Quantitative Methods in the Humanities

and Social Sciences

Matthew L. Jockers

Text Analysis with

R for Students of
Literature

Text Analysis with R for Students of Literature

Quantitative Methods in the Humanities and Social Sciences

Editorial Board
Thomas DeFanti, Anthony Grafton, Thomas E. Levy, Lev Manovich,
Alyn Rockwood

Quantitative Methods in the Humanities and Social Sciences is a book series

designed to foster research-based conversation with all parts of the university campus from buildings of ivy-covered stone to technologically savvy
walls of glass. Scholarship from international researchers and the esteemed
editorial board represents the far-reaching applications of computational
analysis, statistical models, computer-based programs, and other quantitative methods. Methods are integrated in a dialogue that is sensitive to the
broader context of humanistic study and social science research. Scholars,
including among others historians, archaeologists, classicists and linguists,
promote this interdisciplinary approach. These texts teach new methodological approaches for contemporary research. Each volume exposes readers
to a particular research method. Researchers and students then benefit from
exposure to subtleties of the larger project or corpus of work in which the
quantitative methods come to fruition.

For further volumes:

http://www.springer.com/series/11748

Matthew L. Jockers

Text Analysis with R for

Students of Literature

123

Matthew L. Jockers
Department of English
University of Nebraska
Lincoln, Nebraska, USA

ISBN 978-3-319-03163-7
ISBN 978-3-319-03164-4 (eBook)
DOI 10.1007/978-3-319-03164-4
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014935151
Springer International Publishing Switzerland 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publishers location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

For my mother,
who prefers to follow the instructions

Preface

This book provides an introduction to computational text analysis using the open
source programming language R. Unlike other very good books on the use of R for
the statistical analysis of linguistic data1 or for conducting quantitative corpus linguistics,2 this book is meant for students and scholars of literature and then, more
generally, for humanists wishing to extend their methodological toolkit to include
quantitative and computational approaches to the study of text. This book is also
meant to be short and to the point. R is a complex program that no single textbook can demystify. The focus here is on making the technical palatable and more
importantly making the technical useful and immediately rewarding! Here I mean
rewarding not in the sense of satisfaction one gets from mastering a programming
language, but rewarding specifically in the sense of quick return on your investment.
You will begin analyzing and processing text right away and each chapter will walk
you through a new technique or process.
Computation provides access to information in texts that we simply cannot gather
using our traditionally qualitative methods of close reading and human synthesis.
The reward comes in being able to access that information at both the micro and
macro scale. If this book succeeds, you finish it with a foundation, with a broad
exposure to core techniques and a basic understanding of the possibilities. The real
learning will begin when you put this book aside and build a project of your own. My
aim is to give you enough background so that you can begin that project comfortably
and so that youll be able to continue to learn and educate yourself.
When discussing my work as a computing humanist, I am frequently asked
whether the methods and approaches I advocate succeed in bringing new knowledge
to our study of literature. My answer is strong and resounding yes. At the same time,
that strong yes must be qualified a bit; not everything that text analysis reveals is a
breakthrough discovery. A good deal of computational work is specifically aimed
1

Baayen, H. A. Analyzing Linguistic Data: A Practical Introduction to Statistics Using R. Cambridge UP, 2008.
2 Gries, Stefan Th. Quantitative Corpus Linguistics with R: A Practical Introduction. New York:
Routledge, 2009.
vii

viii

Preface

at testing, rejecting, or reconfirming the knowledge that we think we already possess. During a lecture about macro-patterns of literary style in the nineteenth century
novel, I used an example from Moby Dick. I showed how Moby Dick is a statistical
mutant among a corpus of 1,000 other nineteenth century American novels. A colleague raised his hand and pointed out that literary scholars already know that Moby
Dick is an aberration, so why, he asked, bother computing an answer to a question
we already know?
My colleagues question was good; it was also revealing. The question said much
about our scholarly traditions in the humanities. It is, at the same time, an ironic
question: as a discipline, we have tended to favor a notion that literary arguments
are never closed. Do we really know that Moby Dick is an aberration? Maybe Moby
Dick is only an outlier in comparison with the other 20 or 30 American novels that
we have traditionally studied alongside Moby Dick. My point in using Moby Dick
was not to pretend that I had discovered something new about the position of the
novel in the American literary tradition, but rather to bring a new type of evidence
and a new perspective to the matter and in so doing fortify (in this case) the existing
hypothesis.
If a new type of evidence happens to confirm what we have come to believe
using far more speculative methods, shouldnt that new evidence be viewed as a
good thing? If the latest Mars rover returns more evidence that the planet could
have once supported life, that new evidence would be important. Albeit it would not
be as shocking or exciting as the first discovery of microbes on Mars, or the first
discovery of ice on Mars, but it would be an important evidence nevertheless, and
it would add one more piece to a larger puzzle. So, computational approaches to
literary study can provide complementary evidence, and I think that is a good thing.
The approaches outlined in this book also have the potential to present contradictory evidence, evidence that challenges our traditional, impressionistic, or anecdotal
theories. In this sense, the methods provide us with some opportunity for the kind of
falsification that Karl Popper and post-positivism in general offer as a compromise
between strict positivism and strict relativism. But just because these methods can
provide contradiction, we must not get caught up in a numbers game where we only
value the testable ideas. Some interpretations lend themselves to computational or
quantitative testing; others do not, and I think that is a good thing.
Finally, these methods can lead to genuinely new discoveries. Computational text
analysis has a way of bringing into our field of view certain details and qualities of
texts that we would miss with just the naked eye.3 Using computational techniques,
Patrick Juola recently discovered that J. K. Rowling was the real author of The
Cuckoos Calling a book Rowling wrote under the pseudonym Robert Galbraith.
Naturally, I think Joulas discovery is a good thing too.

3 See Flanders, Julia. Detailism, Digital Texts, and the Problem of Pedantry. TEXT Technology,
2:2005, 4170.

Preface

ix

This is all I have to say regarding a theory for or justification of text analysis.
In my other book, Im a bit more polemical.4 The mission here is not to defend the
approaches but to share them.
Lincoln, NE
January 2014

Matthew L. Jockers

Jockers, Matthew. Macroanalysis: Digital Methods and Literary History. UIUC Press, 2013.

Acknowledgments

For many years I taught text analysis courses using a combination of tools and different programming languages. For text parsing, I taught students to use Perl,
Python, php, Java, and even XSLT. For analysis of the resulting data, we often
used Excel. In about 2005, largely at the prompting of Claudia Engel and Daniela
Witten who both offered me some useful, beginner level advice, I began using R,
instead of Excel. For too long after that I was still writing a lot of text analysis code
in Perl, Python, or php and then importing the results into R for analysis. In
2008 I decided this workflow was unsustainable. I was spending way too much time
moving the data from one environment to another. I decided to go cold turkey and
give up everything in favor of R. Since making the transition, Ive rarely had to look
elsewhere.
Luckily, just as I was making the transition to R so too were thousands of other
folks; the online community of R programmers and developers was expanding at
exactly the moment that I needed them. Todays online R-help resources are outstanding, and I could not have written this book without them. There are some phenomenal R programmers out there making some incredibly useful packages. Only
a small handful of these packages are mentioned in this book (this is, after all, a
beginners guide) but without folks such as Stefan Th. Gries, Harald Baayen, and
Hadley Wickham, this book and the R community would be very much the poorer.
Im amazed at how helpful and friendly the online R community has become; it
wasnt that way in the early years, so I want to thank all of you who write packages
and contribute code to the R project and also all of you who offer advice on the R
forums, especially on the R-help list and at stackoverflow.com.
This book began as a series of exercises for students I was teaching at Stanford;
they saw much of this content in a raw and less polished form. There are too many
of them to thank individually, so a collective thanks to all of my former and current
students for your patience and feedback. This book, whatever its faults and shortcomings, is a far better book than it might have been without you.
I first compiled the material from my classes into a manuscript in 2011, and since
then I have shared bits and pieces of this text with a few colleagues. Stfan Sinclair
test drove an early version of this book in a course he taught at McGill. He and his
xi

xii

Acknowledgments

students provided valuable feedback. Maxim Romanov read most of this manuscript
in early 2013. He provided encouragement and feedback and ultimately convinced
me to convert the manuscript to LaTeX for better typesetting. This eventually led
me to Sweave and Knitr: two R packages that allowed me to embed and run R
code from within this very manuscript. So here again, my thanks to Maxim, and
also to the authors of Sweave and Knitr, Friedrich Leisch and Yihui Xie.5 I also
want to acknowledge and thank those who downloaded the draft of this manuscript
that I posted on my website in August 2013. Those who wrote back to me with
feedback and corrections are noted on the contributors page that follows. 6 Finally, I
thank my 14 year old son who entered and executed every line of code in this book.
While this text may not be so easy that a caveman can use it, I do know that a sharp
middle-schooler can follow along without too much coaching.

Friedrich Leisch. Sweave: Dynamic generation of statistical reports using literate data analysis.
In Wolfgang Hrdle and Bernd Rnz, editors, Compstat 2002 - Proceedings in Computational
Statistics, pages 575580. Physica Verlag, Heidelberg, 2002. ISBN 3-7908-1517-9. Yihui Xie.
Dynamic Documents with R and knitr. Chapman and Hall/CRC. ISBN 978-1482203530, 2013
6 The pre-publication draft of this text that was posted online for open review included the following two paragraphs:
Finally, I want to thank you for grabbing a copy of this pre-print e-text and working your way
through it. I hope that you will offer feedback and that you will help me make the final print
version as good as possible. And when you do offer that feedback, Ill add your name to a list of
contributors to be included in the print and online editions. If you offer a substantial contribution,
Ill acknowledge it specifically.
I did not learn R by myself, and there is still a lot about R that I have to learn. I want to
acknowledge both of these facts directly and specifically by citing all of you who take the time to
contribute to this manuscript and to make R world a better place.

Contributors

A draft of this book was downloaded over 1,000 times after I posted it to my website
in August 2013. The readers listed below provided valuable feedback about the
manuscript, and I thank each of them for their contributions to the final manuscript.
All of their specific contributions are archived online at http://www.
matthewjockers.net/text-analysis-with-r-for-studentsof-literature/. The greatest number of contributions came from Charles
Shirley, who provided 133 comments. The most important contribution to the code
came from Carmen McCue, who discovered a nasty little bug in Chap. 4. My thanks
to you all.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.

Brotnov, Mikal
Francom, Jerid
Hawk, Brandon
Huber, Alexander
Johnson, Paul
Kumari, Ashanka
Laudun, John
Maenner, Matthew J.
McCue, Carmen
McMullen, Kevin
Pentecost, Stephen
Shirley, Charles
Tedrow, Kimberley
Wehrwein, Austin
Wolff, Mark
Xie, Yihui

xiii

Contents

Part I Microanalysis
1

R Basics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

First Foray into Text Analysis with R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

Accessing and Comparing Word Frequency Data . . . . . . . . . . . . . . . . . . 25

Token Distribution Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

Part II Mesoanalysis
6

Measures of Lexical Variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Hapax Richness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

Do It KWIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

Do It KWIC (Better) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

10

Text Quality, Text Variety, and Parsing XML . . . . . . . . . . . . . . . . . . . . . . . 89

Part III Macroanalysis

11

Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

12

Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

13

Topic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

xv

xvi

Contents

Variable Scope Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161

The LDA Buffet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Start up Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

R Resources for Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

Practice Exercise Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

Part I

Microanalysis

Chapter 1

R Basics

Abstract This chapter explains how to download and install R and RStudio.
Readers are introduced to the R console and shown how to execute basic commands.

1.1 Introduction
There is a tradition in the teaching of programming languages in which students
write a script to print out (to their screen) the words hello world. Though this book
is about programming in R, this is not a programming book. Instead this text is
designed to get you familiar with the R environment while engaging with, exploring,
and even addressing some real literary questions. If you are like me, you probably
launched R and started typing in commands a few hours ago. Maybe you got stuck,
hit the wall, and are now turning to this book for a little shove in the right direction.
If you are like me, you probably headed to the index first and tried to find some
function or keyword (such as frequency list or count word occurrences) to get
you rolling. You are ready to jump in and start working, and if youve ever done any
coding before, you may be wondering (maybe dreading) if this is going to be another
one of those books that grinds its way through all sorts of data type definitions and
then finally teaches you how to write an elegant little snippet of code with a tight
descriptive comment.
This is not that type of booknot that there is anything wrong with books that
are like that! This book is simply different; it is designed for the student and scholar
of literature who doesnt already know a programming language, or at the very least
does not know the R language, and more importantly is a person who has come to
R because of some literary question or due to some sense that computation might
offer a new or particularly useful way to address, explore, probe, or answer some
literary question. You are not coming to this book because you want to become a
master programmer. You are a student or scholar in the humanities seeking to learn
just enough to gain entry into the wide world of humanities computing.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_1,
Springer International Publishing Switzerland 2014

1 R Basics

If you want to become a master R programmer, or someone who delivers shrinkwrapped programs, or R packages, then this is not the book for you; there are other
books, and good ones, for that sort of thing.1 Here, however, Ill take my cues from
best practices in natural language instruction and begin with a healthy dose of full
immersion in the R programming language. In the first section, Microanalysis, I
will walk you through the steps necessary to complete some basic text analysis of a
single text. In the second part of the book, Mesoanalysis, well move from analysis
of one or two texts to analysis of a small corpus. In the final section, Macroanalysis,
well take on a larger corpus. Along the way there will be some new vocabulary and
even some new or unfamiliar characters for your alphabet. But all in good time. For
now, lets get programming. . . .

1.2 R and RStudio

If you thrive on the command line and like the Spartan world of a UNIX-like environment, download the current version of R: instructions follow below under the
heading Download and Install R. If youd like to work in a more developed GUIlike environment, follow the same instructions and then go directly to the section titled Download and Install RStudio. Throughout this text Ill assume that you are
working in the RStudio environment. You can work directly from the R console
or even from your computers terminal if you prefer, but RStudio is my recommended platform for writing and running R code, this is especially true for newbies.

1.3 Download and Install R

Download the current version of R (at the time of this writing version 3.0.2) from the
CRAN website by clicking on the link that is appropriate to your operating system
(see http://cran.at.r-project.org):
If you use MS Windows, click on base and then on the link to the executable
(i.e., .exe) setup file.
If you are running Mac OSX, choose the link to latest version that is compatible
with your system.2
If you use Linux, choose your distribution and then the installer file.
Follow the instructions for installing R on your system in the standard or default directory. You will now have the base installation of R on your system.
1 See, for example Venables, William N. and David M. Smith. An Introduction to R. Network
Theory Ltd., 2004; Braun W. John and Duncan J. Murdoch. A First Course in Statistical Programming with R. Cambridge University Press, 2008; or any of the books in Springers Use R Series:
http://www.springer.com/series/6991?detailsPage=titles
2 To find what OS you are running, choose About this Mac from the Apple menu.

1.5 Download the Supporting Materials

If you are using a Windows or Macintosh computer, you will find the R application in the directory on your system where Programs (Windows) or Applications (Macintosh) are stored. If you want to launch R, just double click the icon
to start the R GUI.
If you are on a Linux/Unix system, simply type R at the command line to
enter the R program environment.

1.4 Download and Install RStudio

The R application is fine for a lot of simple programming, but RStudio is an
application that offers an organized user environment for writing and running R
programs. RStudio is an IDE, thats Integrated Development Environment for
R. RStudio runs happily on Windows, Mac, and Linux. After you have downloaded R (by following the instructions above in Sect. 1.3) you can (and probably
should) download the Desktop version (i.e., not the Server version) of RStudio
from http://www.rstudio.com. Follow the installation instructions and then
launch RStudio just like you would any other program/application.3

1.5 Download the Supporting Materials

Now that you have R or R and RStudio installed and running on your system,
you will also need to download the directory of files used for the exercises and examples in this book. The materials consist of a directory titled TextAnalysisWithR
that includes the following sub-directories: a directory titled data containing a
sub-directory of sample texts in plain text and another for texts in XML format, a
directory labeled code for saving your R code, and a directory titled results for
saving your derived results. You can download the supporting materials as a compressed zip file from the companion website:
http://www.matthewjockers.net/wp-content/uploads/2013/
08/TextAnalysisWithR.zip.
Unzip the file and save the resulting directory (aka folder) to a convenient
location on your system: If you are using a Mac, the file path to this new directory
might look something similar to the following:
~/Documents/TextAnalysisWithR

It does not matter where you keep this new directory as long as you remember
where you put it. In the R code that you write, you will need to include information
that tells R where to find these materials. All of the examples in this book place the
TextAnalysisWithR directory inside of the main /Documents directory.
3

See
http://www.matthewjockers.net/text-analysis-with-r-forstudents-of-literature/ for a short introductory video about RStudio.

1 R Basics

1.6 RStudio
In case it is not already clear, Im a big fan of RStudio. When you launch the
program you will see the default layout which includes three quadrants or panes
and within each of the panes you can have multiple tabs.4 You can customize this
pane/tab layout in RStudios preferences area. I set my layout up a bit different
from the default: I like to have the script editing pane in the upper right and the R
console pane in the lower right. You will discover what is most comfortable for you
as you begin to spend more time in the program.
The important point to make right now is that RStudios four window panes
each provide something useful and you should familiarize yourself with at least
two of these panes right away. These are the script editing pane and the console
pane. The former provides an environment in which you can write R programs. This
pane works just like a text editor but with the added benefit that it offers syntax
highlighting and some other shortcuts for interacting with R. As you become a more
experienced coder, you will learn to love the highlighting. RStudios script editor
understands the syntax of the R programming language and helps you read the code
you write by highlighting variables in one color and literal characters, comments,
and so on in other colors. If this does not make sense to you at this point, that is
fine. The benefits of syntax highlighting will become clear to you as we progress.
A second point about the script editing pane is that anything you write in that pane
can be saved to file. When you run commands in the R console, those commands do
not get saved into a file that you can reuse.5 When you write your code in the script
editor, you intentionally save this code as a .R file. You can then close and reopen
these files to run, revise, copy, etc.
Along with the scripting pane, RStudio provides a console pane. If you were
simply running R by itself, then this is all you would get: a simple console.6 In
RStudio you can interact with the console just as you would if you had only
launched R. You can enter commands at the R prompt (represented by a > symbol
at the left edge of the console), hit return, and see the results.
Because the scripting and console panes are integrated in RStudio, you can
write scripts in one pane and run them in the other without having to copy and
paste code from the editor into the console. RStudio provides several shortcuts
for running code directly from the script editing pane. Well discuss these and other
shortcuts later. For now just know that if you are working in the scripting pane, you
can hit control and return to send the active line of code (i.e., where your

Actually, the first time you launch RStudio you will only be able to see three of the panes. The
R scripting or Source pane will likely be collapsed so you will not see the word Source until you
either create a new script (File > New > R Script) or un-collapse the Source window pane.
5 This is not entirely true. RStudio does save your command history and, at least while your
session is active, you can access that history and even save it to a file. Once you quit a session,
however, that history may or may not be saved.
6 Console is a word used to refer to a command line window where you enter commands. Sometimes this is called a Terminal or Shell.

1.7 Lets Get Started

cursor is currently active) directly to the console. To run several lines of code, select
them in editing window and use the same key sequence.
Throughout this book, I will assume that you are writing all of your code in the
script editing pane and that you are saving your scripts to the code sub-directory
of the main TextAnalysisWithR directory you downloaded from my website.
To help get you started, Ill provide specific instructions for writing and saving your
files in this first chapter. After that, Ill assume you know what you are doing and
that you are saving your files along the way.7

1.7 Lets Get Started

If you have not already done so, launch RStudio (or launch R via the console
application or by opening your terminal application and issuing the command R
at your terminal prompt).
The first thing to do is to set TextAnalysisWithR as the working directory.
If you are using RStudio, you can save yourself some typing by going to the
Session menu in the menu bar and selecting Set Working Directory and then
Choose Directory (as in Fig. 1.1). You can then navigate to the TextAnalysis
WithR directory using your computers file system browser. Once you have set the
working directory in this manner, you can avoid entering full paths to the various
directories and resources in the main directory.

Fig. 1.1 Setting the working directory

If you decide to avoid using the RStudio IDE, you can use the setwd
function to tell R where to look for the files we will use in this book. Open
the R console application (Fig. 1.2) and type the expression below (replacing my
/Documents/TextAnalysisWithR with the actual path to your directory)
into the R console.8

7 If you do forget to save your file, you can always grab the textbook code from my online code
repository or copy from Appendix C.
8 Paths in R use forward slashes whether you are using Windows, Mac, or Linux.

1 R Basics
> setwd("~/Documents/TextAnalysisWithR")

The > symbol you see here is the R prompt, the place where you will enter
your code (that is, you dont actually enter the >). In future code examples, Ill not
include the R prompt >, which will make it easier for you to copy (and paste if you
are using the e-book) code exactly as it appears here into your own R files. After
entering the command, hit the return key.
If you are working directly from the command line in your computers terminal, see Fig. 1.3. From here on out all instructions and images will be relative to
RStudio.

Fig. 1.2 The R GUI console

Even though RStudio makes path setting and other common commands easy,
you should remember that these are just GUI shortcuts that allow you to avoid typing
commands directly into the R console. You can still always type directly into the
RStudio console pane and get the same result. To set a working directory path
in RStudio without using the GUI menu interface, just find the console pane and
look for the > prompt. At the prompt, use the setwd command as described above,
and you will have set the working directory without using RStudios more userfriendly GUI. It is good to know that you can always run R commands directly in
RStudios console pane; RStudio gives you all the functionality of the command
line as well as all the benefits of an IDE. Cool!

1.7 Lets Get Started

Fig. 1.3 R in the terminal

Practice
In order to gain some familiarity with the way that commands are entered into the R
console, open a new R script file. In RStudio go to the File menu and choose New
and then R Script (i.e. File > New File > R Script). Creating a new script file will
add a forth quadrant to the RStudio layout. Save this file to your code directory
as exercise1.R. Begin editing this new file by adding a call to the setwd function
as described above. Now type each of the following exercise problems into your
script. To compute the answers, you can either copy and paste each question into
the console, or use the control and return shortcut to interact with the console pane
from the script editor.
1.1. Simple addition and subtraction
10+5
10-5

1.2. An asterisk is used for multiplication

10*1576

1.3. A forward slash is used for division

15760/10

10

1 R Basics

1.4. R has some values that are preset.

10+pi
10/pi

1.5. R uses the carat () for exponents

10^2

1.6. R uses the less than symbol followed by the hyphen as an assignment operator.
(The two symbols form an icon that looks like a left facing arrow.) R uses the # symbol for comments; anything you enter after a # will be ignored by the R processor.
x <- 10 # Assign 10 to the variable "x"
x - 3 # subtract 3 from x

1.7. Beware of how R implements order in complex mathematical operations. The

following expression correctly evaluates to 12. Can you explain why?
x <- 10
x - 3 + 10 / 2

1.8. If x <- 10 and x - 3 + 10 / 2 = 12 How would you rewrite the expression to = 8.5?
1.9. R has some built-in mathematical functions:
sqrt(12)
abs(-23)
round(pi)

1.10. R can easily create sequences of numbers:

1:10
12:37

Chapter 2

First Foray into Text Analysis with R

Abstract In this chapter readers learn how to load, tokenize, and search a text.
Several methods for exploring word frequencies and lexical makeup are introduced.
The exercise at the end introduces the plot function.

2.1 Loading the First Text File

If you have not already done so, set the working directory to the location of the
supporting materials directory.1
setwd("~/Documents/TextAnalysisWithR")

You can now load the first text file using the scan function2:
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")

Type this command into a new R script file and then run it by either copying and
pasting into the console or using RStudios control and return shortcut.
This is as good a place as any to mention that the scan function can also load files
from the Internet. If you have an Internet connection, you can enter a url in place of
the file path and load a file directly from the web, like this:

1 Programming code is extremely finicky. If you do not type the commands exactly as they appear
here, you will likely get an error. In my experience about 95% of the errors and bugs one sees when
coding are the result of careless typing. If the program is not responding the way you expect or if
you are getting errors, check your typing. Everything counts: capital letters must be consistently
capitalized, commas between arguments must be out side of the quotes and so on.
2 Throughout this book I will use a naming convention when instantiating new R objects. In the
example seen here, I have named the object text.v. The .v extension is a convention I have
adopted to indicate the R data type of the object, in this case a vector object. This will make more
sense as you learn about Rs different data types. For now, just understand that you can name R
objects in ways that will make sense to you as a human reader of the code.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_2,
Springer International Publishing Switzerland 2014

11

12

2 First Foray into Text Analysis with R

text.v <- scan("http://www.gutenberg.org/cache/epub/2701/pg2701.txt",

what="character", sep="\n")

Whether you load the file from your own systemas you will do for the exercises
in this bookor from the Internet, if the code has executed correctly, you should
see the following result:
Read 18874 items
>

Remember that the > symbol seen here is simply a new R prompt indicating that R
has completed its operation and is ready for the next command. At the new prompt,
enter:
> text.v

You will see the entire text of Moby Dick flash before your eyes.3 Now try:
> text.v[1]

You will see the contents of the first item in the text.v variable, as follows4 :
[1] "The Project Gutenberg EBook of Moby Dick;
or The Whale, by Herman Melville"

When you used the scan function, you included an argument (sep) that told
the scan function to separate the file using \n. \n is a regular expression or metacharacter (that is, a kind of computer shorthand) representing (or standing in for)
the newline (carriage return) characters in a text file. What this means is that when
the scan function loaded the text, it broke the text up into chunks according to
where it found newlines in the text.5 These chunks were then stored in what is
called a vector, or more specifically a character vector.6 In this single R expression,
you invoked the scan function and put the results of that invocation into a new
object named text.v, and the text.v object is an object of the type known as a
character vector.7 Deep breath.
It is important to understand that the data inside this new text.v object is
indexed. Each line from the original text file has its own special container inside
the text.v object, and each container is numbered. You can access lines from the
3 Actually, you may only see the first 10,000 lines. Thats because R has set the max.print
option to 10,000 by default. If you wish to increase the default for a given work session, just begin
your session by entering options(max.print=1000000).
4 From this point forward, I will not show the R prompt in the code examples.
5 Be careful not to confuse newline with sentence break or even with paragraph. Also note that
depending on your computing environment, there may be differences between how your machine
interprets \n, \r and \n\r.
6 If you have programmed in another language, you may be familiar with this kind of data structure
as an array.
7 In this book and in my own coding I have found it convenient to append suffixes to my variable
names that indicate the type of data being stored in the variable. So, for example, text.v has a .v
suffix to indicate that the variable is a vector: v = vector. Later you will see data frame variables
with .df extensions and lists with a .l, and so on.

2.2 Separate Content from Metadata

13

original file by referencing their container or index number it within a set of square
brackets. Entering
text.v[1]

returns the contents of the first container, in this case, the first line of the text, that
is, the first part of the text file you loaded up to the first newline character. If you
enter text.v[2], youll retrieve the second chunk, that is, the chunk between the
first and second newline characters.8

2.2 Separate Content from Metadata

The electronic text that you just loaded into text.v is the plain text version of
Moby Dick that is available from Project Gutenberg. Along with the text of the
novel, however, you also get a lot of Project Gutenberg boilerplate metadata. Since
you do not want to analyze the boilerplate, you need to determine where the novel
begins and ends. As it happens, the main text of the novel begins at line 408 and
ends at 18576. One way to figure this out is to visually inspect the output you saw
(above) when you typed the object name text.v at the R prompt and hit return. If
you had lightning fast eyes, you might have noticed that text.v[408] contained
the text string: CHAPTER 1. Loomings. and text.v[18576] contained the
string orphan. Instead of removing all this boilerplate text in advance, I opted to
leave it so that we might explore ways of accessing the items of a character vector.
Lets assume that you did not already know about items 408 and 18576, but
that you did know that the first line of the book contained CHAPTER 1. Loomings
and that the last line of the book contained orphan. We could use this information
along with Rs which function to isolate the main text of the novel. To do this
on your own, enter the following code, but be advised that in R, and any other
programming language for that matter, accuracy counts. The great majority of the
errors you will encounter as you write and test your first programs will be the results
of careless typing.9
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")

In reality, of course, you are not likely to know in advance the exact contents
of the items that scan created by locating the newline characters in the file, and
the which function requires that you identify an exact match. Later on Ill show a
better way of handling this situation using the grep function. For now just pretend
that you know where all the lines begin and end. You can now see the line numbers
8

Those who have programming experience with another language may find it disorienting that R
(like FORTRAN) begins indexing at 1 and not 0 like C, JAVA and many other languages.
9 In these expressions, the two equal signs serve as a comparison operator. You cannot use a single
equals sign because the equals sign can also be used in R as an assignment operator. Ill clarify this
point later, so for now just know that you need to use two equals signs to compare values.

14

2 First Foray into Text Analysis with R

(the vector indices) returned by the which function by entering the following at the
prompt10:
start.v
end.v

You should now see the following returned from R:11

start.v
## [1] 408
end.v
## [1] 18576

In a moment we will use this information to separate the main text of the novel from
the metadata, but first a bit more about character vectors. . .
When you loaded Moby Dick into the text.v object, you asked R to divide
the text (or delimit it) using the carriage return or newline character, which was
represented using \n. To see how many newlines there are in the text.v variable,
and, thus, how many items, use the length function:
length(text.v)
## [1] 18874

Youll see that there are 18,874 lines of text in the file. But not all of these lines,
or strings, of text are part of the actual novel that you wish to analyze.12 Project
Gutenberg files come with some baggage, and so you will want to remove the nonMoby Dick material and keep the story: everything from Call me Ishmael. . . to
. . .orphan. You need to reduce the contents of the text.v variable to just the
lines of the novel proper. Rather than throwing out the metadata contained in the
Project Gutenberg boilerplate, save it to a new variable called metadata.v, and
then keep the text of the novel itself in a new variable called novel.lines.v.
Do this using the following four lines of code:
start.metadata.v <- text.v[1:start.v -1]
end.metadata.v <- text.v[(end.v+1):length(text.v)]
metadata.v <- c(start.metadata.v, end.metadata.v)
novel.lines.v <- text.v[start.v:end.v]

Entering these commands will not produce any output to the screen. The first
line of code takes lines 1 through 407 from the text.v variable and puts them into
a new variable called metadata.v. If you are wondering where the 407 came
from, remember that the start.v variable you created earlier contains the value
408 and refers to a place in the text vector that contains the words CHAPTER 1.
Loomings. Since you want to keep that line of the text (that is, it is not metadata
but part of the novel) you must subtract 1 from the value inside the start.v
variable to get the 407.
10

In R you can enter more than one command on a single line by separating the commands with a
semi-colon. Entering start.v;end.v would achieve the same result as what you see here.
11 In this book the output, or results, of entering an expression will be prefaced with two hash (##)
symbol.
12 Sentences, lines of text, etc. are formally referred to as strings of text.

