Collocations; 3) N-grams; 4) POS tags.
Before we can start, you will need to register with COCA. It is free of
charge, but as a student, you will have more search options and you will
have more opportunities for searching than the very limited number of
daily searches available to those who don’t register. Go to COCA and, at
the upper right hand corner, click on ‘Register’ While registering does not
require you to pay any money, you may still wish to donate to the site to
support the excellent work they do.
3.1 Keyword in Context (KWIC)

Let’s say that we are interested in searching COCA to see how often the
word ‘say’ is used across three registers: spoken discourse, newspaper,
and academic prose.
36 Available Corpora
The keyword that we are searching for is ‘say’. As we see on the chart
in Figure 3.1 (on the left hand side in the figure under ‘ALL’), our key-
word occurs a total of 426,546 times in the entire corpus; that is, on
average 918.611 times in a million words, including all five registers in
the corpus. We can also see the distributional patterns across the regis-
ters. Our interest is in three registers: spoken, newspaper, and academic
prose. It is apparent, perhaps not surprisingly, that our keyword occurs,
on average, most often in spoken discourse (1,924.90 times in a mil-
lion words) under ‘SPOKEN,’ followed by newspapers (872.18) under
‘NEWSPAPER,’ and finally academic prose (239.06) as shown under
‘ACADEMIC.’ Since we have not yet specified what part of speech cat-
egory we are looking at, and ‘say’ can be both a verb (He said he was
going to be late) and a noun (He has the final say in the matter), these
numbers include all of these options.
If we were interested in how our keyword is used in the immediate
textual context, we would look at the bottom part of what is displayed
in Figure 3.1 (see Figure 3.2). Now, as you see, our keyword is not just
an individual word but is placed back into its textual environment. The
program randomly selects examples from the corpus and lists each occur-
rence of our keyword together with a window of text around it. This kind
of listing is called ‘concordance lines’.
When we use concordance lines, we look for individual (or a group
of) pre-selected keywords to see them in context. The most common
form to display a keyword in context (KWIC) is through concordance
Figure 3.1 Distributional patterns of the word ‘say’ in COCA

Searching a Corpus 37
Figure 3.2 Concordance lines for the word ‘say’ in COCA
lines. As mentioned above, concordance lines highlight the word you

pick and provide additional text around it. You can set the size of
the text by selecting the number of characters (see our discussion on
AntConc in Chapter 5) in the window around the keyword, and the
output lists the examples selected from the text. Each time you run the
search, even with the same keyword (by choosing the button ‘KWIC’
in the top left corner at COCA), it will select another set of randomly
selected examples.
Why would the textual environment be interesting? Since you can see
the word in context now, you will be able to see patterns surrounding the
word. Let’s see what kinds of words follow our keyword. More specifi-
cally, let’s see what part of speech categories seem to be following our
keyword.
When you access COCA’s website, the different colors after the key-
word denote different part of speech categories (here everything is black
and white only). You can get the full list via the website, but for now, let’s
pick a few: prepositions, as in ‘say in the paper,’ are marked as yellow;
nouns are turquoise, as in ‘I heard him say abstinence’; and adverbs and
adverbials are in orange, as in ‘They all say basically the same.’
Project 3.1: ‘Say’ Followed by What Part of Speech?

Given the 100 randomly selected examples that you see, create a table
(see Table 3.1) and plug in how many nouns, pronouns, adjectives,
Table 3.1 Distribution of part of speech categories following the word ‘say’
‘SAY’ NOUN PRONOUN ADJECTIVE PUNCTUATION ‘THAT’ OTHER Total
100
‘that’ complementizers, or punctuation marks you see (each will be a

different color) following the verb ‘say.’ (Reminder: The keyword ‘say’
has to be a verb, hence, marked pink.) That is, your research ques-
tion is “What is the distribution of the different parts of speech cate-
gories following the keyword ‘say’ when it is a verb?” Can you state
some general patterns that you see in the dataset based on these examples?
Project 3.2: ‘Say,’ POS, and Register

For this project, your new research questions are these: “What is the dis-
tribution of the different parts of speech categories following the keyword
‘say’ (when it is a verb) in spoken discourse versus academic prose?” “Is
there a difference in use?”
In this case, since you are more interested in how your keyword is used
in different registers, you would want 100 sample sentences randomly
selected from spoken discourse, and 100 sentences randomly selected
from academic prose. To get those sentences for your sample, take your
cursor to the left hand side of the window so the box appears again on the
left. The third area after ‘display’ and ‘search string’ is ‘sections.’ Click
on ‘spoken’ under the box on the left and leave the box on the right on
‘ignore.’ Hit ‘search.’ Now all your examples are randomly selected but
only from spoken discourse. Type in ‘say_spoken’ under “Create new
list” above the sample sentences and save it. Then click on the academic
prose bar to get your second 100 samples just from that register; save it
as ‘say_academic.’ Create a table like Table 3.2. Based on your results,
what can you tell us about how these two registers use ‘say’ in terms of
their syntactic environment?
Project 3.3: ‘Say’ and ‘State’

You also know that ‘say’ and ‘state’ are synonyms, so they could be used
interchangeably—at least in terms of syntax and semantics when they
are verbs. For simplicity, let’s just pick the past tense/participle forms of
these two verbs: ‘stated’ and ‘said.’ You may predict that different reg-
isters probably use these two verbs differently because one is considered
more ‘formal’ than the other. Your research question is this: “What is the
frequency and distributional pattern of ‘say’ and ‘state’ (as verbs in past
Table 3.2 Distribution of part of speech categories following the word ‘say’ in spoken
discourse and in academic prose
‘SAY’ NOUN PRONOUN ADJECTIVE PUNCTUATION ‘THAT’ OTHER Total
Spoken 100
Acad. 100
Prose
Table 3.3 Raw and normed frequency counts for ‘said’ and ‘stated’ in three registers
SAY/SAID STATE/STATED
Raw frequency Normed Raw frequency Normed

frequency to frequency to
1 million words 1 million words
Spoken
Newspaper
Academic Prose
tense and participial forms) in three registers: spoken discourse, newspa-

per, and academic prose?” Create a table like Table 3.3 to present your
results. Can you see any patterns you can report on?
Project 3.4: ‘Said’ and ‘Stated’

Now, have a closer look at the environment to see how these forms are
actually used. Take a sample of 100 randomly selected sentences from
spoken and another 100 from academic prose for each of the two words
(you will have a total of 400 sentences, 200 for ‘said’ and 200 for ‘stated’)
and categorize the functions of the past tense/participle forms, trying to
answer the following research question: “What is the distributional pat-
tern for the following functions:
• To state an action in the past, as in ‘She said that he had gone to the
party’
• To state a completed action with a result relevant in the present (pres-
ent perfect), as in ‘He has stated his rights already’
• To state action completed in the past prior to a past action (past per-
fect), as in ‘I had said that before’
• To modify a noun (adjectival function), as in ‘the stated example’?”
Fill in Table 3.4 below with your numbers. Are there any differences of
the functions across registers?
Table 3.4 Distribution of ‘said’ and ‘stated’ in two registers
Past action Perfect aspect Past perfect Adjectival Other Total
SAID spoken 100

SAID academic 100
STATED spoken 100
STATED academic 100
3.2 Collocates
As mentioned in Chapter 1, collocates are two words that occur together
more often than we would expect by chance. More specifically, the dic-
tionary defines collocate as “be habitually juxtaposed with another
with a frequency greater than chance” (Webster’s). Firth (1951) coined
the term ‘collocation’ to refer to two words that go together. Some
words, even though they may mean roughly the same thing, may not go
together. In the following example, in each instance, we want to express
the fact that something went bad. However, while we often say rancid
butter, we rarely, if at all, say *sour butter, and, in fact, that sounds
odd to some ears. The latter two words do not seem to go together in
English. The same is true for sour milk, which is a collocate, while *ran-
cid milk is not. Another example could be strong tea vs. *powerful tea
and powerful computers vs. *strong computers. There are many further
examples of two words going together. These could be partially or fully
fixed expressions and they are used in particular contexts. Although
these examples show adjective+noun combinations only, it is not the
case that only these two part of speech categories go together. Other
collocate types are, for example, noun+noun (e.g., bus stop) combina-
tions, verb+noun (e.g., spend money), verb+prepositional phrase (e.g.,
account for), and so on.
Project 3.5: Collocates of ‘Spend’

Go to the main page of COCA (corpus.byu.edu/coca/) and click on
‘collocates’ among the list of options. Read the main introductory page
that explains what collocates are. As you see, you can download the
first 2–300 collocates in this corpus. What is even more interesting is
that whenever you search a word, the search engine gives you the col-
locates as well. Go to www.wordandphrase.info and type in ‘spend’;
specify that you are only interested in this word as a verb by clicking
off all other boxes. Look at what collocates this word has. You will see
that the collocates are, for example, spend time, money, day, etc. If you
click on the word ‘money’ from this collocate listing, you will be able
to bring up all the examples where these two words occur together in
texts. Click on ‘time’, and examine the first 100 samples provided to
you. In which register does this collocate spend (v) + time come occur
most frequently?
3.3 N-Grams
Most often, n-grams in linguistics are sequences of words explored as a
unit, where the value of n denotes how many words there are in that unit.
If the basic unit of analysis is a word, then we can call a word a uni-gram
(1-gram). If we have two words to consider as a unit, they are bi-grams
(2-grams), and if we have three words as a sequence in a unit, it will be a
tri-gram (3-gram), and so on.
A special computer program is often designed to process the text and
to look at sequences of words in a window of text as they emerge in a
corpus. Depending on how “big” your unit is (i.e., how many words in a
sequence you want to trace at a given point in time), the window size is
set accordingly. That is, if you want to identify bi-grams, you will capture
each two-word sequence in the corpus. If you are looking for tri-grams,
you will capture each three-word sequence, and so on. As you are doing
so, the already identified sequences are tracked and the new sequences are
constantly checked against what the program had already found. Each
time the same word-sequence is found, the program counts the frequen-
cies of that sequence. We can explore how frequently particular word
combinations (n-grams) occur together in a corpus or how they are dis-
tributed across different registers.
If you know ahead of time what sequences you are looking for, you
can just type the sequence in the search engine. In this case, you are not
looking for the kinds of n-grams there may be emerging in that corpus;
instead, you are just looking for a pre-defined sequence of words (that
may, or may not, have been extracted from a previous corpus. For exam-
ple, if you type in the word baby, you will see that it occurs over sixty
thousand times in the COCA corpus. But you picked the word ahead of
time, so you knew what to look for. If you are interested in the sequence
a baby, it occurs more than ten thousand times in the same corpus, and if
you are interested in the sequence like a baby, you will see that it occurs
more than six hundred times in COCA. At the same time, the four-word
sequence sleep like a baby only appears twenty-five times in the same cor-
pus. In all of these instances, however, you have typed in the words that
you were interested in.
In contrast, if you don’t know ahead of time what you want to look
for, but you are interested in the possibilities a corpus may have, you can
either design a new computer program for yourself and run your own
data, or run the n-gram program already available to you (e.g., through
AntConc, which is a freely available software that we discuss more in

Chapter 5). The COCA site actually provides the lists for you, including
bi-, tri-, four-, and five-grams, and their frequencies in COCA. You need
to agree that you are not using the data for commercial purposes, but you
can do research on or with them.
3.3.1 One-Grams or Uni-Grams and Individual Words

When you search for uni-grams, you are basically interested in individual
words. When you know ahead of time what word you are interested in,
they are often referred to as ‘keywords.’ As we have discussed earlier in
this chapter, in the KWIC section, you could then see your keyword in a
textual context. The purpose of your search for keywords could be that
you analyze a) what kinds of words surround your keyword, b) how they
are positioned in the sentence or in the phrase, or c) what the frequen-
cies are of your keywords. We have illustrated how this may work in the
previous chapter.
Project 3.6: ‘WAY’

Go to corpus.byu.edu/coca/ and search for the word (or 1-gram) way.
Choose the KWIC button (KeyWord in Context) and then also click on
the chart button. Report on the overall frequency of the keyword, its
normed frequency, and its distributional patterns across all five registers
noted in COCA—namely, spoken discourse, fiction, magazine, newspa-
per, and academic prose.
Project 3.7: ‘Way’ and Its Environment

This time, you are interested in looking at what kinds of words come
before the keyword (previously, we looked at the patterns after the key-
word). In order to see patterns better in the samples you are given through
the concordance lines, you can sort the words alphabetically either before
(or after) the keyword. Click on the lines in the three windows marked
with a circle in Figure 3.3 until they show a number. It will sort it by the
first word preceding the keyword, then the second word to the left of
that, and then the third word. This way, you can describe the kinds of
words that precede way. As you see in Figure 3.3, the lines are now sorted
alphabetically by the first word preceding the keyword, and then by the
second word, etc.
Did you find any patterns you can report on concerning way?
Now let’s have a look at the vocabulary characteristics of a text.
This way we can investigate patterns in larger units such as, a
text. Each word in the COCA corpus is classified into frequency bands.
Figure 3.3 Concordance lines for the word ‘way’ in COCA
That is, each word is ranked depending on how often it occurs in

COCA. For example, the third person pronoun he is ranked as the
15th most frequently occurring word in the corpus with a frequency of
3,138,899. The word way, when a noun, and as it is displayed above,
is ranked 84th in COCA with a frequency of 511,585. The most fre-
quently occurring words in general are function words—for example,
articles (a, an, and the) or prepositions or particles (in, on, etc.). In
fact, the definite article the is ranked #1 in the corpus with a frequency
of 23,782,115.
Project 3.8: Frequency Rank of ‘a’ Versus ‘Juxtaposition’

Go to www.wordandphrase.info and click on ‘Frequency lists’ and type
a in the word box on the top. What ranking does the article a (that will
include an, as well) have? What is the frequency? Is it a very frequent
word too? Now compare the rank and frequency of the word juxtaposi-
tion (as a noun) to the indefinite article that you have just searched. What
is the rank and what is the frequency for this word? Is it a frequent word
in the corpus?
Now, as we mentioned above, we explore a text as to how many very
frequent or less frequent words are used in a text. Go back to www.
wordandphrase.info. Click on ‘Input/Analyze texts’ and, from the ‘Sam-
ples’ drop-down menu on the top left-hand side, pick FICT (fiction).
A random sample text will appear in the text box that will then be ana-
lyzed for its vocabulary make-up. Choose ‘word’ versus phrase and click
on the ‘search’ button. On the right hand side, you will find the same text
but now marked up with different colors. The color each word gets will
depend on where the word ranks as a frequently occurring word in the
database. (See Figure 3.4.)
As you see, there are three frequency bands. Words marked with the
color blue are among the top 500 most frequently occurring words in the
corpus. The green ones rank as the top 501–3,000 most frequently occur-
ring words in the corpus, and the yellow ones are marked for those that
are in the rank of less commonly used vocabulary in the corpus (beyond
the rank of 3,000). Be careful not to interpret these as frequencies because
these are rank numbers. The most interesting words, however, are the red
ones that tell you which words are academic words. These words occur
with a particularly high distributional pattern across academic texts.
More precisely, the following explanation is given on the same website:
. . . these are words that occur at least twice as frequently in COCA as

would otherwise be expected. Since COCA is divided almost evenly
between the five genres of Spoken, Fiction, Magazine, Newspaper,
and Academic, if a word occurs 50 times in the corpus, then the
Academic genre should have about 10 tokens. If Academic had 20
tokens or more (twice the expected amount), then that word would
be marked [+ACADEMIC].
(www.wordandphrase.info/analyzeText.asp)
Figure 3.4 Text analysis based on vocabulary frequency in Word and Phrase
While Figure 3.4 is black and white, if you go to the WordandPhrase

website, you will see that the words are colored. In this text, there are a
lot of blue words indicating that. The text shown in Figure 3.4 contains
a lot of frequently occurring vocabulary and only a small number of aca-
demic words as noted by a lot of blue words if you run the texts directly
on the website. In fact, as you may also notice, for each frequency band,
the percentage of the vocabulary items for that band in the text is indi-
cated. In the example above, the text randomly chosen by the program
(an extract from the novel published in 2008 entitled ’Tis Is the Season)
has the following vocabulary characteristics: a total of 429 words, of
which 62% are words that fall in to the first frequency band. That is,
62% of the words in this text are among the most frequent words in the
corpus that ranked among the first 500 most frequently occurring words.
In addition, 12% of the words fall in the second band, and 26% fall in
the third band. We also see that 2% of the words colored in this text
segment (red, if you go to the site) are identified as academic vocabulary.
This means that the words are occurring more frequently in academic
texts than expected. If we click on the red academic link, the text will be
tagged for those words (see Figure 3.5).
If you click on ‘See lists,’ the program lists all the words individu-
ally that would fall into the different bands (Range 1 and Range 2 in
Figure 3.6) and shows how often they occur in this text segment. For
example, the word ‘the’ is among the first 500 ranked words in terms of
frequency (Range 1), and in this text segment it occurs 17 times. Another
example is ‘and,’ which is also in this frequency band (Range 1) and
occurs 10 times in this text segment.
Figure 3.5 Academic words highlighted in a text in Word and Phrase

Figure 3.6 Frequency patterns of individual words in Word and Phrase
Figure 3.7 Text analysis through Word and Phrase
If you click on any of the words in the list, it gives you information about
that one word in terms of its distributional patterns across the different reg-
isters, provides a definition of the word and its collocates, and, finally, pulls
up examples from the COCA corpus in concordance lines. (See Figure 3.7.)
Project 3.9: Vocabulary and Academic Prose

In our example above, we picked a sample from fiction as a register and
looked at different kinds of words in the text. Following the same steps
as above, choose a sample from another register—for example, academic

prose. Follow the same steps to look at how many and what kinds of
words you see in this text. Do you see any difference between the two
registers in terms of the percentage of frequent (or not frequent) words
in the text?
If you want to check any text of your choice, all you have to do is
copy and paste the text in the textbox and the words will be marked up
for your text in the same way as in the examples. The frequencies, once
again, were identified in COCA, a 450-million-word corpus. It is a great
baseline corpus for your own research as well. This is a very powerful and
useful tool to determine the vocabulary characteristics of a text.
3.3.2 Bi-Grams and Collocates

Bi-grams are two words in sequence. The difference between bi-grams
and collocations is the fact that bi-grams are identified based on two
words that happen to be next to each other in a corpus while colloca-
tions are two words co-occurring more frequently than by chance, as
we have seen the definition of collocates above. Collocates are always
two-word combinations, are statistically determined, and are also called
2-grams (or bi-grams). All collocates are bi-grams but not all bi-grams
are collocates.
Project 3.10: ‘Way’ and Its Collocates

Look for the prepositions that are found with the word way. In COCA,
type in way, click on ‘collocates’ and type in the following: [[way]].
[nn*] * [i*]. This string indicates that we are looking for the word way
as a noun, and we are looking for prepositions that follow it whether
immediately after the word or within a window of text to the right up to
four words. The results will look like Figure 3.8.
As part of your qualitative analysis, you may want to see some of the
meaning differences when you use different types of prepositions, or how
a second particle or preposition is attached.
3.3.3 Tri-Grams
Any three words co-occurring together in the same sequence are known as
tri-grams. Some are complete structural or semantic units such as by the
way or in a way, and can be treated as fixed expressions with a specific
meaning. Others may be a semi-structural unit such as by way of, which is
missing one element coming after it in the phrase, or have a gap in meaning
such as what would you. Most recent studies have explored the possibilities
Figure 3.8 Part of speech search in COCA
of what could potentially come before or after a tri-gram in terms of their

structural correlates (Cortes 2014).
Project 3.11: ‘By the Way’ in Registers

Let’s look at COCA again. It is a common belief that by the way is used
only in spoken discourse, and never in academic prose. First, let’s search
for by the way in two registers: spoken and academic prose. Make a fre-
quency table as we did before and report on their distributional patterns.
Do you see anything interesting to report or discuss?
Project 3.12: ‘By the Way’ and ‘In a Way’

It is believed that both by the way and in a way are used as sentence
adverbials. Make a general search for each and see whether this could be
supported by the 100 examples that you have found.
Project 3.13: ‘By Way of ’ What?

Type by way of into the top and the command [nn*]. This string by way
of [nn*] will allow you to search for nouns that come after the tri-gram
by way of. This command will give you the strings and will list the nouns
following the string.
Step 1: Click on ‘chart’ and ‘search’. This will give you the distribu-
tional patterns of this construction (with any nouns) following
the tri-gram. Then click on KWIC and select ‘Spoken’ from the
choices. This will provide you with examples from spoken dis-
course. Once you have 100 randomly selected samples, classify
the nouns into semantic categories (you may want to go to Biber
et al.’s [1999] Longman Grammar of English or some other
descriptive grammar of English for examples of semantic catego-
ries of nouns).
Step 2: Follow the same procedures as in Step 1 but now select ‘Aca-
demic’ discourse to get examples from that register. Classify the
nouns into semantic categories. Is there a difference between spo-
ken and academic discourse in the kinds of nouns that are used
after by way of?
When tri-grams occur with a particular frequency (e.g., 20 times) and

in one specific register in a corpus, they are often called lexical bundles.
However, four-grams are most generally investigated as lexical bundles
and are discussed in the following section.
3.3.4 Four-Grams and Lexical Bundles

Four-grams are sequences of four words occurring together in a corpus.
When we call them four-grams, it does not matter how often they occur
in a corpus; they are still called four-grams. That is, every four-word
sequence in a corpus is a four-gram. However, similar to tri-grams, when
these four-word combinations occur at least 10 or 20 or more times in
a million words (depending on how conservative we want to be) and
appear in at least five different texts (to avoid idiosyncratic—that is,
individualistic—use) in a register, they are referred to as ‘lexical bundles’
in the literature (see the background on lexical bundles in Cortes 2015).
Lexical bundles were first identified and named by Biber and Conrad in
1999, and were described in detail by Biber et al., 1999. Since then, many
studies have reported on various aspects of these bundles. For example,
even though they are not always complete structural or semantic units,
scholars have described them for their structural and functional make-up.
They have been investigated in a broad number of registers (Biber et al.,
2004) and in specific registers such as university language (Biber and Barb-
ieri 2007). Most recently, they were also examined for their position in the
structure of discourse (Csomay 2013) to see relationships between their
functions and their position in texts. Other studies go inside the bundles
to see which elements in the bundles are more “fixed” than others (Biber
2009) or they look at what kind of semantic categorization can be done on
the words immediately surrounding the actual bundle (e.g., Cortes 2014).
Lexical bundles, a special type of four-word sequences, are defined by

the number of times they occur in a million words in a register. As men-
tioned above, they are not necessarily structurally complete units, e.g.,
in the case of. But sometimes, they happen to be units that we recognize
and know quite well, such as if you look at in classroom discourse or on
the other hand in academic prose. The latter is a semantically and func-
tionally complete unit even though that is not a typical characteristic of
bundles by definition, and, therefore, it is not common to find these among
bundles. Given that earlier studies (especially Biber et al. 2004) classified
these four-word sequences, or lexical bundles, based on their grammatical
as well as functional categories, we can add to our understanding of how
lexical bundles are used if we investigated where in the sentence the lexical
bundles are positioned and whether there is a difference in the distribu-
tion of those positions in, let’s say, written and spoken registers. Shall we
find out?
Project 3.14: ‘On the Other Hand’

Go to COCA at corpus.byu.edu/coca/, sign in, and type in on the other
hand. Hit ‘chart’ and ‘search.’ You should be seeing something like
Figure 3.9.
Report on what you see in terms of spoken/academic usage of the lexi-
cal bundle on the other hand in the COCA corpus. Then click on the
Figure 3.9 Distributional patterns of the four-gram ‘on the other hand’ in COCA
column under spoken. This way all your examples will be from the spo-
ken sub-registers in COCA. (See Figure 3.10.)
Take each one of the 100 samples you get this way and classify each
bundle according to its position in the sentence. You can use the follow-
ing categories: a) sentence initially; that is, when the bundle is the first
four words in the sentence or utterance (for spoken); b) sentence finally;
that is, when the bundle is the last four words in the sentence or utterance
(for spoken); and c) neither sentence initially nor sentence finally; that is,
when neither a) or b) applies. Use Table 3.5 to record the results.
As a next step, do the same with academic prose to get your sam-
ples from that register. Then classify each of the 100 examples as one of
the three categories above. Finally, calculate the percent value for each.
(Note: Since you had 100 observations for each register, the percent value
and the raw counts are the same.) Reset the sample size to 200, run it
again, and see whether your results are similar?
As a final note to this section, the longer the n-gram, or the
word-sequence, the less frequently it will occur simply because n-grams
are embedded in one another. For example, in the four-word sequence
on the other hand, the three-word sequences of on the other and the
other hand are both present. These would be counted as two separate
three-word sequences.
Figure 3.10 Concordance lines for the four-gram ‘on the other hand’ in COCA
Table 3.5 Distribution of the sentence position of ‘on the other hand’ in spoken and
written discourse
Sentence-initial Sentence-final raw Other raw Total raw count

raw count count count (percent)
(percent) (percent) (percent)
Spoken 100 (100)

Written 100 (100)
Total (Total 100 (100)
percent)
3.3.5 Five- and More-Grams

When you are extracting n-grams from a corpus, you can imagine that
your sequences could be endless. However, that is not true. First, all
three-word sequences will contain all the two-word sequences in them.
Similarly, if you look at four-word sequences, they will have all the
three-word sequences in them (and the two-word ones as well). There-
fore, the higher the n is in your n-gram, the less frequency you will get
for each sequence. This is particularly relevant in lexical bundles where
it is not just the sequence that is part of the definition of a bundle but the
cut-off point counts as well.
To illustrate this, we ran a lexical bundle search in a corpus of web
texts. This was a one-million word corpus of cyber-texts collected from
five internet-registers: pop-culture news, advertising, forum requests
for advice, blogs, and tweets (Connor-Linton 2012). Csomay and Cor-
tes (2014) reported on the lexical bundles of four to eight words found
in these five registers and examined their functions. They compared the
webtext functions with those found in non-electronic registers (e.g., refer-
ential bundles, discourse organizers, and stance markers) and highlighted
those that they found specific to these electronic registers (e.g., descrip-
tive, narrative). The following four-word bundles were found unique to
the webstexts: what do you think, do you have any, I am trying to figure
out, a recommended eBay seller. And the following five-word bundles
were found unique to the webtexts: what do you think of, I just added
myself to the, made in the US and imported, originally posted by White
Tiger. The only one six-word bundle to meet the frequency requirement
(at least 10 times in a million words and in five different texts) to make
it a bundle was: I just updated my Squidoo page. Finally, the two longest
n-grams meeting the requirement to be a lexical bundle (at least 10 times
in a million words and in five texts) in this register were a seven-word
bundle, This article was originally published on Engineeringtalk, mostly
occurring in advertisements, and an eight-word bundle, A free brochure
or catalogue is available from, also from an advertisement.
Project 3:15: Webtext Bundles

GlobWbe is an additional corpus on the website that contains texts from
the internet. Check whether the webtext bundles reported above are pres-
ent in all and with what frequency.
In COCA, you can download the n-grams found up to five-grams here:
www.ngrams.info/
3.4 POS Tags

Marking the words with part of speech (POS) tags [n, v, j, r] can lead us
to different types of analyses than those we have seen so far. For example,
you can look for specific words and their associated grammatical pat-
terns or you can look for (co-occurring) grammatical patterns indepen-
dent of the actual words. On the one hand, POS tags can help you be
more specific about the words you are searching if you are going through
an already-existing search engine and if you are searching a corpus that
has been tagged for part of speech. On the other hand, POS tags can also
give you more options and more flexibility about your search. Below are
examples of each. There are many other part of speech categories that
could be potentially interesting for any linguistic study, but before going
into some analyses we can do when we work with tags, let’s clarify some
basic grammar.
Each part of speech belongs to one of two basic classes: open or
closed. Those part of speech categories that belong to the open class
contain an unlimited number of members in them. That is, there is not
a set number of members for an open class POS. There can be as many
as there are in a language and we do not know how many there are. In
contrast, those POSs that belong to the closed class have the character-
istic to contain a limited number of members and we know exactly what
they are. As another way to understand this distinction, note that we
frequently have new open class words coming into the language, but the
same is not the case with closed class words—new closed class words
are quite rare.
Examples of POS belonging to the open category are the four main
parts of speech: nouns, adjectives, verbs, and adverbs. Typically, nouns [n]
are modified by adjectives [j], as in, for example, big red car where big is
a characteristic of the noun in terms of its size, and red is a characteristic
of the noun, telling us the color of the noun, and car is a common noun.
Verbs [v] are modified by adverbs [r], as in, for example, he drove fast
where drove is a verb—in past tense—and fast is an adverb modifying the
verb by telling us how the action expressed by the verb is done. Examples
of POS belonging to the closed category are many, and we only mention
a few examples here to illustrate the point. These are: pronouns [pro], as
in he, she, it, their, him, etc.; determiners [d], such as this, that, etc.; or
conjunctions (both clausal and phrasal) such, as, and, but, because, etc.,
and many more.
Whether a POS belongs to the open or closed class, there are endless
possibilities to search for association patterns, as shown in Chapter 2.
As we have also seen in that chapter, the co-occurring patterns of these
categories are the most interesting types of studies from the perspective of
register variation because they are able to provide us with more compre-
hensive and detailed analyses of texts. Now, some corpora (e.g., COCA)
have POS tags attached to each word, and some corpora (e.g., Michigan
Corpus of Academic Spoken English—MICASE, and Michigan Corpus
of Undergraduate Student Papers—MICUSP) do not have that feature
in addition to the actual words in a corpus. Some scholars find it more
difficult to do a search on POS tags, and others write their own com-
puter programs to process and count the different grammatical patterns
through those tags.
3.4.1 Specifying POS for a Word

First, let’s see how specifying part of speech categories with POS tags
can help you be more specific about your search words. In the previous
section, we compared the use of state and say in spoken and written reg-
isters. Go to wordandphrase.info, click on ‘Frequency,’ and type in say.
As you will see (also in Figure 3.11), the word say is tagged in the COCA
corpus as a verb [v] marked by the pink color (on the site), and as a verb,
it ranks 19th on the list of words frequently occurring in the corpus.
However, as you know, the word say could be used as a noun as well, as
in ‘to have a say about something.’ Now, type in state, and you will see
that, similarly, the word state can be used as a verb or a noun. Clearly,
Figure 3.11 Frequency search in Word and Phrase

when you compare the frequencies of these words in spoken and written
discourse, you want to compare the instances in both cases only when
they are used as verbs. Otherwise, you would be comparing apples with
oranges. Luckily, COCA is marked up with POS tags, and so you could
just uncheck the boxes for all other possible POS categories (see ‘Part of
Speech’ buttons) that you are not interested in when searching for the
verb function in your comparison.
Second, let’s have a look at how this would broaden your options of
looking for patterns—that is, how this will give you more options and
more flexibility in your search. These main POS categories identify the
word as you type it into the search box—that is, orthographically only.
Through the tags, however, we are able to look for variation within POS
categories as well. The tags, for example, allow you to look for a given
word in different word classes, such as can as a noun and can as a modal
verb. Another example of what tags can do could be if you would like to
look at a specific verb but you are only interested in the past tense forms,
or if you want to search for exact examples of a particular noun but
only in its plural form. To do this type of search, you can use POS codes,
such as [v?d*] (e.g., ate, drank, looked, etc.) and [*nn2*] (e.g., houses or
backyards), respectively.
Now click on ‘COCA’ in the right hand side of the wordandphrase.info
site, enter, login, and type in [v?d*] in the search window. This search
will provide you with all the past tense verbs in the corpus ranked by fre-
quency. Now if you tried [*nn2*], what do you think you will get? (See
all POS tags used in COCA via the help menu.)
The third possibility is to see how often a specific word comes about
in texts with different word forms. In the previous examples, we looked
at say as a verb. If you did the same search as in ‘WordandPhrase’ but
in COCA, you would be typing the following string into the search box:
say.[v*]. This indicates that you want to search and see the word say in
this form and when it is a verb. What if you want to find out how often
the verb say is used in past tense because you would like to make a com-
parison between the past tense use of say across four registers or see the
syntactic position of the past tense use of say. Type the following into
the search box: [say].[v*] (note the brackets around say). What kind of
information do you get? This time, if the word is in square brackets, you
work with the lemma of the word (i.e., the base form of the word), but
you allow the inflections to be listed as well (such as past tense mark-
ers, third person ‘s’ marker, etc.) What you see then is that the program
divides the frequencies of the lexical verb say into the different verb
forms, including past tense and third person singular, etc., and reports
on the results accordingly. Click on the link ‘SAID’ which will pull up
all the examples of the word with this form, should you need textual
examples.
The fourth option could be that you have an n-gram (e.g., a lexical
bundle) and you want to see what kinds of words precede or follow that
sequence. Let’s try this with if you look at.
Project 3.16: ‘If You Look at’ What?

Go through the following steps in completing this task:
Step 1: Go to COCA. Type in the search box if you look at [nn*], click
on ‘chart,’ and then ‘KWIC.’ This will give you all the possible
nouns following the bundle if you look at in the concordance lines
and the frequency of this sequence in all of the registers in COCA.
Step 2: Go to a descriptive English reference grammar such as, Biber et al.
(1999) Longman Grammar of English or Biber et al. (2002) Stu-
dent Grammar of Spoken and Written English and make a list of
the categories for nouns (pp. 56–64 in the 2002 book). Now, look
at the nouns that follow the lexical bundle you searched for, and
classify the nouns in the categories you have set up. Can you see
any patterns? That is, can you see any one semantic category that
seems to occur more often than others? Report on your results.
Step 3: Take 30 randomly selected samples for the newspaper listing and
30 randomly selected samples from the spoken corpora listing.
Compare the kinds of nouns following the bundle if you look
at. Based on the examples you see, can you draw any conclu-
sion whether, in your sample, there is evidence that one register
may use more of certain types of nouns than the other with this
bundle?
Step 4: Finally, just for fun, type in the search box if you look at [n*], and
click on ‘chart’ and then ‘KWIC.’ How did your list change from
the one you generated by doing Step 1 above?
Note
1 Please note that these numbers are normed counts. Chapter 8 explains how
normed counts are calculated.
References
Biber, D. (2009) ‘A corpus-driven approach to formulaic language in English:
Multi-word patterns in speech and writing’, International Journal of Corpus
Linguistics 14: 275–311
Biber, D. & F. Barbieri (2007) ‘Lexical bundles in university spoken and written
registers’, English for Specific Purposes 26: 263–286
Biber, D. & S. Conrad (1999) ‘Lexical bundles in conversation and academic
prose’, in Hasselgard, H. & S. Oksefjell (eds) 1999: 81–190
Biber, D., S. Conrad & G. Leech (2002) Longman Student Grammar of Spoken
and Written English. New York: Longman
Connor-Linton, J. (2012, March) ‘Multiple perspectives on analysis of webtexts’,
Colloquium presentation at Georgetown University Round Table, Georgetown
Cortes, V. (2014, September). ‘Analyzing the structures, semantic prosodies,
and semantic preferences of lexical bundles in research article introductions’,
Paper presented at the Conference for the American Association of Corpus
Linguistics (AACL), Flagstaff, Arizona
Cortes, V. (2015) ‘Situating lexical bundles in the formulaic language spectrum’,
in Cortes and Csomay (eds) 2015: 197–216
Cortes, V. & E. Csomay (2015) Corpus-based Research in Applied Linguistics.
Studies in Honor of Doug Biber. Amsterdam: Benjamins
Csomay, E. & V. Cortes (2014, September). ‘Lexical bundles in cybertexts’ Paper
presented at the Conference for the American Association of Corpus Linguis-
tics (AACL), Flagstaff, Arizona
Firth, J.R. (1951) Modes of Meaning. Essays and Studies of the English Associa-
tion [NS 4], 118–149
Hasselgard, H. & S. Oksefjell (eds). Out of Corpora. Amsterdam: Rodopi
Chapter 4
Projects Using Publically

Available Corpora
4.1 Word- and Phrase-Based Projects

4.2 Grammar-Based Projects
This chapter provides you with an opportunity to use readily avail-

able corpora to conduct corpus linguistics projects. In this chapter, you
will gain practical experience in many corpus linguistic methodologies,
including KWIC, collocations, use of search terms, and register analysis.
There are a total of 12 projects in this chapter.
The projects in this part of the book focus on a number of different
corpora developed by Mark Davies at Brigham Young University. As
mentioned in Chapter 3, the corpora are available at the following web
address: corpus.byu.edu. Although you need to register in order to get
access to multiple searches, it is free of charge. In addition to being com-
pletely free, these corpora have a number of advantages for researchers,
teachers, and students of language studies. The corpora include differ-
ent varieties of English, including American (Corpus of Contemporary
American English), British (British National Corpus), and Canadian
(Strathy Corpus). This allows for comparisons of language across differ-
ent varieties of English. A more recent corpus of different language variet-
ies includes twenty different varieties of English used around the world
(Global Web-Based English corpus). The varieties included in this corpus
range from the United States, Great Britain, and Australia to countries
such as Hong Kong, Ghana, and Jamaica. Two of the corpora include
texts from the early nineteenth century (The Corpus of Historical Amer-
ican English) or from the earlier part of the twentieth century (TIME
Magazine Corpus) so that historical trends and language can be investi-
gated. All of the corpora also use the same search interface so that once
you learn how to “ask” for information in one corpus, you can conduct
searches in all of the available corpora. A list of the corpora we will use
for the corpus projects is provided in Table 4.1.
Projects 59
Table 4.1 Information on corpora available at byu.edu
Name of Corpus # of words Language/dialect variety Time period
Global Web-Based English 1.9 billion 20 countries 2012–13

(GloWbE)
Corpus of Contemporary 450 million American 1990–2012
American English (COCA)
Corpus of Historical 400 million American 1810–2009
American English (COHA)
TIME Magazine Corpus 100 million American 1923–2006
British National Corpus 100 million British 1980s–1993
(BYU-BNC)
Strathy Corpus (Canada) 50 million Canadian 1970s–2000s
There are a few important points to remember about these corpora

that are especially true when conducting projects that compare features
across different corpora or when looking at register differences. First, the
corpora are not of equal size. The corpus sizes range from the 1.9-billion-
word GloWbE corpus to the 50-million-word Strathy corpus. This means
that if one is comparing a different feature across different corpora, it
is important to used normalized counts. Another important aspect to
keep in mind relates to different registers in the corpora. Some corpora
(GloWbE, COHA, and the TIME Magazine Corpus) are comprised of
a single register; other corpora (COCA, BYU-BNC, and Strathy) con-
tain multiple registers. Even within a single register, there are important
situational differences that need to be taken into account. For example,
in the BYU-BNC corpus, the spoken data include contexts such as oral
histories, meetings, lectures, and doctor-patient interactions. In COCA,
the spoken examples are all taken from television and radio news and
information shows. These different situational variables need to be con-
sidered when comparing language features across the corpora as well as
looking at the distribution of features within a particular register. For
example, there are many differences between news shows and face-to-
face conversation (for one, news shows display grammatical features that
are characteristic of more informational types of discourse while face-to-
face conversation displays features that have been associated with more
involved or interactional types of discourse; cf. Biber 1995) that would
account for differences in language features. Consequently, findings in the
spoken register of these corpora need to be understood in relation to the
spoken contexts used in the corpus. A feature may be related to spoken
language in news and information broadcasts but not found in other spo-
ken contexts. A final factor to consider when comparing the corpora is
related to time. Some language features are likely to be different depend-

ing on the time period under investigation. Some of the projects below
ask you to look at different corpora in order to understand how language
features have changed over time.
4.1 Word- and Phrase-Based Projects
Project 4.1: Lexical Change Over Time

The Corpus of Contemporary American English (COCA) corpus is
divided into five time periods. Find five words that are recent (have a
higher frequency in the most recent time period) and five examples of
words that are more common in the earliest time period. For the more
recent words, is there a steady growth curve or does the word gain popu-
larity fairly rapidly (i.e., in a single time period)? For the declining words,
is there a steady decline or a fairly rapid decline? What are some possible
reasons for these tendencies you have found?
Project 4.2: Meanings of Literally

The word literally was originally used to mean something similar to
exactly, precisely, or actually. In this sense of the word, the meaning is
closely related to the concept of something being free from metaphor.
This meaning is understood in a sentence such as:
I’ve literally spent more holidays with the animals than I have with
my own family. (example taken from COCA)
In addition to this meaning of the word literally, there is another sense

of the word that shows emphasis (with a meaning similar to really). This
sense of the word is sometimes the opposite of ‘being free from metaphor’
as in a sentence such as:
There’s a story about this in this book and it blows my mind, literally.
(example taken from COCA)
Using COCA, determine the variation in use of both the literal and figura-
tive sense of the word literally by completing the following steps:
Step 1: Develop and state your method for determining whether the word
falls into the “literal” category or the “non-literal” category. Do
you find any other senses of the word that do not fit into your
two categories? If so, describe the extra category or categories
you have found and provide examples to support each category.
Projects 61
Step 2: Making reference to at least 200 examples from three different

registers, do you find a preferred sense of the word literally? That
is, can you find evidence that one sense of the word is used more
than another sense?
Step 3: Making reference to at least 100 examples in all five registers, do
you find any register differences in the meanings of literally?
Step 4: Using the Corpus of Historical American English, do you find any
historical trends in the different senses of the word literally?
Project 4.3: –gate as Post Fix for Any Accepted Problem

After the Term Watergate
In addition to its contribution to the demise of the presidency of Richard
Nixon, the so-called Watergate Scandal has contributed to the American
lexicon through the use of the suffix –gate added to describe controversial
political or social incidents. In this project, you will explore the different
ways this suffix has been used in American English and look at other
language varieties to see if this suffix is also found in other varieties of
English. Complete the following steps:
Step 1: Using COCA, identify at least six cases where this suffix is used
with a noun. For each, note its first appearance; then note when
each term was most frequently used and whether it is still used
today.
Step 2: Using both the British National Corpus (BYU-BNC) and the Stra-
thy Corpus, determine whether this suffix is found in American
English only or whether it is also found in British and/or Canadian
English.
Step 3: Interpret your results: What are some possible reasons for your
findings? Are there other examples of prefixes or suffixes that are
specific to varieties of English?
Project 4.4: Clichés

Owen Hargraves has written a book titled It’s Been Said Before: A Guide
to the Use and Abuse of Clichés (2014). In the introduction of this book,
Hargraves states:
While it is true that a vast number of expressions have become tired

through overuse, there is an opportunity to make even the most
worn expressions striking and powerful. How do we decide which
among many expressions might be just right for the occasion, or just
wrong?
(p. xi)
In this project, we will take a close look at some of the clichés Har-
graves mentions in his book. Complete the following steps:
Step 1: Come up with a “working definition” of a cliché. How has the

concept of a cliché been defined? Is there general agreement on
what a cliché is? What differences in definitions do you note?
Step 2: To add to your working definition, use the collocation option in
COCA and look at the words that surround cliché. You can select
up to four words to the right of your search term and four words
to the left. What words precede it? What words follow it? How
would you characterize these words? What does this information
tell you about the meaning of the word cliché?
Step 3: Using both COCA and the British National Corpus (BYU-BNC),
provide the normalized frequency of each of the phrases below:
dizzying array meteoric rise point the finger

perfect storm touch and go totally awesome
What do these frequencies tell you about the relationship between these
clichés and American and British English?
1. Are there specific registers that use these phrases more frequently
than other registers? What might be some reasons for any differences
you find?
2. Given your working definition of a cliché, would you define any or
all of the six examples as clichés? Are there other examples that are
more representative of clichés?
Project 4.5: Collocation of Modifying Elements

In this project, you will look at the most common words that follow the
modifiers below.
categorically deeply entirely far-reaching massively
You will use COCA, the British National Corpus (BYU-BNC), and
GloWbE. Complete the following steps in the project.
Step 1: For both COCA and the British National Corpus (BYU-BNC),
determine the most common word that follows each of the terms
above. You can do this by using the “COLLOCATES” function;
set the span to “0” on the left pull down menu and “1” on the
right pull down menu.
Projects 63
Step 2: What similarities and differences in the two language varieties do

you find in the types of words that follow the modifiers?
Step 3: Use GloWbE to determine the most common collocate of the five
modifiers above. Using the patterns you found for both American
and British English, try to find a language variety that patterns
like American English and a language variety that patterns like
British English for each of the five modifiers. What factors might
influence these language varieties to pattern like American or Brit-
ish English? Do you find any patterns that are unlike both Ameri-
can and British English? What are some possible reasons for any
new patterns that you find?
Project 4.6: Sustainability

According to the Oxford English Dictionary (www.oed.com), the adjec-
tive sustainable originally referred to the ability to endure something. In
this definition it was synonymous with the adjective bearable. Although
this use of the term is now quite rare, there are other meanings of sus-
tainable that are more commonly used in English. These definitions are
provided below (definitions are quoted from www.oed.com):
1. Capable of being upheld or defended as valid, correct, or true

2a. Capable of being maintained or continued at a certain rate or level
2b. Designating forms of human activity (esp. of an economic nature) in
which environmental degradation is minimized, esp. by avoiding the
long-term depletion of natural resources; of or relating to activity of
this type. Also: designating a natural resource which is exploited in
such a way as to avoid its long-term depletion.
In this project, you will use both COHA and COCA to investigate these
different meanings of the word sustainable (and its noun counterpart,
sustainability) over time and across registers. Complete the following
steps:
Step 1: Using COHA, note the first 50 occurrences of the adjective sus-
tainable. For each use, provide the date of occurrence and note
which of the three definitions provided above best fit with the
occurrence of the word. Make sure to provide examples from the
corpus to support your analysis of their meanings. Is one use of
sustainable more prevalent than other uses of it? Is there a ten-
dency for the meaning to change over time?
Step 2: Using COCA, note the register distribution of the adjective sus-
tainable. In which registers is sustainable most common? In which
registers is sustainable less common? Are there specific meanings
of sustainable that are representative of specific registers? Pro-
vide some reasons for any register or meaning differences that
you find. Make sure to support your analysis with examples from
the corpus.
Step 3: This part of the project asks you to look at the meanings and
register distribution of the noun sustainability. According to the
online site “Environmental Leader”:
Sustainability includes sustainable building, design and opera-

tions. Sustainability is the collection of policies and strategies
employed by companies to minimize their environmental impact
on future generations. Ecological concerns, such as the environ-
mental impact of pollutants, are balanced with socio-economic
concerns such as minimizing the consumption of limited natural
resources to maintain their availability for the future.
(http://www.environmentalleader.com/category/sustainability/)
Using COHA, note the first 20 occurrences of the word sustainability.

For each use, provide the date of occurrence and note which of the three
definitions provided above best fit with the occurrence of the word. Make
sure to provide examples from the corpus to support your analysis of
their meanings. In which registers is sustainability most common? In
which registers is sustainability less common? Provide some reasons for
any register distribution differences that you find. Do the meanings of
sustainability all relate to environmental or ecological issues or are there
other senses of the word that are found in COCA?
4.2 Grammar-Based Projects
Project 4.7: Grammatical Agreement in Collective Nouns

English has a grammar rule—called subject-verb agreement—which
states that the subject of a sentence must agree with the main verb in
number (see Table 4.2 below). With some exceptions (such as the verb
be), there are two forms of the verb. In the regular verb example below,
there are two forms of the verb, like for first and second person singu-
lar subjects and all of the plural subjects (we, you, and they) and likes
for third person singular subjects (he, she, and it). One verb that shows
greater variation in subject-verb agreement is with the verb be, for which
there are three different forms of the verb: one for first person singular
(am), one for second person singular and all plural subjects (are), and
Projects 65
one for third person singular subjects (is). Note that for simple past tense,
there is no difference for the regular verbs and only two distinctions for
the verb be (was for first person and third person singular and were for
all other subjects).
Sometimes certain types of nouns—collective nouns—are ambiguous
with respect to the rule of subject verb agreement. Consider, for example,
the two sentences below:
• My family is loving and supportive.

• My family are loving and supportive.
It is sometimes claimed that collective nouns and subject verb agreement

behave differently in British and American English (e.g., Azar and Hagen
2009; Swann 2005). For example, one grammar website (http://www.
onestopenglish.com/grammar/grammar-reference/american-english-
vs-british-english/differences-in-american-and-british-english-grammar-
article/152820.article) states that in British English, either form is possible
but in American English, the strong tendency is to treat collective nouns
as singular subjects. In this project, you will use both the COCA and Brit-
ish National Corpus (BYU-BNC) to determine the preferred agreement
pattern (singular or plural) of four collective nouns:
• audience
• committee
• crowd
• team
• family
Complete all of the four steps in the analysis:
Step 1: Using only the verb be (e.g., is, are), determine if there is dif-
ferent distribution of singular (is) or plural (are) agreement for
each of the five nouns above. Are there general patterns found
in all five collective nouns in subject position or do the nouns
Table 4.2 Example of a regular and irregular verb in English
Regular verb Be verb
I like books. We like books. I am tired. We are tired.

You like books. You like books. You are tired. You are tired.
He/she it likes They like books. He/she/it is tired. They are tired.
books
show different patterns? What specific search terms did you use
to obtain your results?
Step 2: Do you find the same patterns with verbs other than be? What
specific search terms did you use to obtain your results?
Step 3: Do you notice any register differences in the agreement patterns?
That is to say, do some registers use one agreement pattern more
or less frequently than another register? Are there register differ-
ences with respect to any of the specific collective nouns? That is
to say, are some collective nouns in subject position more frequent
in specific registers than other collective nouns?
Step 4: How would you interpret your findings? Are there specific reasons
for any similarities or differences you have found?
Remember:
• The collective nouns need to be in subject position.