2.2 Separate Content from Metadata

15

The second line does something similar; it grabs all of the lines of text that appear
after the end of the novel by first adding one to the value contained in the end.v
variable then spanning all the way to the last value in the vector, which can be
calculated using length.
The third line then combines the two using the c or combine function. As it
happens, I could have saved a bit of typing by using a shortcut. The same result
could have been achieved in one expression, like this:
metadata.v <- c(text.v[1:(start.v-1)], text.v[(end.v+1):length(text.v)])

Sometimes this kind of shortcutting can save a lot of extra typing and extra code. At
the same time, you will want to be careful about too much of this function embedding as it can also make your code harder to read later on.
The fourth line of the main code block is used to isolate the part of the electronic
file that is between the metadata sections. To capture the metadata that comes after
the last line in the novel, I use (end.v + 1) because I do not want to keep the
last line of the novel which is accessed at the point in the vector currently stored in
the end.v variable. If this does not make sense, try entering the following code to
see just what is what:
text.v[start.v]
text.v[start.v-1]
text.v[end.v]
text.v[end.v+1]

You can now compare the size of the original file to the size of the new
novel.lines.v variable that excludes the boilerplate metadata:
length(text.v)
## [1] 18874
length(novel.lines.v)
## [1] 18169

If you want, you can even use a little subtraction to calculate how much smaller the
new object is: length(text.v) - length(novel.lines.v).
The main text of Moby Dick is now in an object titled novel.lines.v, but
the text is still not quite in the format you need for further processing. Right now
the contents of novel.lines.v are spread over 18,169 line items derived from
the original decision to delimit the file using the newline character. Sometimes, it
is important to maintain line breaks: for example, some literary texts are encoded
with purposeful line breaks representing the lines in the original print publication
of the book or sometimes the breaks are for lines of poetry. For our purposes here,
maintaining the line breaks is not important, so you will get rid of them using the
paste function to join and collapse all the lines into one long string:
novel.v <- paste(novel.lines.v, collapse=" ")

The paste function with the collapse argument provides a way of gluing together a bunch of separate pieces using a glue character that you define as the value
for the collapse argument. In this case, you are going to glue together the lines
(the pieces) using a blank space character (the glue). After entering this expression,

16

2 First Foray into Text Analysis with R

you will have the entire contents of the novel stored as a single string of words, or
rather a string of characters. You can check the size of the novel object by typing
length(novel.v)
## [1] 1

At first you might be surprised to see that the length is now 1. The variable called
novel.v is a vector just like novel.lines.v, but instead of having an indexed
slot for each line, it has just one slot in which the entire text is held. If you are not
clear about this, try entering:
novel.v[1]

A lot of text is going to flash before your eyes, but if you were able to scroll up in
your console window to where you entered the command, you would see something
like this:
[1] "CHAPTER 1. Loomings. Call me Ishmael. Some years ago..."

R has dumped the entire contents of the novel into the console. Go ahead, read the
book!

2.3 Reprocessing the Content

Now that you have the novel loaded as a single string of characters, you are ready
to have some fun. First use the tolower function to convert the entire text to
lowercase characters.
novel.lower.v <- tolower(novel.v)

You now have a big blob of Moby Dick in a single, lowercase string, and this
string includes all the letters, numbers, and marks of punctuation in the novel. For
the time being, we will focus on the words, so you need to extract them out of the
full text string and put them into a nice organized list. R provides an aptly named
function for splitting text strings: strsplit.
moby.words.l <- strsplit(novel.lower.v, "\\W")

The strsplit function, as used here, takes two arguments and returns what R
calls a list.13 The first argument is the object (novel.lower.v) that you want to
split, and the second argument \\W is a regular expression. A regular expression is
a special type of character string that is used to represent a pattern. In this case, the
regular expression will match any non-word character.14 Using this simple regex,
strsplit can detect word boundaries.
So far we have been working with vectors. Now you have a list. Both vectors and
lists are data types, and R, like other programming languages, has other data types
13

Because this new object is a list, I have appended .l to the variable name.
W IKIPEDIA provides a fairly good overview of regular expression and a web search for regular
expressions will turn up all manner of useful information.

14

2.3 Reprocessing the Content

17

as well. At times you may forget what kind of data type one of your variables is,
and since the operations you can perform on different R objects depends on what
kind of data they contain, you may find yourself needing the class function. Rs
class function returns information about the data type of an object you provide as
an argument to the function. Here is an example that you can try:
class(novel.lower.v)
## [1] "character"
class(moby.words.l)
## [1] "list"

To get even more detail about a given object, you can use Rs str or structure
function. This function provides a compact display of the internal structure of an R
object. If you ask R to give you the structure of the moby.words.l list, youll see
the following:
str(moby.words.l)
## List of 1
## $ : chr [1:253989] "chapter" "1" "" "loomings" ...

R is telling you that this object (moby.words.l) is a list with one item and that
the one item is a character (chr) vector with 253989 items. R then shows you the
first few items, which happen to be the first few words of the novel.15 If you look
closely, youll see that the third item in the chr vector is an empty string. Well deal
with that in a moment. . ..
Right now, though, you may be asking, why a list? The short answer is that the
strsplit function that you used to split the novel into words returns its results as
a list. The long answer is that sometimes the object being given to the strsplit
function is more complicated than a simple character string and so strsplit is
designed to deal with more complicated situations. A list is a special type of object
in R. You can think of a list as being like a file cabinet. Each drawer is an item in the
list and each drawer can contain different kinds of objects. In my file cabinet, for
example, I have three drawers full of file folders and one full of old diskettes, CDs
and miscellaneous hard drives. You will learn more about lists as we go on.
It is worth mentioning here that anytime you want some good technical reading
about the nature of Rs functions, just enter the function name preceded by a question
mark, e.g.: ?strsplit. This is how you access Rs built in help files.16 Be
forewarned that your mileage with R-help may vary. Some functions are very well
documented and others are like reading tea leaves.17 One might be tempted to blame
15 I have adopted a convention of appending a data type abbreviation to the end of my object
names. In the long run this saves me from having to check my variables using class and str.
16 Note that in RStudio there is a window pane with a help tab where the resulting information
will be returned. This window pane also has a search box where you can enter search terms instead
of entering them in the console pane.
17 ?functionName is a shortcut for Rs more verbose help(functionName). If you want to see
an example of how a function is used, you can try example(functionName). args(functionName)
will display a list of arguments that a given function takes. Finally, if you want to search Rs
documentation for a single keyword or phrase, try using ??(your keyword) which is a shorthand
version of help.search(your keyword). I wish I could say that the documentation in R is always
brilliant; I cant. It is inevitable that as you learn more about R you will find places where the

18

2 First Foray into Text Analysis with R

poor documentation on the fact that R is open source, but I think its more accurate to
say that the documentation assumes a degree of familiarity with programming and
with statistics. R-help is not geared toward the novice, but, fortunately, R has now
become a popular platform, and if the built-in help is not always kind to newbies, the
resources that are available online have become increasingly easy to use and newbie
friendly.18 For the novice, the most useful part of the built-in documentation is often
found in the code examples that almost always follow the more technical definitions
and explanations. Be sure to read all the way down in the help files, especially if
you are confused. When all else fails, or even as a first step, consider searching for
answers and examples on sites such as http://www.stackexchange.com.
Because you used strsplit, you have a list, and since you do not need a list
for this particular problem, well simplify it to a vector using the unlist function:
moby.word.v <- unlist(moby.words.l)

When discussing the str function above, I mentioned that the third item in the
vector was an empty character string. Calling str(moby.words.l) revealed the
following:
## List of 1
## $ : chr [1:253989] "chapter" "1" "" "loomings" ...

As it happens, that empty string between 1 and loomings is where the period character used to be. The \\W regular expression that you used to split the string ignored
all the punctuation in the file, but then left these little blanks, as if to say, if Id
kept the punctuation, itd be right here.19 Since you are ignoring punctuation, at
least for the time being, these blanks are a nuisance. You will want now to identify where they are in the vector and then remove them. Or more precisely, youll
identify where they are not!
First you must figure out which items in the vector are not blanks, and for that
you can use the which function that was introduced previously.
not.blanks.v

<-

which(moby.word.v!="")

documentation is frustratingly incomplete. In these cases, the Internet is your friend, and there is a
very lively community of R users who post questions and answers on a regular basis. As with any
web search, the construction of your query is something of an art form, perhaps a bit more so when
it comes to R since using the letter r as a keyword can be frustrating.
18 This was not always the case, but a recent study of the R-help user base shows
that things have improved considerably. Trey Causeys analysis Has R-help gotten meaner
over time? And what does Mancur Olson have to say about it? is available online at http://badhessian.org/2013/04/has-r-help-gotten-meaner-overtime-and-what-does-mancur-olson-have-to-say-about-it/.
19 There are much better, but more complicated, regular expressions that can be created for doing
word tokenization. One downside to \\W is that it treats apostrophes as word boundaries. So the
word cant becomes the words can and t and Johns becomes John and s. These can be especially
problematic if, for example, the eventual analysis is interested in negation or possession. You will
learn to write a more sophisticated regex in Chapter 13.

2.3 Reprocessing the Content

19

Notice how the which function has been used in this example. which performs
a logical test to identify those items in the moby.word.v that are not equal (represented by the != operator in the expression) to blank (represented by the empty
quote marks in the expression). If you now enter not.blanks.v into R, you
will get a list of all of the positions in moby.words.v where there is not a blank.
Try it:
not.blanks.v

If you tried this, you just got a screen full of numbers. Each of these numbers corresponds to a position in the moby.words.v vector where there is not a blank. If
you scroll up to the top of this mess of numbers, you will find that the series begins
like this:
not.blanks.v
[1] 1 2 4

Notice specifically that position 3 is missing. That is because the item in the third
position was an empty string! If you want to see just the first few items in the
not.blanks.v vector, try showing just the first ten items, like this:
not.blanks.v[1:10]
[1] 1 2 4 6 7 8 10 11 12 14

With the non-blanks identified, you can overwrite moby.words.v like this20 :
moby.word.v <-

moby.word.v[not.blanks.v]

Here only those items in the original moby.words.v that are not blanks are retained.21 Just for fun, now enter:
moby.word.v

After showing you the first 99,999 words of the novel, R will give up and return a
message saying something like [[ reached getOption("max.print") omitted 204889 entries ]]. Even though R will not show you the entire
vector, it is still worth seeing how the word data has been stored in this vector object,
so you may want to try the following:
moby.word.v[1:10]
## [1] "chapter"
## [5] "me"
## [9] "ago"

"1"
"ishmael"
"never"

"loomings" "call"
"some"
"years"

20 Overwriting an object is generally not a great idea, especially when you are writing code that
you are unsure about, which is to say code that will inevitably need debugging. If you overwrite
your variables, it makes it harder to debug later. Here Im making an exception because I am certain
that Im not going to need the vector with the blanks in it.
21 A shorthand version of this whole business could be written as moby.word.v <moby.word.v[which(moby.word.v != "")].

20

2 First Foray into Text Analysis with R

The numbers in the square brackets are the index numbers showing you the position in the vector where each of the words will be found. R put a bracketed number
at the beginning of each row. For instance the word chapter is stored in the first
([1]) position in the vector and the word ishmael is in the sixth ([6]) position.
An instance of the word ago is found in the ninth position and so on. If, for some
reason, you wanted to know what the 99,986th word in Moby Dick is you could
simply enter
moby.word.v[99986]
## [1] "by"

This is an important point (not that the 99,986th word is by). You can access any
item in a vector by referencing its index. And, if you want to see more than one
item, you can enter a range of index values using a colon such as this:
moby.word.v[4:6]
## [1] "call"
"me"

"ishmael"

Alternatively, if you know the exact positions, you can enter them directly using the
c combination function to create a vector of positions or index values. First enter
this to see how the c works
mypositions.v <- c(4,5,6)

Now simply combine this with the vector:

moby.word.v[mypositions.v]
## [1] "call"
"me"
"ishmael"

You can do the same thing without putting the vector of values into a new variable. Simply use the c function right inside the square brackets:
moby.word.v[c(4,5,6)]
## [1] "call"
"me"

"ishmael"

Admittedly, this is only useful if you already know the index positions you are interested in. But, of course, R provides a way to find the indexes by also giving access
to the contents of the vector. Say, for example, you want to find all the occurrences
of whale. For this you can use the which function and ask R to find which items in
the vector satisfy the condition of being the word whale.
which(moby.word.v=="whale")

Go ahead and enter the line of code above. R will return the index positions where
the word whale was found. Now remember from above that if you know the index
numbers, you can find the items stored in those index positions. Before entering the
next line of code, see if you can predict what will happen.
moby.word.v[which(moby.word.v=="whale")]

2.4 Beginning the Analysis

21

2.4 Beginning the Analysis

Putting all of the words from Moby Dick into a vector of words (or, more precisely,
a character vector) provides a handy way of organizing all the words in the novel in
chronological order; it also provides a foundation for some deeper quantitative analysis. You already saw how to find a word based on its position in the overall vector
(the word by was the 99,986th word). You then saw how you could use which to
figure out which positions in the vector contain a specific word (the word whale).
You might also wish to know how many occurrences of the word whale appear in
the novel. Using what you just learned, you can easily calculate the number of whale
tokens using length and which together22:
length(moby.word.v[which(moby.word.v=="whale")])
## [1] 1150

Perhaps you would now also like to know the total number of tokens (words) in
Moby Dick? Simple enough, just ask R for the length of the entire vector:
length(moby.word.v)
## [1] 214889

You can easily calculate the percentage of whale occurrences in the novel by dividing the number of whale hits by the total number of word tokens in the book. To
divide, simply use the forward slash character.23
# Put a count of the occurrences of whale into whale.hits.v
whale.hits.v <- length(moby.word.v[which(moby.word.v=="whale")])
# Put a count of total words into total.words.v
total.words.v <- length(moby.word.v)
# now divide
whale.hits.v/total.words.v
## [1] 0.0053516

More interesting, perhaps, is to have R calculate the number of unique word

types in the novel. Rs unique function will examine all the values in the character
vector and identify those that are the same and those that are different. By combining
the unique and length functions, you calculate the number of unique words in
Melvilles Moby Dick vocabulary.
length(unique(moby.word.v))
## [1] 16872
22

In R, as in many languages, there are often alternative ways of achieving the same
goal. A more elegant method for calculating the number of whale hits might look like this:
length(moby.word.v[moby.word.v=="whale"]). For beginners, the explicit use of
which can be easier to understand.
23 In these next few lines of code, I have added some comments to explain what the code is doing.
This is a good habit for you to adopt; explaining or commenting your code so that you and others
will be able to understand it later on. In R you insert comments into code by using a # symbol
before the comment. When processing your code, R ignores everything between that # and the
next full line return.

22

2 First Foray into Text Analysis with R

It turns out that Melville has a fairly big vocabulary: In Moby Dick Melville uses
16,872 unique words. Thats interesting, but lets kick it up another notch. What
we really want to know is how often he uses each of his words and which words are
his favorites. We may even want to see if Moby Dick abides by Zipfs law regarding the general frequency of words in English.24 No problem. Rs table function
can be used to build a contingency table of word types and their corresponding
frequencies.
moby.freqs.t <- table(moby.word.v)

You can view the first few using moby.freqs.t[1:10], and the entire frequency table can be sorted from most frequent to least frequent words using the
sort function like this:
sorted.moby.freqs.t <- sort(moby.freqs.t , decreasing=TRUE)

Practice
2.1. Having sorted the frequency data as described in Sect. 2.4, figure out how to
get a list of the top ten most frequent words in the novel. If you need a hint,
go back and look at how we found the words call me ishmael in
moby.word.v. Once you have the words, use Rs built in plot function to visualize whether the frequencies of the words correspond to Zipfs law. The plot
function is fairly straightforward. To learn more about the plots complex arguments, just enter: ?plot. To complete this exercise, consider this example:
mynums.v <- c(1:10)
plot(mynums.v)

24

According to Zipfs law, the frequency of any word in a corpus is inversely proportional to its
rank or position in the overall frequency distribution. In other words, the second most frequent
word will occur about half as often as the most frequent word.

23

6
2

mynums.v

10

2.4 Beginning the Analysis

10

Index

You only need to substitute the mynums.v variable with the top ten values from
sorted.moby.freqs.t. You do not need to enter them manually!25

25 When generating plots in Rstudio, you may get an error saying: Error in plot.new() : figure
margins too large. This is because you have not given enough screen real estate to the plots pane
of Rstudio. You can click and drag the frames of the plotting pane to resolve this issue.

Chapter 3

Accessing and Comparing Word

Frequency Data

Abstract In this chapter, we derive and compare word frequency data. We learn
about vector recycling, and the exercises invite you to compare the frequencies of
several words in Melvilles Moby Dick to the same words in Jane Austens Sense
and Sensibility.

3.1 Accessing Word Data

While it is no surprise to find that the word the is the most frequently occurring
word in Moby Dick, it is a bit more interesting to see that the ninth most frequently
occurring word is his. Put simply, there are not a lot of women in Moby Dick. In
fact, you can easily compare the usage of he vs. she and him vs. her as follows1 :
sorted.moby.freqs.t["he"]
##
he
## 1876
sorted.moby.freqs.t["she"]
## she
## 114
sorted.moby.freqs.t["him"]
## him
## 1058
sorted.moby.freqs.t["her"]
## her
## 330

Notice how unlike the original moby.word.v in which each word was indexed
at a position in the vector, here the word types are the indexes, and the values
are the frequencies, or counts, of those word tokens. When accessing values in

1 This chapter expands upon code that we developed in Chapter 2. Before beginning to work
through this chapter, clear your workspace and run the Chapter 3 starter code found in Appendix C or in the textbook materials directory labeled start.up.code. For more on clearing
the workspace, see 4.2.1

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_3,
Springer International Publishing Switzerland 2014

25

26

3 Accessing and Comparing Word Frequency Data

moby.word.v, you had to first figure out where in the vector those word tokens
resided. Recall that you did this in two ways: you found call, me, and ishmael
by entering
moby.word.v[4:6]

and you found a whole pod of whales using the which function to test for the
presence of whale in the vector.
With the data in a table object (sorted.moby.freqs.t), however, you get
both numerical indexing and named indexing. In this way, you can access a value
in the table either by its numerical position in the table or by its name. Thus, this
expression
sorted.moby.freqs.t[1]

returns the same value as this one:

sorted.moby.freqs.t["the"]

Both expressions return the same result because the word type the happens to be the
first ([1]) item in the vector.
If you want to know just how much more frequent him is than her, you can use
the / operator to perform division.
sorted.moby.freqs.t["him"]/sorted.moby.freqs.t["her"]
##
him
## 3.206061

him is 3.2 times more frequent than her, but, as youll see in the next code snippet,
he is 16.5 times more frequent than she.
sorted.moby.freqs.t["he"]/sorted.moby.freqs.t["she"]
##
he
## 16.45614

Often when analyzing text, what you really need are not the raw number of
occurrences of the word types but the relative frequencies of word types expressed
as a percentage of the total words in the file. Relativizing in this way allows you
to more easily compare the patterns of usage from one text to another. For example, you might want to compare Jane Austens use of male and female pronouns to
Melvilles. Doing so requires compensating for the different lengths of the novels,
so you convert the raw counts to percentages by dividing each count by a total count
of all of the words in the whole text. These are called relative frequencies.
As it stands you have a sorted table of raw word counts. You want to convert
those raw counts to percentages, which requires dividing each count by the total
number of word tokens in the entire text. You already know the total number of
words because you used the length function on the original word vector.
length(moby.word.v)
## [1] 214889

3.2 Recycling

27

It is worth mentioning, however, that you could also find the total by calculating the
sum of all the raw counts in the tabled and sorted vector of frequencies.
sum(sorted.moby.freqs.t)
## [1] 214889

3.2 Recycling
You can convert the raw counts into relative frequency percentages using division
and then a little multiplication by 100 (multiplication in R is done using an asterisk)
to make the resulting numbers easier to read:
sorted.moby.rel.freqs.t <- 100*(sorted.moby.freqs.t/sum(sorted.moby.freqs.t))

The key thing to note about this expression is that R understands that it needs
to recycle the result of sum(sorted.moby.freqs.t) and apply that result to
each and every value in the sorted.moby.freqs.t variable. This recycling
also works with definite values. In other words, if you wanted to multiply every
value in a vector by ten, you could do so quite easily. Here is a simple example for
you to try.
num.vector.v <- c(1,2,3,4,5)
num.vector.v * 10
## [1] 10 20 30 40 50

Having applied the above calculation to the sorted.moby.freqs.t object, you

can now access any word type and return its relative frequency as a percentage.
Because you have multiplied by 100, this percentage shows the number of occurrences per every 100 words.
sorted.moby.rel.freqs.t["the"]
##
the
## 6.596429

the occurs 6.6 times for every 100 words in Moby Dick. If you want to plot the
top ten words by their percentage frequency, you can use the plot function as you
learned in exercise 2.1. Here Ill add a few more arguments to plot in order to
convey more information about the resulting image, and Ill call the axis function
to reset the values on the x-axis with the names of the top ten words. Notice that the
names function can be used to set, or in this case, get the names of an object. The
shape of the line in this plot is the same as in exercise 2.2, but here the values in the
y-axis have been converted from raw counts to counts per hundred (Fig. 3.1).
plot(sorted.moby.rel.freqs.t[1:10], type="b",
xlab="Top Ten Words", ylab="Percentage of Full Text", xaxt ="n")
axis(1,1:10, labels=names(sorted.moby.rel.freqs.t [1:10]))

3 Accessing and Comparing Word Frequency Data

5
4
3
1

Percentage of Full Text

28

the

of

and

to

in

that

it

his

Top Ten Words

Fig. 3.1 Top ten words in Moby Dick

Practice
3.1. Top Ten Words in Sense and Sensibility
In the same directory in which you found melville.txt, locate austen.txt and produce a relative word frequency table for Austens Sense and Sensibility that is
similar to the one created in Exercise 2.1 using Moby Dick. Keep in mind that
you will need to separate out the metadata from the actual text of the novel just
as you did with Melvilles text. Once you have the relative frequency table (i.e.,
sorted.sense.rel.freqs.t), plot it as above for Moby Dick and visually
compare the two plots.
3.2. In the previous exercise, you should have noticed right away that the top ten
most frequent words in Sense and Sensibility are not identical to those found in
Moby Dick. You will also have seen that the order of words, from most to least
frequent, is different and that the two novels share eight of the same words in the
top ten. Using the c combination function join the names of the top ten values in
each of the two tables and then use the unique function to show the 12 word types
that occur in the combined name list. Hint: look up how to use the functions by
entering a question mark followed by function name (e.g., ?unique):
3.3. The %in% operator is a special matching operator that returns a logical (as in
TRUE or FALSE) vector indicating if a match is found for its left operand in its
right operand. It answers the question is x found in y? Using the which function
in combination with the %in% operator, write a line of code that will compute
which words from the two top ten lists are shared.
3.4. Write a similar line of code to show the words in the top ten of Sense and
Sensibility that are not in the top ten from Moby Dick.

Chapter 4

Token Distribution Analysis

Abstract This chapter explains how to use the positions of words in a vector to
create distribution plots showing where words occur across a narrative. Several
important R functions are introduced including seq_along, rep, grep, rbind,
apply, and do.call. if conditionals and for loops are also presented.

4.1 Dispersion Plots

Youve seen how easy it is to calculate the raw and relative frequencies of words.
These are global statistics that show something about the central tendencies of
words across a book as a whole. But what if you want to see exactly where in
the text different words tend to occur; that is, where the words appear and how they
behave over the course of a novel. At what points, for example, does Melville really
get into writing about whales?
For this analysis you will need to treat the order in which the words appear in the
text as a measure of time, novelistic time in this case.1 If you do not already have
the moby.word.v object loaded, go to the Chapter 4 starter code found in Appendix C (and in the start.up.code directory) and regenerate moby.word.v.
You now need to create a sequence of numbers from 1 to n, where n is the
position, or index number, of the last word in Moby Dick. You can create such a
sequence using the seq (sequence) function. For a simple sequence of the numbers,
one to ten, you could enter:
seq(1:10)
## [1] 1

9 10

Instead of one through ten, however, you will need a sequence from one to the
last word in Moby Dick. You already know how to get this number n using the
1 For some very interesting work on modeling narrative time, see Mani, Inderjeet. The Imagined
Moment: Time, Narrative and Computation. University of Nebraska Press, 2010.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_4,
Springer International Publishing Switzerland 2014

29

30

4 Token Distribution Analysis

length function. Putting the two functions together allows for an expression like
this:
n.time.v <- seq(1:length(moby.word.v))

This expression returns an integer vector (n.time.v) containing the positions

of every word in the book.2 I have titled this object n.time.v because it is a vector
(.v) that will serve to represent narrative time (n.time) in the novel.
Now you need to locate the position of every occurrence of whale in the novel,
or, more precisely, in the moby.word.v object. You have already learned how the
which function can be used to locate items meeting certain conditions, so you can
use which to identify the positions in the vector that are an occurrence of whale
and store them in a new integer vector called whales.v:
whales.v <- which(moby.word.v == "whale")

If you now enter the object name (whales.v) into the console and hit enter, R will
return a list of the numerical positions where it found instances of whale.
Ultimately we want to create a dispersion plot where the x-axis is novelistic time.
You have those x-axis values in the n.time.v object. Another vector containing
the values for plotting on the y-axis is now needed, and in this case, the values
need only be some reflection of the logical condition of TRUE where a whale is
found and FALSE or none found when an instance of whale is not found. In R you
can represent the logical value TRUE with a number 1 and FALSE with a 0. Here,
however, since we are not really counting items but, instead, noting their presence
or absence, Ill introduce a special character sequenceNAas in not available for
places where there is no found match. Begin, therefore, by initializing a new vector
object called w.count.v that will be full of NA values. It needs to be the same
length as the n.time.v object, so you can use the rep or repeat function to repeat
NA as many times as there are items in the w.count.v variable.
w.count.v <- rep(NA,length(n.time.v))

Now you simply need to reset the NA values to 1 in those places in the moby.
word.v where a whale was found. You have those numerical positions stored in
the whales.v object, so the resetting is simple with this expression:
w.count.v[whales.v] <- 1

With the places where each whale was found now set to a value of 1 and everything else set to a value of NA, you can crank out a very simple plot showing the
distribution of the word whale across the novel3:
2 Remember that you can find out the data type of any R object using the class function. E.g.
class(n.time.v)
3 If you get an error saying: Error in plot.new() : figure margins too
large you may not have enough screen real estate devoted to the plot pane of RStudio. You
can solve this problem by increasing the size of the plots pane (just click and drag the frame). Your
plot may also appear a lot taller (or thicker) than the one seen here. I have shrunk the plotting pane
height in RStudio to make the image fit this page better.

4.2 Searching with grep

31

plot(w.count.v, main="Dispersion Plot of `whale' in Moby Dick",

xlab="Novel Time", ylab="whale", type="h", ylim=c(0,1), yaxt='n')

whale

Dispersion Plot of whale' in Moby Dick

50000

100000

150000

200000

Novel Time

Fig. 4.1 Dispersion plot of whale in Moby Dick

This simple dispersion plot (Fig. 4.1) shows that the greatest concentration of the
word whale occurs in what is, roughly, the third quarter of the novel. The first real
concentration of whale begins just before 50,000 words, and then there is a fairly
sustained pod of whale occurrences from 100,000 to about 155,000, and then
there is a final patch at the end of the novel, just after 200,000.
By changing just a few lines of code, you can generate a similar plot (Fig. 4.2)
showing the occurrences of ahab.
ahabs.v <- which(moby.word.v == "ahab") # find `ahab'
a.count.v <- rep(NA,length(n.time.v))
# change `w' to `a' to keep whales and ahabs in separate variables
a.count.v[ahabs.v] <- 1 # mark the occurrences with a 1
plot(a.count.v, main="Dispersion Plot of 'ahab' in Moby Dick",
xlab="Novel Time", ylab="ahab", type="h", ylim=c(0,1), yaxt='n')

ahab

Dispersion Plot of 'ahab' in Moby Dick

50000

100000

150000

200000

Novel Time

Fig. 4.2 Dispersion plot of ahab in Moby Dick

4.2 Searching with grep

Running the analysis for the word ahab shows that when whale is most present,
the appearance of ahab is seemingly decreased. In terms of sheer presence, whale
appears to dominate over ahab in the third quarter of Moby Dick in particular.
Figure 4.3 shows the two plots on top of each other, and while these dispersion plots
can be informative, more often than not novels have their own internal structure of

32

4 Token Distribution Analysis

chapters, and it can be more productive to honor the authors own organization of
the text. We should, therefore, devise a method for examining how words appear
across the novels internal chapter structure.

whale

Dispersion Plot of whale' in Moby Dick

50000

100000

150000

200000

Novel Time

ahab

Dispersion Plot of ahab' in Moby Dick

50000

100000

150000

200000

Novel Time

Fig. 4.3 Dispersion plots for whale and ahab in Moby Dick

4.2.1 Cleaning the Workspace

Recall that you have a variable (novel.lines.v) containing the entire text from
the original Project Gutenberg file as a list of lines. If you have had the same R
session opened for a while, and especially if you have been experimenting with your
own code as you work your way through the examples and exercises in this book,
it might be a good idea to clear your workspace. As you work in a given R session,
R is keeping track of all of your variables in memory. When you switch between
projects during the same session, you can avoid a lot of potential variable conflict
(and headache) by refreshing or clearing your workspace. Since you are about to do
something with chapter breaks instead of looking at the novel as a single string, now
is a good time to get a fresh start. In RStudio you can do this by selecting Clear
Workspace from the Session menu. Alternatively, you can just enter rm(list =
ls()) into the R console. Be aware that both of these commands will delete all of
your currently instantiated objects.
Enter the following expression to create a fresh session.
rm(list = ls())

If you now enter ls() you will simply see:

character(0)

4.2 Searching with grep

33

ls is a list function that returns a list of all of your currently instantiated objects.
This character(0) lets you know that there are no variables instantiated in the
session.4 Now that you have cleared everything, youll need to reload Moby Dick
using the same code you learned earlier:
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")
novel.lines.v <- text.v[start.v:end.v]

You may recall from earlier that you can view the whole text of Moby Dick in
your R console, newline by newline, by entering:
novel.lines.v

If you try this now, it might take a few seconds to load, and the results youll see are
not going to be very pretty.5 One thing you might notice in this long list is that the
beginning of each new chapter follows a specific pattern. Each new chapter starts
with a new line followed by the capitalized word CHAPTER and then a space
character and then one or more digits. For example,
[1] "CHAPTER 1. Loomings."
. . .
[185] "CHAPTER 2. The Carpet-Bag."

Because the Project Gutenberg text uses this CHAPTER convention to mark the
chapters, you can split the text into chapters by using this character sequence
(CHAPTER) as delimiter in a manner similar to the way that you split the text into
words using in the vector using the grep function.

4.2.2 Identify the chapter break positions in the vector using

the grep function
In text analysis grep and its related functions are your ever-loyal friends. Be sure to
access the grep help file by typing ?grep at the R prompt. And if you really want
to get rolling, do a web search for regular expressions or what are known as regex
for short. grep is an R function for performing regular expression pattern matching.
Using the regular expression CHAPTER \\d will allow grep to identify lines in
the vector that begin (the start of a line is marked by use of the caret symbol ) with
the capitalized letters CHAPTER followed by a space and then any digit (digits are
represented using an escaped d, as in \\d). Here is the full expression.
chap.positions.v <- grep("^CHAPTER \\d", novel.lines.v)

4 If you are using RStudio, you can simply check the environment window pane. After running
rm(list = ls()) or clearing the workspace from the session menu, it will be blank.
5 Do not be alarmed if you see a series of backslash characters in the text. These are escape
characters that R adds before quotation marks and apostrophes so that they will not be treated as
special characters and parsed by R.