• The corpora may not be of equal sizes, so you will need to normalize
the counts. Do you see any differences between British English and
American English? If so, what are the differences?
• There are different ways to use the corpus to search for agreement
patterns. Make sure that you include the exact search terms you used
in your investigation.
Project 4.8: Variation in the Passive Voice

There are two main “voices” in English. In grammar terms, voice refers
to the relationship between the verb of a sentence and the other par-
ticipants, such as the subject and the object of a sentence. In a sentence
such as The boy saw the ghost (an active voice sentence), we know who
sees the ghost (the subject of the sentence, the boy) and who is seen
(the object of the sentence, the ghost). In this sense, the subject noun
phrase the boy serves as the “actor” of the sentence (the one doing
the action of the verb) and the noun phrase object the ghost serves as
the “patient” or “recipient” of the action (the receiver of the action).
The second type of voice in English is called the passive voice. For
example, in a sentence such as The boy is seen by the ghost, even though
the boy is still in the subject position of the sentence, it is the ghost who
is doing the seeing in this type of sentence. Consequently, the voice of
the sentence provides information on who or what is doing the action
and who or what is affected by the action expressed by the verb. The
passive voice is formed by adding an auxiliary verb (be) to the passive
voice sentence and substituting the “regular” form of the verb with the
past participle. You will be asked to look at this rule a bit closer in the
project below.
Projects 67
Another noteworthy aspect of the passive voice includes variation in

the extent to which the original actor of the sentence (the subject of the
active voice sentence) is present in the passive voice sentence. For exam-
ple, compare the active voice The girl broke the window with its passive
voice counterpart, The window was broken (by the girl). In this pas-
sive voice sentence, there are two options, one that states who broke the
window and another that doesn’t state the subject by the deletion of the
entire “by phrase” (e.g., The window was broken). We can call the first
of these passive types the long passive (a passive sentence that includes
the “by phrase”) and the second type the short passive (a passive sentence
without the “by phrase”).
In addition to the verb be, the passive voice can also be expressed by
the auxiliary verb get. See Table 4.3 for examples of the two auxiliaries.
Using COCA, compare the two forms of the passive, initially concen-
trating on the two auxiliary verbs was and got (hint: try using the search
strings was [v?n*] and got [v?n*]).
1. Which of these two types of passive is more common overall? Are

there register differences in the two passive types? What possible rea-
sons might there be for any differences you have found?
2. Is there a difference in the verbs used in was passives vs. got passives?
How would you describe these differences?
3. Choose five different was passives and five different got passives.
For each type, determine whether there is a preference for the “by
phrase.” Does your data suggest that the “by phrase” is dependent
on the verb, on the auxiliary verb, or is it due to some other factor?
Project 4.9: Going to as a Modal Verb

In the two sentences below, the underlined words going to are different
types of grammatical constructions. In the first, going is the main verb of
the clause and to is a preposition that is followed by a noun phrase (the
loo); in the second sentence, going to precedes the verb go. One way to
show this difference is to determine when going to can be contracted or
simplified to gonna. It is possible to use gonna in the second sentence but
not the first sentence.
Table 4.3 Passive voice
Subject Auxiliary verb Main verb By phrase
John was stopped by the police

John got stopped by the police
(a) My father is up all night going to the loo which keeps both him
and my mother awake.
(b) Yet not everyone is going to go to college, develop software skills,

or become an entrepreneur.
(examples from COCA)
The difference between these two examples illustrates the descriptive

observation that in cases where a contraction is permissible, the going to
is functioning as a modal verb; in cases where gonna is not possible, the
construction is comprised of a main verb (going) followed by a preposi-
tional phrase. In this project, you will examine the modal verb going to
in detail. Complete the following steps in your analysis of the modal verb
going to (gonna).
Step 1: Determine the search term(s) you will use to find examples of
going to followed by a verb (e.g., going to [v*]).
Step 2: Use COHA and determine when this use came into the language.
What patterns of development do you find? Are there certain
verbs that tend to occur with going to? Have these verbs changed
over time?
Step 3: Are there differences in the way the two forms of going to and
gonna are used in COHA? What are some possible reasons for
any differences you may find?
Step 4: Using COCA, do you see any register differences in the use of
going to and gonna as a modal verb? If so, what are the differ-
ences you see? Try to provide some explanation for any register
differences you may find.
Project 4.10: Grammatical Constructions Following

Begin, Continue, and Start
In English, there is a good deal of variation in the forms of non-finite
grammatical clauses that follow (i.e., complement) verbs. The non-finite
clauses in this project are the gerund (–ing) clause (a) and the infinitive
(to) clause (b):
(a) Chekesha was third at district last year and barely missed going
to regionals.
(b) The study subjects were asked to estimate the surface normal at
many points on the drawings.
Projects 69
Note that the infinitive clause is not a possible complement of miss (*. . .
missed to go to regionals) and the gerund clause is not a possible comple-
ment of ask (*. . . asked estimating the surface normal at many points on
the drawings). There are other verbs that allow both gerund and infinitive
clauses as complements. The verb start, for example, allows both, as seen
in (c) and (d).
(c) When you start looking for things in love, there’s nothing left
to find.
(d) Hair can start to lose its luster and look dull and ashy. . .
In this project, you will look at three verbs (begin, continue, and start)
that allow both gerund and infinitive clauses in COCA. Complete all
steps in the analysis.
Step 1: Come up with search terms that will allow you to find both of the
complementation patterns. (The “POS” function will be helpful
for this exercise.) What search terms did you use?
Step 2: Using COCA, report on the complementation patterns of the
three verbs (begin, continue, and start). How do the three verbs
compare in their complementation patterns?
Step 3: For each of the three verbs, determine whether there are register
differences in the patterns.
Step 4: What reasons account for the variation of complementation pat-
terns in these three verbs?
In your answer, you should also include a discussion of begin and start.
One might expect that because begin and start are virtually synonymous,
they might pattern the same. Is this true? If they do not pattern the same,
what are some possible reasons for the difference?
Project 4.11: Grammatical Categories and Coordination

This exercise is concerned with the distribution of grammatical categories
that are coordinated by and, but, and or. The four categories you will
consider are:
Noun coordinator Noun: e.g., girls and boys
Verb coordinator Verb: e.g., make or break
Adjective coordinator Adjective: e.g., true or false
Adverb coordinator Adverb: e.g., back and forth
Complete all steps in the analysis.
Step 1: Using COCA, provide the overall frequency of all three coordina-
tors with all four grammatical categories (there will be a total of
12). You can use the “POS List” drop down menu for your search
terms. For example, to find “Verb and Verb” you would use the
search term “[v*] and [v*]”.
Step 2: Rank the grammatical categories with each coordinator from
most frequent to least frequent. Do all coordinators pattern the
same or do they pattern differently?
Step 3: For each of the 12 coordinator and grammatical category types,
do you find any register differences? What are some potential rea-
sons for any register differences you find?
Project 4.12: Frugal, Cheap, and Thrifty

In this project, we will consider different connotations of the adjectives
cheap, frugal, and thrifty. We will also look at how these words may dif-
fer in their syntactic positions. There are a group of adjectives in English
that can occur in both pre-noun and post-verbal positions. For example,
the adjective little can be used in the sentence The little house is painted
blue as well as in the sentence The house is little. In the first sentence,
the adjective little is called an “attributive” adjective (i.e., it occurs in
the attributive [pre-noun] position); in the second sentence, the adjec-
tive is called a “predicative” adjective (i.e., it occurs in the predicative
[post-verbal] position). Not all adjectives have such a freedom of move-
ment to these different positions. For example, the adjective upset gener-
ally is found in predicative position (The man is upset) and may sound
odd in attributive position (The upset man left the library).
This project will consider the connotations of a group of adjectives that
can occur in both attributive and predicative positions. We will start this
project by considering the following letter that appeared in “Dear Abby”
on March 20, 2015:
DEAR ABBY: My wife, “Tina,” was very hurt by a friend recently.

Her friend “Sally” called her “cheap” during a
conversation (“she’s cheap like you”). Sally didn’t
intend it to be hurtful, just an illustration—but my
wife is very upset about it. We use coupons when we
grocery shop or dine out; we also watch our thermo-
stats, recycle, etc. On the other hand, we have sent
our children to university without loans, our mort-
gage is paid off, we have traveled extensively and our
net worth is north of a million dollars with no debt.
Projects 71
How do I make Tina realize that Sally’s comment

should not upset her so?— THRIFTY IN TEXAS
DEAR THRIFTY: What happened was unfortunate because the prob-
lem may be that Sally simply chose the wrong word.
What she probably meant was that your wife is fru-
gal. The difference between “frugal” and “cheap” is
that being frugal is a virtue. Because Sally hurt your
wife’s feelings, Tina needs to tell her how it made her
feel so Sally can apologize to her before it causes a
permanent rift.
This letter (as well as Abby’s response to it) illustrates the social impact
that synonymous words may have on some people. “Thrifty in Texas”
seeks advice to help his wife understand that their friend Sally did not
intend to offend her. Abby’s response indicates that Sally merely choose
the incorrect word and perhaps should have selected a synonym of
“cheap” that indicated the positive aspects of saving money. This project
will consider 1) what a corpus might tell us about the meanings of the
adjectives cheap, frugal, and thrifty; and 2) whether these three synonyms
share similar syntactic characteristics. Complete all steps in the analysis.
Step 1: Using COCA, report on the frequency and distribution of the

adjectives cheap, frugal, and thrifty. Which of these words is the
most frequent, which of these words are less frequent? Are there
any register differences in the distribution of these words? If so,
what are some possible reasons for any register differences?
Step 2: Using the “KWIC” and “COLLOCATES” functions, explain any
differences in meaning among these three words. Do some words
have a more positive or negative connotation than other words? If
so, what evidence can you provide to support your answer? Make
sure to use examples to back up your analysis.
Step 3: Using the “POS” function in COCA, determine whether each of
these three adjectives is more common in attributive or predicative
position. One possible search string for the attributive position is:
[n*] [vb*] cheap; one possible search string for the predicative
position is: [n*] [vb*] cheap. Do all three adjective have similar
syntactic distributions? Are there differences in meaning when the
same word is in a different syntactic position? Make sure that you
use examples to support your analysis. Also, make sure that you
include all of the search terms that you have used.
Step 4: Given what you now know about the different meanings and syn-
tactic positions of the adjectives cheap, frugal, and thrifty, write
a response to “Thrifty in Texas” that might provide some helpful
advice for how to address his problem.
References
Azar, B. & Hagen, S. (2009) Understanding and Using English Grammar (4th
ed.). New York: Pearson Longman.
Biber, D. (1995) Dimensions of Register Variation: A Cross-linguistic Perspective.
Cambridge: Cambridge University Press.
Hargraves, O. (2014) It’s Been Said Before: A Guide to the Use and Abuse of
Clichés. New York : Oxford University Press.
Swann, M. (2005) Practical English Usage (3rd ed.). Oxford: Oxford University
Press.
Part 3
Building Your Own Corpus,

Analyzing Your Quantitative
Results, and Making Sense
of Data
Chapter 5
Building Your Own Corpus
5.1 Do-It-Yourself Corpora

5.2 Deciding on a Corpus Project
5.3 Building a Corpus
5.4 Software Programs and Your Corpus
This chapter will take you through the steps to complete a corpus proj-
ect. By reference to a specific research question, you will learn how to
build your own corpus and then analyze it using both the register analysis
approach covered in Chapter 2 and the corpus software programs cov-
ered in Chapter 3. You will also learn how to use AntConc in this chapter.
5.1 Do-It-Yourself Corpora

In the previous chapter, you were exposed to readily available corpora
through the projects using the suite of corpora at Brigham Young Uni-
versity. These readily available corpora can be used to explore language
variation by reference to different situations of use, such as newspaper
writing, fiction, and spoken language from news talk shows. The BYU
corpora can also be used to understand variation by reference to language
variety (American, British, and Canadian English, for example) as well as
historical variation. These corpora are not, however, designed to under-
stand language variation in other contexts that may also be of interest.
For example, there is no easy way to determine information about the
gender or age of those who produced the texts. If you were interested
in looking at gender or age differences in language use, these corpora
would not be of much use. Certain research questions require “special-
ized” corpora that are built for specific purposes. Sometimes researchers
need to build their own corpora. Corpus building not only allows you
to answer a specific research question, but it also gives you experience
in corpus construction. There is likely no better way to learn about the
issues in corpus design and to appreciate the larger corpora built by other
76 Your Own Corpus
researchers than to try to build one on your own. Building a good corpus
involves a number of steps that are described below.
Before covering the steps in corpus building, we should acknowledge
potential copyright issues. Chances are, your corpus will use the World
Wide Web for the texts you will use for your corpus. In order to do
this, you will need to carefully consider your selection of materials and
the potential copyright infringement issues that relate to compiling and
storing digital texts. Additionally, it is important to take into account
the country in which the corpus materials are used. Different countries
have different copyright rules. What might be considered a copyright
infringement in one country may not be considered so in another coun-
try. If you are using the corpus for educational purposes and do not
plan on selling the corpus or any information that would result from
an analysis of the corpus (e.g., in publications), the likelihood of being
prosecuted as a copyright violator is usually small. Nevertheless, you
should take into account the following guidelines when building your
own corpora:
• Make sure that your corpus is used for private study and research for
a class or in some other educational context.
• Research presentation or papers that result from the research should
not contain large amounts of text from the corpus. Concordance
lines and short language samples (e.g., fewer than 25 words) are pref-
erable over larger stretches of text.
• When compiling a corpus using resources from the World Wide Web,
only use texts that are available to the public at no additional cost.
• Make sure that your corpus is not used for any commercial purposes.
For those interested in more information on corpus building and copy-

right laws, there are some sources to consult at the end of this chapter.
5.2 Deciding on a Corpus Project

Corpus research projects take a good deal of time commitment to com-
plete. A worthy research project has a number of different components,
including providing a motivation of the significance of the topic, a clear
description of the corpus and the methods used in the study to analyze
your results, presentation of results, a discussion of the results, and a
conclusion that provides a summary and “takeaway message” of the
research. In Chapter 8, you will learn some more about how to present
your research as both a written report as well as an oral presentation.
However, before embarking on this project, it is valuable to spend some
Building Your Own Corpus 77
time thinking seriously about what you want to research and the reasons
for conducting the research; i.e., the motivation for the study. Selecting
an appropriate research issue is not a trivial matter. A well-motivated
research project not only contributes to the knowledge of the field but it
will also hold your interest for the duration of the project (and perhaps
even beyond the duration of the project). The corpus you will build will
depend on your research topic. If you are not clear what you want to
investigate, it will be difficult to build a relevant corpus for the purpose
of your research!
When deciding on a topic, you should choose a topic that is not only
interesting to you but also potentially relevant to others. In other words,
if you were to tell someone about your research topic and they were to
say “So what?” you should have a well-reasoned response. To answer
this question, you should choose a topic that you feel will contribute to
the understanding of how and why language forms may vary in certain
contexts. Some possible general topics could be:
• How does language vary in different varieties of language? (e.g.,

British, American, Canadian, Sri Lankan, Indian, Chinese, Japanese
forms of English)
— Do some forms of English (such as Sri Lankan English or Japa-
nese English) show similarities with other forms of English (such
as British or American English)?
• How does language use vary for people who have different political
or religious beliefs?
— How does language vary within a specific belief or political sys-
tem over time or in those with different genders?
• How do song lyrics vary in different types of music (rock, soul,
hip-hop, country & western)?
— How do song lyrics vary within each type over time or in artists
of different genders?
• How does the language of social media differ from written and/or
spoken language?
— Are posts of social media sites such as Facebook or Twitter simi-
lar to spoken or written language?
• How does language used in “authentic” contexts compare to pre-
scriptive views of language?
— Does newspaper language or academic writing follow the pre-
scriptive rules found in usage manuals?
• Are there gender differences in language use?
— Do male and female romance novels vary in their use of lan-
guage? Do articles written for male audiences differ from articles
written for female audiences?
78 Your Own Corpus
This list is not intended to be exhaustive; instead, it should be viewed

as illustrative. Each of the topics and subtopics described above address
issues that are not specific to the field of corpus linguistics but lend them-
selves to corpus research quite well. For example, a topic examining gen-
der differences could involve creating a corpus of fitness articles with
females as the intended audience and compare this to a corpus of fitness
articles with males as the intended audience. A project looking at the
language of social media posts could be achieved by creating a corpus of
Facebook posts and comparing the language used in the posts with writ-
ten or spoken language found in existing corpora.
In addition to identifying a research issue, you should also write a
research question or set of research questions that you seek to answer in
your research. Because this book uses register analysis as a framework for
interpreting your research, the research questions in your projects all share
the similarity of investigating the extent to which situational variables result
in different linguistic features for some functional reason. In this sense,
all research questions are framed from a particular perspective by refer-
ence to a specific methodology (corpus linguistics). The specific research
question(s) of an individual study depend on the specific variables under
investigation. Note that all of the research issues described above are in the
form of questions. Each research topic has a corresponding question (or set
of questions) or a hypothesis that will be answered in the research study.
Whatever topic you select, you should have a convincing explanation
of your reason for conducting the research. The first questions you can
ask about your project is: “What is my research question?” and “Why
was it important to conduct the research?” If you find, for example, that
song lyrics have different linguistic characteristics in different types of
music, what is the relevance of this finding? Does it say something about
the possible socio-cultural aspects of the consumers of the music or does
it say something about the music genre in general? Clear and convincing
reasons for choosing a research topic will not only help you in motivat-
ing your research, but it will also help you in interpreting the results of
your research. Worthy research topics do not need particular outcomes
to be interesting or relevant. To use the example of song lyrics and musi-
cal varieties again, it would be just as interesting to find little difference
in the linguistic characteristics of musical varieties as it would be to find
strong differences.
A final consideration relates to the type of corpus that you will build to
conduct your project. A vital part of your corpus project is, obviously, the
corpus itself! Before deciding on a final topic, you should determine the
availability of texts that will enable you to address the issue you propose
to research. You will need to make sure that the types of texts you need
to carry out your project are available free of charge (so as to decrease the
chance of a copyright infringement).
Giving careful thought and consideration to the importance and rel-

evance of your research topic (including a strong justification for your
selection of a research topic, i.e., the motivation for your study) is more
likely to result in a project that you are proud of and that contributes to
an understanding of language variation. Taking the time to consider the
significance of your project and its potential application to the field of
applied linguistics (or other fields of study such as sociology, business,
or art and music) is time well spent. Another vital aspect of your topic
selection involves making sure that there are enough suitable texts to
build a relevant corpus. In the following section, you will learn the steps
of building your corpus.
5.3 Building a Corpus

Once you have used the criteria explained in section 5.2 and chosen an
adequate research topic and corresponding research question (or set of
questions), the next step is to build a relevant corpus so that you can
use corpus techniques to identify relevant linguistic features for further
analysis. The corpus that you will likely be building for your project will
not be a large general corpus such as the corpora at Brigham Young Uni-
versity; instead, the corpora you will build will be smaller, “specialized”
corpora that are designed to answer the specific research question(s) you
have decided to investigate. One difference between specialized corpora
and larger, more general corpora relates to their purpose: Specialized cor-
pora are normally designed to address specific research questions while
general corpora are intended for a larger audience and are designed to
answer a larger set of research questions posed by multiple researchers.
This is not to say that specialized corpora are never used to answer dif-
ferent research questions, but they generally are designed to investigate
a restricted set of questions, and therefore, are less likely a representa-
tive of language use in general terms. As you will see further in the next
chapters, you can only draw conclusions in your dataset rather than
generalize the results to larger contexts. Even though smaller, special-
ized corpora are used for more restricted research purposes than gen-
eral corpora, adopting a sound set of guidelines to build the corpus is
still important. A well-designed corpus includes texts that address the
research question(s) of the study, saves the texts into a file format that
allows different software programs to analyze the texts, and identifies
the texts with relevant contextual material so that the different contexts
are easily identifiable in the corpus. We will take a closer look at each of
these below.
The selection of the texts to include in your corpus depends on their
suitability and their availability. Clearly, the texts need to share relevant
characteristics (or variables) that meet your selection criteria for inclusion
80 Your Own Corpus
in the corpus. A project that considers how news writing changes in dif-
ferent time periods would, obviously, require a corpus that includes
newspaper articles written at different periods of time. In order to build
a corpus to address this issue, you would need to make sure that there
is an adequate number of newspaper articles that have been written at
different time periods. Additionally, a corpus addressing this issue would
need to have sub-corpora of relatively equal size. As illustrated in some of
the corpus projects that compared COCA with the BYU-BNC corpus, the
unequal sizes of these two corpora did not allow for straight frequency
comparisons between the two corpora. Thus, corpus “balance” is a key
aspect of reliable corpus building. Note that the balance should consider
not only the number of texts in each sub-corpus but should also consider
the word count of the sub-corpora.
Frequency comparisons are done on the basis of the number of
words, not by the number of texts. If your specialized corpus also con-
tains a sub-corpus component, then you should take care to make the
sub-corpora of fairly equal sizes. Another issue related to corpus balance
in your corpus relates to text types. A news writing corpus would need to
include the various types of news texts—sports and lifestyle news as well
as state, local, national, and international news. If only one of these types
of texts is included then the sample might not account for variation in
the different types of news texts. A balanced news writing corpus would
either include texts of sports, lifestyle, and general news texts or would
select only one of these text types for analysis.
Table 5.1 below examples of projects and the type of corpora that
would need to be built to address specific research topics. Notice that
some of the corpora consist of sub-corpora that are investigated sepa-
rately in order to determine possible variation. Corpora looking at gen-
der differences in song lyrics, fitness articles, or romance novels include
sub-corpora of texts written by or intended for different genders. The
same sub-corpora approach is also seen in projects that investigate lan-
guage variation in blog posts by Christians and Atheists and news talk
shows from different political perspectives or newspaper articles written
in different countries. Other types of corpora do not have sub-corpora
attached to them. For example, a project that compares the language of
Facebook posts to different registers would compare the linguistic fea-
tures of the “Facebook Corpus” with different registers of use in existing
corpora.
Once you have located relevant texts that can be used to build a bal-
anced specialized corpus to address a specific research issue, you will need
to prepare the text to be read by a software program such as AntConc.
Different types of texts have different types of character encoding associ-
ated with them. If you use texts from the World Wide Web, the texts will
likely be in Hypertext Mark-up Language (HTML). A text that is read
Table 5.1 Examples of corpus projects
Project Type of corpus
Gender and song lyrics Song lyrics written by men and women
Facebook as a register? Facebook posts (compared to register in
COCA)
Gender and romance novel authors A corpus of romance novels written by
men and women
Gender differences in fitness articles A corpus of fitness articles written by men
and by women
Variation in English: Newspaper A corpus of English newspaper articles
language in the United States and written in the United States and in
in China China (in English)
Language differences in religious and A corpus of blogs written by Christians
non-religious blogs and Atheists.
Simplification of daily news reports A corpus of newspaper articles written
for the ESL learner for the general public and for second
language learners of English
Comparison of language in Pakistani, A corpus of newspaper articles in
British, and American newspaper American, British, and Pakistani English
English
Linguistic bias in the media A corpus of news talk shows by Rachel
Maddow and Glenn Beck
on a web browser such as Internet Explorer or Google Chrome looks

like this:
You are not very good at parking. You’re just not. Unless you happen
to be a middle-aged gentleman from China called Han Yue. If you are
indeed Mr Yue, then (a) welcome to TopGear.com, and (b) congratu-
lations for obliterating the record for the tightest parallel park . . . in
the wooooorld. Again.
(TopGear 2014)
If this file were to be a part of your corpus and you were to save this text
in an HTML format, the HTML code would be a part of the text and
would include information related to web and text design that is not seen
in the text. The text that is not part of the actual text to examined is often
called “meta-text,” and it is typically put in between brackets < . . . >. The
file would look something like this:
You are not

very good at parking. You’re just not. Unless you happen to be a
middle-aged gentleman from China called Han Yue. If you are indeed
Mr Yue, then (a) welcome to TopGear.com, and (b) congratulations
82 Your Own Corpus
for obliterating the record for the tightest parallel park . . . in the
wooooorld. Again. (source:http://www.topgear.com/uk/car-news/parking-
world-record-video-2014–20–11)
As you can see from the HTML example above, in addition to the actual
text information in the file, there is also extra material that is not rel-
evant to the text to be analyzed. A similar type of extra information is
also present in other types of files, such as Microsoft Word documents.
In order to take this superfluous information out of the text, you will
need to convert any text that you collect into a text file (a file with the
extension “.txt”). The “.txt” format removes all of the mark-up language
found in many other file extensions and allows a software program such
as AntConc to find textual patterns instead of other patterns related to
format or font type. There are different ways you can convert a file into
a text file. If you are collecting texts from the World Wide Web, you can
cut and paste each text into a text file and then use the ‘save as’ option in
order to ensure the file is saved in a plain text format. This same method
works if you have a Microsoft Word document. If you are dealing with
many texts (as we hope you are, in order to build an adequate and repre-
sentative corpus), saving each file into a different format can be tedious
and time consuming. Alternatively, there are a number of file conversion
programs available on the World Wide Web free of charge, such as, for
example, AntConverter. The conversion programs will allow you to con-
vert HTML,. doc/docx, or. pdf files into. txt files using a simple conver-
sion program.
A final and extremely important aspect of corpus building involves
naming and placing your files so you are able to identify them. To return
to the fictitious news writing corpus project described above, a corpus
used to analyze this issue may have 600 total texts, with 200 from three
different time periods. For the purposes of analysis, you would want to
have the 200 texts available as sub-corpora so you could load each of the
sub-corpora separately to see if the files in one sub-corpora varied from
those in another. This would mean that you would need a way to know
which file goes with which sub-corpora. One way to go about doing this
is to come up with a simple coding scheme that allows you to clearly iden-
tify each text as a separate text but also as part of a larger group of texts.
As a simple example, a proposed coding scheme would use a six-number
system in which the first two numbers provide information on the time
period of the text and the last four numbers provide the number of the
text in each time period.
This coding scheme would allow you to clearly identify the text with
“010001” being the first text in time period A, “010002” being the sec-
ond text in time period A, and so on. In order to ensure that the text
Table 5.2 Coding scheme
Sub corpus Text number
Time period A: 01 0001

Time period B: 02 0001
Time period C: 03 0001
numbers relate to the characteristics of each text, each text will also have
the relevant header information described above. Note that this entire
corpus—call it A Historical Corpus of American News Writing—would
consist of three sub-corpora related to each of the three time periods. All
of the files in a single time period would be available in a single folder so
that each sub-corpus could be loaded separately. Depending on different
research questions, the corpus could also be loaded with all three time
periods. Note that if the files followed a consistent labeling practice, you
would be able to determine the time periods by reference to the file
name easily.
An alternative way to name files would be to use transparent file names
with ‘word strings’ instead of numbers. This way, the file names are trans-
parent immediately, and information about the extra-textual features of
the files can be accessed easily. If you choose to do this, you will need
to make sure that the filename length is the same even though you may
not have information in a particular category (for easier processing). For
example, “news_15_election_00” would mean a news text from 2015
about an election in 2000.
Depending on your research question and the design of your corpus,
you may also want to include specific information on the situational char-
acteristics of each text in individual text files in your corpus. Specific
information such as the length of each text (by number of words), the
topic, or the type of each text (if you included different types within a
specific genre) can also be included in individual text files. This type of
information is not a part of the text analysis but it can provide important
interpretive information used in the functional analysis of your results. In
a corpus of general song lyrics, you would want to include lyrics from dif-
ferent types of music (rock, rap, country, popular music, etc.) in order to
achieve balance in your corpus. To be able to identify these different types
of song lyrics, you could either come up with a system of naming each
file (as described above) or you could include some of this information in
each text file. Because you do not want this information counted as part
of the linguistic characteristics of your text, you can put this or any other
relevant information that you do not want to be a part of the linguistic
analysis into angled brackets (< >). This type of information is included
84 Your Own Corpus
in the “headers” of the text but will not be read by the concordance
software. Thus, each individual text file can include relevant header and
other extra-textual information as well the text itself.
Figure 5.1 below is taken from a corpus of argumentative writing by
university students who do not speak English as a first language. There are
six different headers that specify (in order): 1) gender; 2) age; 3) native lan-
guage; 4) degree program; 5) location of university; and 6) topic of essay.
In order to ensure that the header information is not included in the
text analysis, the software program needs to know that anything in
the angled brackets should be ignored in the text analysis. This can be
achieved by indicating that anything in the angled brackets should be con-
sidered extra information, or “tags.” An example of this is provided for
AntConc (see more on this in the next section) as seen in Figure 5.2 below.
The left bracket is identified as the start of the tag information and the
right bracket is used as the end of the tagged information. This essentially
tells the software program to ignore all the information that is included
between the brackets as it does the textual processing and analysis.
In this section, we have outlined the procedure for building a special-
ized corpus to answer a specific research concern. We have outlined some
basic issues in designing your corpus so that it addresses a specific research
issue by constructing a specialized corpus. We have also covered methods
to save the text in a format that is readable by the software program.
Additionally, we have illustrated how separate text files can be uniquely
identified so that different parts of the corpus—the sub-corpora—can be
loaded and compared to each other. Finally, we have considered ways
of potentially adding extra information to each file so that information
on specific characteristics of each text in the corpus can be identified.
<Female>
<21>
<Thai>
<BA>
<LRU>
<Thai women>
In the past, the rights of Thai women were limited by Thai

men. Because of the men believed that the women were not
enough strong for saving the . . .
Figure 5.1 Example of a text file with header information

Figure 5.2 Embedding header tags in AntConc
How ever you go about constructing your corpus, you should use the fol-
lowing questions to guide the process:
1. Do the texts in my corpus allow for an investigation of a specific

research issue?
2. Is the corpus constructed in a balanced manner?
3. Are the texts in a file format that will allow for analysis by the corpus
software you will use in your analysis?
4. Does each text have a specific code and/or header information so that
specific information in each file is identifiable?
5. If relevant to your research question, is the corpus constructed so
that specific sub-corpora are readily retrievable?
5.4 Software Programs and Your Corpus

Laurence Anthony works at Waseda University in Japan. He devel-
ops software programs that are extremely useful for corpus linguistic
analyses and he makes them freely available. To date, he has eleven
software programs available for both PCs and Macs here: www.
laurenceanthony.net/software.html. While it is worth finding out about
86 Your Own Corpus
each one of these as they are very useful, we will mainly focus on two
here, AntWordProfiler for lexical analyses and AntConc for lexical as
well as grammatical analyses, as the most pertinent for your use when
analyzing your corpus.
5.4.1 AntWordProfiler
The function of Anthony’s word profiler is very similar to what we saw
with WordandPhrase, except for two main differences: 1) You can use as
many texts as you want at once for an analysis; and 2) instead of using
COCA as the background or monitor corpus, this one uses two other
word lists (General Service List by Michael West, 1953, and Nation’s
academic word list) on which vocabulary frequency bands are based. (See
Figure 5.3.)
Download the Help file from www.laurenceanthony.net/software/
antwordprofiler/releases/AntWordProfiler141/help.pdf. After reading it,
answer the following questions:
Figure 5.3 AntConc using three-word lists for vocabulary frequency comparison
1. What kind of information can you get to know about your text(s)
through the Vocabulary Profile Tool?
2. What kind of activities can you do through the File Viewer and
Editor tool?
3. What would the different menu options do?
Project 5.1: Let’s say you are interested in finding out about the differ-
ences in the way vocabulary is used in a Wikipedia page and your own
term paper on the same topic. Take one of the papers that you have writ-
ten for another class and save it as a text file. Then search for the same
topical area on Wikipedia, and copy the text, saving it into a text file.
Read both texts into the AntWord Profiler, and run the program twice,
once on each individual file. Note the differences you see between your
text and the Wikipedia text in the following areas:1
a. Number of lines
b. Number of word types
c. Number of word tokens
d. Number of word tokens that fall into each of the three
vocabulary bands
e. Percentage of word coverage from the various vocabulary bands
If you remember, at the WordandPhrase site, each text was marked with
a different color depending on the frequency band it belonged to. Can
we achieve that kind of marking with this tool? If so, how? Would
you modify your text to use less or more frequent words? In what
situation(s) do you think modifying words in a text this way could be
a useful tool? What other kinds of information can you obtain through
this tool?
5.4.2 AntConc
The function of Anthony’s concordance program is very similar to what
we saw at the main interface of COCA. Using your own corpus, you
should be able to do KWIC searches through the concordance lines, and
other types of lexical as much as grammatical analyses in your own texts.
Once again, download the Help file to get an overview of what is possible
with this particular program (Laurence Anthony 2014).
Clearly, this program is capable of facilitating some of the same kind
of analyses COCA did but with your own texts. Among those analyses
are: KWIC (keyword in context), n-grams, collocates in a particular text,
and word lists. In addition, and this is something that we are unable to
do with already existing corpora in general simply because the tool is not
88 Your Own Corpus
available with those corpora, this program is able to show you how the
word (or collocate or lexical bundle or any n-gram) spreads within each
of your texts (a concordance plot).
Read in (i.e., upload) your corpus through the ‘File’ menu (‘Open
files’), and type any search word in the search box that you would like
to find out about in your text(s) and hit the ‘start’ button to get a KWIC
concordance line. (See Figure 5.4.)
If you press the ‘Sort’ button, the words following the keyword will be
in alphabetical order. It is important to keep in mind that the colors in
AntConc do not denote part of speech categories as they do in COCA;
they simply show first and second and third place after the keyword. (See
Figure 5.5.)
If you click on your keyword in the concordance lines, it will show you
the larger textual window (see Figure 5.6).
As mentioned above, if you click on the ‘Concordance plot’ tab, you
will get a view of the spread of your keyword in each of the files you
uploaded as part of your corpus (see Figure 5.7).
Figure 5.4 Searching your own corpus: Concordance lines in AntConc for the word ‘and’
Figure 5.5 Sorting in AntConc
Figure 5.6 File view in AntConc

Figure 5.7 Keyword distribution in full texts
Figure 5.8 Keyword ‘and’ and its collocates

It is also possible to identify collocates of your keyword. In Figure 5.8,

we specified that the collocate that we are looking for should come either
one word to the right or one word to the left of the keyword specified in
the window-span at the bottom.
If you click on any word on the list, it will bring you to a concordance
line listing the instances of that collocate; by clicking on the keyword, you
can take it from here for a larger textual span, as you have seen above.
(See Figure 5.9.)
You can also generate an n-gram list based on the texts you have.
Click on the ‘Clusters/N-grams’ tab on the top and click on ‘N-grams’
under the search term on the bottom, and also specify how big the win-
dow size should be under ‘N-gram size.’ If you are interested in lexical
bundles, you should also specify what the minimum cut-off is under
‘minimum frequency’ and ‘minimum range’ just below. In Figure 5.10,
we set the n-gram size anywhere between 2 and 4 words, and the mini-
mum frequency at 10.
Finally, you can also generate simple word lists with frequency counts
and ranking from your own dataset (Figure 5.11).
When you finish your analysis, be sure to clear all your files and tools,
unless you want to reuse them when you open the program again.
Figure 5.9 Collocates in concordance lines

Figure 5.10 Running your own n-grams in AntConc
Figure 5.11 Running a word list in your corpus in AntConc

Note
1 Word tokens are each word in a text, while word types are each type of word
in a text. For example, in the following text (two sentences), there are eight
tokens (eight word tokens) and six word types (the and cat are repeated, so
they count as one type each): He saw the cat. The cat was black.
References
Anthony, Laurence (2014) AntConc (Windows, Macintosh OS X, and Linux)
http://www.laurenceanthony.net/software/antconc/releases/AntConc343/
help.pdf
TopGear (2014, November 20) http://www.topgear.com/uk/car-news/parking-
world-record-video-2014-20-11
Chapter 6
Basic Statistics
6.1 Why Do Statistical Analyses?

6.2 Basic Terms, Concepts, and Assumptions
6.3 How to Go About Getting the Statistical Results
6.1 Why Do Statistical Analyses?

When doing register analyses, researchers look for patterns of language
use and their associations with the texts’ situational characteristics. We
need empirical measures to see what these associations are, and we need
quantitative measures (e.g., the frequency of a particular language fea-
ture) to see how commonly these patterns occur. We can then look at
how the frequency of that measure is distributed across the two or more
situations that we are interested in.
Descriptive statistics will give us averages through which we can com-
pare typical uses of the features in question. However, this only gives us
an impressionistic view of the difference for our dataset. If we rely solely
on descriptive statistics, we cannot tell how generalizable those differ-
ences may be. To be able to generalize about the ‘typicality’ of patterns of
use, we need to use other statistical procedures.
Generalizability means that the results in our sample can be predicted to
be true, with a high level of certainty, to samples outside of our own dataset
as well. That is, if we were to conduct a new study under the same condi-
tions we are reporting on, we could be 95% or 99% certain to get the same
results. In order to have generalizable results, we need to make sure that a
set of assumptions about our data is met (see later in this chapter).
6.2 Basic Terms, Concepts, and Assumptions

In this section, we outline the basic terms and concepts that are used when
doing any kind of statistical analysis. First, we discuss variable types and
levels that are critical to know before any test can be done. Second, we
Basic Statistics 95
introduce measures of central tendency (“typicality” in a dataset) and

measures of variability or dispersion (“spread”).
6.2.1 Variables and Observations

Variables are typically classified based on a) the type of variable and how
they function in the research design, and b) the range of values and levels
they can have. It is crucial to think about this ahead of time is because
the validity and reliability of our research depends on how we define our
variables and observations. Also, the variable scale and type determines
the types of statistical analyses that can be done.
Variable Types and Functions

“Regular variables” are variables that you either manipulate or want to
see change in your design. They can have a range of values (numeric) or
levels (non-numeric). Numeric values are relevant if you have frequencies
of a linguistic feature (e.g., the number of personal pronouns in a text).
Non-numeric values (or levels) are relevant when you refer to a variable
in terms of categories, e.g., class size (‘small’ or ‘large’). Instead of calling
them small or large, you could also give numbers to these categories (e.g.,
small = 1 and large = 2); however, they are not values, just numeric codes.
That is, there is nothing inherently primary or secondary in the number
they are assigned to.
Further classification of regular variables is based on the function they
have in the design. We distinguish between two types: dependent vari-
ables and independent variables. Dependent variables are the variables
that you are most interested in for your research because you think that
the values of the variable (e.g., frequency) will change (or not) as you are
manipulating some external factors around it. The change that will occur
(or not) depends on your manipulation of other variables around it. Inde-
pendent variables are the variables that you manipulate in order to see
whether there is change in the dependent variable. Dependent variables
are often called “outcomes” and independent variables are often called
“predictors” (of change).
Example
You read in an educational journal article that lectures in small classes
are more ‘personable’ than in large classes. As there is no linguistic evi-
dence provided in the article for this claim, you want to find out yourself.
You decide that you will use first person pronouns (I, we) as a measure
of ‘personable.’ You are interested in whether the frequency of first per-
son pronouns (I, we—and all their variants) changes at all when you
96 Your Own Corpus
are attending a lecture in a large class with 200 students or in a small,

seminar-like class with 20 students. You hope to see that the frequency
of first person pronouns will change depending on which class you will
attend. That is, the use of first person pronouns will depend on the class
size (lecture versus seminar-type). The dependent variable in this design
is first person pronouns (the frequency of which will change) and the
independent variable is class size (the one that you manipulate to see the
change). So, your predictor for change in the “outcome” (pronoun use)
is class size.
“Moderator variables” are referred to as other “predictors” or other
independent variables in your design (if you have more than one). Mod-
erator variables are viewed as an independent variable potentially inter-
acting with other independent variables. In our example, let’s say you
want to see whether the instructor’s gender also has an effect on the use
of first person pronouns, in small or large classes. Your independent vari-
able is class size and the moderator (or other independent variable) is
gender. In this design, you may be interested in whether it really is class
size alone, or gender alone, or the two independent variables together
(class size moderated by gender) that will cause the changes in the use of
first person pronouns.
“Control variables” are not real variables in the way we have been
describing variables so far. You are not able to measure a control variable
in your study; instead, it is just something you control for.
“Intervening variables” are variables that you should have measured
in your study but you realize later that you didn’t. Typically, these are
the variables that are mentioned in the discussion section of an article or
report when calling for further research.
Variable Scales
“Nominal scales” (also called categorical, discrete, discontinuous scales)
are variables measuring categories. They are used in naming and cat-
egorizing data in a variable, usually in the form of identity groups, or
memberships. The variable could occur naturally (e.g., sex, nationality)
or artificially (experimental, control groups) or any other way, but in all
cases, it is a limited number of categories. They represent non-numeric
categories (e.g., religion, L1, ethnicity). When they are assigned to num-
bers, they carry no numeric value. Instead, they are only a category iden-
tifier (e.g., there are two sexes: 1 = male, and 2 = female).
“Ordinal scales” are used to order or rank data. There is no fixed inter-
val, or numeric relationships in the data other than one is “greater than”
or “lesser than” the other. No fixed interval means that we don’t know
whether the difference between 1 and 2 is the same as between 4 and
Basic Statistics 97
5 (i.e., no fixed interval as is the case for interval scales). Examples of

ordinal scales are holistic scoring, Likert scales, questionnaires. They are
numeric in that the numbers represent one being more—or less—than the
other, but they do not say how much more.
“Interval scales” reflect the interval or distance between points of rank-
ing. They are numeric, continuous scales, and are the same as ordinal but
with fixed intervals. That is, while with ordinal scales we do not know
whether the difference between 2 and 3 is the same as between 4 and 5,
with interval scales we do. For example, the difference between 18 and 19
milliseconds is the same as between 22 and 23—that is, one millisecond.
The difference between 2 and 3 meters is the same as between 4 and 5
meters—that is, one meter, 100 centimeters, 1000 millimeters (no matter
how we measure it, the difference is exactly the same). This means that
we always know how much more or less distance there is between the
two measures. Sometimes, frequencies, test grades, or evaluation are con-
sidered interval variables; however, it is not really fixed. The best way to
deal with frequencies, for instance, is to put them under a scale, at which
point they become interval scores. We can do this by norming1 frequency
counts, for example, or by calculating percentages.
“Ratio” only tells us the relationship between two measures. It is not a
very good measure for register studies. Let’s say we want to compare two
texts to see which one has more nouns.
Text 1: noun/verb ratio = .27

Text 2: noun/verb ratio = .32
We are unable to tell which text has more nouns because it is only
in relation to the verbs that we might have more nouns. That is, ratios
measure how common one thing is but only in relation to a potentially
unrelated other thing.
Variable Values (Levels)

Variables can have multiple values or levels. For example, if participant
age is a variable, the (numerical) values can be counted between 0 and
120. If ethnicity is a variable, we can list what ethnicities we would want
to include, and give each a nominal value. For example, African Ameri-
can = 1, Native American = 2, Asian American = 3, etc.
Observations
Observations are individual objects that you are characterizing. They
provide the unit of analysis that will make up your data. For register
98 Your Own Corpus
studies, an observation is typically each text that you enter into your
database. For other linguistic studies, it could be the individual linguistic
feature you are considering.
Example
Let’s assume you are interested in how complement clauses are used by
younger and older generations and also how they are used by people with
different educational backgrounds. You are using a corpus to look for pat-
terns. Take 100 instances of complement clauses and mark each for who
uses them in terms of age and educational background (hopefully, this
information will be available in the corpus). You can use other contextual
variables as well, but the focus should be on the two variables you iden-
tified. Instead of listing your findings in a table exemplified by Table 6.1
below, you list the individual cases in a table exemplified in Table 6.2 below.
Table 6.2 is preferable because, given the way the information is pre-
sented in Table 6.1, we are unable to distinguish between the two vari-
ables. That is, we cannot measure one independent variable from another.
A research is confounded when the variables measure the same thing;
that is, based on the data in Table 6.1 above, we can’t say whether the
frequency of complements is due to age or level of education.
6.2.2 Measures of Central Tendency and

Measures of Variability
Central tendency describes typical for the values for a variable; that is, it
is the central point in the distribution of values in the data. Dispersion,
on the other hand, is how much variation you get within your data. Both
of these are important measures to see patterns.
Table 6.1 Frequency of complement clauses
18–22 22–30 31–40 41+
Undergrad 50 10 2 2
MA 5 20 10 10
PhD 0 3 20 3
Table 6.2 Complement clauses
Clause Complement type Age Education
I told her that he likes me Verb 20 BA

I like the idea that we go by boat Noun 30 MA
The claim that he was happy turned Noun 30 PhD
out to be untrue
Basic Statistics 99
Measures of Central Tendency

Measures of central tendency tell us the most typical score for a dataset.
There are three types: mode, median, and mean.
“Mode” works for any variable scale (nominal, ordinal, or interval). It
is the most frequent/common value (whatever value occurs with highest
frequency) in your dataset. If you draw a frequency polygon, it will show
the most frequently occurring point best. For example, in the following
dataset, what would be the mode?
2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 5
Yes, the mode is 3, because that is the score that occurs most frequently
(six times), versus 2 (twice), 4 (twice), and 5 (once).
There are, however, problems with using mode as a measure of central
tendency, namely:
• If there is not one most frequent score (but more than one—for
instance, two, just like 2 and 4 above occur with the same frequency,
so if those two were the most frequent scores, we could not tell what
the mode is), there is no mode.
• If each ranked score in the dataset only occurs once, i.e., no score
receives the frequency of higher than one (i.e., every score in the data-
set occurs just once), there is no mode.
• The mode is too sensitive to chance scores (when a mistake is made
in entering the scores).
“Median” works for any numeric variable (ordinal or interval). It is

the 50th percentile (i.e., the middle observation). To calculate the median,
rank order all scores and the observation in the middle is the median. If
you have an even number of scores, the median will be in between the
two middle scores; if you have an odd number of scores, the median is
the middle score. The quartiles are located as well in the ranking, and the
number of observations that go with each score with the same number of
observations from both sides. Let’s say we have the average scores for the
use of hedges in our corpus of nine texts.
17 18 | 19 20 20 20 21 | 22 23
25th 50th 75th
The quartile gives us distributional patterns in the data. In this example,

the 25th percentile means that 25% of the texts display a score of 18.5 or
less; the 50th percentile means that half of the texts display a score of 20
or more and half of the texts display a score of 20 or less, and finally, the
75th percentile means that 75% of the texts display a score of 21.5 or less.
100 Your Own Corpus
Median is often used as a measure of central tendency when:
• the number of scores is relatively small

• the data have been obtained by rank order measurement (e.g., a
Likert scale)
• the mean is not appropriate (because the variable is not interval—see
below)
Boxplots are typically used as visuals to show the range of scores (mini-
mum and maximum), the 25th, the 50th (median), and the 75th percen-
tile. A boxplot is also able to show outliers in the dataset. In the example
below, we display the use of nouns by teachers and students in the corpus.
25th percentile (on box plot: lowest line where the box starts)
50th percentile (median—on box plot: thick black line in box)
75th percentile (on box plot: highest line where the box finishes)
The range, the percentile figures, and the outlier are in Figure 6.1. Box-
plots do not tell you the individual scores.
“Mean” = (x bar) only works for interval scale. Add up all the values
(x) and divide by the # of cases or observations (N) to get the arithmetic
average of all scores.
X =
∑x
N
X (X bar) = sum of x divided by N
Outlier (observaon # 159)
Maximum score
75th percenle
50th percenle
25th percenle
Minimum score
Figure 6.1 Boxplot of the use of nouns by teachers and students

Basic Statistics 101
While the mean is the best measure of central tendency for interval
scores, it is at times problematic because it is too sensitive to extreme
scores. If extreme scores enter the equation, it throws the mean off so
much that it cannot be the measure of tendency any more.
Example
Let’s say you are interested in finding out whether undergraduate stu-
dents majoring in natural sciences use fewer ‘hedges’ than students in
the humanities or in the social sciences. You look at a corpus of student
presentations that includes nine presentations from each area. The data
below shows the mean scores for each of the presenter in each of the three
areas. It illustrates how much one score in the dataset can change the
value of the mean, which is the measure of central tendency.
In group A, you have the mean scores for social science students.
A: 0 1 2 20 20 20 97 98 99 (357/9)
Group A results Mode: 20; Median: 20; Mean: 39.6
In group B, the mean scores are listed for natural sciences students.
B: 17 18 19 20 20 20 21 22 23
Group B results Mode: 20; Median: 20; Mean: 20 (180/9)
Group C contains is the mean scores for humanities students.
C: 17 18 19 20 20 20 21 22 99 (256/9)
Group C results Mode: 20; Median: 20; Mean: 28.4
As you can see, the mode (most frequent scores) and median (the cen-
tral score after rank ordering all scores) are the same for all three groups.
However, the mean becomes vastly different depending on the actual
scores in the dataset. In Group A, the scores vary a great deal. Social sci-
ences students use hedges in an idiosyncratic way; that is, it really depends
on the individual. Some students use none or very few, and some use a
lot! When this is true, the mean is relatively high (especially in compari-
son with the others). In Group B, the scores are not going into extremes.
Instead, they are pretty evenly distributed. That is, the students in this
group more or less use hedges the same way, or at least very similarly. In
Group C, students overall use hedges similarly but there is one student
102 Your Own Corpus
who hedges a lot. That one student changes the mean score dramatically.
The two median scores in group B and C are very different (20 in Group
B and 28.4 in Group C) and that difference is due to only one score. In
the mean score, one outlier makes a big difference.
The characteristic of normal distribution is that the mode, the median,
and the mean are identical. Group B above has that example. And if you
look at the variability of the scores, you can see that it is steadily in order.
There are no outliers or extreme scores in the dataset. The scores are
simply normally distributed.
Measures of Variability and Dispersion