34

4 Token Distribution Analysis

To check your work, enter the next R expression:

novel.lines.v[chap.positions.v]

If grep and the regex did their job, you will now see a character vector containing
all 135 of the chapter headings. Here is an abbreviated version showing only the
first and last few items:
[1] "CHAPTER 1. Loomings."
[2] "CHAPTER 2. The Carpet-Bag."
[3] "CHAPTER 3. The Spouter-Inn."
...
[133] "CHAPTER 133. The Chase--First Day."
[134] "CHAPTER 134. The Chase--Second Day."
[135] "CHAPTER 135. The Chase.--Third Day."

The object chap.positions.v holds the positions from the novel.

lines.v where the search string CHAPTER\\d was found. You must now find
a way to collect all of the lines of text that occur between these positions: the chunks
of text that make up each chapter.
That sounds simple, but you do not yet have a marker for the ends of the chapters,
you only know where they begin. To get the ends, you can subtract 1 from the known
position of the following chapter. In other words, if Chap. 10 begins at position
1,524, then you know that Chap. 9 ends at 1524 - 1 or 1523.
This technique works perfectly except for the last chapter where there is no following chapter! There are several ways you might address this situation, but a simple
solution is to just add one more line to the novel.lines.v object and then add
the position of this new line to the chap.positions.v vector. You will find that
last position easily enough with the length function.
But lets slow down so that you can see exactly what is happening:
1. Enter chap.positions.v at the prompt to see the contents of the current
vector:
chap.positions.v
##
[1]
1
185
##
[8] 1141 1222
## [15] 1996 2099
## [22] 3181 3323
## [29] 3893 3993
## [36] 5023 5273
## [43] 6170 6202
## [50] 7274 7360
## [57] 8656 8742
## [64] 9293 9555
## [71] 9971 10175
## [78] 10876 11016
## [85] 11778 11947
## [92] 13066 13148
## [99] 13701 13900
## [106] 14755 14835
## [113] 15462 15571
## [120] 16095 16113
## [127] 16601 16671
## [134] 17473 17761

301
1524
2572
3357
4018
5315
6381
7490
8828
9638
10316
11097
12103
13287
14131
14928
15631
16164
16790

790
1654
2766
3506
4084
5347
6681
7550
8911
9692
10502
11174
12514
13398
14279
15066
15710
16169
16839

925
1712
2887
3532
4532
5371
6771
7689
9032
9754
10639
11541
12620
13440
14416
15148
15756
16274
16984

989
1785
2997
3635
4619
5527
6856
8379
9201
9854
10742
11638
12745
13592
14495
15339
15798
16382
17024

1062
1931
3075
3775
4805
5851
7201
8543
9249
9894
10816
11706
12843
13614
14620
15377
15873
16484
17160

4.3 The for Loop and if Conditional

35

2. Add a new item to the end of the novel.lines.v object using the c function.
Here I have set the value of that last item to END. You will see later on that this
last item serves to mark the end boundary for the last chapter. Now get the last
position using the length function and add it to the chap.positions.v
vector using the c function:
novel.lines.v <- c(novel.lines.v, "END")
last.position.v <- length(novel.lines.v)
chap.positions.v <- c(chap.positions.v , last.position.v)

3. Enter chap.positions.v at the prompt again, this time to see the entire
vector but now with a new value (18,172) appended to the end:
chap.positions.v
##
[1]
1
185
##
[8] 1141 1222
## [15] 1996 2099
## [22] 3181 3323
## [29] 3893 3993
## [36] 5023 5273
## [43] 6170 6202
## [50] 7274 7360
## [57] 8656 8742
## [64] 9293 9555
## [71] 9971 10175
## [78] 10876 11016
## [85] 11778 11947
## [92] 13066 13148
## [99] 13701 13900
## [106] 14755 14835
## [113] 15462 15571
## [120] 16095 16113
## [127] 16601 16671
## [134] 17473 17761

301
1524
2572
3357
4018
5315
6381
7490
8828
9638
10316
11097
12103
13287
14131
14928
15631
16164
16790
18170

790
1654
2766
3506
4084
5347
6681
7550
8911
9692
10502
11174
12514
13398
14279
15066
15710
16169
16839

925
1712
2887
3532
4532
5371
6771
7689
9032
9754
10639
11541
12620
13440
14416
15148
15756
16274
16984

989
1785
2997
3635
4619
5527
6856
8379
9201
9854
10742
11638
12745
13592
14495
15339
15798
16382
17024

1062
1931
3075
3775
4805
5851
7201
8543
9249
9894
10816
11706
12843
13614
14620
15377
15873
16484
17160

The trick now is to figure out how to process the text, that is, the actual content
of each chapter that appears between each of these chapter markers. For this we will
learn how to use a for loop.

4.3 The for Loop and if Conditional

Most of what follows from here will be familiar to you from what we have already
learned about tokenization and word frequency processing. The main difference is
that now all of that code will be wrapped inside of a looping function. A for loop
allows us to do a task over and over again for a set number of iterations. In this case,
the number of iterations will be equal to the number of chapters found in the text.
As a simple example, lets say you just want to print (to the screen) the various
chapter positions you found using grep. Instead of printing them all at once, like
you did above by dumping the contents of the chap.positions.v variable, you
want to show them one at a time. You already know how to return specific items in
a vector by putting an index number inside brackets, like this

36

4 Token Distribution Analysis

chap.positions.v[1]
## [1] 1
chap.positions.v[2]
## [1] 185

Instead of entering the vector indexes (1 and 2 in the example above), you can
use a for loop to go through the entire vector and automate the bracketing of the
index numbers in the vector. Heres a simple way to do it using a for loop:
for(i in 1:length(chap.positions.v)){
print(chap.positions.v[i])
}

Notice the for loop syntax; it includes two arguments inside the parentheses: a
variable (i) and a sequence (1:length(chap.positions.v)). These are followed by a set of opening and closing braces. These braces contain (or encapsulate)
the instructions to perform within each iteration of the loop.6 Upon the first iteration, i gets set to 1. With i == 1 the program prints the contents of whatever
value is held in the first position of chap.positions.v. In this case, the value
is 1, which can be a bit confusing. When the program gets to the second iteration,
the value printed is 185, which is less confusing. After each iteration of the loop,
i is advanced by 1, and this looping continues until i is equal to the length of
chap.positions.v.
To make this even more explicit, Ill add a paste function that will print the
value of i along with the value of the chapter position, making it all easy to read
and understand. Try this now.
for(i in 1:length(chap.positions.v)){
print(paste("Chapter ",i, " begins at position ",
chap.positions.v[i]), sep="")
}

When you run this loop, you will get a clear sense of how the parts are working
together. With that example under your belt, you can now return to the chapter-text
problem. As you iterate over the chap.positions.v vector, you are going to
be grabbing the text of each chapter and performing some analysis. Along the way,
you dont want to print the results to the R console (as in my example above), so
you will need a place to store the results of the analysis during the for loop. For
this you will create two empty list objects. These will serve as containers in which
to store the calculated result of each iteration:
chapter.raws.l <- list()
chapter.freqs.l <- list()

Remember from Chap. 2 that a list is a special type of object in R. You can think of
a list as being like a file cabinet. Each drawer is an item in the list and each drawer
can contain different kinds of objects.

6 Using i is a matter of convention. You could name this variable anything that you wish: e.g.
my.int, x, etc.

4.3 The for Loop and if Conditional

37

To summarize, the for loop will iterate over each item in the chap.
positions.v vector. When it gets to each item, it will use the chapter position
information stored in the vector to figure out the beginning and the end of each chapter. With the chapter boundaries found, the script will then collect the word tokens
found within those boundaries and calculate both the raw and relative frequencies
of those word types using the table function that you learned about earlier. The
frequencies (both the raw counts and the relative frequencies) will then be stored in
the two list variables that are instantiated prior to the loop. The processing inside
the loop is similar to what was done when plotting the occurrences of whale over
the course of all the words in the text. The difference is that now you are plotting
values by chapter.
Though the tasks are similar, there are one or two complicating factors that must
be addressed. The most problematic of these involves what to do when i is equal
to the length of chap.positions.v. Since there is no text following the last
position, you need a way to break out of the loop. For this an if conditional is
perfect. Below I have written out the entire loop. Before moving on to my line-byline explication, take a moment to study this code and see if you can explain each
step.
for(i in 1:length(chap.positions.v)){
if(i != length(chap.positions.v)){
chapter.title <- novel.lines.v[chap.positions.v[i]]
start <- chap.positions.v[i]+1
end <- chap.positions.v[i+1]-1
chapter.lines.v <- novel.lines.v[start:end]
chapter.words.v <- tolower(paste(chapter.lines.v, collapse=" "))
chapter.words.l <- strsplit(chapter.words.v, "\\W")
chapter.word.v <- unlist(chapter.words.l)
chapter.word.v <- chapter.word.v[which(chapter.word.v!="")]
chapter.freqs.t <- table(chapter.word.v)
chapter.raws.l[[chapter.title]] <- chapter.freqs.t
chapter.freqs.t.rel <- 100*(chapter.freqs.t/sum(chapter.freqs.t))
chapter.freqs.l[[chapter.title]] <- chapter.freqs.t.rel
}
}

Now that youve had a chance to think through the logic of this loop for yourself,
here is a line-by-line explication:
1.

for(i in 1:length(chap.positions.v)){

Initiate a for loop that iterates over each item in chap.positions.v.

2.

if(i != length(chap.positions.v)){

As long as the value of i is not equal to the length of the vector, keep iterating
over the vector. Here I introduce the conditional if. if allows me to set a condition that will evaluate to either TRUE or FALSE. If the condition is found to be
TRUE, then the code inside the curly braces of the if statement will be executed.
This has the effect of saying so long as this condition is met, continue iterating.
The condition here is that i not be equal (!=) to the length of the vector. The reason I must set this condition is because there is no chapter text after the last item
in chap.positions.v. I do not want to keep the loop going once it gets to
the end!

38

4 Token Distribution Analysis

Assuming that the condition stated in the if statement is met, I proceed to

the next line. At this stage the program captures the chapter title which is
found at the place in the novel.lines.v indicated by the value held in the
chap.positions.v.
3.

chapter.title <- novel.lines.v[chap.positions.v[i]]

If this is confusing, try this: In your console, set i to 1.

i <- 1

Now enter:
novel.lines.v[chap.positions.v[i]]

When you hit return, you will see:

[1] "CHAPTER 1. Loomings."

If that is still not clear, you can break it down even further, like this:
i <- 1
chap.positions.v[i]
[1] 1
novel.lines.v[chap.positions.v[i]]
[1] "CHAPTER 1. Loomings."
i <- 2
chap.positions.v[i]
[1] 185
novel.lines.v[chap.positions.v[i]]
[1] "CHAPTER 2. The Carpet-Bag."

4.

start <- chap.positions.v[i]+1

I know that the title of the chapter is at the i-th line in novel.lines.v, so I
can add 1 to i and get the values of the next line in the vector. i+1 will give me
the position of the first line of the chapter text (i.e., excluding the chapter title).
5.

end <- chap.positions.v[i+1]-1

What is done here is a bit more subtle. Instead of adding 1 to the value held in the
ith position of chap.positions.v, I must add 1 to i in its capacity as an
index. Instead of grabbing the value of the ith item in the vector, the program
is going to grab the value of the item in the next position beyond i in the vector.
If this isnt clear, you can break it down like this:
i <- 1
chap.positions.v[i]
[1] 1
chap.positions.v[i+1]
[1] 185

When i==1, the value held in chap.positions.v[i] will be 1 because 1

happens to be the first value stored in the vector. When i == i+1, in this case
2, R will return the value held in the second position in chap.positions.v,
or 185. In the next iteration, i will be 2 and so [i+1] will be 3 and the result
will be 301, which is the third value stored in the vector.

4.4 Accessing and Processing List Items

39

chap.positions.v[i+1] will return the next item in the vector, and the
value held in that spot is the position for the start of a new chapter. Since I do
not want to count the words in the chapter heading, I must subtract 1 from that
value in order to get the line number in novel.lines.v that comes just before
the start of a new chapter. Thus I subtract 1 from the value found in the [i+1]
position.
6.

chapter.lines.v <- novel.lines.v[start:end]

chapter.words.v <- tolower(paste(chapter.lines.v, collapse=" "))
chapter.words.l <- strsplit(chapter.words.v, "\\W")
chapter.word.v <- unlist(chapter.words.l)
chapter.word.v <- chapter.word.v[which(chapter.word.v!="")]
chapter.freqs.t <- table(chapter.word.v)

These lines should be familiar to you from previous sections. With start and
end points defined, you grab the lines, paste them into a single block of text,
lowercase everything and then split it all into a vector of words that is tabulated
into a frequency count of each word type.
7.

chapter.raws.l[[chapter.title]] <-

chapter.freqs.t

This next line is where the resulting table of raw frequency counts is stuffed into
the list that was created before entering the loop. The double bracketing here is
used to assign a name or label to the list item, and here each item in the list is
named with the chapter heading extracted a few lines above. It is not necessary
to assign labels to list items in this way. If you leave out the label, the list will
just be created with numerical indexes. The utility of this labeling will become
clear later on.
8.

chapter.freqs.t.rel <- 100*(chapter.freqs.t/sum(chapter.freqs.t))

chapter.freqs.l[[chapter.title]] <- chapter.freqs.t.rel
}
}

The last two lines simply convert the raw counts to relative frequencies based on
the number of words in the chapter. This relative frequency table is then stuffed
into the other list object that was created before entering the for loop.

4.4 Accessing and Processing List Items

With the two lists now populated with data, a way of accessing the data and putting
it into a usable structure that allows for easy comparisons of word frequencies across
chapters is needed. For this you will learn to use three functions: rbind, lapply,
and do.call, and along the way you will learn something more about vector
recycling.

40

4 Token Distribution Analysis

4.4.1 rbind
rbind is the simplest of the three functions introduced in this section. As the name
suggests, rbind is a function for binding rows of data together. For rbind to
work, the rows being bound must have the same number of columns. Enter the
following R code into your console window:
x <- c(1,2,3,4,5)
y <- c(6,7,8,9,10)

These expressions create two vectors of five numerical values each. If you now
use rbind to combine them, you get a matrix object with two rows and five
columns.
rbind(x,y)
##
[,1] [,2] [,3] [,4] [,5]
## x
1
2
3
4
5
## y
6
7
8
9
10

Notice, however, what happens when I recreate the y vector so that x and y are
not of the same length:
y <- c(6,7,8,9,10, 11)
rbind(x,y)
## Warning: number of columns of result is not a multiple of vector length
(arg 1)
##
[,1] [,2] [,3] [,4] [,5] [,6]
## x
1
2
3
4
5
1
## y
6
7
8
9
10
11

First, R reports a warning message that the vectors were not of the same length.
In R, a warning is just a warning; your script did not fail to execute. In fact, you now
have a sixth column. Take a moment to experiment with this example and see if you
can figure out what R is doing when it has two vectors of different lengths.

4.4.2 More Recycling

What you should have discovered is something called recycling. The recycling occurs because you are binding vectors of differing lengths. At some point R discovers
that the shorter vector is at its end and that the longer vector still has uncombined
elements. So R simply returns to the beginning of the shorter vector and begins recycling its elements. R will keep recycling from the shorter vector until it reaches the
end of the process. The elements of the shorter vector will be reused over and over
again until the process is complete. Sometimes this recycling is particularly useful.
Say you want to multiply every item in one vector by a value held in some other
vector. Here, for example, I multiply each number in the x vector by the number
held in the y vector.7
7

It might seem a bit odd, but in R even objects containing only one item are vectors. So in this
example the y object is a vector of one item. If you simply enter y into the console, youll get a
bracketed number 1 [1] followed by the value 2, which is the value held in the first position of
the y vector.

4.4 Accessing and Processing List Items

41

x <- c(1,2,3,4,5,6)
y <- 2
x*y
## [1] 2 4 6 8 10 12

This recycling can get a bit confusing when you have more complicated vectors.
In the example above, each value in the x vector is multiplied by the value in the y
vector. When the y vector contains more than one item, then the recycling gets a bit
more complicated. Consider this example:
x <- c(1,2,3,4,5,6)
y <- c(2, 3)
x*y
## [1] 2 6 6 12 10 18

Here, the 2 and 3 get recycled over and over again in order such that the first
item in the x vector is multiplied by the first item in the y vector (the number 2),
the second item in the x vector is multiplied by the second item in the y vector (the
number 3). But then when R gets to the third item in the x vector, it recycles the y
vector by going back to the first item in the y vector (the number 2) again. Deep
breath.

4.4.3 apply
lapply is one of several functions in the apply family. lapply (with an l in
front of apply) is specifically designed for working with lists. Remember that you
have two lists that were filled with data using a for loop. These are:
chapter.freqs.l
chapter.raws.l

lapply is similar to a for loop. Like for, lapply is a function for iterating
over the elements in a data structure. The key difference is that lapply requires
a list as a second arguments, and it requires the name of some other function as a
second argument. When run, lapply returns a new list object of the same length
as the original one, but it does so after having applied the function it was given in
its arguments to each item of the original list. Say, for example, that you have the
following list called x:
x

<-

list(a = 1:10, b = 2:25, b=100:1090)

This is a list of three integer objects each containing a series of numbers. Enter x at
the R prompt to look at the contents of the x list. Basically, x is like a file cabinet
with three drawers, and each one of the drawers contains an integer vector. If you
now enter: lapply(x, mean) R will return a new list in which the function
(mean) is applied to each object in the list called x

42

4 Token Distribution Analysis

lapply(x, mean)
## $a
## [1] 5.5
##
## $b
## [1] 13.5
##
## $b
## [1] 595

R has calculated the mean for each of the integer vectors in the x list.
Now consider the construction of the lists you filled up using the for loop.
Each list contains a series of frequency tables. Each item in chapter.raws.l
is a table of raw counts of each word type in a chapter, and each list item in
chapter.freqs.l is a table of the relative frequencies of each word type in
a chapter.
If you want to know the relative frequency of the word type whale in the first
chapter of Moby Dick, you could get the value using bracketed subsetting, like this:
chapter.freqs.l[[1]]["whale"]

This expression tells R that you want to go to the first item in the chapter.
freqs .l list (list items are accessed using the special double bracket [[]] notation), which is a frequency table of all the words from Chap. 1, i.e.,
chapter.freqs.l[[1]]

But you also instruct R to return only those values for the word type whale. Try it
for yourself:
chapter.freqs.l[[1]]["whale"]
##
whale
## 0.1336898

The result indicates that the word whale occurs 0.13 times for every 100 words
in the first chapter. Since you know how to get this data for a single list item, it
should not be too hard then to now use lapply to grab the whale values from the
entire list. In fact, you can get that data by simply entering this:
lapply(chapter.freqs.l, '[', 'whale')

Well, OK, Ill admit that using [ as the function argument here is not the most
intuitive thing to do, and Ill admit further that knowing you can send another argument to the main function is even more confusing. So lets break this down a bit. The
lapply function is going to handle the iteration over the list by itself. So basically,
lapply handles calling each chapter from the list of chapters. If you wanted to do
it by hand, youd have to do something like this:
chapter.freqs.l[[1]]
chapter.freqs.l[[2]]
. . .
chapter.freqs.l[[135]]

4.4 Accessing and Processing List Items

43

By adding [ as the function argument to lapply, you tell lapply to apply

bracketed subsetting to each item in the list. Recall again that each item in the list
is a table of word counts or frequencies. lapply allows us to add another optional
argument to the function that is being called; in this case the function is bracketed
subsetting. When you send the key word whale in this manner, then behind the
scenes R executes code for each item that looks like this:
chapter.freqs.l[[1]]["whale"]
chapter.freqs.l[[2]] ["whale"]
. . .
chapter.freqs.l[[135]] ["whale"]

If you enter a few of these by hand, youll begin to get a sense of where things are
going with lapply.
Here is how to put all of this together in order to generate a new list of the whale
values for each chapter.
whale.l <- lapply(chapter.freqs.l, '[', 'whale')

Instead of just printing out the values held in this new list, you can capture the results
into a single matrix using rbind.
One option would be to rbind each item in the whale.l list object by hand:
something like what follows here (but with more than just the first three list items):
rbind(whale.l[[1]], whale.l[[2]], whale.l[[3]])

While this method works, it is not very scalable or elegant. Fortunately R has another
function for just this kind of problem: the function is do.call and is pronounced
do dot call.

4.4.4 do.call (Do Dot Call)

Like lapply, do.call is a function that takes another function as an argument.
In this case the other function will be rbind. The do.call function will take
rbind as an input argument and call it over the different elements of the list object.
Consider this very simple application of the do.call function: First create a list
called x that contains three integer vectors.
x
x
##
##
##
##
##
##
##
##

<-

list(1:3,4:6,7:9)

[[1]]
[1] 1 2 3
[[2]]
[1] 4 5 6
[[3]]
[1] 7 8 9

To convert this list into a matrix where each row is one of the vectors and each
column is one of the three integers from each of the list items, use do.call

44

4 Token Distribution Analysis

do.call(rbind,x)
##
[,1] [,2] [,3]
## [1,]
1
2
3
## [2,]
4
5
6
## [3,]
7
8
9

Using do.call in this way binds the contents of each list item row-wise.
The list of whale occurrences in Moby Dick is fairly similar to the list (x) that was
used in this example. In some ways whale.l is even simpler than x because each
integer vector only contains one item. You can use do.call, therefore, to activate
the rbind function across the list of whale results. Doing this will generate a matrix
object of 135 chapter rows by 1 column of relative frequency values for the word
whale. A matrix object is like a very simple spreadsheet; a matrix has rows and
columns. One special (or limiting) thing about matrix objects in R, however, is that
they can only contain one type of data. That is, they cannot contain text values in one
column and numerical values in another. R has another object for handling mixed
data, and well cover that later on. For now, just think about a matrix as a simple
spreadsheet. Ill call this new matrix whales.m:
whales.m <- do.call(rbind, whale.l)

After you have entered this expression, look at the results by entering whales.m
at the R prompt. Here is an abbreviated version of how the results should look:
whales.m
CHAPTER 1. Loomings. 0.13368984
CHAPTER 2. The Carpet-Bag. 0.06882312
CHAPTER 3. The Spouter-Inn. 0.10000000
CHAPTER 4. The Counterpane. NA
. . .
CHAPTER 133. The Chase--First Day. 0.76965366
CHAPTER 134. The Chase--Second Day. 0.61892131
CHAPTER 135. The Chase.--Third Day. 0.67127746

Using what you have learned thus far, you can create another matrix of chapterby-chapter values for occurrences of ahab. The only thing you need to change in the
code described already is the keyword: youll use ahab in place of whale:
ahab.l <- lapply(chapter.freqs.l, '[', 'ahab')
ahabs.m <- do.call(rbind, ahab.l)

4.4.5 cbind
With both whales.m and ahabs.m instantiated in memory, you can easily bind
them together column-wise using cbind. As it happens, the individual columns in
a matrix object are individual vectors. In this example, the data in the first column
of the whales.m matrix is numeric vector.
class(whales.m[,1])

4.4 Accessing and Processing List Items

45

Remember that a matrix is like a spreadsheet with rows and columns. You can
access any cell in the matrix by identifying its row and column number. Here is a
simple matrix created by cbind-ing several vectors together:
x <- c(1,2,3,4,5,6)
y <- c(2,4,5,6,7,8)
z <- c(24,23,34,32,12,10)
test.m <- cbind(x,y,z)
test.m
##
x y z
## [1,] 1 2 24
## [2,] 2 4 23
## [3,] 3 5 34
## [4,] 4 6 32
## [5,] 5 7 12
## [6,] 6 8 10

To access the value held in the second row and third column in this matrix, you use
bracketed subsetting similar to what you have been using when accessing values
in a vector. Here, you will need to put both row and column information into the
brackets.
test.m[2,3] # show the value in the second row third column
## z
## 23

Inside the brackets 2 was entered to indicate the row number. This was followed by
a comma and then a 3 to indicate the third column. If you wanted to see an entire
row or an entire column, you would just leave the field before or after the comma
empty:
test.m[2,] # show all values in the second row
## x y z
## 2 4 23
test.m[,1] # show all values in the first column
## [1] 1 2 3 4 5 6

It is also worth knowing that if the columns have names, you can access them by
name. By default cbind names columns based on their original variable name:
test.m[,"y"]
## [1] 2 4 5 6 7 8

Now that you know how to access specific values in a matrix, you can easily pull
them out and assign them to new objects. You can pull out the whale and ahab
values into two new vectors like this:
whales.v <- whales.m[,1]
ahabs.v <- ahabs.m[,1]

You can now use cbind to bind these vectors into a new, two-column matrix.
The resulting matrix will have 135 rows and 2 columns, a fact you can check using
the dim function.
whales.ahabs.m <- cbind(whales.v, ahabs.v)
dim(whales.ahabs.m)
## [1] 135
2

46

4 Token Distribution Analysis

Previously I mentioned that by default cbind titles columns based on the input
variable names. You can reset the column names manually using the colnames
function in conjunction with the c function. To rename the two columns in this
example, use this expression:
colnames(whales.ahabs.m) <- c("whale", "ahab")

Once you have reset the column names, you can plot the results side by side using
the barplot function (Fig. 4.4).

0.0

0.5

1.0

1.5

2.0

barplot(whales.ahabs.m, beside=T, col="grey")

whale

ahab

Fig. 4.4 Bar plot of whale and ahab side by side

Practice
4.1. Write code that will find the value for another word (e.g., queequeg) and then
bind those values as a third column in the matrix. Plot the results as in the example
code.
4.2. These plots were derived from the list of relative frequency data. Write a script
to plot the raw occurrences of whale and ahab per chapter.

Chapter 5

Correlation

Abstract This chapter introduces data frames, random sampling, and correlation.
Readers learn how to perform permutation tests to assess the significance of derived
correlations.

5.1 Introduction
It might be tempting to look at the graphs you have produced thus far and begin
forming an argument about the relative importance of Ahab versus the whale in
Melvilles novel. Occurrences of whale certainly appear to occupy the central portion of the book, whereas Ahab is present at the beginning and at the end. It might
also be tempting to begin thinking about the structure of the novel, and this data
does provide some evidence for an argument about how the human dimensions of
the narrative frame the more naturalistic. But is there, in fact, an inverse relationship?

5.2 Correlation Analysis

Using the frequency data you compiled for ahab and whale, you can run a correlation analysis to see if there is a statistical connection. A correlation analysis
attempts to determine the extent to which there is a relationship, or linear dependence, between two sets of points. Thought of another way, correlation analysis
attempts to assess the way that the occurrences of whale and ahab behave in unison
or in opposition to each other over the course of the novel. You can use a correlation analysis to answer a question such as: to what extent does the usage of whale
change (increase or decrease) in relation to the usage of ahab? R offers a simple
function, cor, for calculating the strength of a possible correlation. But before you
can employ the cor function on the whales.ahabs.m object, you need to deal
M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_5,
Springer International Publishing Switzerland 2014

47

48

5 Correlation

with the fact that there are some cells in the matrix that contain the value NA. Not
every chapter in Moby Dick had an occurrence of whale (or ahab), so in the previous
practice exercise when you ran
whale.l

<-

lapply(chapter.freqs.l, "[", "whale")

R found no hits for whale in some chapters of the novel and recorded an NA, as in
not available or missing. You may recall seeing this NA output when you viewed the
contents of whales.ahabs.m matrix:
whales.ahabs.m[1:16, ]
##
whale
ahab
## [1,] 0.13368984
NA
## [2,] 0.06882312
NA
## [3,] 0.10000000
NA
## [4,]
NA
NA
## [5,]
NA
NA
## [6,] 0.24067389
NA
## [7,] 0.21097046
NA
## [8,]
NA
NA
## [9,] 0.24711697
NA
## [10,]
NA
NA
## [11,]
NA
NA
## [12,]
NA
NA
## [13,] 0.17341040
NA
## [14,]
NA
NA
## [15,]
NA
NA
## [16,] 0.16037063 0.3385602

As you see here, there are no occurrences of whale in Chaps. 4 or 5 and no occurrences of ahab until Chap. 16. Because cor is a mathematical function that requires
numerical data, you need to replace the NA values before running the correlation
analysis. Since the appearance of an NA in these cases is equivalent to zero (there
are exactly zero occurrences of the keyword in the given chapter), you can safely
replace all the occurrences of NA in the whales.ahabs.m matrix with zero. One
way to do this is by embedding the conditional is.na function inside a call to the
which function as in: which(is.na(whales.ahabs.m)). To set the values
to 0, place the entire expression inside the brackets of whales.ahabs.m and
assign a 0 to those items that meet the condition:
whales.ahabs.m[which(is.na(whales.ahabs.m))] <- 0

This is the short and easy way to do it, but for the sake of illustration here it is broken
down with comments added to explain what is going on:
# identify the position of NA values in the matrix
the.na.positions <- which(is.na(whales.ahabs.m))
# set the values held in the found positions to zero
whales.ahabs.m[the.na.positions] <- 0

With the NAs set to zero, the correlation can be run.

cor(whales.ahabs.m)
##
whale
ahab
## whale 1.0000000 -0.2411072
## ahab -0.2411072 1.0000000

5.2 Correlation Analysis

49

Because whales.ahabs.m is a matrix of two columns, the result of calling cor

is a new matrix containing two rows and two columns. The row and column names
are the same and the values held in the cells are the correlation values. Its no surprise
to see that whale is perfectly correlated with whale and ahab with ahab. The positive
1.000 in these cells is not very informative, which is to say that running cor
over the entire matrix as Ive done here results in a lot of extraneous information.
Thats because cor runs the correlation analysis for every possible combination of
columns in the matrix. With a two column matrix like this, its really overkill. The
results could be made a lot simpler by just giving cor the two vectors that you
really want to correlate:
mycor <- cor(whales.ahabs.m[,"whale"], whales.ahabs.m[,"ahab"])
mycor
## [1] -0.2411072

The resulting number (0.2411072) is a measure of the strength of linear dependence between the values in the whale column and the values in the ahab column.
This result, called the Pearson Product-moment correlation coefficient, is expressed
as a number between -1 and +1. A negative one (-1) coefficient represents perfectly negative correlation; if the correlation between ahab and whale were -1,
then we would know that as the usage of whale increases, the usage of ahab decreases proportionally. Positive one (+1) represents perfect positive correlation (as
one variable goes up and down, the other variable does so in an identical way). Zero
(0) represents no correlation at all.
The further the coefficient is from zero, the stronger the correlation; conversely
the closer the result is to 0, the less dependence there is between the two variables.
Here, with whale and ahab a correlation coefficient of 0.2411072 is observed. This
suggests that while there is a slightly inverse relationship, it is not strongly correlated
since the result is closer to 0 than to -1. Having said that, how one interprets the
meaning, or significance, of the correlation indicated by this coefficient is largely
dependent upon the context of the analysis and upon the number of observations
or data points under consideration. Generally speaking a coefficient between -0.3
and -0.1 on the negative side of 0 and between 0.1 and 0.3 on the positive side
of 0 is considered quite small. Strong correlation is usually seen as existing at levels
less than -0.5 or greater than 0.5.
These two data points, for ahab and whale, appear to show only weak inverse
correlation and thus lead us only to further hypotheses and further testing. Unsurprisingly, this correlation test does not lead us to any easy conclusions about the
relationship between occurrences of whale and occurrences of ahab. Obviously,
there is much more to be considered.
Consider, for example, how the use of pronouns complicates these results. When
Ahab is not being referred to by name, he is undoubtedly appearing as either he or
him. The same may be said for the whale and the various appellations of whale that
Melville evokes: monster, leviathan, etc. Using the techniques described above, you
could investigate all of these and more. But before leaving this seemingly weak correlation, it might be useful to run a few more experiments to see just how significant
or insignificant the result really is.