Measures of variability and dispersion only work with interval scale type
data. While range looks at the scores at each end of the distribution, vari-
ance and standard deviation measures look at the distance of every score
from the mean and average them. More specifically, range only takes the
highest and the lowest scores into the computation, and variance and
standard deviation take each score into account.
“Range” tells us the spread of scores. We compute the range by sub-
tracting the lowest score from the highest score.
Range = x highest − x lowest
For example, the range of scores for Group A in the example above is
99, for Group B it is 6, and for Group C it is 82. The problem here is the
same as with the mean scores, as it changes drastically when you have
more extreme scores (as you see in the examples). Since it is unstable,
it is rarely used for statistical reporting, but range could be informative
as a piece of additional information (e.g., to see whether there is an
outlier).
“Quartile” (interquartile) or percentile measures tell us how the scores
are spread in different intervals in the dataset. As outlined above, the
median is a measure of central tendency, and by adding the interquartile
figures (the percentile figures), we are able to see the spread as well. Once
again, the 25th percentile tells us what scores we would get for a quarter
of our data, the 50th percentile tells us what score we would get for half
of the data and the 75th percentile refers to the score we would get for
three-quarters of the data.
“Variance” summarizes the distance (i.e., how far) individual scores
are from the mean. Let’s say our mean is 93.5 (X = 93.5) . If we have
a score of 89 (x = 89), that means our score is 4.5 points away from
the mean, and that is the deviation (value) away from the mean. In this
instance, we just discussed one score only. However, what we want is a
measure that takes the distribution and deviation of all scores in the data-
set into account. This is the variance.
To compute variance, take the deviation of the individual scores from
the mean, square each deviation, add them up (oftentimes called the ‘sum
of squares’) and average them for the dataset dividing it by the number of
observations minus one. As a formula, it looks like this:
∑( x
2
� individual
� score − � X � group� mean )
variance = .
N −1
“Standard deviation” is a measure of variability in the data from the

point of central tendency. Standard deviation tells us the variability of
the scores—i.e., the spread of the scores from the central point—and is
most often used as a measure of dispersion in studies of a variety of fields,
including corpus linguistic studies. But why is this important? Let’s take
an example that illustrates why it is important to know the spread. With
another example, from Csomay (2002), we can illustrate how the mean
score is not enough to know about how the distribution of scores goes
while the dispersion measure would tell us the answer to that question
(see example below for details).
Example
Imagine you would like to find out whether one class is more interactive
than another. As Csomay (2002) did, you define interactivity by the num-
ber of turns taken in a class and by how long those turns are (in terms of
number of words). You look at two lecture segments to compare. Each
has 5 turns, and for simplicity, each segment has a total of 150 words in
them. Here are the numbers:
Lecture #1: 5 turns, a total of 150 words, average turn length 30

words, each turn is of equal length.
Turn 1: 30 words
Turn 2: 30 words
Turn 3: 30 words
Turn 4: 30 words
Turn 5: 30 words
Total = 5 turns, 150 words
Average turn length: 30
X = 30
Lecture #2: 5 turns, a total of 150 words, average turn length 30
words, turn length varies for each turn.
104 Your Own Corpus
Turn 1: 2 words
Turn 2: 140 words
Turn 3: 2 words
Turn 4: 3 words
Turn 5: 3 words
Total = 5 turns, 150 words
Average turn length: 30
X = 30
In both instances, the average turn length is 30 words, which is the mea-
sure of central tendency. But it is clear that one lecture is very differ-
ent from another in terms of turn length measures. By calculating the
standard deviation for each, we are able to tell the spread in the scores;
that is, whether the scores are close to each other or they vary, and if the
latter, how much they vary (in terms of magnitude measured by a single
number). For Lecture #1, the standard deviation is 0, and for Lecture #2,
it is 61.49. A zero standard deviation says that there is no variation in the
scores at all (clearly), and 61.49, being very high, tells us that there is a
great variation in the scores.
Does this tell us which lecture is more interactive? If we think that rela-
tively shorter turns are making the class more interactive, then Lecture
#1 is more interactive. If we think that longer stretches of turns coupled
with two- or three-word turns is more interactive, then Lecture #2 it is.
Lecture #1 looks to be the best candidate simply because the number of
turns and the turn length measure together tell us that people would have
more opportunity to express actual ideas rather than just agree to what is
happening with one or two words at a time (see Csomay 2012 for short
turn content).
In sum, the larger the standard deviation, the wider the range of dis-
tribution is away from the measure of central tendency. The smaller the
standard deviation, the more similar the scores are, and the more tightly
the values are clustered around the mean.
To calculate the standard deviation, all you need to do is to square root
the variance (explained above).
standard deviation = variance .
that is,
∑ (xindividual score − X group� mean )2

s.d. =
N−1
For the scores above in Lecture #2, it would look like this:
2 – 30 = –28 squared: 784

140 – 30 = 110 squared: 12,100
2 – 30 = –28 squared: 784
3 – 30= –27 squared: 729
3 – 30= –27 squared: 729
Sum: 15,126 divided by N – 1 (5 – 1 = 4) = 3,781.5 square root: 61.49

With this, we are able to see that two groups could be very similar in
terms of their means but they can be very different because the distribu-
tion of scores away from the mean may be quite different.
6.2.3 Parametric and Non-Parametric Tests, Research

Questions, and Hypotheses
Parametric and Non-Parametric Tests

Non-parametric tests do not require strong assumptions about the distri-
bution of the data. The observations can be frequencies (nominal scores)
or ordinal scales, and can be rank ordered. They can be used with interval
scales, too, when we are unable to meet the assumptions of parametric tests
(e.g., normal distribution in the data). Non-parametric test results can only
be interpreted in relation to the dataset in question. That is, no projections
or predictions could be made to the population it was drawn from, and the
interpretation can only relate to the dataset investigated. A non-parametric
test, for example, is Chi-square (see details on this in Chapter 7).
Parametric tests, however, do require strong assumptions about the
nature and the distribution of the data. These assumptions are:
1. Dependent variables are interval scales (where means and standard

deviations are the measures of central tendency and dispersion,
respectively) and not frequencies or ordinal data.
2. Dependent variables are strongly continuous (rather than discrete as
ordinal scores are). That is, we know exactly how much difference
there is between two scores and they are always at the same distance.
3. We can estimate the distribution in the population from which the
respective samples are taken. That is, the distribution in the ‘sample’
could be projected to the distribution of the ‘population.’ A small
sample size will make it problematic to do this—a minimum of 30
observations for each variable is needed.
4. Data are normally distributed (sometimes we use fewer than 30
observations—remember that is the minimum to assume normality
in the distribution—the larger the size, the better, of course).
106 Your Own Corpus
5. Observations are independent; otherwise, research is confounded, as

discussed before—that is, there is no relationship between the obser-
vations, or cases.
Why do parametric tests? The reason parametric tests are more powerful
than non-parametric tests is because a) they have predictive power (i.e.,
we can predict that if we followed the same procedures, and did the study
the same way, we will gain the same results) and therefore, b) the results
are generalizable (i.e., we can generalize that the results are true to the
larger population the samples are drawn from—that is, if we repeat the
study, we would get the same results). Therefore, they are very powerful!
Research Questions and Hypotheses

According to Hatch and Lazaraton (1991), the characteristics of research
questions are:
(a) Keywords to define the area

(b) Element of new information
(c) Manageable data
(d) Question that is specific enough to make the study focused
Research questions have three components:
(a) What group are you dealing with? What collection of things (not
only people)?
(b) What happens? What’ are the outcomes? What result are you look-
ing for?
(c) What influences the results?
Hypotheses, on the other hand, are basically the formalization of research

questions into strict statements that could be rejected (or not):
(a) They are phrased in the form of statements (rather than questions).
(b) Their statements show specific outcomes.
(c) They need to be testable.
In other words, a “hypothesis is a statement of possible outcome of

research” (Hatch and Lazaraton 1991: 24). We have a null hypothesis and
alternative hypotheses. The null hypothesis (H0) is a statement—usually a
negative statement, and the alternative hypothesis is the opposite, a positive
statement. Additionally, the alternative hypothesis then could be in a direc-
tional form, taking a (+/-) direction. Our aim is to reject the null hypothesis.
Typically, we are looking for either differences between two or more

groups or we are looking for relationships between two groups (see
Chapter 8 for further explanation).
In looking for differences, our null hypothesis will be stating that there
is no difference between two or more groups (independent variables) with
respect to some measure (dependent variable). (These are typically para-
metric tests.)
For example, we may have the followong n [ull hypothesis;]
H0 There is no difference in the use of nouns across disciplines.
The alternative hypothesis would be:
H1 There is a difference in the use of nouns across disciplines. In

looking for relationships between two or more variables our null
hypothesis will be stating that there is no relationship between
two or more measures. (Typically this is used for non-parametric
tests. As a concrete example, see the null hypothesis below, fol-
lowed by the alternative hypotheses.)
H0 There is no relationship between the use of nouns and first person

pronouns in university classroom talk.
Alternative hypotheses:
H1 There is a relationship between the use of nouns and first person

pronouns in university classroom talk.
H2 There is a positive relationship between the use of nouns and
first person pronouns in university classroom talk. (That is, when
nouns occur, first person pronouns will as well.)
H3 There is a negative relationship between the use of nouns and
first person pronouns in university classroom talk. (That is, when
nouns occur, first person pronouns will not occur.)
We look to reject the null hypothesis of ‘no difference’ or ‘no relation-

ship.’ A p< .05 (probability of 5%) means that we have a 95% chance
of being right in doing so. A p< .01 means that we have 99% chance of
being right in doing so, and a p < .001 means that we have 99.9% chance
of being right in rejecting the null hypothesis. There are two types of
errors that we can commit in rejecting the hypothesis: Type 1 and Type 2.
See their description below.
108 Your Own Corpus
When we reject the null hypothesis, we want the probability (p) to

be very low that we are wrong. If, on the other hand, we must accept
the null hypothesis, we still want the probability to be very low that
we are wrong in doing so.
(Hatch and Lazaraton 1991: 224)
Type 1 error: The researcher rejects the hypothesis when it should not
have been rejected.
Type 2 error: The researcher accepts the null hypothesis when it
should have been rejected.
The probability value (alpha, or p) basically tells you how certain you can
be that you are not committing a Type 1 error. When the probability level
is very low (p< .001), we can feel confident that we are not committing
a Type 1 error described above, and that our sample group of students
differs from other groups who may have taken the test in the past or who
might take it in the future (population). We test whether the data from
that sample ‘fit’ with that of the population. A p< .05 tells us that there
are fewer than 5 chances in 100 that we are wrong in rejecting the H0.
That is, we can have confidence in rejecting the H0.
Two-Tailed Test/Hypothesis
In two-tailed tests, we specify no direction for the null hypothesis ahead
of time (that is, whether our scores will be higher or lower than more
typical scores). We just say that they will not be different (and then reject
that if significant). (See first example above.)
One-Tailed Test/Hypothesis
We have a good reason to believe that we will find a difference between
the means based on previous findings. The one-tailed tests will specify the
direction of the predicted difference. In a positive directional hypothesis,
we expect the group to perform better than the population. (See second
example above.) In a negative directional hypothesis, the sample group
will perform worse than the population.
One crucial remark: We cannot repeat tests as often as we may want
to. The statistical tests that we introduce in this book are not exploratory
statistics, but they are experimental design, testing hypotheses. One-time
deal only. Steps for hypothesis testing:
Step 1: State null hypothesis.

Step 2: Decide whether to test it as one- or two-tailed hypothesis.
Question: Is there research evidence on the issue?

a. NO: Select two-tailed  will allow rejection of null hypoth-
esis in favor of an alternative hypothesis.
b. YES: Select one-tailed  will allow rejection of null hypoth-
esis in favor of directional.
Step 3: Set the probability level (typically p < .05 or lower). Justify your
choice based on the literature.
Step 4: Select appropriate statistical test.
Step 5: Collect data—apply statistical test.
Step 6: Report the results and interpret them correctly.

Several statistical programs are commercially available and are poten-
tially cheaper at a student price, and there are others that are free. While
SAS, STATA and R, for example, are powerful statistical software pro-
grams, we will be showing you how to do descriptive statistics and the
basic statistical methods we outlined above with SPSS (Statistical Package
for the Social Sciences). We consider this program the most user friendly,
as the other three mentioned above require some programming abilities.
Also, SPSS is still the most frequently used program at university cam-
puses and is typically available for the students at computer labs free of
charge through a university license. In this section, we will show you how
to organize your data in SPSS (very different from Excel!) and how to
access descriptive statistical results.
6.3.1 Preparing the Data and Descriptive Statistics

In SPSS, the way we organize the data is very different from the way data
could be entered in Excel. Therefore, we would like you to completely
forget Excel while you are using SPSS. As a start, there are two views you
can have in SPSS: the variable view and the data view. Before we explain
each view a bit more in detail, let’s review one more time the dependent
versus independent variables and what the basic unit of analysis is (obser-
vations) in the example we use.
When we characterize registers based on one or more linguistic fea-
tures, the unit of analysis is a text. That means that each observation is
a text in which we look for the particular variable that we hope to see
variation in (i.e., the dependent variable). In our examples, it has been
an individual linguistic feature, such as nouns, or pronouns, etc. Each
text then will have other, ‘extra-textual’ features as well. An extra-textual
feature, for example, is what register it comes from—that is, whether it is
news, or face-to-face conversation, etc. Another extra-textual feature can
110 Your Own Corpus
be the time the text was produced: whether the text comes from the year
1920 or 2015. These are your independent variables, and depending on
your research question, you will manipulate these to see if there is varia-
tion in the dependent variable.
When we characterize individual speakers’ way of using certain lan-
guage features, the unit of analysis is the text produced by those speakers.
The unit of analysis is still the text (because the language was produced
and transcribed), but it may not be obviously understood in the same
sense as the text above because each text is more associated with indi-
vidual speakers who would have certain characteristics. Yet, it is the text
produced by them, and that will be the basis for comparison.
Finally, when we look at characteristics of individual linguistic fea-
tures (e.g., article type position in subject and object positions), our unit
of analysis is each instance of that feature. Then, we characterize each
observation for its features, which in this case would be syntactic position
and type of article.
Preparing the Data: Entering Data into SPSS

We will now show you the basics in each view, and tell you how to
organize your data in these two settings. Let’s start with the variable
view (see the highlighted tab at the bottom left hand corner). Here, you
will enter the names and characteristics of both of your dependent and
independent variables. Let’s take the example we discussed earlier in this
chapter when explaining mean. Here’s the text again: You are interested
in finding out whether undergraduate students majoring in natural sci-
ences use fewer ‘hedges’ than students in the humanities or in the social
sciences. You look at a corpus of student presentations that includes 9
presentations from each area (a total of 27 texts). The data below shows
the normed scores for ‘hedges’ for each of the presenters in each of the
three areas.
Social science presentations: 0 1 2 20 20 20 97 98 99

Natural sciences presentations: 17 18 19 20 20 20 21 22 23
Humanities presentations: 17 18 19 20 20 20 21 22 99
Your dependent variable is ‘hedges’ (interval scores, as it is normed) and

your independent variable is discipline (nominal) with three levels (the
three disciplinary areas). Because SPSS is not good at processing ‘string’
data, i.e., text, for its variables, we need to give a nominal numeric value
to each discipline. We name Social Sciences 1, Natural Sciences 2, and
Humanities 3. There is no numeric value across these categories.
Your SPSS Variable view will look like Figure 6.2.
We need to enter two variables in this view: one will be the dependent
variable and the other will be the independent one. Remember, all vari-
ables and their characteristics will need to be entered here. The name of
the variables will be ‘hedge’ and ‘discipline’, repectively. So let’s enter
those and see what characteristics each has. (See Figure 6.3.)
We need to focus on some of these headings, but not all. For example,
those that seem less important are ‘Width,’ which determines how wide
the cell is in your data view, and ‘Columns,’ which determines how many
Figure 6.2 Variable view in SPSS
Figure 6.3 Important tabs in Variable view in SPSS

112 Your Own Corpus
columns there are. ‘Align’ is also less important as it sets how you would
like to see the text aligned in the data view (to the left, the middle, or to
the right), and ‘Role’ is what role you assign this variable in the dataset
(it will all be input for us). We really do not need to worry much about
these tabs. However, we do need to know more about all the others:
‘Type,’ ‘Decimals,’ ‘Label,’ ‘Values,’ ‘Missing,’ and ‘Measure.’ We will go
through each of these one by one:
Type: Numeric (whether it has a numeric value or not—see nominal
independent variables above).
Decimals: You can set the number of decimals you want to see. For
interval scores, we typically use two decimal points and for nominal
scores, we use 0 decimal points (since they have no numeric value, they
do not and will not have any fractions).
Label: SPSS takes very short names for variable names and only in one-
word strings. Labels, then, provide you with the opportunity to give lon-
ger names that could be used as the labels for your output results as well.
Values: These are the values that you can assign to the levels. For
hedges, we will not have any values assigned. But for the nominal vari-
ables, as we mentioned above, we have 1 = Social Sciences, 2 = Natural
Sciences, and 3 = Humanities. As you enter each one, make sure you hit
the ‘Add,’ or else it will not be added to the list. (See Figure 6.4.)
Missing: It is good not to have any missing data because that will affect
the calculations and the results or we need to set the software program in
sophisticated ways not to do so.
Measure: In this area, you will need to determine what kind of variable
you have. In our example, since hedges are interval variables, we will
choose ‘Scale,’ and since discipline is a nominal variable, we will choose
‘Nominal.’ (See Figure 6.5.)
Figure 6.4 Value labels in SPSS

Figure 6.5 Measures in SPSS
Figure 6.6 Adding variables in SPSS
Before we turn to our Data View, let’s add one more variable, so we can
keep track of our observations. The filenames will be portrayed as a string
variable called ‘text_number’ (we really are not including this as a variable
in any calculations; it is more like a reference for us to know which text file
the data is coming from). So it will be string, and it will be a nominal type
of data (all strings are nominal). (See Figure 6.6.)
Now that this is all set, let’s turn to our Data view to see how the data
will need to be entered. First, as we see in Figure 6.7, we have both vari-
ables and each in a column. In SPSS, all variables are in columns, and
each observation is a different row. So in our case, the dependent and
independent variables are in the columns, and each text (each represent-
ing a presentation) will be in a different row.
114 Your Own Corpus
Figure 6.7 Data view in SPSS
Now we can start running some descriptive statistics.
Descriptive Statistics
In order to get information in a stratified manner for your levels, you
want to give the following command. On the top bar with ‘File,’ ‘Edit,’
‘View,’ choose the following set: Analyze → Descriptive Statistics →
Explore to get to the window shown in Figure 6.8.
Following our case study, as you see, your dependent variable (hedges)
needs to be under ‘Dependent List’ and your independent variable
(discipline) needs to be under ‘Factor List.’ This way, your descriptive
statistics will be calculated for each level (i.e., for each of your disciplines)
Figure 6.8 Descriptive statistics through ‘Explore’ in SPSS
Table 6.3 Descriptive statistics for ‘hedges’ in three disciplines through SPSS
Descriptives
Discipline Statistic Std. Error

Hedge Social Sciences Mean 39.66 14.84
Median 20.00
Std. Deviation 44.52
Minimum .00
Maximum 99.00
Range 99.00
Natural Sciences Mean 20.00 .62
Median 20.00
Std. Deviation 1.87
Minimum 17.00
Maximum 23.00
Range 6.00
Humanities Mean 28.44 8.83
Median 20.00
Std. Deviation 26.50
Minimum 17.00
Maximum 99.00
Range 82.00
116 Your Own Corpus
versus giving just one mean score of the entire dataset you have. Run the
statistics, and see to what extent the results by SPSS match the descrip-
tive statistics we calculated earlier (they really should!). If you only want
the numbers, click on ‘statistics’ if you want a boxplot (described in this
chapter) and the numbers, click on both. Explore what option you may
have further by clicking on the ‘Options’ button at the upper right-hand
side.
If you only wanted the numbers, it should look like the details in
Table 6.3 listing all the necessary descriptive statistics for each of the
disciplinary areas.
We believe the numbers generated by SPSS match the hand calculations
we made in this chapter. In the next chapter, we will get serious and look
at four different statistical tests that we can apply to our datasets.
Note
1 To calculate the normed counts, we typically take the frequency of the linguis-
tic feature itself, divide it by the total number of words in the given text and
multiply it with a number (typically 1,000) for each observation. That is, let’s
say Text 1 has 45 first person pronouns, and it is 1,553 words long. Then we
will calculate the normed count to 1,000 words (as if the text were that long)
by (45/1,553)*1,000 and we get 28.97. This way, we translated a nominal
score into an interval score. Also, if texts have different lengths, this way we
can actually compare the numbers. If Text 2 has 125 of the feature and the
text is 3,552 words long, then the normed count will be (125/3,552)*1,000,
or exactly 35.19.
References
Csomay, E. (2002) ‘Variation in academic lectures: Interactivity and level of
instruction’, in Reppen, R. et al. (eds) 2002: 203–224
Csomay, E. (2012) ‘A corpus-based look at short turns in university classroom
interaction’, in Csomay, E. (ed) 2012: 103–128
Csomay, E. (ed) (2012) Contemporary Perspectives on Discourse and Corpora.
Special issue of Corpus Linguistics and Linguistic Theory. 8/1
Hatch, E. & A. Lazaraton (1991) The Research Manual. Design and Statistics for
Applied Linguistics. Boston: Heinle & Heinle
Reppen, R., Fitzmaurice, S., & D. Biber (2002) Using Corpora to Explore Lin-
guistic Variation. Amsterdam: Benjamins
Chapter 7
Statistical Tests (ANOVAs,

Chi-square, Pearson
Correlation)
7.1 Difference Tests

7.2 Relationship Tests
7.1 Difference Tests

When doing difference tests (e.g., One-way ANOVA or Two-way ANOVA),
we test whether there is statistically significant difference in the average,
i.e., mean scores between two or more variables. The goal of difference
tests is to see the extent to which the independent variable(s) is/are respon-
sible for the variability in the dependent variable. That is, we are inter-
ested in how one or more variables affect another variable. We can make
claims about cause and effect, i.e., one variable changes because another
variable has a profound (statistically significant) effect on it. We cannot
talk about the results in terms of more or less significant, however. Once
the results are statistically significant, you can investigate where the dif-
ferences are with post-hoc tests and how strong the association is between
the dependent and independent variable with R2.
7.1.1 One-Way ANOVA (Analysis of Variance)

We can do a One-way ANOVA test when we have one dependent vari-
able and one independent variable, the latter with more than two levels
(see example below). The dependent variable has to be an interval score,
and the independent variable has to be nominal. With parametric tests,
like One-way ANOVA, we can generalize to the population that the
sample was drawn from. Conceptually, with a One-way ANOVA we
are interested in identifying the change in the dependent variable. More
specifically, whether the variability in the dependent variable is due to
the variability of the scores within each level of the independent variable
or across the level groups we are comparing. That is, One-way ANOVA
assesses whether the differences in mean scores are attributed to the
118 Your Own Corpus
variability within the groups or across the groups. If the ratio of these
two is small—that is, if the “across” group variation is small relative
to the “within” group variation—there is no statistical difference. If,
however, the “across” group variation is large relative to the “within”
group variation, there is a statistically significant difference across the
groups. That is, the larger this ratio between the “within” group vari-
ability measures and the “across” group variability measures (F score),
the more likely that the difference between the means across the groups
is significant.
Assumptions and requirements with ANOVA:
1. We have one dependent and one independent variable with three or

more levels.
2. Dependent variable must be reported in interval scores (e.g., normed
counts for linguistic features) and must be continuous, and the inde-
pendent variable is nominal.
3. Measured variables must be independent (not repeated).
4. Normal distribution of scores is expected in each group.
5. Number of observations is equal in each group (a balanced design),
although it is only necessary when we do calculations by hand. The
statistical package (e.g., SPSS1) accounts for an imbalance.
6. Values/categories on independent and dependent variables must be
mutually exclusive and exhaustive.
7. Cell values cannot be too small. A minimum of five observations per
cell is necessary.
Example
As an example, let us say you are investigating university classrooms
as your context. You analyze your context for all the situational vari-
ables outlined in Chapter 2, and you realize that discipline may a situ-
ational variables in the academic context that may have an effect on
how language is used in the classrooms. In fact, you have read earlier
that the use of pronouns may vary depending on the discipline. Based
on your readings, you also know that first person pronouns are more
apparent in spoken discourse, and have been associated with situations
where the discourse is produced under more involved kind of produc-
tion circumstances (e.g., where the participants share the same physical
space, allowing for the potential of immediate involvement in interac-
tion). Knowing all of this, you isolate this one pronoun type because you
are interested in the use of first person pronouns (I, me, we, us). More
specifically, you would like to find out whether there is a significant dif-
ference in the use of first person pronouns in different disciplines (more
than two).
Statistical Tests 119
You formulate your research question in one of two ways:
1. How does the use of first person pronouns differ across disciplines?
OR
2. Is there a difference in first person pronoun use across disciplines?
The dependent variable is the composite normed score for the first person
pronouns as listed above (use normed counts—an interval score), and
the one independent variable is discipline with three levels (nominal vari-
able). The three levels are the three disciplinary areas: Business, Humani-
ties, and Natural Sciences.
You formulate your hypothesis:
H0: There is no significant difference in the use of first person pro-

noun use across the three disciplines.
H1: There is a significant difference in the use of first person pronoun
use across the three disciplines.
The statistical test to use is One-way ANOVA (one dependent variable

with interval scores and one independent variable with nominal scores
and with multiple levels, i.e., in this case, three disciplines). The signifi-
cance level is set at .05 level, and to locate the differences, in case the
ANOVA results in a significant difference, we will use Scheffe post-hoc
test.
To illustrate how the statistical program calculates the F score, we
will do a step-by-step demonstration. Table 7.1 below shows the fictive
normed counts for the three disciplines. Once again, as we saw in Chap-
ter 6 (section 6.3.1), these numbers are entered into SPSS, as illustrated
in Table 7.1 below. Each observation (i.e., each text with each normed
count) will be in a different row. The two variables are: First person pro-
noun use in each text normed to 1,000 words (interval variable) and Dis-
cipline (nominal, three levels: 1 = Business; 2 = Humanities; 3 = Natural
Sciences).
Table 7.1 First person pronoun data in SPSS data view format
Text number First person pronoun normed count Discipline
Text 1 3 1
Text 2 7 2
Text 3 8 3
Text 4 10 3
Text 5 12 3
Text 6 . . . 3 1
120 Your Own Corpus
But in this chapter, we will go through the steps of calculating the