50

5 Correlation

As noted above, the number of samples can be a factor in how the importance of
the correlation coefficient is judged, and in this case there are 135 observations for
each variable: one observation for each chapter in the novel.
One way of contextualizing this coefficient is to calculate how likely it is that we
would observe this coefficient by mere chance alone. In other words, assuming there
is no relationship between the occurrences of whale and ahab in the novel, what are
the chances of observing a correlation coefficient of 0.2411072? A fairly simple
test can be constructed by randomizing the order of the values in either the ahabs or
the whales column and then retesting the correlation of the data.

5.3 A Word About Data Frames

Before explaining the randomization test in detail, I want to return to something
mentioned earlier (see Chap. 4, Sect. 4, SubSect. 4) about the R matrix object and
its limitations and then introduce you to another important data object in R: the data
frame.
Thus far we have not used data frames, but as it happens, data frames are Rs
bread and butter data type, and they offer us some flexibility that we do not get with
matrix objects. Like a matrix, a data frame can be thought of as similar to a table
in a database or a sheet in an Excel file: a data frame has rows and some number
of columns, and each column contains a specific type of data. A major difference
between a matrix and a data frame, however, is that in a data frame, one column
may contain character values and another numerical values. To see how this works,
enter the following code to create a simple matrix of three rows by three columns:
x <- matrix(1, 3, 3)
x
##
[,1] [,2] [,3]
## [1,]
1
1
1
## [2,]
1
1
1
## [3,]
1
1
1

If you ask R to return the data type (class) of any one of the values in this matrix,
it will return the class numeric.
class(x[1,2]) # get class of cell in first row second column
## [1] "numeric"

Now change the value of one cell in this matrix so that it contains character data
instead of a number.
x[1,2] <- "Sam I am"
x
##
[,1] [,2]
## [1,] "1" "Sam I am"
## [2,] "1" "1"
## [3,] "1" "1"

[,3]
"1"
"1"
"1"

You will notice right away that all of the values in the matrix are now shown inside quotation marks. This is because the entire matrix has been converted to char-

5.3 A Word About Data Frames

51

acter data. Those 1s are no longer numbers; they are the 1 character. Among other
things, this means that you cannot perform mathematical operations on them any
more! If you check the class, R will report the change:
class(x[1,2]) # get class of cell in first row second column
## [1] "character"
class(x[1,3]) # get class of cell in first row third column
## [1] "character"

To see the difference between a matrix and a data frame, recreate the first matrix
example and then convert it to a data frame, like this:
x <- matrix(1, 3, 3)
x.df <- as.data.frame(x)
x.df
##
V1 V2 V3
## 1 1 1 1
## 2 1 1 1
## 3 1 1 1

You can see right away that a data frame displays differently. Instead of bracketed
row and column numbers you now see column headers (V1, V2, V3) and simple row
numbers without the brackets. You can now repeat the experiment from above and
assign some character data into one of the cells in this data frame.
x.df[1,2] <- "Sam I am"
class(x.df[1,2])# get class of cell in first row second column
## [1] "character"
class(x.df[1,3])# get class of cell in first row third column
## [1] "numeric"
x.df
##
V1
V2 V3
## 1 1 Sam I am 1
## 2 1
1 1
## 3 1
1 1

When using a matrix, the assignment of character data to any one cell resulted
in all the cells in the matrix being converted into character data. Here, with a data
frame, only the data in the column containing the target cell are converted to character data, not the entire table of data. The takeaway is that a data frame can have
columns containing different types of data. This will be especially useful as your
data get more complicated. You may, for example, want a way of storing metadata
(such as author gender or chapter title) alongside the numerical data associated with
these metadata facets.
Another handy thing about data frames is that you can access columns of data
using a bit of R shorthand. If you want to see all the values in the second column of
the x.df variable, you can do so using bracketed index references, just as you have
done previously with matrix objects. To see the entire second column, for example,
you might do this:
x.df[, 2]
## [1] "Sam I am" "1"

"1"

52

5 Correlation

Alternatively, you can use the fact that the data frame has a header to get column
information by referencing the column name, like this:
x.df[,"V2"]
## [1] "Sam I am" "1"

"1"

And, most alternatively, you can use the shorthand ($) to get column data like this:
x.df$V2
## [1] "Sam I am" "1"

"1"

That is a basic overview of data frames. Now we can return to correlating values in
Moby Dick.

5.4 Testing Correlation with Randomization

In this section you will use your new knowledge of data frames. First convert the
matrix object whales.ahabs.m into a data frame called cor.data.df:
# make life easier and convert to a data frame
cor.data.df <- as.data.frame(whales.ahabs.m)

As a gut check, you can use the cor function on the entire data frame, just as you
did with the matrix object. The output should be the same.
cor(cor.data.df)
##
whale
ahab
## whale 1.0000000 -0.2411072
## ahab -0.2411072 1.0000000

The goal now is to determine if that observed correlation coefficient of -0.2411

could have been likely to occur by mere chance. To assess this you are going to take
the values for one of the two columns in the data frame and shuffle them into a random order. Youll then run a new correlation test with the randomized column. In
this way a chance distribution of the values that is independent of the actual structure of the chapters in the book can be simulated. If the correlation of the shuffled
data is similar to the actual (as in unshuffled) data, then youll have to concede that
the relationship between whale and ahab observed in the actual data is really no
different from what might be observed if you threw all the occurrences of whale and
ahab up in the air and then created 135 arbitrary piles.
The first step is to randomize the order of the values (the word frequency measurements) in one of the two columns of data in cor.data.df. Since the columns
contain chapter-by-chapter measurements, this randomizing will have the effect of
shuffling the chapter order for one set of measurements and leaving the other set
in chronological order. R provides a function called sample for generating a random shuffling of data. At its most simple, the sample function requires a vector of

5.4 Testing Correlation with Randomization

53

values to shuffle. So, to get a random ordering of the values in the whale column of
cor.data.df you can simply enter:
sample(cor.data.df$whale)
##
[1] 0.66760365 0.61892131
##
[5] 0.28391557 0.00000000
##
[9] 0.11580776 0.16722408
## [13] 0.61099796 0.77565632
## [17] 0.00000000 0.67127746
## [21] 1.09151973 0.10638298
## [25] 0.56191467 0.06882312
## [29] 0.44247788 0.16037063
## [33] 0.50125313 0.62176166
## [37] 0.88832487 0.00000000
## [41] 0.83275503 0.41793313
## [45] 0.00000000 0.41841004
## [49] 0.94339623 0.00000000
## [53] 0.55865922 0.00000000
## [57] 0.13368984 1.76565008
## [61] 0.60716454 0.83682008
## [65] 0.00000000 0.39761431
## [69] 0.00000000 0.11926058
## [73] 0.82840237 0.10000000
## [77] 2.02788340 0.76965366
## [81] 1.82481752 0.69124424
## [85] 0.98159509 0.29296875
## [89] 0.22271715 0.82987552
## [93] 0.19762846 0.29411765
## [97] 0.64878893 0.30193237
## [101] 0.04448399 2.07452939
## [105] 0.71283096 0.14035088
## [109] 0.00000000 0.89485459
## [113] 0.81168831 0.27548209
## [117] 0.00000000 0.96566524
## [121] 0.39920160 0.87623220
## [125] 0.15723270 1.00767754
## [129] 0.32017076 1.26506024
## [133] 1.29198966 1.07469103

1.04895105
1.35440181
0.23790642
0.07047216
0.41841004
0.08207934
2.06782465
0.87131367
0.76628352
0.00000000
0.00000000
0.69930070
0.58167717
0.15485869
1.51515152
0.00000000
0.00000000
0.96562379
0.00000000
0.15829046
0.17341040
1.03578154
1.24777184
0.06079027
1.02739726
0.00000000
0.11286682
0.24067389
0.15313936
1.05485232
0.89726335
0.00000000
1.26582278
0.21097046

0.06079027
0.00000000
0.10857763
0.13114754
0.38722168
1.15546218
0.00000000
0.00000000
0.00000000
0.54305663
0.00000000
0.46838407
0.07949126
0.80200501
0.24711697
0.00000000
0.16260163
0.85653105
0.00000000
0.18761726
0.69620253
0.35971223
0.24375381
0.00000000
0.78616352
0.64400716
0.08748906
1.70807453
0.17942584
0.00000000
0.34364261
0.21901007
0.00000000

Go ahead and try entering this a few times and you will see that each time
sample randomly shuffles the order of the values from the whale column.
With the ability to randomize the values, you now need to correlate these randomized values against the ordered values in the unshuffled ahab column. Using
the dollar sign to reference the columns in the data frame, the expression can be
written as simply as this:
cor(sample(cor.data.df$whale), cor.data.df$ahab)

In my first test of this code, R returned a correlation coefficient of 0.0713801.1

I copied and pasted the code ten more times and observed the following correlation
coefficients for the various shuffles of the data.

Your results will be different given the sampling.

54
cor(sample(cor.data.df$whale),
[1] 0.122331
cor(sample(cor.data.df$whale),
[1] 0.00818978
cor(sample(cor.data.df$whale),
[1] -0.01610114
cor(sample(cor.data.df$whale),
[1] -0.1289073
cor(sample(cor.data.df$whale),
[1] 0.05115036
cor(sample(cor.data.df$whale),
[1] 0.0443622
cor(sample(cor.data.df$whale),
[1] 0.08513762
cor(sample(cor.data.df$whale),
[1] -0.1019796
cor(sample(cor.data.df$whale),
[1] 0.07842781
cor(sample(cor.data.df$whale),
[1] 0.04410211

5 Correlation
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)
cor.data.df$ahab)

As you see, in this small sample of ten randomizations, the highest positive correlation coefficient was 0.122331 and as lowest negative correlation coefficient
was -0.1289073. Remember that the actual correlation coefficient before we began shuffling anything was -0.2411072. In other words, the actual data seems to
be quite a bit below (i.e., further from 0) what is observed when shuffling the data
and simulating a chance distribution of values. Still, 10 randomizations is not very
many. Instead of copying and pasting the code over and over again, you can develop
a more programmatic way of testing the correlation using a for loop and 10,000
iterations!
With a for loop you can repeat the randomization and correlation test process
multiple times and at each iteration capture the result into a new vector. With this
new vector of 10,000 correlation values, it will be easy to generate some statistics
that describe the distribution of the random data and offer a better way of assessing
the significance of the actual observed correlation coefficient in the unshuffled data.
The code required for this is simple. Begin by creating an empty container variable called mycors.v, and then create a for loop that iterates a set number of
times (10,000 in my example). Within the curly braces of that loop, youll add
code for shuffling and then correlating the vectors. At each step in the loop, you will
capture the correlation coefficient by adding it to the mycors.v vector using the c
function. Here is how I wrote it:
mycors.v <- NULL
for(i in 1:10000){
mycors.v <- c(mycors.v, cor(sample(cor.data.df$whale), cor.data.df$ahab))
}

With this step completed, you can now use some basic R functions such as min,
max, range, mean and sd to get a general sense of the results. Here is what
my randomization tests returned; your results will be similar but not identical:

5.4 Testing Correlation with Randomization

55

min(mycors.v)
## [1] -0.2768103
max(mycors.v)
## [1] 0.3421222
range(mycors.v)
## [1] -0.2768103 0.3421222
mean(mycors.v)
## [1] 0.0001778007
sd(mycors.v)
## [1] 0.08796756

What these descriptive statistics reveal is that our actual observed value is more
typical of the extremes than the norm. A low standard deviation suggests
that most of the values recorded are close to the mean, and here the mean is very
close to zero (1.7780066 104 ), which you will recall from above can be interpreted as meaning very little correlation. A high standard deviation would
indicate that the values are spread out over a wide range of values. So even though
the min value of -0.2768103 is slightly less than our actual observed value of
-0.2411072, that 0.2768103 is very atypical of the randomized data. In fact,
using a bit of additional code that I will not explain here, I can generate a plot showing the distribution of all the values in mycors.v (Fig. 5.1).
The plot reveals, in dramatic fashion, just how much the data clusters around the
mean, which as you recall from above is nearly 0. It also dramatizes the outlier
status of the actual value (-0.2411072) that was observed. In 10,000 random
iterations, only 10 correlation coefficients were calculated to be less than the actual observed value and the actual observed value was nearly 3 (2.8) standard
deviations away from the mean. In short, the probability of observing a random
value as extreme as the actual value observed (0.2411072) is just 0.4 %.2
Here is the code I used to create the histogram:
h <- hist(mycors.v, breaks=100, col="grey",
xlab="Correlation Coefficient",
main="Histogram of Random Correlation Coefficients\n
with Normal Curve",
plot=T)
xfit <- seq(min(mycors.v),max(mycors.v),length=1000)
yfit <- dnorm(xfit,mean=mean(mycors.v),sd=sd(mycors.v))
yfit <- yfit*diff(h$mids[1:2])*length(mycors.v)
lines(xfit, yfit, col="black", lwd=2)

2 Another way to test the significance of a correlation coefficient is to use the cor.test function.
Use ?cor.test to learn about this function and then run it using the method="pearson"
argument. To make more sense out of the results, consider consulting http://en.
wikipedia.org/wiki/P-value on p-values and HTTP:// EN . WIKIPEDIA .ORG/WIKI/T-TEST
on t-tests.

56

5 Correlation

50 100
0

Frequency

200

Histogram of Random Correlation Coefficients

with Normal Curve

0.3

0.2

0.1

0.0

0.1

0.2

0.3

Correlation Coefficient

Fig. 5.1 Histogram plot of random correlation coefficients

Practice
5.1. Add two more columns to the matrix with data for the words i and my and then
rerun the cor function. What does the result tell you about the usage of the words i
and my?
5.2. Calculate the correlation coefficient for i and my and run a randomization test
to evaluate whether the results are significant.

Part II

Mesoanalysis

Chapter 6

Measures of Lexical Variety

Abstract In this chapter well begin to transition from microanalysis to macroanalysis. Well leave behind the study of single terms and begin to explore two
global measures of lexical variety: mean word frequency and type-token ratios.

6.1 Lexical Variety and the Type-Token Ratio

Moby Dick is a complicated book with a complex vocabulary. Readers of the book
inevitably remember Chapter Thirty-Two. This is the cetology chapter in which
Melville offers a zoological and pseudo-scholarly, pseudo-comical account of whale
history and physiology. Students frequently complain that this section of the novel
is more complex or difficult. One way to measure the complexity of the language is
to calculate a measure of vocabulary richness. Such a measure can be represented
as a mean word frequency or as a relationship between the number of unique words
used (i.e., the working lexicon) and a count of the number of word tokens in the
document. Using either measure as a proxy for lexical complexity or variety, you
can compare the lexical variety in the cetology chapter to the other chapters of the
novel.
Vocabulary richness is more commonly expressed as a percentage or ratio of
unique word types to the total number of word tokens. A type-token ratio, or TTR
as it is generally called, is calculated by dividing the total number of unique word
types by the total number of word tokens. The result is then typically multiplied by
100 so as to end with a percentage. As you can surmise, a lower type-token ratio
(TTR) is suggestive of less lexical variety.
In the previous chapters, you learned how to use a for loop to generate two
list objects containing tables of words and their frequencies for each chapter of
Moby Dick. The list titled chapter.raws.l contains the raw counts of each
word type and chapter.freqs.l contains the relative frequencies of each word
type in the given chapter. To calculate the mean word frequency and TTR values for
each chapter of Moby Dick, you will need chapter.raws.l. You can use your
existing code, or go to the code repository for this chapter.
M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_6,
Springer International Publishing Switzerland 2014

59

60

6 Measures of Lexical Variety

6.2 Mean Word Frequency

To calculate mean word frequency on a chapter-by-chapter basis, youll first get
the total number of word tokens in each chapter by summing the raw frequency
counts in each, and then youll calculate the number of unique word types in
each chapter. These are two very simple calculations that you can derive from the
chapter.raws.l list object. It is worth taking a moment to recall just exactly
what this list contains. The first thing youll want to know is the size or length of the
list.
length(chapter.raws.l)
## [1] 135

The length function reveals that there are 135 items in the list. As you will recall,
each list item corresponds to a chapter in the novel. If you want to see the chapter
titles, you can find them in another part of the list object that is accessed via the
names function.
names(chapter.raws.l)
[1] "CHAPTER 1. Loomings."
[2] "CHAPTER 2. The Carpet-Bag."
. . .
[135] "CHAPTER 135. The Chase.--Third Day."

Any item in a list may also be accessed and examined individually. Like so much
in R, such access is facilitated through bracketed subsetting. In this case, since you
have also stored the chapter titles as names for each list item, you can also use the $
shortcut character in the way that you did with data frames in the previous chapter.
The chapter.raws.l contains a series of table objects, which you can check by
calling the class function on one item of the list:
class(chapter.raws.l$"CHAPTER 1. Loomings.")
## [1] "table"

or, if you do not know the exact name, you can achieve the same result using the
numerical index:
class(chapter.raws.l[[1]])
## [1] "table"

As you can see, the first item in the list is a table object. To access the word
frequency table for this first chapter, just remove the call to the class function
using either this expression
chapter.raws.l$"CHAPTER 1. Loomings."

or this
chapter.raws.l[[1]]

to return a table of word types and the corresponding counts of those words in the
first chapter.

6.3 Extracting Word Usage Means

61

You have already learned how to find useful information about R objects using the
class function. For even more detailed information, you saw how to use str, the
structure function. The class function tells us something general about the kind
of data an R object contains. For example, entering class(chapter.raws.l)
shows us that the object is a list. Alternatively, str(chapter.raws.l) returns additional information about the size (135 items) and construction of the list,
including a long string of output containing information about each of the 135 items
in the list.
str also reveals that the first item of the main list is a table object with 854
items. The line of output that begins with the word table shows the counts for the
first 10 items in the table. The int that you see tells you that the items are stored in
an integer vector. Two lines below that is another line showing a series of words: a
abandon etc. These are prefaced with the chr marker. chr indicates that these
values are stored as a character vector. These, of course, are the actual word types
from the chapter, and the integers above them are the counts of the occurrences of
these word types in the chapter. The word a, for example, occurs 69 times in the
first chapter. The fourth word in the vector, about, occurs seven times in the chapter
and so on.
With an understanding of how and where the data are stored, it is a fairly routine
matter to calculate the mean word frequency for each chapter. Summing the integer
values held in each chapter table will give you a count of all word tokens in the
chapter, and you can use the length function to return the total number of word
types.
sum(chapter.raws.l[[1]])
## [1] 2244
length(chapter.raws.l[[1]])
## [1] 854

Using these two figures, the mean word frequency can be calculated by dividing the
total number of tokens by the total number of unique word types.
sum(chapter.raws.l[[1]])/length(chapter.raws.l[[1]])
## [1] 2.627635

The result shows that each word type in the first chapter is used an average of 2.6
times. A much simpler way of doing this is to just use Rs built-in mean function.
mean(chapter.raws.l[[1]])
## [1] 2.627635

6.3 Extracting Word Usage Means

Since the chapters are already in a list object, all that you need now is a method
for extracting the frequency data from all of the chapters at once. For this you can
employ the lapply function. lapply is an alternative to for that in some sense

62

6 Measures of Lexical Variety

simply hides the operations of a for loop. In another sense, lapply simplifies the
code needed by automatically generating a new list object for a result. That is, we
do not need to create an empty list outside of a loop and then fill it with each iteration
of the loop. lapply takes two arguments: a list object and a function to apply to
the items in the list. To get the mean word usage for each chapter in Moby Dick, for
example, you could use lapply (chapter.raws.I, mean) as follows:
lapply(chapter.raws.l,mean)
$CHAPTER 1. Loomings.
[1] 2.628
$CHAPTER 2. The Carpet-Bag.
[1] 2.34
$CHAPTER 3. The Spouter-Inn.
[1] 3.72
. . .
$CHAPTER 135. The Chase.--Third Day.
[1] 3.44

Calling lapply in this way generates a new list object. In the example here,
I have just printed the results to the screen instead of saving the output into a new
object.
Since a list is not very handy for further manipulation, you can wrap this
lapply expression inside a do.call function, which will take the list output
from lapply and apply another function (rbind) to the results. This has the effect
of putting all of the results into neat rows in a matrix object. Since you want to be
able to plot this data, you can direct the results of all of this into a new object called
mean.word.use.m.
mean.word.use.m <- do.call(rbind, lapply(chapter.raws.l,mean))

The dimensions of the resulting matrix can be obtained using the dim function (for
dimensions):
dim(mean.word.use.m)
## [1] 135
1

dim reports that the matrix has 135 rows and 1 column. But there is a bit
more information stored in this matrix object, and you can get a hint of that content by using the str function discussed above. During the creation of this matrix,
the individual chapter names were retained and assigned as rownames. Entering
rownames(mean.word.use.m) returns the names.
At this point, you can plot the values and visualize the mean word usage pattern
across the novel. Calling plot returns a simple barplot (Fig. 6.1), in which chapters
with higher bars are, in one manner of speaking, less rich.
plot(mean.word.use.m, type="h")

In the chapters with high values, individual word types are used more often; there
is more repetition of the same word types. Alternatively, in chapters where the bar is
low, each word type has a lower overall usage frequency. In the chapters with high
bars, the reader can expect to see the same words repeated rather frequently; in the

63

3.0
2.5
1.5

2.0

mean.word.use.m

3.5

6.3 Extracting Word Usage Means

20

40

60

80

100

120

140

Index

Fig. 6.1 Barplot of mean word use

lower bar chapters, the reader is treated to a collection of words that might give the
impression of greater variety for being repeated less often.
To be more interpretable, you may want to consider normalizing these values
across the entire text. R provides a scale function for normalizing or scaling data.
In such scaling, the overall mean for all of the chapters is first calculated and then
subtracted from each individual chapter means. This method has the effect of subtracting away the expected value (expected as calculated by the overall mean) and
then showing only the deviations from the mean. The result is a vector of values that
includes both positive and negative numbers. You can look at the scaled values by
entering:
scale(mean.word.use.m)

Instead of just studying the numbers, however, it might be better to visualize the
results as a barplot similar to Fig. 6.1. In the resulting plot, 0 on the y-axis will
correspond to the mean across the entire novel. You will only see the deviations
from the mean (Fig. 6.2).
plot(scale(mean.word.use.m), type="h")

By this measure of mean word use, the cetology chapter, which readers so often
remember as being one of denser, richer vocabulary, is not exceptional at all. Words
in the cetology chapter are repeated fairly often. In fact, each unique word type is
used an average of 3.5 times.

6 Measures of Lexical Variety

1
0
2

scale(mean.word.use.m)

64

20

40

60

80

100

120

140

Index

Fig. 6.2 Mean word usage plot

6.4 Ranking the Values

To see where the cetology chapter ranks in terms of average word use, you can
employ the order function to arrange the data in decreasing rank order. Beware,
however, that the order function can be confusing. If you enter:
order(mean.word.use.m)

R will return a vector of numbers corresponding to the ranked positions of each item
in the mean.word.use.m vector, like this:
order(mean.word.use.m)
##
[1] 122 111 120 97 30
## [12] 69 25 95 37 43
## [23] 77 117 112 70
7
## [34] 98 121 49 11 62
## [45]
8 12 88 76 26
## [56] 94 50 109 65 127
## [67] 56 125 28 90 27
## [78] 96 68 75 105 40
## [89] 78 128 130 126 86
## [100] 72 44 53 35 34
## [111] 108 99 119 31 134
## [122] 133 87 100 45 48
## [133] 16
3 54

23 131 114 116 39 38

66
6 63 84 14 52
57 59
5 107 115 92
58 67 33 47 129 106
46 102 118 79 104 83
20 124 103 51 93 21
2 80 15 82 29 123
13 24 113 61 132 60
22 10 101 55
4 18
89
1 71 42 110 74
36 41 85 17 91 19
9 64 81 73 135 32

6.5 Calculating the TTR Inside lapply

65

What this vector reveals is that the first item in the mean.word.use.m object,
the mean of the word usage in chapter one of the novel, is the 122nd in rank when
the means are sorted in increasing order, from smallest to largest.
If you want to order them according to decreasing rank, you need to set the
decreasing argument of order to TRUE.
order(mean.word.use.m, decreasing=TRUE)

Again, just to emphasize this point, order does not order the means; it returns
a vector of ranks in which the vector positions correspond to the positions in the
vector being ranked. The vector of means can then be reordered, sorted, using this
new vector of rank positions inside the brackets:
mean.word.use.m[order(mean.word.use.m, decreasing=TRUE),]

Here is an abbreviated look at the results:

CHAPTER 54. The Town-Hos Story.
3.748727
CHAPTER 3. The Spouter-Inn.
3.719777
CHAPTER 16. The Ship.
3.692105
CHAPTER 32. Cetology.
3.477622

After sorting you will see that the cetology chapter has the fourth largest mean of
the 135 chapters. Only three other chapters recycle words at the rate of the cetology
chapter! By this measure, it is not an especially interesting chapter at all.

6.5 Calculating the TTR Inside lapply

The last few sections demonstrated how to use Rs built-in mean function with
lapply to calculate the mean word frequency of each chapter. Mean word usage
is one way of thinking about lexical variety. A Type-Token Ratio provides a similar
value for assessing lexical richness, but since R does not already have a function for
calculating TTR, you will need to modify the arguments given to lapply. Instead
of using mean, you will create your own function.
Above, you saw that you could calculate the mean for one chapter using this
expression:
sum(chapter.raws.l[[1]])/length(chapter.raws.l[[1]])
## [1] 2.627635

You can calculate the TTR using a similar expression in which the numerator and
denominator are reversed.
length(chapter.raws.l[[1]])/sum(chapter.raws.l[[1]])*100
## [1] 38.05704

66

6 Measures of Lexical Variety

To run a similar calculation for all of the chapters as part of a call to lapply
a generalized version of this calculation needs to be provided to lapply as the
function argument. As you have seen, lapply takes a function argument such as
mean or sum, etc. Since there is no function for TTR, you can provide lapply
with your own custom function. In place of an existing function such as mean you
can insert an inline function definition using a variable x to stand in for each item
in the main list.
ttr.l <- lapply(chapter.raws.l, function(x) {length(x)/sum(x)*100})

Within the parentheses of lapply the TTR function is defined as follows:

function(x) length(x)/sum(x)*100. When executed, lapply will treat
each item in the chapter.raws.l list as the value for x (much in the same way
that we have been using i inside of a for loop). The calculations will be performed on each item and the results returned in a new list object. You can then run
do.call with rbind just as you did when calculating the means.
ttr.m <- do.call(rbind, ttr.l)

Now order the results:

ttr.m[order(ttr.m, decreasing=TRUE),]

or plot them:
plot(ttr.m, type="h")

6.6 A Further Use of Correlation

Unfortunately, measures such as the mean word frequency and TTR are not terribly
useful because text length, or chapter length in this case, can be a strong determiner in the rate at which words get recycled. As chapter length increases you can
generally expect more new words to be introduced. At the same time, many of the
existing words will see repeated use because they provide the necessary structure
or scaffolding for the introduction of new words. The practice exercises that follow
provide you an opportunity to test these assertions.1

1 In addition to the two measures of lexical variety offered in this chapter, and another approach
offered in the next, readers may wish to consider Yules K (see Yule G. Udny. The Statistical
Study of Literary Vocabulary. Cambridge University Press, 1994). Yule attempts to compensate
for text length and provide a stable measure of lexical variety in what he called the K characteristic. A function for computing Yules characteristic constant K can be found in the languageR
package.

6.6 A Further Use of Correlation

67

Practice
6.1. To test the assertion that document length is a strong determiner in the rate
at which words get recycled, measure the strength of correlation between chapter
length and TTR. For this you need two vectors of data. You already have the TTR
values in the ttr.m matrix. Convert that data to a vector using as.vector. You
now need another vector of chapter lengths. For this you can use lapply with the
sum function instead of using mean. Once you have the two vectors, run a correlation analysis similar to the correlation you did previously with occurrences of whale
and ahab. Write up your code and an analysis of the result.
6.2. Run a similar correlation test using the values in the mean.word.use.m
instead of the TTR values. Write up your code and an interpretation of the result.
6.3. Use randomization to test the likelihood that the correlation coefficient observed
with the TTR values could have been the result of chance. Explain the outcome of
the test.
6.4. Explain the difference between the results derived in practice Exercises 6.1
and 6.2.

Chapter 7

Hapax Richness

Abstract This chapter expands the analysis of vocabulary by focusing on words that
occur very infrequently. Readers learn how to use sapply and to create another
simple function.

7.1 Introduction
Another way of thinking about vocabulary richness and the experience of reading a
particular text is to consider the fact that many words appear quite infrequently or
even just once. These words that occur just once are sometimes referred to as singletons or even one-zies, but they are more formally called hapax legomena. Hapax
(for short) may provide a different way of assessing the lexical richness of a given
segment of prose. In this chapter you will learn how to calculate the total number of
hapax and see if there is a correlation between the number of hapax and the length
of a chapter. The working hypothesis will be that as chapter length increases, you
would expect to see an increase in the number of hapax legomena.

7.2 sapply
For this analysis, you must return to the chapter.raws.l list. Instead of extracting a count of all word tokens, you will compute a sum of all of the word types that
appear only once in each chapter. To extract a count of the hapax, you can use the
sapply function in combination with an argument that identifies the values that
are equal to 1. sapply is a simplified, or, as the R documentation calls it, a userfriendly version of lapply. The main difference is that sapply returns a vector
instead of a list. The arguments that you provide to sapply are going to be very
similar to those given to lapply, but here you are going to add some additional
conditions in the form of a custom function that calculates a sum of only those
M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_7,
Springer International Publishing Switzerland 2014

69

70

7 Hapax Richness

values that meet the condition of being equal to 1. Your function is going to count
how many words in the vector are only used once.

7.3 A Mini-Conditional Function

Instead of using built-in functions such as mean or sum, for this task you need to
construct your own function using the inline function(x) argument of sapply
followed by a definition, or declaration, of that function. This is similar to what you
did in Chap. 6 in order to compute the TTR of each chapter. In this case the custom
function will enclose a conditional expression that sums all the values in the raw
counts table that are equal to one. It is important to emphasize here that to express
equivalence, R expects the use of two equal signs (==). As you recall, this is done
to avoid confusing the use of a single equals sign, which can be used in R as an
assignment operator.1 The code required for counting the number of hapax in each
chapter, therefore, looks like this2 :
chapter.hapax.v

<-

sapply(chapter.raws.l, function(x) sum(x == 1))

Translating this code block into plain English, we might say something like this:
For each item in chapter.raws.l, return the sum of the values that are equal
to one. Since the values in this case are all one, sum is in essence returning a
count of the words that occur just once in each chapter. If you print the contents
of chapter.hapax.v to the console, you will see the hapax counts for each
chapter.
chapter.hapax.v
CHAPTER 1. Loomings.
605
CHAPTER 2. The Carpet-Bag.
433
. . .
CHAPTER 135. The Chase.--Third Day.
903

This is a start, but now you need to divide the number of hapax in each chapter
by the total number of words in each chapter. As it happens, you already have these
values in the chapter.lengths.m variable from your in practice Exercise 6.1
in the last chapter. Here it is again:
chapter.lengths.m <- do.call(rbind, lapply(chapter.raws.l,sum))

Since R easily facilitates matrix division (that is, R allows you to divide one
matrix of values by the corresponding values in another matrix) the code is simple.
Instead of having to perform the division on one value at a time, like this
1

In R values can be assigned to an object using either <- or =. Throughout this book I use <because it is the most common convention among users of R, and it avoids the whole = vs. ==
confusion.
2 Remember that you can get the start up code for this chapter from Appendix C or the code
repository.

7.3 A Mini-Conditional Function

71

chapter.hapax.v[1] / chapter.lengths.m[1]
CHAPTER 1. Loomings.
0.2696078
chapter.hapax.v[2] / chapter.lengths.m[2]
CHAPTER 2. The Carpet-Bag.
0.2980041

You can do it all at once, like this:

hapax.percentage

<-

chapter.hapax.v / chapter.lengths.m

This expression returns a new matrix containing the chapter names and the percentage of hapax in each chapter. These values can then be plotted, so that you can
visualize the chapter-by-chapter hapax richness (Fig. 7.1).

0.0

0.1

0.2

0.3

0.4

0.5

barplot(hapax.percentage, beside=T,col="grey",
names.arg = seq(1:length(chapter.raws.l)))

1 7 15 24 33 42 51 60 69 78 87 96 106 117 128

Fig. 7.1 Hapax percentage plot

72

7 Hapax Richness

Practice
7.1. First calculate the extent of the correlation between chapter length and the number of hapax. Offer a brief interpretation of the result (similar to what you did in
Exercise 6.1).
7.2. Use order to rank the values in hapax.percentage. How does the rank
of the Cetology chapter compare to the others.
7.3. The correlation statistic found in Exercise 7.1 is not especially useful in and of
itself. It becomes much more interesting when compared to another authors work,
such as Jane Austens Sense and Sensibility that you will find in the same folder
as you found Moby Dick: data/plainText. First write the code necessary to
get the chapter-by-chapter counts of the hapax in Austens Sense and Sensibility.
Save these in a list object titled sense.raws.l. From this object calculate the
number of hapax and the chapter lengths for each chapter of Sense and Sensibility.
Compute the correlation and describe how the correlation results for Melville and
Austen compare and what they tell us about the two writers in terms of vocabulary
usage habits.
7.4. Use what you learned in the previous chapter to test the likelihood that these two
correlation figures could have been the result of chance. To complete this exercise
you will need to code separate randomization tests for both Moby Dick and Sense
and Sensibility.3

Thanks to Carmen McCue for this exercise suggestion.

Chapter 8

Do It KWIC

Abstract In the last chapter a simple function was created within a call to the
sapply function. In this chapter we explore user-defined functions more broadly
and write a function for producing a keyword in context (KWIC) list.

8.1 Introduction
KWIC or Keyword in Context searches are a standard way of addressing Rupert
Firths observation that you will know a words meaning or sense by looking at the
other words around it, that is, by its context.1 In this section (including Exercises 8.1
and 8.2), you will learn how to build a flexible KWIC tool in R. You will also be
introduced to some R functionality that will allow you to access and analyze multiple
texts at once.
Begin as usual by setting a path to a working directory:
setwd("~/Documents/TextAnalysisWithR")

Unlike previous chapters where you loaded a single file using scan, you are now
going to access a collection of files in a directory. Begin by defining a variable that
contains the relative path to the directory of text files.
input.dir <- "data/plainText"

You can now call Rs dir function using this path variable to retrieve the names
of all of the files in the directory. In addition to the path argument, which you have
now stored in the input.dir object, the dir function can take an optional search
pattern string written using a regular expression. Since you only want dir to return
the files in the directory that have a .txt file extension, use a period . followed
by txt and the $ symbol (to mark the end of a string). This create a matching
1

See Firth, John Rupert. A Synopsis of Linguistic Theory, 19301955. In Studies in Linguistic
Analysis. Oxford: Blackwell (1957): 132.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_8,
Springer International Publishing Switzerland 2014

73

74

8 Do It KWIC

pattern (in regular expression speak) that will match any string of characters that
ends with .txt. The expression looks like this:
files.v <- dir(input.dir, "\\.txt$")

Having run this, you should now be able to type files.v into R and return a
vector of file names; in this example, just two files.
files.v
## [1] "austen.txt"

"melville.txt"

8.2 Custom Functions

This console display of the content of the files.v vector is not very pretty and
once you get a corpus containing many files, it can become difficult to read through
this display. To better understand the idea of custom functions, you will write code to
make the contents of the files.v vector display in a more organized and readerfriendly fashion.
Functions are, primarily, reusable chunks of code. You have already learned
about many of Rs built-in functions, and you have used and reused them many
times. You can also create your own custom functions and call them over and over
again. If you were baking a cake for a friends birthday, youd buy some ingredients,
pull out some pots and pans, and bake the cake. If, on the other hand, you decided to
go into the cake baking business, youd probably invest some time (and money) setting up a cake baking system that would take a certain set of ingredients and pump
out a cake on the other end. Thats what functions are; they are ingredient assembly
systems.
In this section you will write a simple function called show.files that you can
use for displaying file names. The specific purpose of the function in this example
will be to display the contents of the files.v vector in an easy to read format. In
R you begin a new function by giving it a name (show.files) and then using the
function declaration (which looks like you are calling the function function).
Inside parentheses, you will define the arguments (ingredients) that the function
requires: in this case, just one argument, a vector of file names.2
show.files

<- function(file.name.v)

Here is the basic outline of the function:

show.files <- function(file.name.v){
# some code goes here
}

It is worth noting that the name(s) we assign to the arguments inside the parentheses have a scope
that is limited to the function; these variable names do not persist or exist outside of the function,
and they will not overwrite a variable with the same name that may exist outside of the function.
That said, it is best to avoid duplication of variable names. Notice that I have named the argument
file.name.v instead of files.v. For an example in code, see Appendix B.

8.2 Custom Functions

75

The parenthetical arguments section is followed by a set of opening and closing

curly braces that surround the inner workings of the function. This inner section,
inside the curly braces, is where the instructions (recipe) for how to assemble the
ingredients will be defined.
The objective of using this function is to provide some sort of pretty printed list
of the files in the files.v variable. You already know that you can use a for
loop to iterate over the items in a vector, so add code for a for loop inside the curly
braces here:
show.files <- function(file.name.v){
for(i in 1:length(file.name.v)){
# some code goes here
}
}

This for-looping should be old hat for you by now, but just to review, the for loop
will be used to iterate over each of the items stored in the vector of file names, one
item at a time.
For printing lines into the R console (do not confuse this sense of the word print
with printing on paper to your printer), R has several functions you could tap, but
lets use the cat function here. cat is a function for concatenation (joining) and
printing. Here you want to join the file name with the index of the file name in the
vector and then add a line return (using the backslash escape character and an n to
mean newline). To achieve this, you will be joining three items: the vector index key,
the contents or value of the item in the vector at that index position, and a newline
character. To join these pieces requires a bit of glue, so cat asks us to define a
separator using the sep argument. You can use a space character for the glue, and
the final function looks like this:
show.files <- function(file.name.v){
for(i in 1:length(file.name.v)){
cat(i, file.name.v[i], "\n", sep=" ")
}
}

Before taking the function for a test drive (cake walk?), there is one more thing
to do. Just as it is very easy to write complicated functions, it is also very easy to
forget what they do! So add a little comment to the function so that when you come
back a week later you can be reminded of what it does:
# Function
# friendly
show.files
for(i in
cat(i,
}
}

to print a vector of file names in user

format
<- function(file.name.v){
1:length(file.name.v)){
file.name.v[i], "\n", sep=" ")

If you have not already done so, set your working directory and enter the following code to load the data and the new function.3
3 Remember that, you can run a whole block of code by selecting it all in the editing window and
then hitting the control and return keys at the same time.

76

8 Do It KWIC

input.dir <- "data/plainText"

files.v <- dir(input.dir, "\\.txt$")
# Function to print a vector of file names in user
# friendly format
show.files <- function(file.name.v){
for(i in 1:length(file.name.v)){
cat(i, file.name.v[i], "\n", sep=" ")
}
}

With all of this done, you can now send the files.v object to the
show. files function and see the numbered result as output.
show.files(files.v)
## 1 austen.txt
## 2 melville.txt

Mission accomplished, your first function!

8.3 A Word List Making Function

In previous chapters, you learned how to generate a word vector from a text file.
Since you are now an expert function maker, youll build a function that will do
this task for a whole bunch of files, and youll write the function to store all of
the results in a list so that you can easily access the word data from any file in
your corpus. Unlike the show.files function from above, this new function will
take two arguments: a vector of file names and a directory path telling the function
where to find the files on your system. It is always a good idea to give your functions
names that make sense; call this one make.file.word.v.l and begin it with a
comment:
# Function takes a vector of file names and a directory path and
# returns a list in which each item in the list is an ordered
# vector of words from one of the files in the vector of file names
make.file.word.v.l <- function(files.v, input.dir){
}

Here I have entered a comment describing what I want this function to do.
Articulating an objective in advance can be a great way to guide your coding.
The definition says that I want to return a list object, so the first thing to do
is to instantiate one inside the function, an empty one in this case. Ill call it
text.word.vector.l:
make.file.word.v.l <- function(files.v, input.dir){
text.word.vector.l <- list()
# more code needed here to iterate over the input vector. For
# each file we need to load the text and convert it into a word
# vector
}

Notice that here again is another comment to help keep track of what is happening at each stage in the process. It turns out that everything required in that

8.3 A Word List Making Function

77

comment is code that you have already written in previous sections and/or exercises.
By recycling code from your prior work you can produce the following:
make.file.word.v.l <- function(files.v, input.dir){
#set up an empty container
text.word.vector.l <- list()
# loop over the files
for(i in 1:length(files.v)){
# read the file in (notice that it is here that we need to know the input
# directory
text.v <- scan(paste(input.dir, files.v[i], sep="/"),
what="character", sep="\n")
#convert to single string
text.v <- paste(text.v, collapse=" ")
#lowercase and split on non-word characters
text.lower.v <- tolower(text.v)
text.words.v <- strsplit(text.lower.v, "\\W")
text.words.v <- unlist(text.words.v)
#remove the blanks
text.words.v <- text.words.v[which(text.words.v!="")]
#use the index id from the files.v vector as the "name" in the list
text.word.vector.l[[files.v[i]]] <- text.words.v
}
return(text.word.vector.l)
}

You now have code that will open each file in the input directory, and for each file
it will create a word vector and store that word vector as an item in a list object.
The only thing that remains to do, once the inner for loop has completed its cycle,
is to return the list object back to the main script. For this, you need to call the
return function, which is a way of telling the function to take all the cakes it has
built and send them back to the head baker in a nice box. So the last line of code
before the closing curly brace reads4 :
return(text.word.vector.l)

If you have not yet done so, write out this function and then load it in R by either
copying and pasting it into the console or using the RStudio shortcut. You can
then call the function with the files.v and input.dir arguments and put the
returned result into the new variable my.corpus.l:
my.corpus.l <- make.file.word.v.l(files.v, input.dir)

If everything worked, you should have obtained the following message from R:
Read 10906 items
Read 18874 items

Assuming that all is well, you will now have a new list object called
my.corpus .l. Use the class and str functions to investigate its contents.

4 In R you do not always have to explicitly call return. By default R will return whatever object
is in the last line of the function. Explicitly calling return, however, often makes it easier to read
and debug your function code.

78

8 Do It KWIC

8.4 Finding Words and Their Neighbors

Now the fun begins. Consider that for each file (each item in the my.corpus.l
list), you now have an ordered vector of the words. If you were to enter
my.corpus.l[[1]][1:100]

you would get the first 100 words of Jane Austens novel, one word at a time.5 At
this point, I hope that you are already one step ahead of me and thinking to yourself,
hey, if I have all the words in order, I can find any word in the text and return its
position in the text using a which statement. You already did this when you found
the occurrences of whale in Moby Dick. It is a little bit trickier here because you are
working with a list, but consider this next R expression in which the double brackets
after the my.corpus.l object provide a way of accessing the first item in the list:
positions.v <- which(my.corpus.l[[1]][]=="gutenberg")

After indicating a desire to access the vector held as the first item in the list using
the double-bracketed number 1, the remainder of the expression seeks out which of
those values is a match for the keyword: gutenberg. This expression will return the
positions of every instance of the word gutenberg in the vector that is indexed as the
first item in my.corpus.l. Go ahead and enter this now, and see what you get.
The result should be something like this:
positions.v
## [1]
3
## [7] 120925
## [13] 121155
## [19] 121281
## [25] 121451
## [31] 121677
## [37] 121804
## [43] 121986
## [49] 122163
## [55] 122343
## [61] 122559
## [67] 122731
## [73] 123205
## [79] 123328
## [85] 123560
## [91] 123927

47
121004
121168
121326
121522
121709
121814
122020
122215
122385
122587
122740
123222
123339
123778
123936

57
121012
121179
121371
121535
121728
121823
122043
122244
122402
122599
122750
123225
123372
123819
123946

86
121018
121191
121401
121549
121740
121882
122091
122278
122445
122618
123133
123289
123378
123833

120883
121110
121206
121421
121564
121750
121918
122134
122302
122481
122641
123154
123298
123438
123857

120896
121126
121259
121434
121594
121758
121940
122158
122327
122542
122648
123192
123313
123555
123869

These are the positions of every occurrence of gutenberg in the file titled
austen.txt. If you can find the position of every occurrence of gutenberg in the word
vector, you can find any other word (i.e., whale or dog). And, if you can find a
words position, you can also find the items that are next to it: before it and after
it. You can do this by simply adding or subtracting values from the position of the
found word. Deep breath.
To summarize, you have used the which statement to find all the instances of
gutenberg and stored those positions in a new vector called positions.v. If you
5 If you try this now, youll see that you are actually getting Project Gutenbergs boilerplate words,
but you get the idea.

8.4 Finding Words and Their Neighbors

79

check the length of this positions.v vector, you will get a count of the number
of times gutenberg occurs in the file:
length(positions.v)
## [1] 93

Now lets say that you want to know the words that come just before and just
after the first instance of gutenberg in this file. You might begin by specifically
identifying the first instance:
first.instance <- positions.v[1]

Which is to say that you could put the value that is held in the first item
in the positions.v vector into a new variable called first.instance.
If you look contents of positions.v, youll see that the first value in the
positions.v vector is 3. The first instance of gutenberg is the third word in
the file. With this last R expression, you have put the number 3 into the variable
called first.instance.
If you want to check your work, just use that new variable in the original word
vector, like this:
my.corpus.l[[1]][first.instance]
## [1] "gutenberg"

Ta Da! Of course, since you already knew that gutenberg is the third word in the
file, you could have also done this:
my.corpus.l[[1]][3]
## [1] "gutenberg"

Ta Da! And, if you want to see the words just before and just after the third word in
the file, you could, of course, just do this:
my.corpus.l[[1]][2:4]
## [1] "project"
"gutenberg" "ebook"

But now consider that another way of getting access to the second and fourth
positions in the vector is to add and subtract 1 from 3. Since 3 is the value
already stored in the first.instance variable, you could subtract 1 from
first.instance. With that in mind, you can use the following expression to
achieve the same result as above, but without hard coding any of the vector positions.
my.corpus.l[[1]][(first.instance-1):(first.instance+1)]
## [1] "project"
"gutenberg" "ebook"

If you want to see the results pretty printed, just use cat:
cat(my.corpus.l[[1]][(first.instance-1):(first.instance+1)])
## project gutenberg ebook

80

8 Do It KWIC

Practice
8.1. Using the functions described in this chapter and what you now know about
vector indexing, write a script that will produce a five-word KWIC list for all occurrences of the word dog in both Moby Dick and Sense and Sensibility.
8.2. KWIC with Cleaner Output
For an even cleaner look, use your new knowledge of the cat function to format
your output so that it looks something like this:
----------------------- 1 ----------------------all over like a newfoundland [dog] just from the water and
----------------------- 2 ----------------------a fellow that in the [dog] days will mow his two
----------------------- 3 ----------------------was seen swimming like a [dog] throwing his long arms straight
----------------------- 4 ----------------------filling one at last down [dog] and kennel starting at the
----------------------- 5 ----------------------not tamely be called a [dog] sir then be called ten
----------------------- 6 ----------------------t he call me a [dog] blazes he called me ten
----------------------- 7 ----------------------sacrifice of the sacred white [dog] was by far the holiest
----------------------- 8 ----------------------life that lives in a [dog] or a horse indeed in
----------------------- 9 ----------------------the sagacious kindness of the [dog] the accursed shark alone can
----------------------- 10 ----------------------boats the ungracious and ungrateful [dog] cried starbuck he mocks and
----------------------- 11 ----------------------intense whisper give way greyhounds [dog] to it i tell ye
----------------------- 12 ----------------------to the whale that a [dog] does to the elephant nevertheless
----------------------- 13 ----------------------aries or the ram lecherous [dog] he begets us then taurus
----------------------- 14 ----------------------is dr bunger bunger you [dog] laugh out why don t
----------------------- 15 ----------------------to die in pickle you [dog] you should be preserved to
----------------------- 16 ----------------------round ahab and like a [dog] strangely snuffing this man s
----------------------- 17 ----------------------lad five feet high hang [dog] look and cowardly jumped from
----------------------- 18 ----------------------as a sagacious ship s [dog] will in drawing nigh to
----------------------- 19 ----------------------the compass and then the [dog] vane and then ascertaining the

Chapter 9

Do It KWIC (Better)

Abstract This chapter expands upon the previous chapter in order to build an interactive and reusable Key Word in Context (KWIC) application that allows for quick
and intuitive KWIC list building. Readers are introduced to interactive R functions
including readline and functions for data type conversion.

9.1 Getting Organized

In the previous chapter, you learned how to find and access a series of index positions in a vector and then how to return values on either side of the found positions.
In the exercise at the end of Chap. 8 you hard coded a solution for finding occurrences of the word dog in Sense and Sensibility and Moby Dick. In this section youll
learn how to abstract that code and how to create an interactive and reusable application that will allow you to repeatedly find key words in context within any directory
of files, all without having to hard code the search terms.
If you have not already done so, now is the time to get organized. You will be
dealing with more and more files as this book continues, and unless you keep your
working spaces well defined and organized things can get complicated. Within your
TextAnalysisWithR directory, you should already have a sub-directory labeled
code. This is where you should store all of your .R files. You will also have another
sub-directory labeled data that contains all of your text/corpus files. If you do not
already have a sub-directory called results, create one now because in the last
exercise in this chapter youll be generating a .csv file that you can save and then
open in R or in a spreadsheet application such as Excel or Open Office.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_9,
Springer International Publishing Switzerland 2014

81

82

9 Do It KWIC (Better)

9.2 Separating Functions for Reuse

In the last chapter you created two functions, and in this chapter you will create
another. Because you can reuse functions in separate projects, it is convenient to
keep them in a separate file so that you can access them from different R scripts
that you write for different projects. You should begin this chapter, therefore, by
copying your functions into a new file that you will title corpusFunctions.R. Save
this new file inside your code sub-directory. Your functions file should include both
show.files and make.file.word.l from the last chapter.
With your functions stored in a separate file, you can now call the functions file
as part of a new R script. Open R, or, if it is already open, start a new session by
clearing your workspace. Open a new document in your code editor and begin by
defining a working directory with the setwd function or by choosing a working
directory using RStudios menu shortcut. Now enter the following expression
that uses Rs source function to show R where to find your functions file.
source("code/corpusFunctions.R")

When your main script is executed, R will load all of the functions in the corpusFunctions.R file.
As in Chap. 8, you need to show R where to find your text files, so next you will
define an input directory with a relative path to the data/plainText directory
input.dir <- "data/plainText"

Since you will also be using R to create derivative data files, which will need to
be saved out to another directory, youll need to tell R where to write these files.
Define an output directory like this:
output.dir

<- "results/"

You can now recycle a few lines of code from the prior chapter to load the corpus
into memory:
files.v <- dir(input.dir, ".*txt")
my.corpus.l <- make.file.word.v.l(files.v, input.dir)

The objective now is to write a KWIC application that will allow you to repeatedly enter different keywords and return the hits for those terms along with some
amount of context on either side of the key term.

9.3 User Interaction

R includes a set of built-in functions that, when invoked, require user feedback to
complete. Thus far we have been hard-coding file paths in R, but we could have
been using Rs file.choose function instead. If you enter file.choose() at

9.5 Building a Better KWIC Function

83

the R prompt, you will be prompted with a new popup window that allows you to
navigate your file system and locate a file. Here is an example that you can try on
your system. Just enter the following expression in the R prompt in the console pane
and then use your computers windowing system to locate the file in the exercise
directory called melville.txt.
mytext <- scan(file.choose(), what="character", sep="\n")

If you did everything correctly, you should see the message:

Read 18874 items

You will now be able to enter

mytext

and see all the lines of Moby Dick.

9.4 readline
There are other functions in R that allow for user interaction as well, and one that
we will use for this section is called readline. readline is a function that will
print information to the R console and then accept input from the user. Enter this
expression into the console and hit return:
myyear <- readline("What year was Moby Dick published? \n")

When prompted, enter a number (e.g., 1851) and hit return. If you now type
myyear at the R prompt and hit return, youll find that R has stored the value, that
you entered in the myyear variable. Here is how it should look:
> myyear <- readline("What year was Moby Dick published? \n")
What year was Moby Dick published? 1851
> myyear
[1] "1851"

9.5 Building a Better KWIC Function

Using the readline function, you can write a KWIC list function that asks the
user (you) for a file to search, a keyword to find, and an amount of context to be
returned on either side of the keyword. When you call this function, you will send
it your existing my.corpus.l list object as the sole argument. Name this new
function doitKwic: once you write the function, you will be able to call it like
this:

84

9 Do It KWIC (Better)

doitKwic(my.corpus.l)

The only argument that you will need to send is the list object you created and
stored in my.corpus.l. This object already contains the name of every file in the
corpus directory as well as all the word vectors for each file. Open your corpusFunctions.R file and begin writing this new function like this:
doitKwic <- function(named.text.word.vector.l){
# instructions here will ask user for a file to search
# a keyword to find and a "context" number for context
# on either side of the of the keyword
}

Keep in mind that the argument name used inside the parentheses of the function does not have to be the same as the name used outside of the function.
So here I am defining a function that takes an argument called named.text.
word.vector.l. This is the same type of object as the my.corpus.l object
you have already instantiated, but when inside the workings of the function, it can
be referred to by another name.
You do not have to write your code this way (i.e., using different names when
inside or outside of the function), but I find it useful to name my function arguments
in a way that is descriptive of their content and a bit more abstract than the names I
give my objects within the main script. I may decide to use this function on another
project, and several months from now I may have forgotten what my.corpus.l
means. Using named.text.word.vector.l is verbose, but at least it gives
me some clues about what kind of object I am dealing with.
As the commented sections of the code suggest, you want the new function
to ask the user for input. First it needs to ask which file in the corpus to search
in, then what keyword to search for, and finally how much context to display.
For the first item, the function should display a list of the files that are inside
the named.text.word.vector.l object and ask the user to choose one.
Luckily you already have a function called show.files that does exactly that.
You can call this function from inside the new function! Remember too that the
show.files function is expecting to get a vector of file names as its argument.
You can get a vector of file names from the my.corpus.l object by calling the
names function. You can try this by entering
show.files(names(my.corpus.l))
## 1 austen.txt
## 2 melville.txt

Inside the doItKwic function, you will be able to achieve the same result with:
show.files(names(named.text.word.vector.l))

Having called show.files, the readline function can then be invoked

to ask the user for a number corresponding to a file (i.e., 1 for Austen and 2
for Melville). This user-entered information can be stored in a new object called
file.id.

9.6 Fixing a Problem

85

file.id <- as.numeric(readline(

"Which file would you like to examine? Enter a file number: \n"))

Notice how I have also wrapped the result inside a call to as.numeric. Even
though the user enters a number, input received from the user with the readline
function is treated as character data and needs to be converted to numeric for use in
this function. Something very similar is done for the context number that is needed:
context <- as.numeric(readline(
"How much context do you want to see? Enter a number: \n"))

Finally, you will want R to ask the user for a keyword, and youll make sure it is
changed to lowercase1 :
keyword <- tolower(readline("Enter a keyword: \n"))

With these three bits of information, you are now ready to run a KWIC search.
You will use which to find and return the position indexes of the hits for the users
keyword. The search results can be stored in a new object called hits.v.
hits.v <- which(named.text.word.vector.l[[file.id]] == keyword)

Notice that to create and fill the new hits.v object, I had to use the value held
in both the file.id and keyword variables that I just got by asking the user to
provide them! The rest of the function is simple; it will iterate over the positions in
the hits.v object and return context words to the left and right of the position.
Here is the most obvious and simple solution. . .
for(h in 1:length(hits.v)){
start <- hits.v[h]-context
end <- hits.v[h]+context
cat(named.text.word.vector.l[[file.id]][start:end])
}

9.6 Fixing a Problem

Unfortunately, this simple solution cannot handle all of the possible scenarios that
might occur. What if the very first word in the file you are searching is a hit? In
this case the first position in the hits.v vector would be 1 and that would cause
start to be set to 1 - (minus) context: that is one minus what ever number the user entered for context. The result of that subtraction would be a negative
number and R would choke trying to access a value held at a negative vector index!
You cant have that, so you need to add some code to deal with this possibility. Here
is one way to deal with the problem using an if conditional:
1

Obviously you could create a more complex function that would give the user flexibility in terms
of capitalization. For simplicity, we are working with a lower-cased text with all of the punctuation
stripped out.

86

9 Do It KWIC (Better)

for(h in 1:length(hits.v)){
start <- hits.v[h]-context
if(start < 1){
start <- 1
}
end <- hits.v[h]+context
cat(named.text.word.vector.l[[file.id]][start:end])
}

The if conditional tests to see if the value of start is set to less than 1. If it is,
then start gets reset to 1. A similar problem occurs if you happen to search for a
keyword that does not exist. If the word doed not exist, the hits.v variable will
be empty and you will get an error. You can deal with this situation in a similar
fashion, using another if conditional to test to see it hits.v has a length greater
than 0. Assuming you implement this solution, the whole function should now look
like this:
# Function take a list containing word vectors
# from text files and then allows for
# interactive user input to produce KWIC lists
doitKwic <- function(named.text.word.vector.l){
show.files(names(named.text.word.vector.l))
# ask the user for three bits of information
file.id <- as.numeric(readline(
"Which file would you like to examine? Enter a file number: \n"))
context <- as.numeric(readline(
"How much context do you want to see? Enter a number: \n"))
keyword <- tolower((readline("Enter a keyword: \n")))
hits.v <- which(named.text.word.vector.l[[file.id]] == keyword)
if(length(hits.v)>0){
for(h in 1:length(hits.v)){
start <- hits.v[h]-context
if(start < 1){
start <- 1
}
end <- hits.v[h]+context
cat(named.text.word.vector.l[[file.id]][start:end], "\n")
}
}
}

The result of calling this function and looking for the keyword dog in Sense and
Sensibility looks like this:
doitKwic(my.corpus.l)
1 austen.txt
2 melville.txt
Which file would you like to examine? Enter a file number:
1
How much context do you want to see? Enter a number:
5
Enter a keyword:
dog
a fellow such a deceitful dog it was only the last

9.6 Fixing a Problem

87

Practice
9.1. In prior exercises and lessons, you have learned how to instantiate an empty
list object outside of a for loop and then how to add new data to that result
object during the loop. You have learned how to use cbind to add columns of
data and rbind to add rows. You have also learned how to use paste with the
collapse argument to glue together pieces in a vector of values and how to use
cat to concatenate items in a vector. And you have used colnames to get and set
the names of columns in a data frame.
Using all of this knowledge, modify the function written in this chapter so that
the results of a KWIC search are put into a dataframe in which each row is a
single KWIC result. Your data frame should have four columns labeled as follows:
position, left, keyword, and right. The position column will contain the index value
showing where in the file the keyword was found. The left column will contain the
words in the file vector that were found to the left of the keyword. The keyword
column will contain the keyword, and the right column will contain the context that
was found to the right of the keyword. Here is an example of results generated using
the keyword dog with two words of context in the file melville.txt.
position left keyword right
[1,] "14555" "a newfoundland" "dog" "just from"
[2,] "16376" "in the" "dog" "days will"
[3,] "27192" "like a" "dog" "throwing his"
[4,] "51031" "last down" "dog" "and kennel"
[5,] "51107" "called a" "dog" "sir then"
[6,] "51565" "me a" "dog" "blazes he"
[7,] "73930" "sacred white" "dog" "was by"
[8,] "107614" "in a" "dog" "or a"
[9,] "107700" "of the" "dog" "the accursed"
[10,] "137047" "and ungrateful" "dog" "cried starbuck"
[11,] "137077" "way greyhounds" "dog" "to it"
[12,] "147004" "that a" "dog" "does to"
[13,] "167296" "ram lecherous" "dog" "he begets"
[14,] "170197" "bunger you" "dog" "laugh out"
[15,] "170577" "pickle you" "dog" "you should"
[16,] "171104" "like a" "dog" "strangely snuffing"
[17,] "202940" "high hang" "dog" "look and"
[18,] "206897" "ship s" "dog" "will in"
[19,] "206949" "then the" "dog" "vane and"

Make a copy of the doitKwic function from this chapter and rename it
doitKwicBetter. Now modify the function to produce a result like that seen
above.
9.2. Copy the function you created in Exercise 9.1 and modify it to include a feedback loop asking the user if the results should be saved as a .csv file. If the user
answers yes, generate a file name based on the existing user input (keyword,
file name, context) and write that file to the results directory using a call to the
write.csv function, as in this example:
write.csv(your.file, paste(output.dir, "your.file.csv", sep=""))

Save a copy of this new function in your corpusFunctions.R file as doitKwicBest.

Chapter 10

Text Quality, Text Variety, and Parsing XML

Abstract This chapter introduces readers to parsing XML in R with an emphasis on

TEI encoded XML.