One-Way ANOVA by hand. For this reason, and for hand-counting the
ANOVA, we will use a different type of organization, as it is easier to see
what is happening within the groups when listed by group.
In calculating the F score (the ratio for the mean sum of squares between
and across groups), we need to take several steps. Conceptually, we are
looking for the mean score for each group and then the variation as to
how the scores are dispersed or spread. This way we can tell whether the
variance can be attributed to variation inside each group or the differ-
ences (variation) across the groups.
We take eight steps to do the calculations and determine where the dif-
ferences lie. The following are the eight steps:
Step 1: Calculate the mean score for each group and for the entire dataset.
Step 2: Calculate distances across scores (and square them).
Step 3: Calculate degrees of freedom.
Step 4: Calculate mean sum of squares.
Step 5: Calculate F score.
Step 6: Determine whether the F score is significant.
Step 7: Calculate strength of association.
Step 8: Locate the differences with post-hoc tests.
Step 1: Calculate the mean score for each group and for the entire dataset.
Business Humanities Natural sciences
3 7 8
3 8 10
4 9 12
4 7 8
5 8 16
5 9 18
XBusiness = 4 X Humanities� = 8 X Nat .sci. = 12 XTotal = 8
Step 2: Calculate distances across scores.
Before we get into details, however, it is necessary to make the differ-

ence between two notions: a) a score being x mean away from another
score, and b) a score being x value away from the mean. The following is
the explanation for the difference between these two notions:
a) If a score is x mean away from another score, it means that we are

measuring the distance in the value of the mean score. Let’s assume,
for example, that Mean = 4; Score(1) = 2; and Score(2) = 10. In this
case, Score(2) is two means away from Score(1) because the differ-
ence between Score(1) and Score(2) is eight, and that is exactly two
times the mean.
b) If a score is x value away from the mean, it means that we simply
calculate the difference between the mean and the given score and
get a value as a result. Let’s assume that Mean = 4; Score(1) = 2; and
Score(2) = 10. In this case, Score(2) is 6 “points/values” away from
the mean, and Score(1) is –2 points or values away from the mean
because the difference between Score(1) and the mean is –2 (2–4),
and between Score(2) and the mean is 6 (10–4).
In our calculations, we will mostly use the second type of distance mea-
sure. In looking at how the scores are dispersed, we need to calculate
a) the distance between the individual score and its own group’s mean,
b) the distance between the group mean and the mean for the grand total,
and c) the distance between the individual score and the mean for the
grand total.
We will work with the following terminology: within sum of squares
(SSW) (the sum of squares within each group), between sum of squares
(SSB) (the sum of squares across groups), total sum of squares (SST) (the
sum of squares for the entire dataset), degree of freedom within (DfW)
(degree of freedom within each group) and degree of freedom between
(DfB) (degree of freedom across groups).
a) Within each group: How far is each score from its own group’s
mean?
To calculate the within group sum of squares (SSW or group variance),

take each individual score (x) minus the mean for its group (X group),
and square it. Add values gained this way for each group; then add each
group together. You will get the within sum of squares (SSW), or group
variance (see Table 7.2).
SSW =� ∑ ( x � − X
� group )
2
XBusiness = 4 XHumanities = 8 X Nat .sci. = 12
SSW = 96
122 Your Own Corpus
Table 7.2 Calculating sum of squares within groups
Business Humanities Natural sciences Total

_ _ _
Raw x–Xgroup Squared Raw x–Xgroup Squared Raw x–Xgroup Squared
score value score value score value
3 3–4 = –1 1 7 7–8 = -1 1 8 8–12 = -4 16

3 3–4 = –1 1 8 8–8 = 0 0 10 10–12 = -2 4
4 4–4 = 0 0 9 9–8 = 1 1 12 12–12 = 0 0
4 4–4 = 0 0 7 7–8 = -1 1 8 8–12 = -4 16
5 5–4 = 1 1 8 8–8 = 0 0 16 16–12 = 4 16
5 5–4 = 1 1 9 9–8 = 1 1 18 18–12 = 6 36
Total 4 4 88 96
b) Between groups: How far is each group from the total mean?
To calculate the between group sum of squares (SSB) (between group

variance), take each group mean ( X group ) minus the mean for the
total ( X total ), square it and multiply by the number of observations
in the given group (Ngroup). Do the same for each group and add up all
results (see Table 7.3).
XTotal = 8 Ngroup = 6 NTotal = 18 (6 in each group x 3 groups)
∑ � N� ( X group� − X � total)
2
SS B =�
Between sum of squares:
SSB = 192
c) Total sum of squares
To calculate
_ the total sum of squares, take each score (x) minus total
mean (X ), square it and sum it up (see Table 7.4).
∑ ( x � − X� total)
2
SST =�
SST = 288
An easier way to calculate this score is by adding up the ‘within’

and ‘between’ sum of square values calculated before: SSW + SSB = SST
(96 + 192 = 288).
Table 7.3 Calculating sum of squares between (across) groups
Business Humanities Natural Sciences Total
6 x (4–8)2 6 x (8–8)2 6 x (12–8)2

6 x 16 6x0 6 x 16
Total 96 0 96 192
Table 7.4 Calculating sum of squares total
Business Humanities Natural sciences Total

_ _ _
Raw score x-X total Squared Raw x-X total Squared Raw x-X total Squared
value score value score value
3 3–8 = –5 25 7 7–8 = –1 1 8 8–8 = 0 0

3 3–8 = –5 25 8 8–8 = 0 0 10 10–8 = 2 4
4 4–8 = –4 16 9 9–8 = 1 1 12 12–8 = 4 16
4 4–8 = –4 16 7 7–8 = –1 1 8 8–0 = 0 0
5 5–8 = –3 9 8 8–8 = 0 0 16 16–8 = 8 64
5 5–8 = –3 9 9 9–8 = 1 1 18 18–8 = 10 100
Total 100 4 184 288
Step 3: Calculate degrees of freedom.
Conceptually, the degree of freedom “refers to the quantities that can

vary if others are given” (Hatch and Lazaraton 1991: 254). For example,
if you know A and B in the following equation, there is no degree of
freedom for the value to change for C: A + B = C; but if you do not know
B, then there is one degree of freedom for you to change that will also
change the results for C. The degree of freedom values apply to datas-
ets as well as groups. Here, we have to calculate two types of degree of
freedom.
DfW (degree of freedom within groups) = all observations minus #

of groups
DfW = N - Ngroup
For our dataset: 18 – 3 = 15, so our DfW = 15.

Six observations in three disciplinary areas (6 x 3) minus three groups.
DfB (degree of freedom between groups) = total # of groups minus 1
DfB = Ngroup - 1
124 Your Own Corpus
For our dataset: 3 – 1 = 2, so our DfB = 2.
This will be important in looking up whether our F score is significant.
Step 4: Calculate mean sum of squares.
As an intermediary step between the distance calculations and the

degree of freedom, we need an average of the squares (see Table 7.5
below). We will use the mean squares within group (MSW), and the mean
squares between groups (MSB) as a final step before being able to arrive
at the F score. The mean square within the group is the within sum of
squares divided by within degree of freedom.
For our dataset: 96/15 = 6.4
The mean square is between sum of squares divided by degree of free-

dom between groups.
SSB
MSB =
DfB
For our dataset: 192/2 = 96
Step 5: Calculate F score.
The F score equals MSB divided by MSW .

MSB
F=
MSW
For our study: F = 96/6.4 = 15, so our F = 15.
Step 6: Determine whether the F score is significant.
We can determine whether the F score is significant by looking up the

value in a chart in juxtaposition with the degree of freedom. We looked
up the critical value for significance for the two degrees of freedom. In
such tables, on the side are the degree of freedom values for within groups
and on the top are the degree of freedom values for across groups.
The critical value for .05 level is 3.68 and our F score ended up at a
value of 15. Clearly, this is more than 3.68. Therefore, the F value rep-
resents statistical significance. In fact, if we look at the p< .01 level, it is
significant there as well. The critical value for that is 6.36, and our F score
is higher than that (15).
This means that the variation of scores across the groups is larger than
the variation within the groups. That is, we can reject the null hypothesis,
which states that “there is no statistically significant difference across
the disciplines in the use of first person pronouns.” Hence, we accept the
alternative hypothesis, which states that “there is a statistically significant
difference in the use of first person pronouns across disciplines.” What
we do not know yet, however, is how strong the association is between
the two variables (dependent and independent) and where the differences
lie. Below is Table 7.5, showing how the statistical program SPSS reports
on the Descriptive statistics and on the One-way ANOVA results.
Table 7.6 above has all the numbers we have calculated in the process
by hand previously.
Step 7: Calculate the strength of association.
The strength of association tells us the “proportion of the variability

in the dependent variable that can be accounted for by the independent
variable” (Hatch and Lazaraton 1991: 330).
To calculate the R-squared for the One-way ANOVA, we take the sum
of squares between groups divided by the total sum of squares.
SSB
R2 =
SST
Table 7.5 Descriptive statistics for first person pronoun
N Mean Std. Deviation Minimum Maximum
Business 6 4.00 .894 3 5

Humanities 6 8.00 .894 7 9
Natural 6 12.00 4.195 8 18
Sciences
Total 18 8.00 4.116 3 18
Table 7.6 One-way ANOVA results for first person pronouns
Sum of Squares df Mean Square F Sig.
Between Groups 192.00 2 96.00 15.00 .000

Within Groups 96.00 15 6.40
Total 288.00 17
126 Your Own Corpus
In our example,
R2 = 192/288 = .666
R2 = .666 means that 66% of the variance in the first person pronoun
use can be accounted for by the discipline. That is, if you know the
discipline, you can predict the use of pronouns more than half the time.
Or, by knowing the first person pronoun score, we are able to predict
which discipline it comes from with quite good certainty—more than
half the time.
Step 8: Locate the differences with post-hoc tests.
With identifying the F score’s significance, we can only say that there
is a statistically significant difference in the use of, in our case, first per-
son pronouns. What we cannot say is where the statistically significant
differences are exactly. In order to do so, we can use a range of post-
hoc tests, including Scheffe, Tukey, Bonferroni, Duncan, or LSD. We are
using Scheffe for the current question and dataset to illustrate how this
works. Table 7.7 was, again, created by SPSS (Statistical Package for
Social Sciences).
As Table 7.7 shows, the mean difference across each of the three dis-
ciplines is statistically significant. To determine where the significant dif-
ferences actually lie and which direction they go, we need to look at each
pair from this table to say the following:
Table 7.7 Post-hoc test (Scheffe) results
(I) Discipline (J) Discipline Mean Std. Error Sig. 95% Confidence
Difference (I-J) Interval
Lower Upper Bound

Bound
Business Humanities –4.00(*) 1.461 .048 –7.96 –.04

Natural Sciences –8.00(*) 1.461 .000 –11.96 –4.04
Humanities Business 4.00(*) 1.461 .048 .04 7.96
Natural Sciences –4.00(*) 1.461 .048 –7.96 –.04
Natural Business 8.00(*) 1.461 .000 4.04 11.96
Sciences
Humanities 4.00(*) 1.461 .048 .04 7.96
*The mean difference is significant at the .05 level (as we also see in the Sig. column, all values are
below p<.05)
1. Business—Humanities: The Business mean is 4 values lower that the

Humanities mean (hence the negative number). Looking at our mean
scores, it is true, since the Business mean was 4 and the Humanities
mean was 8. Throughout the analysis, this was considered to be a
statistically significant difference.
2. Business—Natural Sciences: The Business mean is 8 values lower
than the Natural Sciences mean (hence the negative number). Look-
ing at our mean scores, it is true, since the Business mean was 4 and
the Natural Sciences mean was 12. Throughout the analysis, this was
considered to be a statistically significant difference.
3. Humanities—Natural Sciences: The Humanities mean is 4 values
lower than the Natural Sciences mean (hence the negative number).
Looking at our mean scores, it is true, since the Humanities mean
was 8 and the Natural Sciences mean was 12. Throughout the analy-
sis, this was considered to be a statistically significant difference.
The rest of the information in the table is a repetition of this but with
reversed direction. If we look at our original mean scores, it is true that
the Business mean was 4, the Humanities mean was 8, and the Natural
Sciences mean was 12. And now we know that these differences are, in
fact, statistically significant.
Interpretation
Based on previous readings we know that first person pronouns are
typically associated with a communicative context where language is
produced in a shared physical space, and under involved production
circumstances allowing for the potential of interaction. We also know
through the situational analysis that disciplines may differ in the way the
material is presented, and so we want to know to what extent first person
pronouns would be an indicator of such difference. The statistical results
in our mini-study showed that there is a statistically significant differ-
ence across disciplines, and that significantly more first person pronouns
are used in Natural Sciences than in any of the other two disciplines. In
addition, it also shows that when compared to Humanities, Business also
uses significantly fewer first person pronouns. These results indicate that
in Natural Sciences classrooms language features seem to be similar to
those in spoken discourse (rather than written), which then is associated
with a discourse produced under involved production circumstances sug-
gesting interaction. The fact that Business showed the least number of
personal pronouns may be attributed to less interaction in the classroom,
and more teacher talk perhaps.
128 Your Own Corpus
7.1.2 Two-way ANOVA (Analysis of Variance)

We do a Two-way ANOVA test when we have one dependent variable, and
at least two independent variables that could have two or more levels each.
As with other parametric tests, the dependent variable has to be an interval
score (as the mean has to be the best measure of central tendency and the
standard deviation has to be the best measure of dispersion), and the inde-
pendent variables have to be nominal. With parametric tests, like the Two-
way ANOVA, we can generalize to the population that the sample was
drawn from. Conceptually, the Two-way ANOVA helps us identify which
one of the two (or both) independent variables is responsible for the vari-
ability in the dependent variable. More specifically, the question is whether
the variability in the dependent variable is due to one or both independent
variables. Again, we are looking at the variability of the scores within each
level group versus across the levels but not only for one independent vari-
able as we did with the One-Way ANOVA but two. That is, Two-way
ANOVA assesses whether the differences in mean scores are attributed to
the variability within or across the level groups when it comes to two dif-
ferent variables and their combinations. Again, if the cross-group variation
is large relative to the within group variation, there is a statistically sig-
nificant difference across the groups. That is, the larger this ratio between
within group variability measures and across group variability measures
(F score), the more likely that the difference between the means across the
groups is significant. Although we would be using the same calculations if
we wanted to calculate a Two-way ANOVA as we have just done with the
One-way ANOVA, since the computation becomes rather complex with
two variables we will not do that by hand. We will, however, rely on the
statistical package to give us the results (as we see it is pretty reliable!).
What we need to be careful of here is the interpretation of the results.
Example
As you have attended classes at the university, you noticed that teachers
talk differently in classes not only from different disciplines (as we have
seen the example before), but also in classes with different educational
levels. Your primary investigation is discipline but it seems that level of
instruction may also be a variable that could intervene in the variability
of the data, and you are hoping that it does not affect your previous
findings. You do not know whether the language change is attributed to
only one of the variables or the two together. In your situational analysis
then, you take discipline as your main variable, and level of instruction
as another, intervening variable. As for the teacher ‘talking differently’,
you continue to believe that, based on your previous readings, first person
pronoun use is what makes the difference.
First, you formulate your research question in one of two ways:
1. How does the use of first person pronouns differ across disciplines
and levels of instruction? OR
2. Is there a difference in first person pronoun use across disciplines or
across levels of instruction?
The dependent variable is first person pronouns (use normed counts as

discussed before an—interval score), and the two independent variables
are discipline (with three levels) and level of instruction (with three lev-
els). The three levels for discipline are Business, Education, and Natural
Sciences, and the three levels of instruction are lower division undergrad-
uate, upper division undergraduate, and graduate.
Second, you formulate your hypothesis:
H0: There is no effect on first person pronoun use for discipline or

level instruction and there is no effect for the interaction.
H1: There is an effect on first person pronoun use for discipline.
H2: There is an effect on first person pronoun use for level of
instruction.
H3: There is an interaction effect on first person pronoun use.
Number of observations in each cell is summarized in Table 7.8.

The Two-way ANOVA results provided by SPSS are in Table 7.9. As
we see from Table 7.9, the significance value is lower than p< .05 (Sig
.011) only for discipline. Hence, we can conclude with confidence that
discipline alone is a significant factor in the varied use of first person pro-
nouns. That is, level of instruction is not a significant factor alone, nor
does it have an effect on the variation in the dependent variable (first per-
son pronoun use). Since the interaction effect is also not significant (see
line with Discipline * Level), the effect on the variation in the data was
not moderated by the effect of level of instruction. It is clearly discipline
alone that is responsible for the variation in the dataset.
Table 7.8 Distribution of cases in the dataset
Value Label N
Discipline Natural sciences 10

Education 10
Business 10
Level Low division 13
Upper division 6
Graduate 11
130 Your Own Corpus
Table 7.9 Two-way ANOVA results for first person pronouns
Source Type III Sum of Squares df Mean Square F Sig.
Corrected Model 2245.306a 8 280.663 2.706 .032

Intercept 31131.297 1 31131.297 300.136 .000
Discipline 1154.974 2 577.487 5.568 .011
Level 77.000 2 38.500 .371 .694
Discipline * Level 893.091 4 223.273 2.153 .110
Error 2178.206 21 103.724
Total 46476.120 30
Corrected Total 4423.512 29
a. R Squared = .508 (Adjusted R Squared = .320)
Interpretation
We set out to investigate whether discipline accounts for the variability in
our data. At the same time, we also realized that in the academic context,
an intervening variable, i.e., level of instruction, may have an effect on
this variability as well. We tested this by running a Two-Way ANOVA to
see what effect level of instruction may have. Based on the results above,
we can conclude that discipline alone accounts for the variability in the
data, and level has no effect on this variability. This means, that there is
a consistent change in the use of first person pronouns across disciplines,
irrespective of whether the class is an undergraduate lower or upper divi-
sion class, or a graduate class. Post hoc tests could identify where exactly
the differences are. It turns out that first person pronouns are used most
frequently in Education classes, indicating a more interactive class than
in the other two disciplines.
A word of warning: It must be noted that if the interaction effect is sig-
nificant (that is, if p <.05 or lower in the line with the two variables jux-
taposed Variable 1 * Variable 2), we cannot isolate any of the variables
as significant even if they show a significant result on their own. Consider
the following results (see Table 7.10) in terms of significance level and
instead of what we had before (we are just doing this for the sake of the
exercise—the numbers are not true results):
In this dataset, each of the independent variables is significantly mark-
ing the variation in the first person pronoun use. At the same time, the
interaction measure (Discipline * Level) is also significant with a p< .05.
This means that neither discipline nor level of instruction alone is respon-
sible for the variability in the use of first person pronouns. Instead, the
two variables together cause the change in the dataset. In other words,
we cannot say that, for example, Natural Sciences consistently use more
first person pronouns than Humanities, because their use of pronouns
Table 7.10 Two-way ANOVA results with significant interaction
Source F Sig.
Corrected Model 2.706 .032

Intercept 300.136 .000
Discipline 5.568 .011
Level 5.371 .000
Discipline * Level 5.153 .00
Error
Total
Corrected Total
depends on the level of instruction. Apparently, they use more in their

lower and graduate classes, but not in the upper division undergraduate
classes. This variation is also true for the other two disciplines, and so
discipline alone is not a factor for the change in the dependent variable.
It also depends on the level of education at least as robustly. All in all,
the interaction effect, if significant, overrides the effect of the individual
independent variables.
7.2 Relationship Tests

When doing relationship tests (e.g., Chi-square and Pearson correla-
tion), we test the relationship between two or more variables. That is,
we test how well they go together. We are not interested in how one
variable affects another one, as that is the goal of a test of difference
seen in previous sections. Therefore, we also cannot make claims of
cause and effect with relationship tests. We can only talk about the
results in terms of strong or weak relationship between two or across
many variables.
7.2.1 Chi-square
With Chi-square tests, both the dependent and the independent variables
can be nominal data. The results of non-parametric tests, like Chi-square,
cannot be generalized to the population the sample was drawn from but
we can ask questions related to the given dataset. Namely,
• Is there a relationship between two variables in the dataset?

• How strong is the relationship in the data?
• What is the direction and shape of the relationship in the data?
• Is the relationship due to some intervening variable(s) in the data?
132 Your Own Corpus
Conceptually, we typically want to know whether there is a relationship

between two variables (and their levels). Chi-square compares the actual
observed frequencies of some phenomenon with the frequencies we would
expect if there were no relationship at all between the two variables in
the sampled dataset. That is, Chi-square tests our actual results against
the null hypothesis (i.e., no relationship) and assesses whether the actual
results are different enough to overcome a certain probability that they
are due to sampling error. The further apart the observed and expected
values are, the more likely it is to be a significant Chi-square.
Assumptions and requirements:
1. The sample must be randomly drawn from the population.

2. Data must be reported in raw frequencies (not in scales, e.g., as
percentages would be).
3. When frequencies of a phenomenon are counted, the frequency of
non-occurrence will also have to be counted.
4. Measured variables must be independent.
5. Values/categories on independent and dependent variables must be
mutually exclusive and exhaustive.
6. Observed frequencies cannot be too small; the expected cell
frequency has to be at least 5.
With Chi-square, we also have One-way and Two-way designs. The dif-
ference between the two is simply how many levels each variable has.
A One-way design (see Table 7.11 below) would have a dependent nomi-
nal variable with no levels, and an independent variable with potentially
two or more levels. For example, if you wanted to compare the raw fre-
quency scores of a particular type of relative clause across registers, you
would use a one-way design, as shown in Table 7.11.
All in all, a one-way Chi-square design would have one variable with
no levels (e.g., relative clauses) and one variable with two or more levels
(registers). On the other hand, a two-way design would have a dependent
nominal variable with two or more levels and an independent nominal
variable with two or more levels. Table 7.12 shows a two-way Chi-square
design: one variable with two or more levels (e.g., articles: a, an, the, zero
article) and one variable with two or more levels (position: subject or
object).
Table 7.11 One-way Chi-square table
News Fiction Spoken Academic
Wh-relative 15 5 10 25
We rarely use one-way designs in our studies, and therefore, we will

focus on a two-way design.2
Example
Imagine that you would like to find out about the relationship between
article type (a, an, the, zero article) and position (subject or object). Here
are the steps you need to take:
Step 1: Formulate your research question: Is there a relationship between

type of article use and clause position?
Step 2: State your null hypothesis: H0 — There is no relationship between
type of article use and clause position.
Step 3: Create a cross tab of frequencies of two nominal variables, article
and position.
Each cell reports on how many observations produced that combina-

tion of independent and dependent values (see Table 7.12 below). For
example, 54 definite articles were found in nominal clauses in subject
position, etc. Although this dataset looks good, we need to be aware of
the cell sizes. Here are some of the rules:
• For 1x2 or 2x2 table, expected frequency values in each cell must be
at least 5.
• For a 2x3 table, expected frequencies should be at least 2.
• For a 2x4 or 3x3 table, if all frequencies but one are at least 5 and if
the one small cell is at least 1, chi-square is still a good approximation.
If you are worried about the frequencies in the cells, you could collapse
categories that make sense. In the example above, the two types of indefi-
nite articles (a/an) can be collapsed since it’s use is dependent on the word
following it and will not affect the syntactic position. Table 7.13 shows
the crosstab of frequencies.
If the article distribution were the same in subject and object positions,
we would get an equal number of them across article types. So the ques-
tions are: How far is this off? Can we say that they are really off, and
whether there is a difference, or not? That is the real question. In other
Table 7.12 Two-way Chi-square table
the a an 0
subject 54 27 8 19
object 92 49 3 44
134 Your Own Corpus
words, if there were no relationship between the article type and the posi-
tion, we would get an even distribution of the frequencies. Considering
this, the questions are: Is there a relationship between the article type
and clause position? We calculate what we would expect if there were no
relationship, and compare that with the existing dataset.
First, we calculate the row and column totals.
Second, we calculate the expected value for each cell by taking the row
total and the column total, multiply the two, and divide it by the grand
total. Below is the formula.
Row total × Column total
fexpected =
N
For example, to calculate the expected value for the first cell (‘the’ in
subject position) take 108 (Row total) times 146 (column total) divided
by 296 (N) = 53.27. Do the same for each cell. (Figure 7.15)
Finally, we can calculate the Chi-square ( χ 2): Deduct the expected
value from the observed value, square it, and divide it by the expected
value, and add it all up (Figure 7.16).
Table 7.13 Revised Chi-square table
the a/an 0 Total
Subject 54 35 19 108
Object 92 52 44 188
Table 7.14 Chi-Square calculations (row and column totals)
Observed values the a/an 0 Row total
Subject 54 35 19 108

Object 92 52 44 188
Column total 146 87 63 296
Table 7.15 Calculating expected values
Expected values the a/an 0
Subject 53.27 31.74 22.99

Object 92.73 55.26 40.01
Table 7.16 Calculating Chi-square value
The a/an 0
O—E squared sq/E O—E squared sq/E O—E squared sq/E

Subject 0.73 0.53 0.01 3.26 10.63 0.34 -3.99 15.92 0.69
Object -0.73 0.53 0.01 -3.26 10.63 0.19 3.99 15.92 0.40
Total 0.02 0.53 1.09
(f )
2
− � f expected
( ) ∑
2
� χ � �=
Chi � square
observed �
f expected �
χ 2 = 1.64
Degree of freedom:
(# of Rows—1) times (# of Columns—1)