10.1 Introduction
If you have ever downloaded a digital text from the Internet, you already know that
there is a great variety when it comes to quality. Some digital texts are available in
what is referred to as dirty OCR. This means that the texts have been scanned and
run through an optical character recognition (OCR) process but not subsequently
hand checked and corrected or cleaned up by a human editor (hence the term dirty).
On the other end of the spectrum, there are digital texts that have been carefully
created by double keying and human correction. Double keying involves the use
of two typists who each key the entire text into a computer. Once the two versions
are completed, they are compared to identify discrepancies. Double keying is not
perfect, but it is one of the more reliable methods for deriving a high quality digital
version of a text. Somewhere in between double keying and dirty OCR lies corrected
OCR. In this case an original document is scanned and then cleaned by a human
editor. While this method is still prone to errors, it is a considerable step beyond
dirty OCR and frequently good enough for processing and analysis tasks that involve
generating global statistics, which is to say a big picture perspective where a single
mis-keyed word will have little impact on the overall result.
Anyone working with digital text must at some point assess his or her corpus
and form an opinion about its quality and in what ways the quality of the material
will impact the analysis of the material. Promising research by Maciej Eder has examined the extent to which OCR errors impact stylometric analysis.1 This research
1 See Eder, Maciej. Mind your corpus: systematic errors in authorship attribution. in Conference
Abstracts of the 2012 Digital Humanities Conference, Hamburg, Germany. http://www.
dh2012.uni-hamburg.de/conference/programme/abstracts/mind-yourcorpus-systematic-errors-in-authorship-attribution/.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_10,
Springer International Publishing Switzerland 2014

89

90

10 Text Quality, Text Variety, and Parsing XML

gives us hope of being able to quantify the margin of error caused by OCR problems.
And, make no mistake, this is a very big problem. As the scanning efforts of Google
continue and as projects such as the Internet Archive and HathiTrust continue to
make more and more dirty OCR text available online, an algorithmic method for
dealing with dirty OCR becomes more and more important. Some, myself included,
have argued that at the large scale these OCR issues become trivial. That is a hypothesis, however, and one born out of frustration with the reality of our digital corpora.
If we want to mine the digital library as it exists today, we need to have a fairly high
tolerance for error.
But alongside these large and messy archives there are a good number of digital
collections that have been carefully curated, and, in some cases, enriched with detailed metadata. Two very fine examples are the Chadwyck Healey and Alexander
Street Press collections. Both of these content providers offer carefully corrected,
XML (or SGML) encoded digital texts. The high quality of these texts does, however,
come at a price: access to these corpora is available for a fee, and the fee is beyond
the budget of a typical scholar. If your institution does not subscribe to one of these
collections, you are more or less out of luck.
Somewhere in between the high quality products of vendors such as Chadwyck
Healey and Alexander Street Press and the dirty OCR of free resources such as
Google Books and the Internet Archive is Project Gutenberg. The texts in Project
Gutenberg tend to be of fairly high and fairly consistent quality. Having said that,
they lack detailed metadata and text provenance is often unclear. If your research
does not demand the use of a particular edition, and if you can tolerate some degree
of textual error, then Project Gutenberg may be a suitable source for digital texts.
Project Gutenberg texts are frequently available in multiple formats: plain text, html,
epub, etc. In many cases, it is possible to convert files in one format into another,
and in my own work I have developed scripts for converting Gutenbergs plain text
into TEI XML.

10.2 The Text Encoding Initiative (TEI)

The Text Encoding Initiative (TEI) offers a document-encoding standard that is
commonly used by humanities scholars. The TEI markup scheme provides a way
of storing an original text file alongside an almost infinite amount of metadata. Since
the files are extensible and editable, the amount of metadata available is only limited
by the encoders willingness to modify the documents. Say for example, you are collecting novels written by Irish and GermanAmerican authors. For this project you
might have a metadata field in your document where you can indicate the authors
national origins. You may have another field where you indicate the authors gender,
or birth date, or race, or sexual orientation. Once metadata of this sort is added to
the XML files, it can be easily accessed by computer scripts and used, for example,
as a sorting facet for a particular type of analysis.

10.4 Installing R Packages

91

In the rest of this book, you will be working with a corpus of texts that are encoded in TEI compliant XML. Unlike the plain text files (Moby Dick and Sense and
Sensibility) that you have processed thus far, these TEI-XML files contain extratextual information in the metadata of the <teiHeader> element. To proceed,
you must be able to parse the XML and extract the metadata while also separating
out the actual text of the book from the marked up apparatus around the book. You
need to know how to parse XML in R.

10.3 Parsing XML with R

An in-depth discussion of XML and of the TEI standard is beyond the scope of this
book. To understand the way that R parses XML, readers should be familiar with the
basic construction of an XML document as an ordered hierarchy of content objects
(OHCO) and should have some general familiarity with the structure of a TEI document: its primary divisions into <teiHeader>, <text>, <front>, <body>,
and <back>. The R package you will be using in this section (named simply XML)
can be used to load an XML file as DOM-tree object based on the standard Document
Object Model (DOM).
The DOM is a language-neutral set of objects for representing the nodes (or elements) and structure of an XML document. The topmost element in the DOM tree
is called the document object (or sometimes the root). The beauty of XML and
TEI-XML in particular is that it provides us with a structured environment for
storing information. In some sense, an XML document can be viewed as a kind of
database in which different types of information is stored in different fields. In a
database, for example, you might have an author table containing fields for last
name, first name, birthplace, gender, etc. You might then have a paragraph table
that stores each paragraph of a document associated or linked with a specific author.
This same kind of information can be easily stored in XML, and the TEI standard
gives us some specific rules, or standards, for how to organize such documents. Because TEI is an agreed upon set of standards for the encoding of literary texts, you
can develop programs in R (and many other languages) for processing these TEI
based documents.
To get your feet wet parsing XML in R, you will begin by loading the XML library.
In order to do that, however, youll first need to install the XML package.

10.4 Installing R Packages

R packages are collections of functions developed by programmers in the open
source R community. These packages add functionality to R that is not part of the
base installation.

92

10 Text Quality, Text Variety, and Parsing XML

There are several ways to install packages in R. The simplest and most direct way
is to type the command install.packages(XML) at the R prompt. The first
time you install a package, you will be asked to select a mirror, a server location
from which to download the software. When asked to select a mirror, choose one
that is geographically close to your location. You will then be prompted with a list of
all the available packages. If you are using the R GUI, you will find a menu option
for Packages & Data. Under this menu option, you can select Package Installer.
In RStudio you will find an Install Packages option under the main Tools menu.
Once the installer window pops up, type XML (caps sensitive) in the Packages field
and check the box to install dependencies. Checking this box ensures that you get
any other software that the XML package requires.2

10.5 Loading and Using the XML Package

Once the XML package (or any package for that matter) is installed, you must call it
into the active R session. For this you use the expression:
library(XML) # note that "XML" here is caps sensitive.

Unlike the simple scan function that you used to read text files of Moby Dick
and Sense and Sensibility, with XML files youll need some more sophisticated function that can understand the structure of XML. If you are just opening R be sure to
clear your workspace and reset your working directory. Begin by loading the XML
version of Moby Dick3 :
doc <- xmlTreeParse("data/XML1/melville1.xml", useInternalNodes=TRUE)

When working with the plain text version of Moby Dick, you found the chapter
breaks using grep; finding the chapter breaks in the XML file is a lot easier because the chapters are all marked up in the XML using a <div1> element and a
chapter attribute. You can gather the chapters easily using an XPath expression
2 While it is possible to download all of the available packages, doing so would certainly take a long
time and would clog up your installation with way too many irrelevant features. R is a multipurpose
platform used in a huge range of disciplines including: bio-statistics, network analysis, economics,
data-mining, geography, and hundreds of other disciplines and sub-disciplines. This diversity in the
user community is one of the great advantages of R and of open-source software more generally.
The diversity of options, however, can be daunting to the novice user, and, to make matters even
more unnerving, the online R user community is notoriously specialized and siloed and can appear
to be rather impatient when it comes to newbies asking simple questions. Having said that, the
online community is also an incredible resource that you must not ignore. Because the packages
developed for R are developed by programmers with at least some amount of ad hoc motivation
behind their coding, the packages are frequently weak on documentation and generally assume
some, if not extensive, familiarity with the academic discipline of the programmer (even if the
package is one with applications that cross disciplinary boundaries).
3 Notice the different path here. The XML version of Moby Dick is located in a different subdirectory
of the main TextAnalysisWithR.

10.5 Loading and Using the XML Package

93

such as this: "/tei:TEI//tei:div1[@type=chapter]". XPATH is a

language for representing and selecting XML nodes, or elements in an XML document. XPATH uses forward slashes to represent the ordered hierarchy of nodes in the
document much in the same way that R and UNIX and other systems and languages
use forward slashes to represent the structure of the directories (or folders) in your
computer.
When you send the XML document object (now in a variable named doc) along
with this XPath expression to the getNodeSet function, it will return a new
Node Set object (see below). The third argument, or parameter, required by the
getNodeSet function is a bit tricky to understand because it has to do with XML
namespaces, which is not so much about R as it is about XML. The getNodeSet
function expects us to identify an XML namespace as an item in a vector, so in
what follows I have called that vector tei. The tei prefix is used as a prefix in each
part of the XPath expression.
chapters.ns.l <- getNodeSet(doc,
"/tei:TEI//tei:div1[@type='chapter']",
namespaces = c(tei = "http://www.tei-c.org/ns/1.0"))

If, after executing the command, you enter class(chapters.ns.l), youll

see that chapters.ns.l is an XMLNodeSet object which is a subclass of an R
list object (which is why I have used the .ns.l extension). chapters.ns.l is
a special kind of list in that each item in the list is an XML node. What this means
is that as you iterate over the list, you must employ XML-based functions to further
refine the operations. For example, each XML chapter node encloses a <head> node
as a child.4 This <head> node is where the title of the chapter is stored. Enter the
following expression to examine the contents of the first list item.
chapters.ns.l[[1]]

If you scroll up in the R console, youll see the beginning of the chapter:
chapters.ns.l[[1]]
<div1 type="chapter" n="1" id="\_75784">
<head>Loomings</head>
<p rend="fiction">Call me Ishmael. Some years agonever mind how long precisely- having little. . .

Notice that the chapter title, Loomings, is inside the <head> element. If you
enter class(chapters.ns.l[[1]]), you will see that the first item of this R
list is a XML node:
class(chapters.ns.l[[1]])
## [1] "XMLInternalElementNode"
## [2] "XMLInternalNode"
## [3] "XMLAbstractNode"

One way to extract the <head> node of the chapter is to use the
xmlElementsByTagName function. To get the <head> node from the first
chapter, you might write:
4

A node inside of another node is often referred to as a child node.

94

10 Text Quality, Text Variety, and Parsing XML

chap.title <- xmlElementsByTagName(chapters.ns.l[[1]], "head")

chap.title
## $head
## <head>Loomings</head>

But what you really want is not the entire node as an XML object, but rather the
content of the node. To get the content, that is, the chapter title Loomings, you also
need the xmlValue function.5
xmlValue(chap.title[[1]])
## [1] "Loomings"

With a little understanding of R lists from the first part of this book and with
some sense of how TEI XML files are structured, you can put all of this together and
generate a chapter-by-chapter analysis of Moby Dick exactly as you did previously
using the grep function.
In TEI the textual data of each chapter is stored inside of <p> elements. What
this means is that for each chapter (<div1>) you want to extract both the title of
the chapter (found inside of <head>) and the paragraphs (found inside <p>) as
two separate items. Ultimately you want to produce a new list in which each item in
the list is named with the chapter title and the value of the named list item is a table
of words. This is exactly what you did with the plain text files earlier in this book;
from the XML file, you create a list object identical to the one created from the plain
text version of Moby Dick. And, just as you did earlier, here again you will use a
for loop to iterate over the items in the chapters.ns.l object.
First create two list objects in which to store the results:
chapter.freqs.l <- list()
chapter.raws.l <- list()

And now the for loop where all the work gets done. Comments have been added
to explain the workings of the loop.
for(i in 1:length(chapters.ns.l)){
# first get the chapter title from the head element
chap.title <- xmlValue(xmlElementsByTagName(chapters.ns.l[[i]],
"head")[[1]])
# get only the contents of the paragraph tags
paras.ns <- xmlElementsByTagName(chapters.ns.l[[i]], "p")
#combine all the words from every paragraph
chap.words.v <- paste(sapply(paras.ns, xmlValue), collapse=" ")
# convert to lowercase
words.lower.v <- tolower(chap.words.v)
# tokenize
words.l <- strsplit(words.lower.v, "\\W")
word.v <- unlist(words.l)
word.v <- word.v[which(word.v!="")]
# calculate the frequencies
chapter.freqs.t <- table(word.v)
chapter.raws.l[[chap.title]] <- chapter.freqs.t
chapter.freqs.l[[chap.title]] <- 100*(chapter.freqs.t/sum(chapter.freqs.t))
}

Notice that the chap.title object is another type of list, which is why the further bracketed
sub-setting is required in order to get at the text contents.

10.5 Loading and Using the XML Package

95

After execution, you will have two list objects derived from a TEI-XML version
of Moby Dick. These new lists will be almost identical to the lists created from the
plain text version of the novel.6 You can now run the exact same analysis of whale
and ahab that you ran when working with the plain text:
whales <- do.call(rbind, lapply(chapter.freqs.l, '[', 'whale'))
ahabs <- do.call(rbind, lapply(chapter.freqs.l, '[', 'ahab'))
whales.ahabs <- cbind(whales, ahabs)
whales.ahabs[which(is.na(whales.ahabs))] <- 0
colnames(whales.ahabs) <- c("whale", "ahab")

And naturally, just as you did with the plain text versions, you can run a correlation
test to see if whales and ahabs behave in a correlated manner (Fig. 10.1):

0.0

0.5

1.0

1.5

2.0

barplot(whales.ahabs, beside=T, col="grey")

whale

ahab

Fig. 10.1 Barplot showing occurrences of whale and ahab

They wont be exactly the same because they come from slightly different sources.

96

10 Text Quality, Text Variety, and Parsing XML

whales.ahabs.df <- as.data.frame(whales.ahabs)

cor.test(whales.ahabs.df$whale, whales.ahabs.df$ahab)
##
## Pearson's product-moment correlation
##
## data: whales.ahabs.df$whale and whales.ahabs.df$ahab
## t = -2.7982, df = 131, p-value = 0.005915
## alternative hypothesis: true correlation is not equal to 0
## 95 percent confidence interval:
## -0.39186935 -0.07009107
## sample estimates:
##
cor
## -0.2374843

10.6 Metadata
The real power of processing XML files will become apparent when you are dealing
with multiple files. For now, however, it is worth pointing out that with TEI-XML,
you have access to a great deal of human-added metadata that is not part of the raw
text of the novel.
Here are five examples that use the xpathApply function to find information
contained in specific places inside the TEI-XML hierarchy. XPath allows you to
use doubles slashes // to indicate the stepping down from one level to several
levels deeper. Enter the following expression, for example, to find some additional
details about the title of Moby Dick:
xpathApply(doc,
"/tei:TEI//tei:fileDesc//tei:titleStmt//tei:title",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"))

This expression asks R and the XML parser to begin at the root </TEI> and
then descend into the document until a <fileDesc> element is found. From
there the parser should continue to descend the tree looking for a path that contains a <titleStmt> element. The parser should continue looking deeper for a
<title> element that is a descendant of <titleStmt>. After you run this code,
the result, an xmlNodeSet, will look like this:
##
##
##
##
##

[[1]]
<title type="main">Moby Dick; Or The Whale</title>
attr(,"class")
[1] "XMLNodeSet"

To refine the result further and get that title out of the node set, modify the code
to reference the item held inside the first ([[1]]) position:
xpathApply(doc,
"/tei:TEI//tei:fileDesc//tei:titleStmt//tei:title",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"))[[1]]
## <title type="main">Moby Dick; Or The Whale</title>

10.6 Metadata

97

And you can even assign the result to a new variable called title:
title <- xpathApply(doc,
"/tei:TEI//tei:fileDesc//tei:titleStmt//tei:title",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"))[[1]]
title
## <title type="main">Moby Dick; Or The Whale</title>

Notice how I first put a result into a variable called title and then issued a
second command to print the contents of title. You can save yourself some time
by taking advantage of a handy R shortcut in which you wrap any R expression inside parentheses. Doing so instructs R not only to perform the instructions inside the
parentheses but also to then print the result to the console. Here are a few examples
using parentheses to both perform and print a given expression:
(title <- xpathApply(doc,
"/tei:TEI//tei:fileDesc//tei:titleStmt//tei:title",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"))[[1]])
## <title type="main">Moby Dick; Or The Whale</title>
(author <- unlist(xpathApply(doc,
"/tei:TEI//tei:author//tei:name",
namespaces=c(
tei = "http://www.tei-c.org/ns/1.0")))[[1]])
## <name type="main">Herman Melville</name>

In addition to providing access to the contents of the elements themselves, the

XML parser gives access to the attributes of the elements. Notice, for example, that
the <note> element has two attributes, one called nation with a value American
and another called gender with a value of M.
xpathApply(doc,
"/tei:TEI//tei:teiHeader//tei:note",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"))
## [[1]]
## <note nation="American" gender="M"/>
##
## attr(,"class")
## [1] "XMLNodeSet"

To get these attribute values, you can add an xmlGetAttr function to the expression. Below are three examples for you to try.
(nation <- unlist(xpathApply(doc,
"/tei:TEI//tei:teiHeader//tei:note",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"),
xmlGetAttr, "nation")))
## [1] "American"
(gender <- unlist(xpathApply(doc,
"/tei:TEI//tei:teiHeader//tei:note",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"),
xmlGetAttr, "gender")))
## [1] "M"
(pubdate <- unlist(xpathApply(doc,
"/tei:TEI//tei:teiHeader//tei:creation/tei:date",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"),
xmlGetAttr, "value")))
## [1] "1851"

98

10 Text Quality, Text Variety, and Parsing XML

Practice
10.1. Using what you have learned about accessing metadata contained in XML
files, write a single line R expression to extract the information contained in the
<respStmt> element of the TEI header (this is a very simple exercise).

Part III

Macroanalysis

Chapter 11

Clustering

Abstract This chapter moves readers from the analysis of one or two texts to a larger
corpus. Machine clustering is introduced in the context of an authorship attribution
problem.

11.1 Introduction
This chapter introduces you to document clustering using a rather small corpus. It
might be good to think about this experiment as a prototype, or model, for a much
larger experiment. Many of the basic tasks will be the same, but instead of working
with 4,300 books youll hone your skills using just 43. Much of the processing done
in this section will be familiar to you from the first half of the book where you
developed code to compare the vocabulary richness of Moby Dick on a chapter-bychapter basis. Here, instead of chapters, youll have entire books to work with. The
raw materials and the basic R objects will be the same.

11.2 Review
At the R prompt, use setwd to change the directory path to match the location of
the TextAnalysisWithR directory associated with this book. If you prefer to
leverage the RStudios menu system, just select Set Working Directory from the
Session menu. You can then use your computers windowing system to navigate to
the location of the TextAnalysisWithR directory.1
1 If you have been working on other sections of this book or on R projects of your own, it might
be a good idea to either restart R or to clear the R workspace. To do the latter, just click on the
Session menu of the RStudio GUI and select Clear Workspace. This will remove all R objects
and functions that you may have been using, wiping the R slate clean, as it were.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_11,
Springer International Publishing Switzerland 2014

101

102

11 Clustering

setwd("~/Documents/TextAnalysisWithR")

Once this command is executed, the R prompt > will return and you will be able
to enter a new command. Check to see that you set the working directory correctly
using:
getwd()
## [1] "/Users/mjockers/Documents/TextAnalysisWithR"

If all is well, you will get a response that looks something like the above.

11.3 Some Oddities in R

You may have noticed here, or elsewhere in this book, that R has prefixed its returned
result with a bracketed 1. The fact is that everything in R is vectorized. Even getwd
returns its result as a vector. What you are seeing here with the [1] is that the result
of calling getwd is a vector with one item. This behavior can feel especially strange
to those who are coming to R from another language. The truth is that R has a number
of funky behaviors. I have already noted that R indexes vectors beginning with 1,
whereas many other languages begin with 0. One of the strangest R features that I
am aware of involves the non-value NA. NA is used in R to represent missing values
(not available). So suppose you had some data about the nationality of different
authors:
nations <- c(Joyce="Irish", Twain="American", Dickens="English")

Now suppose that you found a new author and were not sure of the authors
nationality. You could enter that authors nationality as NA:
nations <- c(nations, Smith=NA)

Notice here that NA is entered without quote marks around it. This is to indicate
to R that it is not being entered as a character string but a special set of reserved
characters indicating no value or missing value. You can check on this using the
is.na function that tests to see if a value is NA or not:
is.na(nations)
##
Joyce
Twain Dickens
##
FALSE
FALSE
FALSE

Smith
TRUE

As you can see, the value for Smith is indeed NA and not a character string. The
surprising thing about this is that when you then ask R to compute a count of the
number of characters in each of the strings in the nations vector, R returns a count
of 2 for the NA value for Smith!

11.4 Corpus Ingestion

nchar(nations)
##
Joyce
Twain Dickens
##
5
8
7

103

Smith
2

Irish = 5, yes; American = 8, yes; English = 7, yes; but NA = 2? How could this
be since NA is not a string? It is definitely not because there are two characters in
the string NA because NA is specifically not a string in this example! The answer
turns out to be an oddity in the way R was designed, and it is only by consulting
the help documentation for nchar (e.g., ?nchar) that you learn that currently
nchar always returns 2 for missing values! Why R was designed in this way is not
made clear, but the use of currently suggests that at least somewhere along the
line someone figured this behavior might be something worth changing.2

11.4 Corpus Ingestion

The files you will use in this clustering experiment are all stored in the directory
located at data/XMLAuthorCorpus. The first thing you will require is a bit of
R that will go to this directory and survey its contents. To keep things neat, put the
path to this directory into an R object and call it input.dir.
input.dir <- "data/XMLAuthorCorpus"

You can now use the dir function to generate a vector containing the names of
all the files contained inside input.dir.
files.v <- dir(path=input.dir, pattern=".*xml")

Notice how in addition to the path argument, I have added a pattern argument.
This pattern, a regular expression, tells dir to return only those files with names
matching the regular expression.3 The files.v variable now contains a vector of character strings from the file names of the 43 XML files found inside the
XMLAuthors folder. Here they are:

Patrick Burns has written a 125 page book documenting many of Rs unusual behavior. The book
is informative and entertaining to read. You can find it online at http://www.burns-stat.
com.
3 Enter ?regex at the prompt to learn more about regex in R.

104
files.v
## [1]
## [3]
## [5]
## [7]
## [9]
## [11]
## [13]
## [15]
## [17]
## [19]
## [21]
## [23]
## [25]
## [27]
## [29]
## [31]
## [33]
## [35]
## [37]
## [39]
## [41]
## [43]

11 Clustering

"anonymous.xml"
"Carleton10.xml"
"Carleton12.xml"
"Carleton14.xml"
"Carleton3.xml"
"Carleton5.xml"
"Carleton7.xml"
"Carleton9.xml"
"Donovan2.xml"
"Driscoll2.xml"
"Edgeworth1.xml"
"Jessop2.xml"
"Kyne1.xml"
"LeFanu1.xml"
"LeFanu3.xml"
"LeFanu5.xml"
"LeFanu7.xml"
"McHenry1.xml"
"Norris1.xml"
"Norris3.xml"
"Polidori1.xml"
"Quigley2.xml"

"Carleton1.xml"
"Carleton11.xml"
"Carleton13.xml"
"Carleton2.xml"
"Carleton4.xml"
"Carleton6.xml"
"Carleton8.xml"
"Donovan1.xml"
"Driscoll1.xml"
"Driscoll3.xml"
"Jessop1.xml"
"Jessop3.xml"
"Kyne2.xml"
"LeFanu2.xml"
"LeFanu4.xml"
"LeFanu6.xml"
"Lewis.xml"
"McHenry2.xml"
"Norris2.xml"
"Norris4.xml"
"Quigley1.xml"

Notice that the very first of these files is titled anonymous.xml. This is the file
whose authorship is uncertain. In this chapter, you will use text analysis and unsupervised clustering to compare the word frequency signal of the anonymous novel
to the signals of the others, and then, based on that comparison, you will take a guess
at which author in the corpus is the most likely author of this anonymous novel.
With the file names in the files.v variable, you must now write a bit of code
to iterate over all of these values and at each value, pause to load and process the
text corresponding to the value; for each file name in the files.v vector, you want
the script to perform some other bit of processing related to that file. This is very
similar to the code that you have developed in other chapters. You begin with a for
loop and a new variable called i (for integer):
for (i in 1:length(files.v)){
# Some code here
}

This expression tells R to begin by setting i equal to 1 and to iterate over all
of the elements of the vector files.v and to stop only when it has finished with
the 43rd item, which, in this case, is what is returned by the call to the length
function.
Remember from Chap. 10 that to parse an XML file, you used the xmlTree
Parse function. This XML function takes several arguments, the first of which is
the path to the location of the file on your computer. The first question you must
address, therefore, is how to give the program the information it requires in order to
figure out the file path to each of the files in the corpus directory.
You already know how to access one of the items in the files.v object via
sub-setting; so now consider the following expression that uses the file.path
function to join together the two objects you have instantiated:
file.path(input.dir, files.v)

11.4 Corpus Ingestion

105

Used in this way, file.path returns a series of file path expressions that look
like this:
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##

[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]

"data/XMLAuthorCorpus/anonymous.xml"
"data/XMLAuthorCorpus/Carleton1.xml"
"data/XMLAuthorCorpus/Carleton10.xml"
"data/XMLAuthorCorpus/Carleton11.xml"
"data/XMLAuthorCorpus/Carleton12.xml"
"data/XMLAuthorCorpus/Carleton13.xml"
"data/XMLAuthorCorpus/Carleton14.xml"
"data/XMLAuthorCorpus/Carleton2.xml"
"data/XMLAuthorCorpus/Carleton3.xml"
"data/XMLAuthorCorpus/Carleton4.xml"
"data/XMLAuthorCorpus/Carleton5.xml"
"data/XMLAuthorCorpus/Carleton6.xml"
"data/XMLAuthorCorpus/Carleton7.xml"
"data/XMLAuthorCorpus/Carleton8.xml"
"data/XMLAuthorCorpus/Carleton9.xml"
"data/XMLAuthorCorpus/Donovan1.xml"
"data/XMLAuthorCorpus/Donovan2.xml"
"data/XMLAuthorCorpus/Driscoll1.xml"
"data/XMLAuthorCorpus/Driscoll2.xml"
"data/XMLAuthorCorpus/Driscoll3.xml"
"data/XMLAuthorCorpus/Edgeworth1.xml"
"data/XMLAuthorCorpus/Jessop1.xml"
"data/XMLAuthorCorpus/Jessop2.xml"
"data/XMLAuthorCorpus/Jessop3.xml"
"data/XMLAuthorCorpus/Kyne1.xml"
"data/XMLAuthorCorpus/Kyne2.xml"
"data/XMLAuthorCorpus/LeFanu1.xml"
"data/XMLAuthorCorpus/LeFanu2.xml"
"data/XMLAuthorCorpus/LeFanu3.xml"
"data/XMLAuthorCorpus/LeFanu4.xml"
"data/XMLAuthorCorpus/LeFanu5.xml"
"data/XMLAuthorCorpus/LeFanu6.xml"
"data/XMLAuthorCorpus/LeFanu7.xml"
"data/XMLAuthorCorpus/Lewis.xml"
"data/XMLAuthorCorpus/McHenry1.xml"
"data/XMLAuthorCorpus/McHenry2.xml"
"data/XMLAuthorCorpus/Norris1.xml"
"data/XMLAuthorCorpus/Norris2.xml"
"data/XMLAuthorCorpus/Norris3.xml"
"data/XMLAuthorCorpus/Norris4.xml"
"data/XMLAuthorCorpus/Polidori1.xml"
"data/XMLAuthorCorpus/Quigley1.xml"
"data/XMLAuthorCorpus/Quigley2.xml"

Once wrapped inside a loop, you will be able to easily iterate over these file paths,
loading each one in turn. Before you start looping, however, set the new variable i
equal to 1. Doing so will allow you to do some prototyping and testing of the code
without running the entire loop.
i <- 1

With i equal to 1, reenter the previous expression with the value i inside the
brackets of the files.v variable, like this:
file.path(input.dir, files.v[i])
## [1] "data/XMLAuthorCorpus/anonymous.xml"

Instead of returning the paths for all of the files, you get just the first file (alphabetically) in the files.v object. If you have not already done so, enter the
following expression to load the XML package:

106

11 Clustering

library(XML)

With the package loaded and this path name business sorted out, you can now
use the xmlTreeParse function to ingest, or load, the first XML file. Rather than
assigning the result of the file.path command to another variable, you can just
embed one function inside the other, like this4 :
doc.object <- xmlTreeParse(file.path(input.dir, files.v[i]),
useInternalNodes=TRUE)

If you enter this expression, and then type doc at the R prompt, youll see the entire contents of the file titled anonymous.xml. You must now embed this expression
within a for loop so that you can iterate over all of the XML files in the files directory. Do this as follows:
for(i in 1:length(files.v)){
doc.object <- xmlTreeParse(file.path(input.dir, files.v[i]),
useInternalNodes=TRUE)
# some more code goes here. . .
}

This code will consecutively load all of the XML files into an object called doc.
Each time the program starts a new loop, the previous contents of the doc object will be overwritten by the new XML file. So, before the next iteration begins,
you need to process the contents of the doc object and store the results in some
other variable that will persist beyond the loop. For this, create an empty list called
book.freqs.l before the for loop even begins. This list will serve as a container for the results that will be generated during the processing that takes place
inside the loop. Putting it all together, your code should now look like this:
book.freqs.l <- list()
for(i in 1:length(files.v)){
doc.object <- xmlTreeParse(file.path(input.dir, files.v[i]),
useInternalNodes=TRUE)
#some more code goes here. . .
}

11.5 Another Function

Now that you have code for handling the iteration and for loading the XML files,
you need to process the loaded files to extract the word frequencies. You have already seen how Rs built-in functions work, and you built some simple functions for
making KWIC lists in earlier chapters. Anytime you know you are going to repeat
a task multiple times, it is handy to have a function. For each of the XML files in
the corpus, you need to extract, count, and calculate the relative frequency of every
4

You can learn more about the useInternalNodes argument in the documentation for the
xmlTreeParse function. Basically, setting it to TRUE avoids converting the contents into R
objects, which saves a bit of processing time.