In our study: (2–1) × (3–1)
Critical value on the p< .05 for 2 df is 5.991
Our Chi-square value is 1.62, which is below the critical value—
therefore, not significant.
Interpretation
What does this mean? It means that we cannot reject the null hypothesis
stating that there is no relationship between the article type and position.
In other words, any of the articles could pretty much randomly occur in
any position, as there is no relationship between the position and the type
of article used. While this example may not have direct relevance to reg-
ister studies, we could follow up with a register study. Instead of looking
at the potential relationship between article type and syntactic position,
the focus of the investigation would be to see whether one type of article,
when in a certain position, occurs more often in one register over another,
and versus in another position.
7.2.2 Correlation
Among the three different types of correlations (Pearson, Spearman Rank
Order, and Point-biserial), Pearson correlation is the most frequently used
statistical procedure in corpus studies. With Pearson, we need interval
data for both the dependent and independent variables.
136 Your Own Corpus
Conceptually, we are looking for relationships between two or more

variables in the dataset. Again, as with Chi-square, we do not look at how
one variable affects the other but how they relate to each other. There-
fore, the research question also aims at looking for relationships (whether
strong or weak), and not difference (whether there is an effect or not).
The null hypothesis for difference studies (e.g., One-Way ANOVA) is
like this: “There is no difference between the two variables with respect
to some measure.”
The null hypothesis for relationship studies is like this: “There is no
relationship between two measures.”
Example
You noticed that I mean and ok often come as a package in spoken dis-
course. You also noticed that both teachers and students use it, but you
don’t know whether it’s the same when they are presenting in front of
an audience. Let’s assume you would like to find out whether there is a
relationship in the use of “I mean” and “ok.” You have a small corpus of
presentations comprising of two sub-corpora: teacher presentations and
student presentations. Let’s say, the mean score for I mean used for teach-
ers is 39.1, and for students, it is 42.5.
There is too much overlap between the two types of presentations in
terms of “I mean” use. That is, if we did a difference type of test (e.g.,
like ANOVA), there would not be a significant difference between teacher
and student presentations in terms of the use of “I mean.” In other words,
we would not be able to predict whose presentation it is by knowing the
“I mean” count.
In contrast, the mean score for ok use for teachers is 150, and for
students, it is 328. There is no overlap between the two types of presenta-
tions in terms of the use of “ok.” That is, in a difference type of test (e.g.,
One-way ANOVA), this would show a significant difference between
teacher and student presentations in terms of “ok” use. That is, we could
predict who gives the presentation by knowing the “ok” count.
“I Mean” and “Ok” Use in Two Settings: Teacher

Presentation and Student Presentation
If the uses of the two expressions consistently overlapped, seeing a cor-
relation may be nice. That is, if we knew one feature, we could know that
the other feature count will also be high. So the relationship between the
two features would be strong.
With two interval variables, you want to see what the strength is
between the two variables so you can predict the occurrence of one by
knowing the occurrence of the other. If there is no correlation, there is no
relationship. If there is a correlation, then that means there is a relation-

ship between the two variables. The questions to ask then are: a) what
kind of relationship it is, and b) how strong the relationship is. There are
two kinds of relationships: positive and negative. In positive relation-
ships, if one score is high, the other score is also high. Translated to our
question, if the “I mean” score is high, then the “ok” score will also be
high. In a negative relationship, if one score is high, the other score is low.
In our case, if the “I mean” score is high, the “ok” score would be low
and it is consistently that way.
The correlation coefficient (r) is between 0 and 1 (whether positive or
negative depending on the direction of the correlation explained above),
where zero means no correlation, and + 1 means perfect correlation with
a 100% overlap.
+1 . . . . . . . . . . . . . 0 . . . . .. . . . . . .-1 0 = absolutely no relationship
In terms of strength, we need to see at what percentage can we predict

one over the other. The direct measure of strength is r2, and we are look-
ing at the percent overlap between 0 and 100%.
Let’s have a visual about a potential dataset. Look at Figure 7.1. The
data-point in light grey tells us that this text has 20 “ok” and 30 “I mean”
scores. The data point in darker grey tells us that this text has 17 “ok” and
27 “I mean.” There seems to be a consistent relationship between the use
of “I mean” and “ok” in this dataset (as the dotted line indicates the pat-
tern); that is, if one occurs, the other one also occurs, and it seems to be
a positive relationship. The positive relationship means that if one occurs,
the other one occurs too. The question is how consistent and how strong
the relationship is between these two features in the entire dataset.
40
Counts: "I mean" per 1,000 words
35
30 20, 30
10, 27 17, 27
25 15, 25
12, 22
20 10, 20
7, 18
15 5, 15
7, 12
10
5
2, 3
0
0 5 10 15 20 25
Counts: "ok" per 1,000 words
Figure 7.1 Correlation between ‘ok’ and ‘I mean’

138 Your Own Corpus
We can calculate the Person product r through Excel or SPSS. Below

Table 7.17 produced by SPSS, where r = .885, and that the correlation is
statistically significant at the p< .01 level. While the significance does not
matter that much in correlation studies (see below for more discussion
on that), the bigger the r value (i.e., the closer it is to 1), the stronger the
relationship is between the two variables. So the magnitude tells us how
well the two variables go together. As it is a positive value here, we can
be certain that our dataset shows that the higher the number of the “ok”
counts, the higher the number of the “I mean” counts as well. What the
cut-off point is (i.e., what counts as an acceptably strong correlation)
depends on the field of study.
Next we calculate the r2, which is the direct measure of strength
between the two variables (or variance overlap, which is the variability of
the data around the mean, if you recall). In our example, r2 = .783, which
means that about 78% of the data has shared variation. Visually, it looks
like Figure 7.2.
This means that 78% of the time we can predict that if one feature
occurs, the other one occurs too. That is, “the magnitude of r2 indicates
Table 7.17 Correlations calculated by SPSS
ok I_mean
Ok Pearson Correlation 1 .885**

Sig. (2-tailed) .000
N 11 11
I_mean Pearson Correlation .885** 1
Sig. (2-tailed) .000
N 11 11
** Correlation is significant at the 0.01 level (2 tailed).
I mean
78 % overlap
ok
Figure 7.2 Overlap between ‘I mean’ and ‘ok’
the amount of variance in” one variable “which is accounted for by the
other variable” or the other way round (Hatch and Lazaraton 1991:
441). To translate this to our study here, it sounds like this: The amount
of variance in the use of “I mean” is accounted for by the use of “ok.”
What is a strong overlap and what is a weak one is hard to tell without
knowing the question. If you wish to show that one text is very similar
to another, the higher the overlap the better. In general though, in social
science research if the overlap is over 25%, it is considered very high. But
since it is genuine continuous data, there is no need for a cut-off point.
The degree will depend on how the disciplines regard this as strong or
not, and that is why no significance level is necessary.
Some useful hints:
1. Forget the groups here. Whether there is a relationship between the

two variables is the question.
2. Only use words like strong and weak and not significant when talk-
ing about correlations.
3. Be aware that there is a relationship between sample and correla-
tion. Every correlation has a strength part and a significance part in
terms of correlating or not (as we have seen). The closer we get to
100% relationship, the stronger the correlation is, but that increasing
strength does not affect the significance of the correlation.
In this section, we only looked at how two linguistic variables may relate
to one another (or what relationship they may have) but we can look at
more than two at once. It is almost like going to a party where you try to
figure out who is hanging out with whom and what characteristics they
have. In any case, if you look at correlations of more linguistic variables
at once, you can start characterizing texts for their comprehensive lin-
guistic make-up. We will briefly discuss this and point you in that direc-
tion in the last chapter of this book.
Interpretation
The interpretation here is simple: when one language feature occurs, the
other one does as well. That is, there is a positive relationship between
the two variables and so we can predict that if there is a high number of
one, there will be a high number of the other as well. This kind of study
becomes more interesting is when we look at more than just two linguis-
tic features co-occurring with one another. When we are able to detect
how a number of features, when thrown in the same pot and having an
effect on one another, will behave. This will be discussed in Chapter 9
further as we are looking ahead.
140 Your Own Corpus
7.3 How to Go About Getting the

Statistical Results
As in the previous chapter, we will show you how to get the results in
SPSS for the four different tests you set out to investigate.
7.3.1 Difference Tests
One-Way ANOVA
To run a One-Way ANOVA test in SPSS, from the tabs select Analyze →
Compare Means → One-Way ANOVA. Again, your ‘Dependent List’ will
contain the dependent variable. Although you could only have one depen-
dent variable to test a One-Way ANOVA, if you want to run the test on
more than one dependent variable at the same time (e.g., you want to see
variation in hedges and also in noun use), instead of opening the window
for each individually, you can list all of them under the dependent list. The
program will take them one by one, and run the test on each separately.
Your independent variable with multiple levels will go into the “Factor”
window.
Two-Way ANOVA
To run a Two-Way ANOVA test in SPSS, from the tabs select Analyze →
General Linear Model → Univariate. Again, you will put your dependent
variable in the ‘Dependent Variable’ field, and the independent variable
will come under ‘Fixed Factor(s)’ and your intervening variable (your sec-
ond independent variable) will come under ‘Random factor(s).’ You can
also determine what ‘Posthoc’ test you may want to use in case only one
variable significantly accounts for the variability of the data.
7.3.2 Relationship tests
Chi-Square
To obtain Chi-square results, go to Analyze → Non-parametric tests →
Legacy Dialogs → Chi-Square. Again, your variables will go into the
appropriate boxes and you can hit “run”.
Pearson Correlation
To obtain Pearson correlation results, go to Analyze → Correlate →
Bivariate Correlations, click on ‘Pearson.’ Again, your variables will go
into the appropriate boxes, and you need to decide whether you want a
one-tailed or a two-tailed analysis (see Chapter 6). Again, hit “run”.
Notes
1 Statistical Package for the Social Sciences.
2 We must be careful as to how we select them in the corpus, though—and the
best way to do so is to get a set of randomly selected relative clause sentences
and continue our classification based on that.
Reference
Hatch, E. & A. Lazaraton (1991) The Research Manual. Design and Statistics for
Applied Linguistics. Boston: Heinle & Heinle.
Chapter 8
8.1 Doing a Register Analysis of Your Project

8.2 Situational Analysis
8.3 Linguistic Analysis
8.4 Functional Interpretation
8.5 Reporting on Your Project
The previous chapters of this book have provided you information on 1)

how to do a register analysis (Chapter 2); 2) the types of software that you
can use in corpus analysis (Chapter 3); 3) how to do corpus projects using
existing corpora (Chapter 4); 4) how to build and structure your corpus
for analysis (Chapter 5); and 5) how to apply and interpret basic statistical
techniques (Chapters 6–7). This chapter will guide you through the steps
and procedures to actually put the corpus to use and to report on your
research findings. In the following section, we will provide some guiding
principles for how to go about answering your research question(s) using
a register analysis framework and corpus methods. Then, we will describe
the different parts of a research study and provide some guidelines for writ-
ing up and presenting your research project as well as provide some infor-
mation on how your corpus project can be assessed.
8.1 Doing Register Analysis of Your Project

As illustrated in Chapter 2, a register analysis is comprised of three com-
ponents: 1) describing situational characteristics of texts; 2) identifying
frequent linguistic characteristics of these texts; and 3) providing a func-
tional interpretation of why these frequent linguistic features are found
in the texts. In the following sections we will cover the steps in complet-
ing your project. In order to illustrate some of the concepts in doing a
situational and linguistic analysis along with a functional interpretation,
we will refer to a small corpus of English as Second Language writers
who were asked to produce problem solution paragraphs in two different
Doing Corpus Linguistics 143
Table 8.1 A corpus of problem solution essays written by Thai speakers of English
Individual Collaborative
# of texts 102 51
# of words 14,124 9,284
# of words/text 138.47 182.03
conditions. The research question for this study is: Do collaborative and
individual texts differ in terms of their lexico-grammatical features? The
texts from this corpus were collected under two different conditions:
First, the students were placed into pairs and asked to write a problem
solution paragraph collaboratively. Later in the semester, each student
had to write a problem solution paragraph as part of an in-class examina-
tion. The essays were then typed and saved as text files with headers and
codes to show the authors and topics of the essays. The corpus was also
divided into two sub-corpora: one consisting of collaboratively written
essays and another consisting of individual essays. A description of the
corpus is shown in Table 8.1.
This table provides the number of individual and collaborative essays,
the total number of words in each sub-corpus, and the average number
of words for the individual and collaborative texts. We will return to this
table when we discuss the analysis of linguistic features in section 8.3.
However, before looking at how to search for linguistic features, we will
first describe the method for completing a situational analysis.
8.2 Situational Analysis

As you recall in our discussion of register analysis in Chapter 2, the first
step in a register analysis requires a description of the situational charac-
teristics of your corpus. We follow the work of Biber (1988) and Biber &
Conrad (2009) in our description of the situational characteristics. As
discussed in Chapter 2, there are seven main components used to describe
the context of a language text or “event.” These are described in Table 8.2
below. The situational variables include information on the roles and rela-
tionships of the participants; the type of language used (whether it is writ-
ten or spoken; whether it is edited or unplanned); the setting; the purpose
of the communication; and the topic. In a sense, register analysis predicts
that variation in (at least some of) these situational characteristics will
result in variation in language use. In the table below, we provide some
questions that are intended to guide you in your situational analysis.
Depending on the structure of your corpus as well as your research
questions, projects that include sub-corpora should provide separate situ-
ational analyses for each sub-corpus. That way, any variation in language
144 Your Own Corpus
Table 8.2 Situational characteristics of registers
Situational variables Questions to consider when identifying situational

variables
Participants Addressor(s): Are the texts produced by a single
writer or speaker or are there multiple writers/
speakers in a single text?
Addressee(s): Who is the intended audience? A single
person, multiple people?
Relations among participants Is there linguistic interaction between participants?
Are there power or status differences among the
participants?
What is the nature of the personal relationships among
participants (e.g., friends, co-workers, strangers)?
Do the participants share different types of
information or knowledge (e.g., personal
information, topic-specific information)?
Channel Is the language written or spoken? Is the language
permanent (recorded or written) or is the language
temporary (face-to-face conversation, television,
telephone)?
Production circumstances Is the language revised and edited, planned or
unplanned?
Setting Do the participants share the same place and time?
Is the language used in a private or public setting?
Does the language take place in the present or in the
past (including the historical past)?
Communicative purposes What is the general purpose of the language (e.g.,
to report, describe, tell a story, inform, explain,
persuade, entertain)?
What is the specific purpose of the language (e.g.,
describe methods, summarize sources, teach)?
Is the language factual, hypothetical, imaginative, or
opinion?
Topic What is the general and specific topic of the language
used?
use can potentially be accounted for by reference to the situational analy-

sis. Should your project necessitate this, you can make a single table with
multiple columns to illustrate the shared characteristics of the corpora as
well as the characteristics that are different. A sample situational analysis
of the problem solution corpus is shown in Table 8.3 below. As we can
see from this table, the texts differ in the relations among participants,
the production circumstances, and topic. A potential research question
could be: Do individual and collaborative texts have different linguistic
features? As discussed in Chapter 5, your corpus has already been built
with a research question or set of research questions in mind; however,
the situational analysis can be very helpful in describing and motivating
your research question(s).
Table 8.3 Situational analysis of problem solution essay corpus
Sub-corpus A Sub-corpus B
Participants Second language writers of English Second language writers of

Native language is Thai English
Intermediate level of English Native language is Thai
proficiency Intermediate level of English
proficiency
Relations among One writer; teacher as audience Two writers of a single
participants essay; teacher as
audience
Channel Written Written
Production Timed exam Timed In-class activity
circumstances
Setting Classroom Classroom
Communicative Describe problem and propose a Describe problem and
purposes solution propose a solution
Topic Lack of English in Thai Shortage of teachers in
teachers; World Hunger; Thailand;Videogame
Protecting natural addiction among
resources in parks students
8.3 Linguistic Analysis

Once the situational analysis is completed, you can use software tools
to search the corpus for relevant linguistic features. When searching for
linguistic features, there are two basic approaches that you can take. As
we mentioned in earlier chapters, the first method involves searching for
language patterns in your corpus without having any preconceived idea
of what you might find. This type of method is sometimes referred to as
“corpus-driven” because it utilizes tools to search the corpus for language
features that you may not necessarily be aware of. Two useful tools in this
approach are word lists and n-grams. To generate a word list, all of the
individual texts can be loaded into AntConc (see Chapter 5 for details)
and the word list function is used to provide a list of the most frequent
words in the corpus. This same procedure can also be used with the col-
laborative texts to see if there are specific words that are frequent in one
sub-corpus that may not be as frequent in another sub-corpus. Figure 8.1
provides a screen shot of the word list for the individual texts in the
example corpus (also referred to a unigrams in Chapter 3). We will look
more closely at the word list findings below.
The “Clusters/N-Grams” function is another tool in AntConc that
is useful when employing a corpus-driven method. (The concept of an
n-gram has been discussed in section 3.3.) The size of n-grams can range
from unigrams (as discussed in Chapter 3 and illustrated in the word
146 Your Own Corpus
Figure 8.1 Word list in a corpus of problem solution essays written by Thai speakers of
English
list function above) to longer sequences. Figure 8.2 shows the most fre-
quent four-word n-grams in the individual problem solution texts. These
n-grams are found by using the “Clusters/N-gram” tab and then specify-
ing the n-gram size (4) as well as the minimum frequency (5), using the
tools found at the bottom of the screen (as shown in Figure 8.2). Note
that you can change the size of the n-grams to look at bigrams, trigrams,
and even longer sequences of words if you choose to do so.
The second general method to use is the “corpus-based method.” If
you choose to take a “corpus-based” approach, you already have an idea
of the linguistic features that will be the focus of your research. This
approach is used when previous corpus information determines the
linguistic features used in the analysis. For example, previous corpus
Figure 8.2 Four-grams in a corpus of problem solution essays written by Thai speakers
of English
research has illustrated that face-to-face conversation has many first and
second person pronouns. The high number of these pronouns can be
understood by reference to the interactional nature of face-to-face con-
versation where speakers often make reference to themselves as well as
to the other participants of the conversation. This corpus-informed fact
can serve as a reason to look for the frequency of first and second person
pronouns in corpora that do not involve face-to-face conversation but
that also include examples of spoken language in other contexts (such
as academic lectures, where the purpose of communication is primarily
informational). If one were to find a high number of first and second per-
son pronouns in academic lectures, this may be indicative of the interac-
tional nature or style of academic lectures. In this sense, corpus-informed
searches can be attributed to the fact that certain linguistic features have
148 Your Own Corpus
specific functions associated with them (as illustrated in previous research)

so that the occurrence in another corpus with different situational charac-
teristics can be related to this same functional interpretation.
In one sense, the distinction between corpus-driven and corpus-based
research methods can be misleading. At least with respect to a register
analysis approach, any feature that is identified through a corpus-driven
approach will still merit closer scrutiny and analysis in the corpus. If,
for example, a word list shows that one particular word is more fre-
quent in one sub-corpus than in another (a corpus-driven method), then
the researcher will still need to look at the distribution and use of this
feature more closely in the corpus by investigating its use in some more
detail. One might argue that this second step can be seen as corpus-based
because it identifies features in the corpus that merit further attention.
One the other hand, it might also be argued that the feature was not
identified by previous corpus-based research so it is, by definition, not
corpus-based. We see the merits of both approaches in trying to under-
stand language use and would encourage the use of both methods, espe-
cially in the smaller corpora that serve as the basis for your projects.
In Chapter 5, we mentioned the importance of building sub-corpora of
fairly equal sizes (see section 5.3). Sometimes this is not possible, as in the
case of the problem solution corpus described above. Since the design of
this study was focused on writing paragraphs, the researchers had no con-
trol over the length of the texts. Furthermore, because the same writers
produced texts both individually and collaboratively, it was not possible
to simply add more collaborative texts to make the corpora equal. If your
project involves comparing the frequency findings of sub-corpora in your
corpus (or to another existing corpus), you will need to employ a simple
statistical procedure known as “normalization” to ensure that your results
are comparable (see a brief mention of this in Chapter 6). Normalization
allows frequency counts taken from corpora of different sizes to be com-
pared by providing a count of the frequency of the feature in a similar
number of words.
Table 8.4 provides the five most frequent words in both the individual
and collaborative essays. The most frequent word in the individual corpus
is “the” and occurs 726 times; in the collaborative corpus, “the” occurs
425 times. The most frequent word in the collaborative corpus is “to”
and occurs 432 times; in the individual corpus, “to” occurs 524 times.
Although the frequency of both “the” and “to” is higher in the individual
texts, the sub-corpora are not of equal length (see Table 8.1). If we want
to compare the frequency in two corpora that do not contain equal words,
we can normalize the counts and then use the normalized (or normed)
counts as the basis for comparison. To do this, we divide the frequency of a
given feature by the total number of words in the corpus and then multiply
this by a reference number that will tell us how frequently this word (or
feature) occurs per x number of words. In this case, we will use 1000 as
the reference number. The results would show that individual “the” occurs
51.40 times per 1000 words (726/14,124 x 1000 = 51.40) and collabora-
tive “the” occurs 45.77 times per 1000 words (425/9,284 x 100 = 45.77).
When considering the frequency comparison of the word “the” in the two
sub-corpora, we can then compare 51.40 (individual) with 45.77 (collab-
orative) and note that the frequency differences between these two corpora
are not so great. If we look at “to,” we would calculate the normed count
in individual texts at 37.09 (524/14,124 x 1000 = 37.09) and collaborative
“to” would be 46.53 (432/9,284 x 1000 = 46.53). In this case, the normed
count in collaborative texts is actually higher than in the individual texts
even though the raw count is lower.
In a linguistic analysis, it is also worthy to note not only the potential
frequency differences in shared words across the two types of texts but also
the use of words that are different in the texts. In this small sample, we not
only see differences in nouns (“English” is the third most frequent word
in the individual corpus and “students” is the fourth most frequent word
in the collaborative corpus) but also differences in function words (“in” is
the fourth most frequent word in the individual texts and “of” is the fifth
most frequent words in the collaborative texts). Some differences might be
related to topic (as with the nouns) but other differences might be related
to the production circumstances or relations among participants. Only a
closer examination of these features in the corpus can provide us with evi-
dence to support the analysis. Furthermore, Table 8.4 only shows the five
most common words; looking at longer word lists would likely show other
differences in the corpus that are worthy of analysis and interpretation.
In addition to word lists, the n-gram function can show us potential
variation in corpora. As discussed Chapter 3, n-grams are contiguous
sequences of words that can vary in length depending on the interest of
the researcher. In Table 8.5, we provide the five most frequent four-grams
in the individual and collaborative texts. We have included both the raw
frequency counts as well as the counts that are normed to 1000 words in
parentheses.
Table 8.4 The five most frequent words in a corpus of problem solution essays written
by Thai speakers of English (both raw and normed counts to 1000 words)
Individual 1-gram Frequency Collaborative 1-gram Frequency
the 726 (51.40) to 432 (46.53)

to 524 (37.09) the 425 (45.77)
English 411 (29.09) and 236 (25.42)
in 372 (26.33) students 212 (22.83)
and 342 (24.21) of 182 (19.60)
150 Your Own Corpus
Table 8.5 The five most frequent four-grams in a corpus of problem solution essays
written by Thai speakers of English (both raw and normed counts to 1000 words)
Individual four-gram Frequency Collaborative four-gram Frequency
to solve this problem 27 (1.91) of teachers in Thailand 31 (3.33)

the first solution is 20 (1.41) shortage of teachers in 31 (3.33)
there are many 18 (1.27) addiction among CMU 25 (2.69)
solutions students
the second solution is 17 (1.20) games addiction among 19 (2.04)
CMU
to solve the problem 17 (1.20) teachers in Thailand 17 (1.83)
problem
The following patterns can be observed in this dataset:
1. No four-grams are shared between the two groups.