11.5 Another Function

107

word type. This is exactly the process used in the first part of this book, but here you
are dealing with XML, so things are a bit different. Since you want to be able to do
this task over and over again with any XML file, you can wrap all of this processing
in a reusable function called getTEIWordTableList. This function will take
just one argument: an XML document object that has been derived by calling the
xmlTreeParse function (as we just did above). The function definition begins
like this:
getTEIWordTableList <- function(doc.object){
# some code here
}

Within the function you will need to add code to extract and process all of the
words in the XML files. First grab all of the paragraph nodes using the getNodeSet
function from the XML package.
paras <- getNodeSet(doc.object, "/tei:TEI/tei:text/tei:body//tei:p",
c(tei = "http://www.tei-c.org/ns/1.0"))

Notice that the second argument of the function is an XPath expression telling
the function to only get paragraphs that occur as children of the <body> portion
of the XML file. An XPath expression, as you will recall from Chap. 10, provides a
way of digging down into the structure of an XML file. This expression is used because there might be <p> tagged content that is wrapped in <front> and <back>
tags that is not part of the core text of the novel.5
Next you need to join all of the paragraph nodes found by this XPath expression together into a single character string. The getNodeSet function returns an
XMLNodeSet object. This object is similar to a list object. The next expression,
therefore, combines the sapply and paste functions in order to join all the paragraph node content (found with the xmlValue function) together using a simple
blank space as the joining glue. You have done this sort of thing before.
words <- paste(sapply(paras,xmlValue), collapse=" ")

Next you will need to convert all the words to lowercase and then split the entire
string using the \\W regular expression. As you saw earlier, this regex breaks the
text on any non-word characters and chunks, or breaks, the string by word boundaries, splitting the text into word tokens based on any white space or punctuation
characters.
words.lower <- tolower(words)
words.l <- strsplit(words.lower, "\\W")

The result of strsplit is a list, but you only need a simple vector of words,
so use unlist to scrap the list elements.
word.v <- unlist(words.l)

See Chapter 10, section 10.5 for an explanation of the namespace argument.

108

11 Clustering

Now you can do a little spring-cleaning. Inevitably there are blank items that get
stuck in the vector where punctuation marks would have appeared. You can identify
these using a which function to locate items with no character content.
word.v[which(word.v!="")]

This little slight-of-hand expression can be embedded inside another function so that
you do not have to instantiate another variable as a temporary storage place for its
result. In other words, embed all of this inside a table function which will return
the final word frequency data you require:
book.freqs.t <- table(word.v[which(word.v!="")])

With this done, it is simple to calculate the relative word frequencies by dividing
the raw counts by the sum of words in the entire book. Optionally, you can multiply
the result by 100 to make the percentages easier to read:
book.freqs.rel.t <- 100*(book.freqs.t/sum(book.freqs.t))

The last step of this function is to return this final data back to the main script.
For that employ the return function:
return(book.freqs.rel.t)

The whole function now looks like this:

getTEIWordTableList <- function(doc.object){
paras <- getNodeSet(doc.object,
"/tei:TEI/tei:text/tei:body//tei:p",
c(tei = "http://www.tei-c.org/ns/1.0"))
words <- paste(sapply(paras,xmlValue), collapse=" ")
words.lower <- tolower(words)
words.l <- strsplit(words.lower, "\\W")
word.v <- unlist(words.l)
book.freqs.t <- table(word.v[which(word.v!="")])
book.freqs.rel.t <- 100*(book.freqs.t/sum(book.freqs.t))
return(book.freqs.rel.t)
}

Save this function to your corpusFunctions.r file in your code directory and add
a line to the top of your main working script to call this supporting file:
source("code/corpusFunctions.r")

You can now embed a call to this function into the for loop that you began
building up previously. Your main script should now look like this:
source("code/corpusFunctions.r")
book.freqs.l <- list() # a list object to hold the results
for(i in 1:length(files.v)){
doc.object <- xmlTreeParse(file.path(input.dir, files.v[i]),
useInternalNodes=TRUE)
worddata <- getTEIWordTableList(doc.object)
book.freqs.l[[files.v[i]]] <- worddata
}

11.6 Unsupervised Clustering and the Euclidean Metric

109

When run, the values returned from processing each text with the function are
added to the book.freqs.l list that was created before entering the loop. After
processing, all of the data necessary for continuing the clustering experiment will be
contained in the single list. Before you go on, however, you might want to inspect
this new object. Try a few of these commands, and be sure that the results all make
sense to you:
class(book.freqs.l)
names(book.freqs.l)
str(book.freqs.l)

11.6 Unsupervised Clustering and the Euclidean Metric

Many years of authorship attribution research have taught us that the most effective
way to distinguish between the text of one author and another is by comparing the
different usages of high frequency features in their writing. High frequency features
include words such as the, of, and, and to, as well as, in some studies, marks of
punctuation and even common bigrams, such as of the. Here I will assume some
familiarity with the concept of distinct stylistic signals and jump right into describing a process for comparing the word usage patterns of the writers in the sample
corpus.6
The technique that I describe here involves a measurement known as Euclidean
distance. Using the Euclidean metric, or what is sometimes called the Pythagorean
metric, you can calculate each single books distance from every other book in a
corpus. Books with a closer distance will have more in common in terms of their
feature usage habits, and books with a greater relative distance will be dissimilar.
For the sake of illustration, assume that you have just three books and only two
features for each book. Call the three books a, b, and c, and the two features f1 and
f2. Assume further that the measurements of the two features in each of the book
are frequencies per 100 words, as follows:
my.m
##
f1 f2
## a 10 5
## b 11 6
## c 4 13

That is, in book a, feature f1 occurs 10 times per 100 words and feature f2 occurs
5 times per 100 words. In book b feature f1 is found 11 times for every 100 words
and so on. You can represent this information in an R matrix using this code:
a <- c(10, 5)
b <- c(11,6)
c <- c(4,13)
my.m <- rbind(a,b,c)
colnames(my.m) <- c("f1", "f2")

For a brief overview of how this work is conducted, See Jockers, Matthew L. Macroanalysis:
Digital Methods and Literary History. University of Illinois Press, 2013. Pages 6367.

110

11 Clustering

These feature measurements can in turn be represented as x and y coordinate

values and plotted in a two-dimensional space, as in Fig. 11.1. Once plotted, you

f2

10

12

b
a
4

10

11

f1

Fig. 11.1 Two-dimensional plotting

can measure (as with a ruler) the distance on the grid between the points. In this
case, you would find that books a and b are closest (least distant) to each other.
Naturally, you dont want to actually plot the points and then use a ruler, so instead
you can employ the dist function in R.
dist(my.m)
##
a
## b 1.414214
## c 10.000000

b
9.899495

The result reveals that the standard or ordinary distance between points a and b
is 1.4, the distance between a and c is 10, and the distance between b and c is 9.9.
These distances provide a way of describing the relative nearness of the points, and,
therefore, the similarity of the documents from which these values were extracted.
For convenience, you can think of these distances as meters, feet, or miles; it does
not ultimately matter since you are only concerned with the relative closeness of the
points. In this example, using only two features (f1 and f2) you would conclude that
book a and book b are the most similar.
When there are only two dimensions (or features) as in this example, the plotting
and measuring is fairly simple and straightforward. It becomes more complex when
thought of in terms of fifty or five hundred features and twenty or forty books. Nevertheless, the closeness of items in this high dimensional space can still be calculated

11.6 Unsupervised Clustering and the Euclidean Metric

111

using the Euclidean metric (which is the default method employed by Rs dist
function). The metric is expressed like this:

d(p, q) = (p1 q1)2 + (p2 q2)2 + ... + (pi qi )2 + (pn qn )2 .
where d is the distance and p and q are two books.
p1
is the measure of feature one in book p and
q1
is the measure of feature one in book q, and so on through all of the features.
Assume you have a new data set in which there are four features:
a <- c(10, 5, 3, 5)
b <- c(11,6, 5, 7)
c <- c(4,13, 2, 6)
my.m <- rbind(a,b,c)
colnames(my.m) <- c("f1", "f2", "f3", "f4")

Using the Euclidean metric, the distances d between books (a, b, c) are calculated
as follows:

d(a, b) = (10 11)2 + (5 6)2 + (3 5)2 + (5 7)2 = 3.162278

(10 4)2 + (5 13)2 + (3 2)2 + (5 6)2 = 10.09950

d(b, c) = (11 4)2 + (6 13)2 + (5 2)2 + (7 6)2 = 10.39230

d(a, c) =

To get the same results in R, you simply enter:

dist(my.m)
##
a
b
## b 3.162278
## c 10.099505 10.392305

You see that the distance between a and b (3.162278) is much smaller than the
distance between a and c. This indicates that a and b are more similar to each other
in terms of these four features. Using R it is a trivial matter to calculate the distances
between every book and every other book in the example corpus. Everything you
need for doing this calculation is already stored inside the book.freqs.l list
object. You simply need to get that data out of the list and into a data matrix in
which each row is a book and each column is one of the word features.

112

11 Clustering

11.7 Converting an R List into a Data Matrix

Before you can apply the Euclidean metric to the authorship data, you need to
get the word frequency information out of the book.freqs.l list and into a
data matrix in which each row is a book and each column is a word feature. The
cells in this matrix will contain the relative frequency values that were calculated
using getTEIWordTableList. The first step in this process involves converting the book.freqs.l into an R data.frame. For this, the mapply function is
handy. Here you will create a new list object called freqs.l that is similar to
book.freqs.l except that each list item is converted from a table object to a
data.frame object. Ill show the code first and then explain it.
freqs.l

<-

mapply(data.frame,
ID=seq_along(book.freqs.l),
book.freqs.l, SIMPLIFY=FALSE,
MoreArgs=list(stringsAsFactors=FALSE))

In this first step, the word frequency tables held in the list items of
book.freqs .l are converted into individual data.frame objects. To compare the difference, type the following code at the R prompt.
class(freqs.l[[1]])
## [1] "data.frame"
class(book.freqs.l [[1]])
## [1] "table"

The first item in freqs.l is a data.frame object, whereas the first item in
the original book.freqs.l is a table object. The mapply function used here
is in the same family of functions as lapply and sapply that you have been using elsewhere. mapply differs in a couple of ways. First, mapply takes a function
name (whereas sapply and lapply have the list object as the first argument).
In this case, the function is data.frame. Next is a sequence of ID numbers corresponding to the books in the corpus. For this, you use seq_along to construct
the id numbers (ID=seq_along(book.freqs.l)) from the information contained in the book.freqs.l object.7 In this case, the numbers are 1 through 43.
The next argument is the actual list object to which the data.frame function will
be applied. The SIMPLIFY argument is set to FALSE, so that mapply will not try
to convert the result into simpler data type (i.e., a vector or matrix). Finally, use the
MoreArgs argument to let mapply know that when building the data frames, you
do not want to convert any of the columns into factors.8
The result of running mapply is that each item in the resulting list (titled
freqs.l) is a data frame with some number of rows and three columns. The three

7 seq_along is a simple R function for generating a sequence of numbers. Check the

R-help documentation for details. In this example, I could have just as easily used 1:43 or
1:length(book.freqs.l).
8 Factors are explained in a later section.

11.7 Converting an R List into a Data Matrix

113

columns are automatically labeled as ID (which is taken from the ID element given
to mapply), Var1 (which is the column of unique word types), and Freq (which
is the frequency calculated for that word type in the given text). To see the first ten
rows of the data frame contained inside first list item (in this case the anonymous
text which has been assigned an ID of 1), use the following expression.
freqs.l[[1]][1:10,]
##
ID Var1
Freq
## 1
1
_i_
0.0009794223
## 2
1
02
0.0009794223
## 3
1
03
0.0009794223
## 4
1
05
0.0009794223
## 5
1
10
.0039176893
## 6
1
10
0.0019588447
## 7
1
11
0.0009794223
## 8
1
12
0.0029382670
## 9
1
15
0.0009794223
## 10 1
1850 0.0009794223

In this case the word types happen to all be digits because our regex expression
treats numbers as word characters. To see some more interesting information, try
this expression
freqs.l[[1]][100:110,]
##
ID
Var1
## 100 1
ace
## 101 1
aces
## 102 1
ache
## 103 1
achieve
## 104 1
acknowledge
## 105 1
acknowledged
## 106 1
acknowledges
## 107 1
acknowledgment
## 108 1
acquaintance
## 109 1
acquainted
## 110 1
acquiesced

Freq
0.0039176893
0.0009794223
0.0019588447
0.0009794223
0.0039176893
0.0019588447
0.0009794223
0.0009794223
0.0078353787
0.0029382670
0.0009794223

Remember that in order to use the dist function, you need to get all of this
data into a matrix object. The next step, therefore, will be to use do.call with
an rbind argument to bind all the data in all the lists into a single three-column
data.frame.
freqs.df

<-

do.call(rbind,freqs.l)

If you use the dim function you can check the size of your data.frame.
dim(freqs.df)
## [1] 300447

You can also inspect the data contained in the data frame using sub-setting. Entering
the following expression, for example, will show the data in rows 100 through 110:

114

11 Clustering

freqs.df[100:110,]
##
ID
Var1
Freq
## 100 1
ace
0.0039176893
## 101 1
aces
0.0009794223
## 102 1
ache
0.0019588447
## 103 1
achieve
0.0009794223
## 104 1
acknowledge
0.0039176893
## 105 1
acknowledged
0.0019588447
## 106 1
acknowledges
0.0009794223
## 107 1
acknowledgment 0.0009794223
## 108 1
acquaintance
0.0078353787
## 109 1
acquainted
0.0029382670
## 110 1
acquiesced
0.0009794223

You will notice when you run this expression that R provides the column names
for this data frame: ID, Var1, and Freq. Under the column titled ID, you are
seeing the unique identifiers for the files in the original files.v vector object. In
other words, the rows that have a 1 under the ID column are rows derived from the
anonymous.xml file, which happens to be indexed at the first position in files.v.
The items in the Var1 column are the word types, and the Freq column contains
the relative frequencies.
What you now have in the freqs.df object is a long form table. The next step
is to reshape the data frame to a wide format so that there are only 43 rows (one
for each text) and 52,048 columns, one for each unique word type in the whole
corpus.
This reshaping can be achieved in a number of different ways. Ill use the xtabs
function since it is ideally suited to the conversion task and relatively intuitive.9
result

<-

xtabs(Freq ~ ID+Var1, data=freqs.df)

xtabs is a function specifically designed for creating contingency tables, or

cross tabulations. You might be more familiar with this concept as a pivot table in
Excel. The function takes as its first argument a formula with the cross-classifying
elements joined together using a plus (+) sign on the right-hand side of a tilde ()
sign. The items joined by the plus sign are the cross-classifying elements that will
become the row and column names in the resulting table. The item to the left of the
tilde is the element that will be entered into the cells of the resulting table. Used
in this way, xtabs converts the long, three column, data frame into a wide format
data frame in which each row is a book and each column is a word feature. If you
run the previous expression and check the dimensions, youll see that the data.frame
is now 43 by 53233.
dim(result)
## [1]
43 53233

And, if you enter

colnames(result)

9 Other options include using reshape and expressions that leverage the apply family of functions.

11.8 Preparing Data for Clustering

115

you will get a very long list of all of the word features. For an easier going peek,
you might try
colnames(result)[100:110]
## [1] "ace"
## [3] "ache"
## [5] "acknowledge"
## [7] "acknowledges"
## [9] "acquaintance"
## [11] "acquiesced"

"aces"
"achive"
"acknowledged"
"acknowledgement"
"acquainted"

You are almost there! Before you can perform the final analysis, you need to
clean up the data types. Right now youll see that the result variable is an object
of the xtabs table variety.
class(result)
## [1] "xtabs" "table"

To be useful in the next step, you need to convert result into a matrix object. As with most things in R, there are many roads to the same destination. For
efficiency, I use apply to convert the columnar data to numeric format and by extension into a numerical matrix object. The 2 in this expression is a reference to
the columns (a 1 here would serve as a reference to rows.)
final.m <- apply(result, 2, as.numeric)

This generates a numeric matrix of 43 by 53233.

11.8 Preparing Data for Clustering

While it is certainly the case that you could apply the Euclidean metric to this huge
matrix, doing so does not make a lot of sense in the case of authorship attribution.
The goal is to figure out which of these texts is most stylistically similar to the
anonymous text, and you dont want to bias the results by clustering the texts based
on the similarity of their themes or content. Say, for example, that two of the books
in this corpus were about horses. These two books would likely be drawn together
in the clustering because they shared a similar subject and not necessarily because
they shared a similar style. Therefore before clustering it is useful to winnow the
data to just those features that are extremely frequent.
There are many ways to do this winnowing; you could, for example, sort the
data and keep only the 100 most frequent words in the corpus. I prefer to use a
winnowing method based on setting a frequency threshold. In other words, limit the
feature list to only those words that appear across the entire corpus with a mean
relative frequency of some threshold. For this example, I will set the threshold to
0.25, one-quarter of one percent. Here again the apply function becomes useful,
and in a single line, I can calculate the column means and generate a subset of
final.m that consists of only those columns that have a combined mean of at least

116

11 Clustering

0.25.10 In the following R expression, the number 2 is used to tell R to perform

the calculations over columns rather than rows. Rows, as noted above, would be
referenced with a 1.
smaller.m

<-

final.m[,apply(final.m,2,mean)>=.25]

You can now compare the size of the original matrix to the new, smaller one.
dim(final.m)
## [1]
43 53233
dim(smaller.m)
## [1] 43 53

Using 0.25 reduced the feature set to the 53 most frequent word features in the
corpus. With the data matrix reduced in this way, you can now run the clustering
very efficiently.

11.9 Clustering Data

With the hard work of data preparation over, you move to the simple and rewarding
work of extracting some results. R has a set of wonderful functions for clustering
data such as this. In this exercise you will use dist (a function that employs the
Euclidean metric to create a distance matrix) and hclust (which clusters the data
in a distance matrix).11 As you may have guessed, the function names are shorthand
for distance and hierarchical clustering. Each of these methods has a variety of
arguments that you can enter to fine-tune the way you want to run the analysis. For
our purposes the default values are adequate.
Using the following code
# Create a distance object
dm <- dist(smaller.m)
# Perform a cluster analysis on the distance object
cluster <- hclust(dm)
# Get the book file names to use as labels.
cluster$labels <- names(book.freqs.l)
# Plot the results as a dendrogram for inspection.
plot(cluster)

you can produce a cluster dendrogram (Fig. 11.2) and visually inspect the tree to
identify the known authors and texts that are most similar to anonymous.xml. If everything went well, you should have found anonymous.xml nestled comfortably between Kyne1.xml and Kyne2.xml. Peter B. Kyne is the author of the anonymous.xml!

10

Remember that the getTEIWordTableList function that we built multiplies all the relative
frequencies by 100.
11 For details, consult the documentation for the dist and hclust functions.

11.9 Clustering Data

117

4
Quigley1.xml
Quigley2.xml
Carleton5.xml
Lewis.xml
McHenry1.xml
McHenry2.xml
Carleton10.xml
Carleton12.xml
Carleton13.xml
Carleton14.xml
Carleton1.xml
Carleton11.xml
Kyne2.xml
anonymous.xml
Kyne1.xml
Carleton8.xml
Carleton3.xml
Carleton7.xml
Carleton9.xml
LeFanu1.xml
LeFanu6.xml
Carleton4.xml
Edgeworth1.xml
Driscoll1.xml
Driscoll2.xml
Driscoll3.xml
Jessop3.xml
Jessop1.xml
Jessop2.xml
Carleton6.xml
Polidori1.xml
Norris2.xml
Norris4.xml
Norris1.xml
Norris3.xml
Carleton2.xml
Donovan1.xml
Donovan2.xml
LeFanu3.xml
LeFanu2.xml
LeFanu5.xml
LeFanu4.xml
LeFanu7.xml

Height

Cluster Dendrogram

dm
hclust (*, "complete")

Fig. 11.2 Cluster dendrogram

Practice
11.1. Now that you have the correct answer, go back to the line of code in which
you generated the smaller.m matrix. Experiment with different threshold values.
Examine how the attribution result changes, or does not, depending upon the number
of features that you keep. What is the smallest number of word features you could
use in this clustering experiment and still arrive at the same answer?
11.2. As a final experiment, write some code to see what happens if you select a
random collection of features. In other words, instead of selecting from among the
most high frequency features, write code that uses the sample function to grab a
random sample of 50 or 100 word features and then see if you still get accurate
author clustering.

Chapter 12

Classification

Abstract This chapter introduces machine classification in the context of an authorship attribution problem. Various methods of text pre-processing are combined here
to generate a corpus of 430 text samples. These samples are then used for training
and testing a support vector machines supervised learning model.

12.1 Introduction
The clustering described in the last chapter is not ideally suited to authorship attribution problems. In fact, clustering is more often used in cases in which the classes are
not already known in advance. Clustering is often employed in situations in which a
researcher wishes to explore the data and see if there are naturally forming clusters.
When the classes are known in advance (i.e., when there is a closed set of possible
classes or authors in this case) supervised classification offers a better approach.
In addition to providing more information about feature level data, a supervised
approach can also provide probabilistic data about the likelihood of a given document being written by one author versus another within the closed set of candidates.
Though I will use an authorship attribution example here again, consider that any
category of metadata can be inserted into the place held by author. For example, if
you wished to gauge the extent to which Irish style differs from British style, you
could use nationality in place of author as the target class.

12.2 A Small Authorship Experiment

For this chapter, you will use the same corpus of novels that was used in the clustering chapter, and you will be able to recycle much of the code you have already
written. If you have not already done so, clear your R workspace and set the working

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_12,
Springer International Publishing Switzerland 2014

119

120

12 Classification

directory to the location of your TextAnalysisWithR directory. You can now

start a new R script with the following code from Chap. 11:
setwd("~/Documents/TextAnalysisWithR")
library(XML)
input.dir <- "data/XMLAuthorCorpus"
files.v <- dir(input.dir, ".*xml")

12.3 Text Segmentation

Instead of treating every novel as a single text, as was done in the last chapter,
here you will first break each text into segments. Instead of having 43 texts for
training and testing a classification algorithm, youll create 430 by first breaking
each text into ten equal portions. In the last chapter, you wrote a for loop to
load a series of XML files and then send each XML document object to a function
(getTEIWordTableList) that would return a table of frequencies. Here you
will write a similar function, but before calculating the frequencies and returning
the list of tables, this function will first segment each text file into multiple chunks.
This new function will take two arguments: an XML document object and a chunk
size parameter. Lets call the new function getTEIWordSegmentTableList.
Just like the previous function you wrote (getTEIWordTableList), this one
will also extract the paragraph level content from the XML and then generate a vector
of words. Instead of making a table of that entire word vector, however, this new
function will first cut the vector into a set number of slices. To get started, you can
copy the first few lines from the getTEIWordTableList function and begin a
new function:
getTEIWordSegmentTableList <- function(doc.object, chunk.size=10){
paras <- getNodeSet(doc.object,
"/d:TEI/d:text/d:body//d:p",
c(d = "http://www.tei-c.org/ns/1.0"))
words <- paste(sapply(paras,xmlValue), collapse=" ")
words.lower <- tolower(words)
words.list <- strsplit(words.lower, "\\W")
word.v <- unlist(words.list)
# . . . some new code here
}

The new code that you will now add must calculate the length of the word.v
object and divide it by the value of the chunk.size argument, which has been set
to 10 as a default.1
You already know how to get the length of a vector, so all you need to do here is
add a further bit of division:

1 Inside a function definition, you can define default values for the different arguments. When the
function is called from the main script, the default will be used unless you specifically set a value
in the function call.

12.3 Text Segmentation

121

length(word.v)/chunk.size

Of course, division like this is never perfect; there are remainders and decimals
to deal with. There are several ways you can deal with this, but here is one way
that uses the built-in seq_along, split and ceiling functions to split the
word.v vector into a series of chunks:
max.length <- length(word.v)/chunk.size
x <- seq_along(word.v)
chunks.l <- split(word.v, ceiling(x/max.length))

The resulting chunks.l object will be a list in which each item in the list is
a character vector. These new character vectors are just slices of the full word.v
object. They can now be converted into frequency tables in the same way that a table
can be generated from the full vector (as you did in Chap. 11).
Before you table them, however, you will need to remove those pesky blank
characters. If you were to peek inside the function at the structure of the chunks.l
object for a given text, you would see something like this:
str(chunks.l)
List of 10
$ 1 : chr [1:7421]
$ 2 : chr [1:7422]
$ 3 : chr [1:7421]
$ 4 : chr [1:7422]
$ 5 : chr [1:7421]
$ 6 : chr [1:7422]
$ 7 : chr [1:7421]
$ 8 : chr [1:7422]
$ 9 : chr [1:7421]
$ 10: chr [1:7422]

"there" "are" "one" "hundred" ...

"little" "over" "a" "year" ...
"" "for" "" "600" ...
"therefore" "predict" "for" "him" ...
"there" "is" "no" "good" ...
"arrival" "" "he" "engaged" ...
"" "if" "such" "a" ...
"is" "a" "journal" "of" ...
"easily" "made" "in" "california" ...
"" "" "" "" ...

Notice, for example, the first item in the third list item: . You need to remove
the blanks before making the table, and while there are several ways you could do
this (e.g., we used which in previous lessons), here Ill show you how to write a
very small, reusable function that can be called iteratively using lapply:
removeBlanks <- function(x){
x[which(x!="")]
}

Save this function to your corpusFunction.R file, and you can add a line to the
main getTEIWordSegmentTableList function that will send the chunk list
object to this new removeBlanks function. In this case, since all you are doing is
removing the blanks from each character vector, you can overwrite the chunks.l
object with the new, cleaner version:
chunks.l <- lapply(chunks.l, removeBlanks)

As seen in previous lessons, the lapply function is well suited for use in combination with the table function. You can now table each word vector in the
chunks.l list using another call to lapply.
freq.chunks.l <- lapply(chunks.l, table)

122

12 Classification

These raw frequencies may now be converted to relative frequencies, and by

now you should know that you can convert a single table of raw counts to relative
frequencies like this:
freq.chunks.l[[1]]/sum(freq.chunks.l[[1]])

Since we do not want to convert each table in the list one at a time, having another
function that we can use in a call to lapply is handy. Naturally R has just such a
function: prop.table.
rel.freq.chunk.l

<- lapply(freq.chunks.l, prop.table)

In the resulting rel.freq.chunk.l object, you will now have a list object
containing the relative frequency data for 10 segments of the text. Now simply add
a final line of code to the function that will return the results back to the main script.
Here is how your full function should look.
getTEIWordSegmentTableList <- function(doc.object, chunk.size=10){
paras <- getNodeSet(doc.object,
"/d:TEI/d:text/d:body//d:p",
c(d = "http://www.tei-c.org/ns/1.0"))
words <- paste(sapply(paras,xmlValue), collapse=" ")
words.lower <- tolower(words)
words.list <- strsplit(words.lower, "\\W")
word.v <- unlist(words.list)
max.length <- length(word.v)/chunk.size
x <- seq_along(word.v)
chunks.l <- split(word.v, ceiling(x/max.length))
chunks.l <- lapply(chunks.l, removeBlanks)
freq.chunks.l <- lapply(chunks.l, table)
rel.freq.chunk.l <- lapply(freq.chunks.l, prop.table)
return(rel.freq.chunk.l)
}

The last line of this function returns a list object back to the main script. If the
chunk.size argument is left at the default value of 10, then the resulting list
will contain 10 word frequency tables derived from a single book. Save this new
function to your corpusFunctions.R file.
As you did in Chap. 11, here again you will instantiate a list object outside of a
for loop as a place to hold the results returned from the new function.
book.freqs.l <- list()
for(i in 1:length(files.v)){
#. . . Some code here
}

With each iteration of the loop, an XML file will be loaded into a new document
object and sent to getTEIWordSegmentTableList. The function will return
a new list object for each XML file. You will then need to add each of these list
objects to the larger book.freqs.l object. While you could do this all in one
single line of code:
book.freqs.l[[files.v[i]]] <- getTEIWordSegmentTableList(doc.object, 10)}

Your code will be easier to read if you break it into two steps. First

12.3 Text Segmentation

chunk.data.l <-

123

getTEIWordSegmentTableList(doc.object, 10)

and then
book.freqs.l[[files.v[i]]] <-

chunk.data.l

The chunk.data.l list is a temporary storage container. The contents of chunk.

data.l get inserted into the larger container list book.freqs.l. At the same
time, you can use double brackets to insert a name for the list that you derive from
the file name in the (files.v) vector. Remember that the (files.v) vector is
populated with the file names of the XML files inside the input.dir. When the
script is run, the book.freqs.l list will contain 43 list objects (one for each
text) and each of these will contain another list of 10 items, each one of these 10
will hold a frequency table corresponding to the one-tenth of the original text. In
other words, you will have a list of lists of tables. Deep breath.
book.freqs.l <- list()
for(i in 1:length(files.v)){
doc.object <- xmlTreeParse(file.path(input.dir, files.v[i]),
useInternalNodes=TRUE)
chunk.data.l <- getTEIWordSegmentTableList(doc.object, 10)
book.freqs.l[[files.v[i]]] <- chunk.data.l
}

Run this code now and then examine the main list with the length function,
youll see that it contains 43 records.
length(book.freqs.l)
## [1] 43

Each of these records, however, then contains 10 items, and you can access any
single list item via double-bracketed sub-setting:
length(book.freqs.l[[1]])
## [1] 10

With all the necessary data collected from the texts, you now require a way to
munge it all into a single data frame such that each row is a text chunk and each
column is a different word feature. Unlike the clustering experiment in Chap. 11,
you now have 430 rows, ten for each novel or one row for each novel segment.
It might help at this point to explore the book.freqs.l list object a bit
deeper. You can begin with str(book.freqs.l) where you will see that
book.freqs.l is a list of 43 items. Note that each of these 43 items is a named
list of ten table objects. The names, you will recall, got assigned based on the original file names in the files.v object. Since the items in the main list are named,
you can inspect them by numeric index
str(book.freqs.l[[37]])

or by name
str(book.freqs.l$Norris1.xml)

124

12 Classification

You will see that the item named Norris1.xml in the main book.freqs.l list
is another list containing ten items (the ten text segments). You will also see that
each of these ten items is a table object containing a named vector of words and a
corresponding set of values for the relative frequencies.
If you enter book.freqs.l$Norris1.xml[[1]], you will see the entire
table of values for the first chunk of the book in the file titled Norris1.xml. Here Ill
show only the first three words and their corresponding relative frequencies.
book.freqs.l$Norris1.xml[[1]][1:3]
##
##
a
ability
able
## 0.0282745826 0.0000742115 0.0000742115

If you want to see the first three values in the second text segment, enter
book.freqs.l$Norris1.xml[[2]][1:3]
##
##
a
abjuring
able
## 2.681674e-02 7.554011e-05 7.554011e-05

As in Chap. 11, you now need a way of converting this list into a matrix. This time,
however, things are a bit more complicated because you will need to keep track of
ten separate chunks for each text.