2. All of the four-grams produced by the first group contain verbs.
3. All of the four-grams produced by the second group contain at least
one (at times two) prepositions (“of,” “in,” and “among”).
4. None of the four-grams in the first group have prepositions.
5. None of the four-grams in the second group contain verbs.
6. Overall, the second group seems to have used more four-grams.
There are many different types of searches that you can do with your cor-
pus. As mentioned in Chapter 5, you should feel encouraged to explore
the AntConc program (as well as related literature such as the “read me”
files in the AntConc website) to learn about the program. New tools
are frequently available so you should visit the site from time to time to
learn about the new functions and to download the latest version of the
program.
For example, an important consideration in understanding the use
of any feature (including both word lists and n-grams) relates to the
distribution (or dispersion) of a feature in the corpus. A given feature
may be frequent, but it is important to make sure that the feature is
not used in a few texts at a very high frequency. Although not in a sta-
tistical sense as we described dispersion in Chapter 7, you can check
the visual of the distribution in AntConc by using the ‘Concordance
Plot’ option (see Chapter 5 for details). This function will show you
how many different files the given feature occurs in as well as how
many times the feature occurs in a single file. This is an easy method
to provide a visual representation of the distributional patterns of the
feature that you are looking at. Distributional patterns like these can
be very helpful in interpreting your results. Seeing the distributional
patterns can also help in examining whether your findings for a given
feature are, in fact, spread in your corpus or are found in a limited
number of texts only. If the latter, you may need to be aware that that
language feature is probably used in an idiosyncratic way; that is, it
is used only by one or two participants or in only a few of the texts
(depending on what your unit of analysis is).
8.4 Functional Interpretation

As Biber and Conrad (2009) indicate, the final step in a register analy-
sis is the functional interpretation of the results. We mentioned this in
Chapter 2; the basic principle here is that language is used in situations
and for communicative purposes. Put in another way, language forms
have functions and are used to meet the communicative purpose in a
given situation. Aspects of different situations can vary, and so do the
language forms associated with those situational aspects. According to
Biber and Conrad (2009: 64), after we describe the patterns (see above
in our mini-project), we need to interpret why the patterns exist the way
they do. This step requires reference to both the situational and linguistic
characteristics of a text.
For the specific dataset above, we have described the patterns that we
observed. The fact that no four-grams are shared by the two groups (i.e.,
in the two types of texts) could be attributed to certain differences in
the situational characteristics; namely the topic, the production circum-
stances, and relations among participants. Although both groups wrote
problem solution paragraphs, the topics they chose were different; it is
possible that these four-grams are topic related. However, the n-grams
in the individual texts do not seem to be topic related; they are more
focused on providing solutions to the problem. The collaborative group
did have four-grams that mentioned specific problems, but since the
individual group did not do this, it is difficult to see how the topic might
have influenced the individual writers to mention solutions and the col-
laborative writers to mention specific problems. Since the time was the
same in both the individual and collaborative assignments, the fact that
the collaborative group used more types of four-grams cannot be attrib-
uted to the time they may have had to complete the work. The difference
might be attributed to another production circumstance: the difference
between an exam and an in-class activity. While the first group wrote the
essays individually in an exam condition, the collaborative group com-
pleted the work as an in-class activity. This is a possible explanation, but
it seems more likely that the differences have to do with the production
circumstances related to individual and collaborative writing.
The production circumstances, as one of the situational variables,
may also be the reason that one group used verbs, while the other used
152 Your Own Corpus
prepositions. In previous register variation studies (Biber 1988, 1995),

particular linguistic features were associated with involved discourse,
and others were associated with informational discourse. For example,
and among many others, features such as pronouns and contractions
were found to be characteristic to discourse that is produced in real
time (i.e., under time constraints), as in face-to-face conversations (one
of the most involved types of discourse). In contrast, nouns, preposi-
tions, and attributive adjectives were found to be characteristic to dis-
course where the text could be revised and reformulated, as in academic
prose (one of the most informational types of discourse). In addition,
results of researching lexical bundles (see definition in Chapter 3) show
that conversation has more bundles that contain verbs, while academic
prose packages information through more bundles that contain prepo-
sitional phrases (Biber et al. 1999). This pattern is present in our data-
set above as well, but instead of being attributed to time constraints,
it may be attributed to the fact that collaborative writing influences
writers to adopt more forms found in academic writing. That is, when
the participants produce language that is the product of collaboration
(where they need to negotiate aspects of the text such as motivating a
problem, structuring their essay, and working on sentence-level gram-
mar and word choice), they seem to use more prepositional phrases.
When these same writers produce written texts by themselves, they use
more verbs.
8.5 Reporting on Your Project

The previous three sections of this chapter have led you through the
steps of analyzing your corpus using a register perspective. At this point,
you have a well-structured corpus that is guided by a well-motivated
research question (see chapter 5, section 5.2). You also have a clear
methodology for searching for and interpreting your results. It is now
time to package your research project so that others can learn about
your work. Below, we provide template for your research paper and
include some questions that you can use to guide your research. There
are five general parts to a research paper: 1) establishing the research
context and significance of the study; 2) introduction and explanation
of your data and methodology you used in the study; 3) your results;
4) a discussion of your results; and 5) your conclusion. (Sometimes the
results and discussion are found in a single section of the paper but
we place them in different sections here.) When writing a paper of this
type, you may want to think about writing a small paper for each of
these five sections and then putting these sections together for the final
product. At the end of this chapter, we also provide you with a sample
rubric that can guide your project and give you an idea of how your
project can be assessed (see Table 8.6).
8.5.1 Parts of a Research Paper (and Guiding Questions)

1. Research Context and Significance
What is the research issue?
• What is the rationale of the current research?
• Why was it important to conduct the research?
• Is the statement of the problem adequate and convincing?
What other research studies were conducted in the same area?
• What were the main findings?
• What are the research questions in your study?
What are your situational variables?
2. Data and methodology
Description of the corpus
• Number of texts; number of words
• Source of the texts
• Sub-corpora description (if applicable)
• File naming system
Type of corpus research (corpus-driven; corpus-based; combination)
and methods used
• Software programs used
• Search terms
• Other methods (e.g., n-grams; word lists)
• Other coding processes (hand coded/counted)
• Normalization
Linguistic features analyzed
3. Results
What are the main findings of the study?
• Charts/graphs/prose presenting what you found in the corpus
4. Discussion of results
What do you conclude from the findings?
How do the findings relate to your research issue?
What are the implications of the findings?
5. Conclusion
Are the results logically drawn from the analysis?
Are the conclusion, implications, and recommendations justified by
the results?
What are the limitations of the study and why do you think that they
are limitations?
154 Your Own Corpus
8.5.2 Research Presentation

In addition to the research paper, you may also be asked to give an oral
presentation of your research project. For example, you may be required to
present your work in 10 minutes, leaving 5 minutes for questions at the end.
Your presentation should be accompanied by a visual aid such as a Power-
Point slide presentation and/or a one-page handout. Should you choose to
do a PowerPoint presentation, do not use too much text in your slides. Each
PowerPoint slide should have no more than six lines and each line should
contain no more than six words. The slides serve as an outline for your pre-
sentation. You should expect that your presentation will be evaluated using
the following criteria:
Description of your problem/research issues: Explanation of why

your issue is important/a real world problem
A description of your corpus: Size, number of texts, how it is structured
A description of how you analyzed the corpus: Search terms, com-
mands to the software program you have used
Some results and analysis
A (tentative) conclusion
Format and clarity of your visual aid
Table 8.6 Rubric to evaluate student presentations
Criteria Level 1 Level 2 Level 3 Points
Introduction / Weak introduction Adequate Proficient /4

thesis of topic, thesis and introduction introduction
subtopics that states topic, that states
Thesis is weak and thesis, and some topic, thesis,
lacks an arguable of the subtopics and all
position Thesis is somewhat subtopics in
clear and arguable proper order
Thesis is a clear
and arguable
statement of
position
Quality of Limited information Some aspects of Corpus is clearly /7
information on corpus with lack the corpus are described
on corpus of details or lack of well-described with detailed
representativeness but others are and accurate
lacking (# of description in a
texts, # of words, table format
description of
sub-corpora, # of
words in corpus/
sub corpus)
Criteria Level 1 Level 2 Level 3 Points

Support of Limited connections Some connections Consistent /10
ideas / made between made between connections
analysis evidence, subtopics, evidence, made between
and thesis/topic subtopics, and evidence,
Lack of analysis thesis/topic subtopics, and
Showing analysis thesis/topic
Showing good
analysis
Organization / Paper lacks clear and Somewhat clear Clear and logical /10
development logical development and logical subtopic
of ideas of ideas with development order that
weak transition of subtopics supports thesis
between ideas and with adequate with good
paragraphs transitions transitions
between between
paragraphs paragraphs
Conclusion Lack of summary of Adequate summary Good summary /4
topic/thesis and of topic/thesis of topic/
subtopics with and some thesis and
weak concluding subtopics with all subtopics
ideas some final with clear
concluding ideas concluding
ideas
Language Inconsistent grammar, Paper has some Paper is clear, /8
conventions spelling, and errors in with proper
paragraphing grammar, spelling, grammar,
throughout paper and paragraphing spelling, and
paragraphing
Citation style Lack of proper Some errors in APA Proper APA /2
format and limited format with most format used
details with sources shown in alphabetical
sources missing or order with all
incomplete sources shown
Total /45
References
University Press
Biber, D. (1995) Dimensions of Register Variation: A Cross-linguistic Compari-
son, Cambridge: Cambridge University Press
University Press
Biber, D., S. Johansson, G. Leech, Conrad, S. & E. Finegan (1999) Longman
Grammar of Spoken and Written English, New York: Longmans
Chapter 9
A Way Forward
Beyond simply illustrating how searches can be done with a corpus, the
purpose of this book is to show how a complete corpus-based project can
be carried out, including some of the technical aspects and some basic sta-
tistical analyses. As we have discussed many times throughout the book,
one can follow a) a corpus-driven or b) a corpus-based approach to the
linguistic analysis. In this book, we made an attempt to illustrate both.
We showed how we can do a corpus-based study with already-identified
language features, whether doing a lexical study or searching for the use
of particular grammatical patterns. We have also illustrated the notion
of a corpus-driven study, as we extracted lexical items (n-grams) from a
small corpus. While it is relatively easy to carry out lexical studies with
corpus-driven approaches (whether you rely on existing corpora or ana-
lyze your own corpus), as available tools allow you to extract lexical pat-
terns from the texts, it is quite difficult to apply corpus-driven approaches
to do a full lexico-grammatical analysis of texts. The main reason for this
difficulty related to the fact that texts need to be grammatically tagged so
that grammatical categories can be extracted from corpora in the same
way that specific lexical items are. Tagged corpora cannot only include
specific types of grammatical items (such as nouns, verbs and adjectives)
but also sub-categories in these different word types (such as concrete
or abstract nouns, private and suasive verbs, attributive and predicative
adjectives). Some tagging software is available, but there is an increasing
need for corpus researchers to gain computational and statistical skills to
carry out more in-depth analyses. If you don’t have such skills (yet), per-
haps the best solution is to continue doing corpus-based studies. You can
continue to look for lexico-grammatical patterns that you find interest-
ing, or you can carry out corpus-based studies that rely on the results of
previous, corpus-driven studies. For the latter, you would use the findings
and apply them to new datasets. Without access to tagged corpora and
advanced programming and statistical knowledge, your corpus-driven
research will be limited to focusing on word lists and n-grams.
A Way Forward 157
Should you choose to expand your corpus linguistic skills, we pres-

ent below some influential register studies that a corpus-driven approach
can offer with the goal to provide comprehensive linguistic characteriza-
tions of texts in different registers. The purpose of the brief description
is to point you to a way forward if you became interested in this type of
research.
Unlike the corpus-driven approaches illustrated in this book, more
advanced corpus-driven register studies have identified co-occurring
linguistic features that have emerged through corpus analyses. While
researchers do rely on earlier work to identify functional categories and
their associated features before a corpus investigation (Biber and Conrad
2009: 63), when we are investigating the linguistic profile of a text, it is
a mistake not to rely on the constellation of linguistic features. Only this
approach would provide us with a comprehensive analysis (as mentioned
in Chapter 2). Typically, a large number of lexico-grammatical features
are counted at once, and through sophisticated, multivariate statistical
methods we are able to see correlations across all of them. Through this
methodology we are able to see what linguistic features tend to co-occur
and then trace how they are marking different types of texts from differ-
ent registers.
When we do this kind of research from the beginning (instead of
using patterns already identified), the analytical framework apply-
ing this empirically based, statistical method to provide comprehensive
linguistic descriptions was developed by Biber (1988), and is coined as
“multi-dimensional analysis” of linguistic variation. Among many other
works (e.g., Biber 1995, Biber et al. 2002, Biber et al. 2004, etc.) in which
you can read more about the details of this methodology, we highly rec-
ommend Biber and Conrad’s 2009 book titled Register, Genre and Style.
Many studies have applied a multi-dimensional analytical framework
(Biber 1988, Biber 1995, Conrad and Biber 2001, Biber and Conrad
2009) as they describe language variation across registers. Typically,
and depending on the motivation for the study, researchers have used a
multi-dimensional analytical framework in two ways. On the one hand,
researchers use their own corpus to run a new factor analysis (a multi-
variate statistical method) to identify dimensions of linguistic variation in
their own datasets. Examples of this approach include studies that iden-
tify dimensions of linguistic variation across registers in English (Biber
1988) and in languages other than English (e.g., Somali and Korean by
Biber 1995), variation in student and adult speech and writing (Reppen
2001), or dimensions of variation in language use within just one register
(e.g., university classroom discourse in Csomay 2005).
On the other hand, researchers may also use an already-existing
model where the dimensions (and associated linguistic features and
158 Looking Ahead
communicative functions) have already been identified prior to the given

study. The purpose of these studies is to investigate how their own texts
place on the existing continuum of variation (and in this sense have a
corpus-informed component to them). Examples of these types of studies
include those that use an existing dimensional framework, typically using
Biber’s (1988) study, to look at, for example, “register evolution from a
historical perspective (Atkinson 2001; Biber and Finegan 2001), varia-
tion in language use as it relates to specialized domains such as, author’s
style (Connor-Linton 2001), disciplinary language use (Conrad 2001),
intra-textual patterns in medical writing (Biber and Finegan 2001), or
dialect variation (Rey 2001; Biber and Burges 2001; Helt 2001)” (Cso-
may, 2015: 6).
The “way forward” that we have outlined here is most reflective of our
own interests and biases in corpus studies and we should end this book
by acknowledging many other corpus approaches that are not directly
involved in register variation. Some examples can be found in the areas
of Learner Corpora (http://www.learnercorpusassociation.org/), transla-
tion studies (http://corpus.leeds.ac.uk/), forensic linguistics (http://www.
forensiclinguistics.net/research.html), and the use of corpora in Natural
Language Processing (http://www-nlp.stanford.edu/links/statnlp.html).
These are just a few examples of the many different applications of cor-
pora. We invite the reader, armed with the basic information introduced
and practiced in this book, to pursue the areas of corpus linguistics that
meet their areas of interest.
References
Atkinson, D. (2001) ‘Scientific discourse across history: A combined multidimen-
sional / rhetorical analysis of The Philosophical Transactions of the Royal
Society of London’, in Conrad, S. & D. Biber (eds) 2001: 45–65
University Press
Biber, D. (1995) Dimensions of Register Variation: A cross-linguistic comparison,
Cambridge: Cambridge University Press
Biber, D. & J. Burges (2001) ‘Historical shifts in the language of women and
men’, in Conrad, S. and D. Biber (eds) 2001: 21–37
Biber, D. & E. Finnegan (2001) ‘Intra-textual variation within medical research
articles’, in Conrad, S. & D. Biber (eds) 2001: 108–123
University Press
Biber, D., S. Conrad, R. Reppen, P. Byrd & M. Helt (2002) ‘Speaking and writing
in the university: A multidimensional comparison’, TESOL Quarterly 36: 9–48
A Way Forward 159
Connor-Linton, J. (2001) ‘Author’s style and worldview: a comparison of texts

about American nuclear arms policy’, in Conrad, S. & D. Biber (eds) 2001:
84–93
Conrad, S. (2001) ‘Variation among disciplinary texts: A comparison of text-
books and journal articles in biology and history’, in Conrad, S. & D. Biber
(eds) 2001: 94–107
Conrad, S. and D. Biber (eds) (2001) Variation in English: Multidimensional
Studies, London: Longman
Cortes, V. & E. Csomay (eds) (2015) Corpus-based Research in Applied Linguis-
tics. Studies in Honor of Doug Biber. Amsterdam: Benjamins
Csomay, E. (2005) Linguistic variation within university classroom talk:
A corpus-based perspective. Linguistics and Education 15/3: 243–274
Csomay, E. (2015) ‘A corpus-based analysis of linguistic variation in teacher and
student presentations in university settings’ in Cortes, V. & E. Csomay (eds)
2015: 1–24
Helt, M. (2001) ‘A multi-dimensional comparison of British and American spo-
ken English’, in Conrad, S. & D. Biber (eds), 2001: 171–183
Reppen, R. (2001) ‘Register variation in student and adult speech and writing’, in
Conrad, S. & D. Biber (eds), 2001: 187–199
Rey, R. (2001) ‘Changing gender roles in popular culture: dialogue from Star
Trek episodes from 1966 to 1993’, in Conrad, S. and D. Biber (eds) 2001:
138–156
Index
Note: page numbers in italics indicate figures
academic writing 14 35 – 40; n-grams and 41 – 53;

active voice 66, 67 part of speech (POS) tags 53 – 6;
adjectives, connotations of 70 – 1 research projects, deciding on
alternative hypotheses 107 76 – 9; searching 35 – 56; software
analysis of variance 117 – 31 programs and 85 – 92; tri-grams and
AntConc 87 – 92, 145; concordance 47 – 9; uni-grams and individual
lines in 88; file view in 89; n-grams words 42 – 7; vocabulary and
in 92; sorting in 88, 89; tags in 84; academic prose 46 – 7; Webtext
word list in 92 bundles and 53
Anthony, Laurence 85 corpus-based method 9, 146–8, 156 – 8
AntWord Profiler 86 – 7 corpus-driven method 11, 145, 148,
156 – 8
Biber, Douglas 16 corpus linguistics 5 – 11, 142 – 55;
boxplots 100 characteristics of 8 – 9; corpus-based
British National Corpus (BYU-BNC) method 146 – 8, 156 – 8;
58, 62, 80 corpus-driven method 145, 148,
156 – 8; functional interpretation
central tendency, measures of 98, 151 – 2; language variation and 5;
99 – 102 linguistic analysis 145 – 51; register
Chi-square tests 131 – 5 analysis and 142 – 3; research paper
clusters/n-grams function 145 – 6 and guiding questions 153 – 4;
COCA see Corpus of Contemporary research presentation 154 – 5;
American English situational analysis and 143 – 5;
collective nouns 65 spoken language and 6; written
collocates: bi-grams and 47; in language and 5 – 6
concordance lines 91; corpus and Corpus Linguistics at Work
40 – 1 (Tognini-Bonelli) 9
control variables 96 Corpus Linguistics: Investigating
corpus: bi-grams and 47; building Language, Structure and Use (Biber,
79 – 85; collocates and 40 – 1; Conrad and Reppen) 8
definition of 6; do-it-yourself Corpus of Contemporary American
75 – 6; five- and more-grams 52 – 3; English (COCA) 6, 7, 9, 35, 41 – 2,
four-grams and lexical bundles 44, 58 – 60, 62, 80
49 – 52; juxtaposition, frequency of Corpus of Historical American
43 – 6; keyword in context (KWIC) English (COHA) 58, 63 – 4
162 Index
degree of freedom 121, 123, 124, 135 measures of variability 102 – 5

dependent variables 95 median 99 – 100
descriptive statistics 94, 114 – 16 Michigan Corpus of Academic Spoken
dispersion, measures of 102 – 5 English (MICASE) 28
do-it-yourself corpus 75 – 6 Michigan Corpus of Upper-level
Student Papers (MICUSP) 28
editorial news writing 20 modal verb 67 – 8
mode 99
face-to-face conversation 19 – 20 moderator variables 96
F score 124 – 5
functional interpretation, register n-grams, corpus and 41 – 53
analysis and 24 – 8 nominal scales 96
non-finite grammatical clauses 68 – 9
generalizability 94 non-parametric tests 105
gestural systems 19 normal distribution 102
Global Web-Based English corpus 58 null hypothesis 107 – 8
GloWbE corpus 59
grammar-based projects, corpora observations 97 – 8
64 – 71; adjectives, connotations of one-tailed test/hypothesis 108 – 9
70 – 1; collective nouns, grammatical one-way ANOVA 117 – 28, 140;
agreement 64 – 6; grammatical assumptions and requirements
categories and coordination 69 – 70; with 118
modal verb 67 – 8; non-finite ordinal scales 96
grammatical clauses 68 – 9; passive
voice, variation in 66 – 7 parametric tests 105 – 6
part of speech (POS) tags 53 – 6;
Hargraves, Owen 61 specifying for word 54 – 6
Historical Corpus of American News passive voice, variation in 66 – 7
Writing 83 Pearson correlations 136 – 41
Hypertext Mark-up Language probability value 108
(HTML) 80 – 1 publically available corpora:
hypotheses 106 – 9 grammar-based projects 64 – 71;
projects using 58 – 71; word- and
independent variables 95 phrase-based projects 60 – 4
interval scales 97
It’s Been Said Before: A Guide to quartile measure 102
the Use and Abuse of Clichés
(Hargraves) 61 range 102
ratio 97
keyword in context (KWIC) 35 – 40 register 14 – 15
keywords: definition of 42; register analysis 16 – 17, 142 – 3;
distribution 88, 90 components of 16; definition of
KWIC see keyword in context 13; functional interpretation 24 – 8;
genre perspective and 16 – 17;
language 3 – 5; in corpus 8; grammar situational characteristics of
rules and 4; and rules/systems 3 – 5 17 – 24; social factors and 15; style
linguistic variable 16 perspective and 17; units of analysis
and 28 – 9; variables, identifying
mean 100 and 17 – 24; variation 15; vs.
measures of central tendency 98, sociolinguistics 15
99 – 102 Register, Genre and Style (Biber and
measures of dispersion 102 – 5 Conrad) 157
Index 163
regular variables 95 tabs in 111; measures in 113; value

research questions 106 labels in 112; variable view in 111
Romeo and Juliet (Shakespeare) 17 statistical tests 117 – 41; Chi-square
tests 131 – 5, 140; correlations
situational analysis 143 – 5 136 – 41; difference tests 117 – 31,
situational analysis, register and 140 – 1; one-way ANOVA 117 – 28,
17 – 24; classroom vs. symposium 140; Pearson correlations 136 – 41;
presentations 22 – 4; email to friend relationship tests 131 – 41; two-way
and boss 21; news writing and news ANOVA 128 – 31, 140
talk 21 – 2 Strathy Corpus 58
situational variable 16, 18, 143, 144 strength of association 125 – 6
sociolinguistics 14, 15, 29 subject-verb agreement 64, 65
software programs, corpus and sum of squares 121 – 2
85 – 92; AntConc 87 – 92; AntWord
Profiler 86 – 7 TIME Magazine Corpus 58, 59
specialized corpora 79 T2KSWAL corpus 28
speech situation 17 Tognini-Bonelli, Elena 9
split infinitives 7 two-tailed test/hypothesis 108
spoken language 6, 8 – 9 two-way ANOVA 128 – 31, 140
standard deviation 103 – 4
statistical analyses 94; basic terms, variability, measures of 102 – 5
concepts, and assumptions 94 – 109; variables 95 – 8; scales 96 – 7; types
central tendency, measures of and functions 95 – 6; values
98, 99 – 102; data and descriptive (levels) 97
statistics 109 – 16; observations variance 102 – 3; analysis of 117 – 31
and 97 – 8; parametric and
non-parametric tests 105 – 6; research word- and phrase-based projects,
questions and hypotheses 106 – 9; corpora 60 – 4; clichés 61 – 2;
statistical package for social sciences collocation, of modifying elements
109 – 14; variability and dispersion, 62 – 3; gate suffix addition 61;
measures of 102 – 5; variables 95 – 7 lexical change, time and 60;
statistical package for social sciences literally, meaning of 60 – 1;
(SPSS) 109 – 14; adding variables in sustainability 63 – 4
113; data view in 114; descriptive written forms of language 19
statistics 114 – 16, 115; important written language 5 – 6, 14

Doing Corpus Linguistics

Uploaded by

Copyright:

Available Formats

Doing Corpus Linguistics

Uploaded by

Document Information

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

Doing Corpus Linguistics

Uploaded by

Copyright:

Available Formats

Doing Corpus Linguistics

Doing Corpus Linguistics offers a practical step-by-step introduction to

William J. Crawford is an Associate Professor in the Department of Eng-

Eniko Csomay is a Professor in the Department of Linguistics and Asian/

William J. Crawford and

List of Tables vii

1 Linguistics, Corpus Linguistics, and Language Variation 3

4 Projects Using Publically Available Corpora 58

5 Building Your Own Corpus 75

7 Statistical Tests (ANOVAs, Chi-square, Pearson Correlation) 117

8 Doing Corpus Linguistics 142

9 A Way Forward 156

1.1 Common right collocates of equal and identical in the

6.3 Descriptive statistics for ‘hedges’ in three disciplines

2.1 Situational variables 18

6.4 Value labels in SPSS 112

In our experiences teaching introductory corpus linguistics classes, we

1.1 Language and Rules/Systems

1.1 Language and Rules/Systems

A prescriptive rule can provide useful guidelines for effective commu-

1.2 What Is Corpus Linguistics?

would expect, therefore, some degree of language variation in these differ-

be able to make such an observation. This is one example of how corpus

First word right Number First word Number

Equal OPPORTUNITY 950 Identical TWINS 417

We have already seen a few examples of what corpus information can

• it is empirical, analyzing the actual patterns of use in natural

In our discussion above, we have already seen examples of the first

(Part of ) a newspaper text A corpus of newspaper texts

The site appeared to be working again late

1.3 Outline of the Book

2.1 Why Register?

As we have discussed in Chapter 1, language variation is one prevalent

2.1 Why Register?

2.2 What Is Register Analysis?

2.3 Describing Situational Characteristics

Figure 2.1 provides an overview of the situational framework and pro-

Figure 2.1 Situational variables

following short description of each component should help you to under-

them. In addition to the participants themselves, another relevant situ-

events. A face-to-face conversation involves a shared physical space but

2.3.1 Email to a Friend and Email to a Boss

Email to a friend Email to a boss

Participants One author addressing One author addressing

2.3.2 News Writing and News Talk

News writing News talk show

Participants One addressor, multiple One addressor, multiple

(potentially new) information to an audience (under Communicative Pur-

2.3.3 Classroom Versus Symposium Presentations

point in time during the teacher’s presentation, presentation settings have

Classroom presentation Symposium presentation

Participants One addressor, multiple One addressor, multiple

Classroom presentation Symposium presentation

All participants have some All participants have some

In sum, there are enough situational differences in some of the aspects

2.4 Providing a Functional Interpretation

requires a description of the context. In fact, from this perspective, lan-

2.4.1 News Writing and News Talk

News writing News talk show

2.4.2 Classroom Versus Symposium Presentation