12.4 Converting an R List into a Matrix

The goal now is to convert this list, or more precisely, list of lists of tables into a
large matrix in which each row is a text chunk (or segment) and each column is
a single word feature. The values in the corresponding cells should be the relative
frequencies for the given word column in a given row chunk. Along the way, you
will need to keep track of which rows are from which texts.
Some of the R code for doing this will be familiar from the clustering chapter, but
with the addition of one more step that allows you to dig one level deeper into the
book.freqs.l list. Recall from the clustering chapter that you used mapply to
convert each table of word frequencies into a data frame:
mapply(data.frame,ID=seq_along(book.freqs.l),
book.freqs.l,SIMPLIFY=FALSE,
MoreArgs=list(stringsAsFactors=FALSE))

You now have ten tables for each list item, so you can use lapply to iterate over
the primary list (book.freqs.l) and send each of the ten table objects to another
function that will stitch them all together. For this you will write a short function
called my.mapply.
Begin with code for calling lapply with the new function:
freqs.l

<-

lapply(book.freqs.l, my.mapply)

Now you just need to write my.mapply; lapply will do the work of calling
the new function for each item in the book.freqs.l list. The inner workings of

12.5 Organizing the Data

125

this custom my.mapply function will look very similar to the code you wrote in
the clustering chapter. my.mapply will map each of the ten tables together and
then convert them into data frames using do.call and rbind. The only real
change here is that you have now encapsulated this process inside a function that
gets iteratively called using lapply. Here is the function:
my.mapply <- function(x){
my.list <- mapply(data.frame, ID=seq_along(x),
x, SIMPLIFY=FALSE,
MoreArgs=list(stringsAsFactors=FALSE))
my.df <- do.call(rbind, my.list)
return(my.df)
}

As you see, my.mapply returns a data frame in which each row contains a
chunk reference, a feature type, and a frequency value. You can now call the function
and rbind the resulting list object into a long form data frame.
freqs.l <- lapply(book.freqs.l, my.mapply)
freqs.df <- do.call(rbind,freqs.l)

The resulting data frame is 770818, 3, and using the head function, you can preview
the first few rows of the results:
head(freqs.df)
##
ID
Var1
Freq
## anonymous.xml.1 1
11 9.458948e-05
## anonymous.xml.2 1
1850 9.458948e-05
## anonymous.xml.3 1
a 2.724177e-02
## anonymous.xml.4 1
abandon 9.458948e-05
## anonymous.xml.5 1 abandoned 1.891790e-04
## anonymous.xml.6 1
aboard 1.891790e-04

12.5 Organizing the Data

Examining this output, you will notice that each row begins with a row name corresponding to the original file name in the corpus. You will also see that there are three
columns of data with the column names ID, Var1, and Freq. The ID is the chunk
reference number, Var1 is the word type (in this case the first two word types are
the numbers 11 and 1850 because digits were not stopped out during tokenization),
and Freq is the relative frequency of the word type in the particular chunk of the
particular text. Before you can cross tabulate this data and create the wide form data
frame you need for the analysis, you will first need to extract some metadata from
the row names and ID columns. You can do this using a new regular expression and
the gsub function.
Like grep, gsub is a pattern matching function. More specifically, gsub is
a function for finding patterns and replacing them with something else: a global
find and replace on steroids. First, to tidy things up a bit, you can use gsub to
generate a character vector of file names with the .xml stripped off (replaced). For

126

12 Classification

arguments, gsub takes a pattern to search for, a replacement value, and a object
to search within. You can use the regular expression \\..* to find the period (.)
character followed by any number of other characters in the row names vector.2 You
can replace those matches with nothing () and store the resulting strings in a new
variable called bookids.v.
bookids.v <- gsub("\\..*", "", rownames(freqs.df))

Using the paste function, you can then glue each of these new strings to the
corresponding segment/chunk reference value in the ID column.
book.chunk.ids <- paste(bookids.v, freqs.df$ID, sep="_")

This allows you to create a unique book-with-chunk identification string. To make

these easy to read, I have glued them using an underscore character as the value for
the sep argument. Now you can replace the existing values in the ID column with
the new values.
freqs.df$ID <- book.chunk.ids

If you look at the first few rows, youll see that the ID column values have changed
to values expressing both the file name and the chunk ID.
head(freqs.df)
##
## anonymous.xml.1
## anonymous.xml.2
## anonymous.xml.3
## anonymous.xml.4
## anonymous.xml.5
## anonymous.xml.6
##
## anonymous.xml.1
## anonymous.xml.2
## anonymous.xml.3
## anonymous.xml.4
## anonymous.xml.5
## anonymous.xml.6

ID
Var1
anonymous_1
11
anonymous_1
1850
anonymous_1
a
anonymous_1
abandon
anonymous_1 abandoned
anonymous_1
aboard
Freq
9.458948e-05
9.458948e-05
2.724177e-02
9.458948e-05
1.891790e-04
1.891790e-04

12.6 Cross Tabulation

All of this was necessary in order to cross tabulate the data. As you saw in Chap. 11,
cross-tabulation works much like a pivot table in Excel. You want to create a single
row of data for each text chunk. In the current long form matrix you have a row
for each word type. There are several ways of doing cross tabulation in R, but here
again I will use the xtabs function.
2 In regex the period character is used as a special wild card. So in this expression the first period
must be escaped using the double backslashes. This tells the regex engine to find the literal period.
The second period in the expression is the period being used as a wild card metacharacter. The
asterisk is another special character that is used as a multiplier. So here the asterisk repeats the
wild card character indefinitely, until the end of the search string is reached.

12.7 Mapping the Data to the Metadata

result.t

<-

127

xtabs(Freq ~ ID+Var1, data=freqs.df)

The xtabs function is given a formula: on the right side of the tilde character,
the variables for cross classification: ID and Var1, on the left side, the word frequency values held in the Freq column. In building this new wide format matrix, R
will treat ID and Var1 as the row name and column name values, respectively. It will
then insert the Freq value into the corresponding cells in the new matrix. The new
xtabbed object is 430 x 53233. To make the data easier to work with, convert this
xtab table object into a data frame:
final.df <- as.data.frame.matrix(result.t)

If you would like to examine the values for any specific word type, you can do
that easily. Here is how to look at the frequencies for the words of and the in the
first ten rows:
final.df[1:10, c("of", "the")]
##
of
the
## anonymous_1 0.02989028 0.06356413
## anonymous_10 0.02172390 0.04765242
## anonymous_2 0.02339697 0.05041605
## anonymous_3 0.02262488 0.05631734
## anonymous_4 0.02081288 0.05105046
## anonymous_5 0.02227040 0.05764683
## anonymous_6 0.02654010 0.04726850
## anonymous_7 0.02180132 0.05031074
## anonymous_8 0.02395151 0.05601721
## anonymous_9 0.02011694 0.05569319

Before you can use any of this in a classification test, however, you still have
a bit more preprocessing to do. Since this is an authorship attribution experiment,
you probably want to reduce the data frame to include only the very high frequency
features, and you will also need a way of keeping track of the metadata, specifically
which texts belong to which authors.

12.7 Mapping the Data to the Metadata

First and foremost, you need a way of mapping the word frequency data not just
to specific text samples (i.e., the specific chunks) but also to the specific authors.
In this corpus you have multiple texts from multiple authors. In fact, excluding the
anonymous book, there are 12 authors, 42 books, and 420 book chunks. What you
need right away is an author column. Because these files were named with the authors last name, you can extract the necessary metadata from what is now the row
name in the final.df object.
Begin by deriving a new matrix object (metacols.m) by splitting the row
names using that underscore character that was inserted during the paste command
above.

128

12 Classification

metacols.m <- do.call(rbind, strsplit(rownames(final.df), "_"))

head(metacols.m)
##
[,1]
[,2]
## [1,] "anonymous" "1"
## [2,] "anonymous" "10"
## [3,] "anonymous" "2"
## [4,] "anonymous" "3"
## [5,] "anonymous" "4"
## [6,] "anonymous" "5"

To keep things organized and human readable, reset the column names to something
that makes more sense. Then, using head you can inspect the first few rows for the
anonymous text, but remember that for some authors there are multiple books.
colnames(metacols.m) <- c("sampletext", "samplechunk")
head(metacols.m)
##
sampletext samplechunk
## [1,] "anonymous" "1"
## [2,] "anonymous" "10"
## [3,] "anonymous" "2"
## [4,] "anonymous" "3"
## [5,] "anonymous" "4"
## [6,] "anonymous" "5"

If you want to see all of the unique values in the sampletext column, the
unique function is handy:
unique(metacols.m[,"sampletext"])
## [1] "anonymous" "Carleton1"
## [4] "Carleton11" "Carleton12"
## [7] "Carleton14" "Carleton2"
## [10] "Carleton4" "Carleton5"
## [13] "Carleton7" "Carleton8"
## [16] "Donovan1"
"Donovan2"
## [19] "Driscoll2" "Driscoll3"
## [22] "Jessop1"
"Jessop2"
## [25] "Kyne1"
"Kyne2"
## [28] "LeFanu2"
"LeFanu3"
## [31] "LeFanu5"
"LeFanu6"
## [34] "Lewis"
"McHenry1"
## [37] "Norris1"
"Norris2"
## [40] "Norris4"
"Polidori1"
## [43] "Quigley2"

"Carleton10"
"Carleton13"
"Carleton3"
"Carleton6"
"Carleton9"
"Driscoll1"
"Edgeworth1"
"Jessop3"
"LeFanu1"
"LeFanu4"
"LeFanu7"
"McHenry2"
"Norris3"
"Quigley1"

As you can see, there are 43 unique texts. You can also see that some texts are by
the same authors. You need a way to identify that both books by Quigley (Quigley1
and Quigley2) are by the same author, and likewise for the other authors from whom
there are multiple samples. Use gsub again, and another regular expression, that
will find instances of one or more digits (i.e., the 1 or 2 in Quigley1 and Quigley2)
followed by the end of a character string, which you indicate using the dollar ($)
symbol. When a match is found, gsub will replace the matched string with nothing,
which has the effect of deleting the digits. The result can be saved into a new object
called author.v.
author.v <- gsub("\\d+$", "", metacols.m[,"sampletext"])

12.8 Reducing the Feature Set

129

You can then check your work using unique.

unique(author.v)
## [1] "anonymous"
## [4] "Driscoll"
## [7] "Kyne"
## [10] "McHenry"
## [13] "Quigley"

"Carleton"
"Edgeworth"
"LeFanu"
"Norris"

"Donovan"
"Jessop"
"Lewis"
"Polidori"

With a new vector of author names, you can now create a final data frame that
binds this vector as a new column along with the two columns in the metacols
variable to the existing final.df:
authorship.df <- cbind(author.v, metacols.m, final.df)

12.8 Reducing the Feature Set

At 430 by 53236, this new data frame contains way too many features for an
authorship attribution test. You will need to reduce the number of columns to just
those that contain the high frequency features. In Chap. 11, I showed one way of
achieving this using apply with a conditional expression (see Sect. 11.7). Here I
show an alternative approach using colMeans to select only those feature with a
mean relative frequency across the corpus of 0.005.3
The task here is to calculate the overall mean of each word type column in the
authorship.df object. R gives us colMeans for doing just this. Remember
though that the first three columns in authorship.df are metadata (containing
the author and text information), so you only want to get the means for the columns
containing frequency data. To access just these columns, you can use bracketed
sub-setting and a sequence vector running from 4 through the number of columns
in the authorship.df object. To determine that end point, use ncol (number of
columns), another R function that is similar to length but specific to data frames
and matrices.
freq.means.v <- colMeans(authorship.df[,4:ncol(authorship.df)])

You can now identify which of these column means is greater than or equal to
0.005 using which. Ill save these to a vector called keepers.v
keepers.v <- which(freq.means.v >=.005)

and since there are not going to be too many of them, Ill inspect the entire vector
in the console.

Note that in this chapter the function has not been written to multiply the relative frequency
values by 100.

130

12 Classification

keepers.v
##
a and
as
at
be but for had
he her
##
3
83 116 122 171 285 806 940 970 993
## him his
i
in
is
it not
of
on
s
## 1008 1010 1050 1072 1110 1113 1424 1441 1448 1765
## she that the
to was with you
## 1849 2102 2103 2149 2282 2357 2397

From this vector of values that met the condition, you can grab the names of the
word types using names(keepers.v). You could then use those names to identify the subset of columns in the authorship.df object that you want to retain
for analysis:
smaller.df <- authorship.df[, names(keepers.v)]

Unfortunately, the line of code above does not include the metadata columns about
the authors and texts that you were so careful to preserve and organize. While you
could just cbind those meta columns back in to the new smaller.df like this:
smaller.df <- cbind(author.v, metacols, smaller.df)

a simpler solution would be to identify the columns you want right from the start
but combining the names from the keeper.v vector with the column names of the
first three columns in the main authorship.df data frame.
smaller.df <- authorship.df[, c(names(authorship.df)[1:3],
names(keepers.v))]

12.9 Performing the Classification with SVM

With all of the data preparation done, you are finally ready to perform the classification analysis and see if you can figure out who wrote that anonymous book! Begin
by identifying the rows in the new data frame belonging to the anonymous author.
anon.v <- which(smaller.df$author.v == "anonymous")

Next identify the data that will be used to train the model by telling R to take only
the rows of smaller.df that do not include those identified in the anon vector.
This negation is done using the - operator before the object name. It has the effect
of saying all the rows except for these or less these. For this classification, you
do not want to include the first three columns where the metadata is stored, so use
4:ncol(smaller.df) to grab only the fourth through last columns.
train

<-

smaller.df[-anon.v,4:ncol(smaller.df)]

Now identify a class column that the classifier will use to organize the data. That
is, you need to give the classifier a vector of values for the classes that are already
known. In this case, the true author names are stored in the column headed author.v.

12.9 Performing the Classification with SVM

class.f

<-

131

smaller.df[-anon.v,"author.v"]

This new vector is of a special type that R calls a factor.4 With the classes identified, you just need to pick a classifier and run the classification.5 To keep things
simple and to avoid having to load a lot of complex classification packages, well
use a comparatively familiar algorithm, SVM or Support Vector Machines which is
part of the e1071 package.
If you need help installing a package, see Chap. 10, Sect. 10.4 (Installing R Packages) for instructions. Otherwise, load the library:
library(e1071)

## Loading required package: class

You can now generate a model using the svm classifier function and the data contained in the train and class.f objects:
model.svm

<-

svm(train, class.f)

Once the model is generated, you can examine the details using the summary
function.
To test the accuracy of the model, use the predict function with the model
.svm and the training data in the train object.
pred.svm

<-

predict(model.svm, train)

The pred.svm object will now contain a vector of text labels and the machines
guesses. If you examine the contents of pred.svm
as.data.frame(pred.svm)

youll see that the text sample labeled Carleton5_1 was incorrectly assigned to
Quigley and that McHenry1_5 was incorrectly assigned to Carleton. For the most
part, however, youll see that the model has done very well. To see a summary in
the form of a confusion matrix you can use table:
table(pred.svm, class.f)

4 Factors are very similar to vectors except that in addition to storing the vector data, in this case a
set of character strings referring to authors, the factor also stores levels. Factors provide an efficient
way of storing repetitive character data because the unique character values are actually only stored
once and the data itself is stored as a vector of integers that refer back to the single character strings.
5 There are many good classification algorithms that can be used for authorship attribution testing,
and in 2010 Daniela Witten and I published a bench-marking study of five well known algorithms.
See Jockers, Matthew L. and Daniela M. Witten. A Comparative Study of Machine Learning
Methods for Authorship Attribution. Literary and Linguistic Computing, 25.2, (2010): 215224;
doi: 10.1093/llc/fqq001. We concluded that the Nearest Shrunken Centroids was especially good,
but frankly, the others we tested also performed quite well. Interested readers should also look at
the work of Jan Rybicki and Maciej Eder found at the Computational Stylistics Group website:
https://sites.google.com/site/computationalstylistics/.

132

12 Classification

When you look at the Carleton column in the results, youll see that 139 of the
Carleton samples were assigned correctly to Carleton and only one was incorrectly
assigned to Quigley. Youll see that all 20 of the Donovan samples were correctly
assigned to Donovan, and so on. This model has performed very well in terms of
accurately classifying the known authors.
Based on this validation of the models accuracy in classifying the known authors, you can classify the anonymous text with a good deal of confidence. First
isolate the test data:
testdata

<-

smaller.df[anon.v,4:ncol(smaller.df)]

and then send the test data to the model for prediction. View the results using
as.data.frame.
final.result <- predict (model.svm, testdata)
as.data.frame(final.result)
##
final.result
## anonymous_1
Kyne
## anonymous_10
Kyne
## anonymous_2
Kyne
## anonymous_3
Kyne
## anonymous_4
Kyne
## anonymous_5
Kyne
## anonymous_6
Kyne
## anonymous_7
Kyne
## anonymous_8
Kyne
## anonymous_9
Kyne

The results of this svm classification confirm what was observed in the clustering
test in Chap. 11; Kyne has been identified as the most likely author of every single
segment of the anonymous book!

Practice
12.1. Now that you know the author and have seen how the classifier correctly
guesses the author of each of the ten samples, increase the number of features
the model uses by decreasing the feature mean used to determine the number of
features retained in the keepers.v object. In the example in this chapter, 27
high-frequency features were retained using a mean relative frequency threshold
of .005. Decrease this number in order to observe how the attributions change
with the addition of context sensitive words. When using the 681 highest frequency
word features, for example, the classifier gets every single attribution wrong. Why?
12.2. For the example in this chapter, I used a corpus wide mean relative frequency
threshold to select the high-frequency features to keep for the analysis. In practice
Exercise 12.1 you saw how increasing the number of features can lead to incorrect results because author style gets lost in context. Another winnowing method
involves choosing features based on the restriction that every selected feature must
appear at least once in the work of every author. Write code that will implement such

12.9 Performing the Classification with SVM

133

winnowing in order to generate a new set of values for the keepers.v object. Here
is some sample code for you to consider. Notice that feature f1 is not found in either
of the samples from author C and feature f5 is not found in any of the samples from
author A. Features f1 and f5 should, therefore, be removed from the analysis.
authors <- c("A","A","B","B","C", "C")
f1 <- c(0, 1, 2, 3, 0,0)
f2 <- c(0, 1, 2, 3, 0,1)
f3 <- c(3, 2, 1, 2, 1,1)
f4 <- c(3, 2, 1, 2, 1,1)
f5 <- c(0, 0, 1, 2, 1,1)
author.df <- data.frame(authors, f1,f2,f3,f4, f5)
author.df # Show the original data frame
##
authors f1 f2 f3 f4 f5
## 1
A 0 0 3 3 0
## 2
A 1 1 2 2 0
## 3
B 2 2 1 1 1
## 4
B 3 3 2 2 2
## 5
C 0 0 1 1 1
## 6
C 0 1 1 1 1
author.sums <- aggregate(author.df[, 2:ncol(author.df)],
list(author.df[,1]), sum)
reduced.author.sums <- author.sums[,
colSums(author.sums==0) == 0]
keepers.v <- colnames(
reduced.author.sums)[2:ncol(reduced.author.sums)]
smaller.df <- author.df[, c("authors", keepers.v)]
smaller.df # show the new data frame
##
authors f2 f3 f4
## 1
A 0 3 3
## 2
A 1 2 2
## 3
B 2 1 1
## 4
B 3 2 2
## 5
C 0 1 1
## 6
C 1 1 1

Chapter 13

Topic Modeling

Abstract This chapter introduces topic modeling using the mallet package, part
of speech tagging with openNLP, and topic-based word cloud visualization using
the wordcloud package. (In this chapter I assume that readers are already familiar
with the basic idea behind topic modeling. Readers who are not familiar may consult
Appendix B for a general overview and some suggestions for further reading.)

13.1 Introduction
I think it is safe to say that topic modeling is one of the hottest trends in digital
humanities research today. I was introduced to topic modeling sometime in 2004 or
2005. At the time I was naively working on some code for a tool that I was calling the canonizer. Basically, my canonizer was built to do a kind of comparative
corcording. I needed software that would compare word frequency lists and collocates across different metadata facets such as year, author gender, nationality, and so
on. With this code I hoped to isolate thematic elements in each text by identifying
frequently collocated word clusters. Id then use these feature clusters to compare
texts and explore the question of canonicity and whether there were marked differences between books in the traditional canon and those that have been historically
marginalized.
Sometime during my work on the canonizer, I bumped into latent semantic
analysis (LSA). I jumped into that rabbit hole and spent a good deal of time trying
to apply LSA to my canonizer problem. LSA seemed complicated enough, but then
in 2006 I read about the work of David Newman and his group at UCI.1 It was then
that I got formally introduced to and seduced by Topic Modeling.

Newman, Smyth, Steyvers (2006). Scalable Parallel Topic Models. Journal of Intelligence
Community Research and Development and Newman and Block (2006). Probabilistic Topic
Decomposition of an Eighteenth Century Newspaper. In JASIST, March 2006.

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4_13,
Springer International Publishing Switzerland 2014

135

136

13 Topic Modeling

Topic modeling offered a way of sifting through the noise in my canonizer

application; it was an approach that would better isolate the gold nuggets and better
sift out all the sand. If the method worked, it would offer a way of tracking collocates on a grand scale. I began a correspondence with David and eventually invited
him to Stanford to give a lecture on topic modeling.
In advance of the meeting, I gave David a small corpus of books by Jane Austen,
Dickens, and Melville. I asked him to run the models and give us a tour of the results.
For me and a few others in the room, it was a dazzling experiment. David described,
for example, a cluster of words the machine had identified that seemed to be about
sentiment. It was a cluster of words related to expressions of feeling and emotion.
Not only was it present in this corpus, but using the output of the model, David
could track and visualize its presence across the various books in the corpus.
Almost immediately, some in the room saw that there was something similar in
the charts David was showing for Emma, Mansfield Park, and Northanger Abbey,
something that was different from the patterns of sentiment use seen in the other
three Austen novels. Of course, the question on all our minds, whether we agreed
that there was a pattern on not, was whether this new sort of quantitative data could
be usefully interpreted. Did these thematic trends mean something and, more important, if they did mean something was it something new?
Since that meeting, Ive spent a good deal of time trying to disambiguate the
complexities of topic modeling while applying the method in ways that I believe
generate new knowledge.2 In the later chapters of Macroanalysis, I use topic modeling as a way to study thematic trends in a corpus of nineteenth century fiction,
3,346 books in total. Over a period of four years, I modeled and remodeled this
data, mostly using Davids Mimnos implementation of LDA in MALLET. Along
the way, I learned a few tricks about how to derive what I consider to be satisfactory
themes. In this chapter, I show you a few of these tricks.

13.2 R and Topic Modeling

At the time of this writing there are three topic modeling packages for R. These
include topicmodels from Bettina Grn and Kurt Hornik, lda by Johnathan
Chang, and mallet by David Mimno.3 Though the mallet package for R is a
relative newcomer, the Java package upon which it is based is not. I have chosen to
2 Readers seeking a user-friendly introduction to how topic modeling actually works should consult
Appendix B.
3 The topicmodels package provides an implementation of (or interface to) the C code developed by LDA pioneer David Blei. See Blei, David M., Ng, Andrew Y., and Jordan, Michael I.
Latent Dirichlet Allocation. Journal of Machine Learning Research, 3 (2003) 9931022.
Johnathan Chang is a researcher at Facebook who has worked with Blei and with whom he
has coauthored several papers including the influential topic modeling paper: J Chang, S Gerrish,
C Wang, JL Boyd-Graber, DM Blei. Reading tea leaves: How humans interpret topic models.
Advances in neural information processing systems, 2009 http://machinelearning.
wustl.edu/mlpapers/paper_files/NIPS2009_0125.pdf.

13.3 Text Segmentation and Preparation

137

use the mallet package here because Mimnos implementation of topic modeling
in the MALLET Java package is the de facto tool used by literary researchers.4 In
short, MALLET is the most familiar topic modeling package in the humanities, and
it makes the most sense to work with it here.5

13.3 Text Segmentation and Preparation

Topic modeling treats each document as a bag of words in which word order is disregarded. Since the topic model works by identifying words that tend to co-occur, the
bigger the bag, the more words that will tend to be found together in the same bag.
If novels, such as those we will analyze here, tended to be constrained to only a very
small number of topics or themes, then treating each entire novel as one bag might
be fruitful. In reality, though, novels tend to have some themes that run throughout and others that appear at specific points and then disappear. In order to capture
these transient themes, it is useful to divide novels (and other large documents) into
chunks or segments and then run the model over those segments instead of over the
entire text.6 You must, therefore, begin by pre-processing the novels in the corpus
into segments.
Unlike the previous chapter where you segmented texts based on percentage (i.e.,
each book was chunked into ten equal sized portions), here you will write a function
that allows for chunking based on a set number of words. That is, you will be able to
set a specific chunk size, such as 1000 words, and then divide each text into some
number of 1000 word segments.
Begin by loading the XML package, referencing the corpus directory, and then
generating a vector of file names. This should be familiar from previous chapters.
library(XML)
inputDir <- "data/XMLAuthorCorpus"
files.v <- dir(path=inputDir, pattern=".*xml")

David Mimno, a professor at Cornell, is the developer and maintainer Of the Java implementation of LDA in the popular MAchine Learning for LanguagE Toolkit (MALLET) developed at
the University of Massachusetts under the direction of Andrew McCallum: McCallum, Andrew
Kachites. MALLET: A Machine Learning for Language Toolkit. 2002. See http://mallet.
cs.umass.edu.
4 Mimno released (to CRAN) his R wrapper for the MALLET Topic modeling package on August
9, 2013.
5 I have used all three of these packages to good effect and prior to the release of the mallet
package I taught workshops using both topicmodels and lda. Each one has its advantages and
disadvantages in terms of ease of use, but functionally they are all comparable.
6 There appears to be no conventional wisdom regarding ideal text-segmentation parameters. David
Mimno reports in email correspondence that he frequently chunks texts down to the level of
individual paragraphs. Until new research provides an algorithmic alternative, trial and experimentation augmented by domain expertise appear to be the best guides in setting segmentation
parameters.

138

13 Topic Modeling

Now create and set a variable called chunk.size.

chunk.size <- 1000 # number of words per chunk

Now you need to write a function to handle the chunking. You wrote a similar function in Chap. 12 and you can reuse some of that code here. Find your
getTEIWordSegmentTableList function from Chap. 12 and copy and paste
it with the new name: makeFlexTextChunks. The modified function you will
write now will be designed to do percentage-based chunking by default (as in the
original from Chap. 12) but also able to accommodate word-count-based segmentation. Unlike the prior function, this one will not be returning lists of tables. For
topic modeling, we do not need to pre-calculate the word frequencies. The new
makeFlexTextChunks function will take three arguments: an XML document
object, a chunk size value (expressed as either a percentage or as a number of words),
and a third argument called percentage that will be set to TRUE by default but
can be reset to FALSE at run time in order to allow for word-count-based chunking.
Here is the initial framework for the function:
makeFlexTextChunks <- function(doc.object, chunk.size=10, percentage=TRUE){
paras <- getNodeSet(doc.object,
"/d:TEI/d:text/d:body//d:p",
c(d = "http://www.tei-c.org/ns/1.0"))
words <- paste(sapply(paras,xmlValue), collapse=" ")
words.lower <- tolower(words)
. . . # some new code here
}

As long as we are improving the way this function handles chunking, we will go
ahead and make a small change to the regular expression you have been using for
tokenizing the words in the file. In the older version of the function, you used the
\\W regular expression to break the files into words based on word boundaries. With
this regular expression, all alphanumeric characters (alphabetic letters and digits) are
retained, and all of the punctuation is stripped out. One of the by-products of this
stripping is that contractions and possessives that use an apostrophe are split into
two tokens: Cant becomes Can and t, and Bobs becomes Bob and s.7
There are a variety of ways that one might deal with this apostrophe situation
and there are many regular expression recipes for doing fairly precise tokenization,
but here things will be kept simple and straightforward. First you will use Rs
gsub function to find all punctuation marks excepting the apostrophe and replace
them with a blank spaces. Then you will tokenize the file using strsplit and the
regular expression s+ that will break the text string apart based on places where it

7 Another problem involves hyphens. Hyphens can appear at the end of lines as a printing convention but also in compound adjectives. Well not deal with that trickier problem here.

13.3 Text Segmentation and Preparation

139

finds one or more (indicated by the +) blank space characters.8

words.lower <- gsub("[^[:alnum:][:space:]']", " ", words.lower)
words.l <- strsplit(words.lower, "\\s+")

Recall from prior chapters that strsplit returns a list that you do not need.
The next line in the function removes the list. The line after that simply creates a
vector of word positions that you will need for figuring out where to chunk the text.
word.v <- unlist(words.list)
x <- seq_along(word.v)

With the words in a simple vector, the function must now either chunk by percentage (as in the default) or chunk by words. To handle this forking path, you can use
an if/else conditional expression that examines the value of the percentage
argument to make a decision. If percentage is TRUE (the default), the function
will execute the code exactly as it did in the original function from Chap. 12. If
percentage is FALSE, then the function will execute new code for doing wordcount-based segmentation. At this point the function looks just as it did when you
wrote it for Chap. 12, except that you have improved the tokenization and added the
internal if/else conditional. Here it is:
makeFlexTextChunks <- function(doc.object, chunk.size=1000, percentage=TRUE){
paras <- getNodeSet(doc.object,
"/d:TEI/d:text/d:body//d:p",
c(d = "http://www.tei-c.org/ns/1.0"))
words <- paste(sapply(paras,xmlValue), collapse=" ")
words.lower <- tolower(words)
words.lower <- gsub("[^[:alnum:][:space:]']", " ", words.lower)
words.l <- strsplit(words.lower, "\\s+")
word.v <- unlist(words.l)
x <- seq_along(word.v)
if(percentage){
max.length <- length(word.v)/chunk.size
chunks.l <- split(word.v, ceiling(x/max.length))
} else {
}
chunks.l <- lapply(chunks.l, paste, collapse=" ")
chunks.df <- do.call(rbind, chunks.l)
}

Unlike the percentage-based method where you divide the length of the word
vector (word.v) by the chunk.size variable, here you want to split the word
vector into chunks equal in length to the value of the chunk.size variable. The
word-count-based segmentation code is, therefore, very similar to the code in the

8 The regular expression used here may appear complicated compared to the simple W that has
been used thus far. In this case, the expression simply says: replace anything, except for an apostrophe, that is not an alphanumeric character with a blank space. ). [:alnum:] matches any
alphabetic or numeric character and [:space:] matches any blank space. The (apostrophe
character) then matches any apostrophes. The (caret character) at the beginning of the expression
serves as a negation operator, in essence indicating that the engine should match on anything that
is not a character, digit, space, or apostrophe: i.e. match all other characters!

140

13 Topic Modeling

percentage method, but it does create one added complication. When chunking a file
using percentages, each chunk is almost exactly the same size. When you split a text
into 500 or 1000 word chunks, however, the last chunk will typically be something
smaller than the chunk size you have set. Thus, you need a way of dealing with these
remainder chunks.
A simple way to deal with this situation is to add the remainder chunk onto the
second to last chunk, but you might not always want to do this. Say, for example, that
the chunk.size variable is set to 1000 words, and the last chunk ends up being
950 words long. Would you really want to add those 950 words to the previous
chunk? The answer is, of course, a subjective one, but a chunk of 950 words is
probably close enough to 1000 to warrant full chunk status; it should remain a
chunk of its own. But what if the last chunk were just 500 words, or 100 words;
those samples are getting fairly small. Since you must pick a cutoff value, it is
convenient to set a condition such that the last chunk must be at least one-half the
number of words as the value inside the chunk.size variable. You can code this
exception easily using the length function and some simple division wrapped up
inside another if conditional, like this:
if(length(chunks.l[[length(chunks.l)]]) <= chunk.size/2){
chunks.l[[length(chunks.l)-1]] <- c(chunks.l[[length(chunks.l)-1]],
chunks.l[[length(chunks.l)]])
chunks.l[[length(chunks.l)]] <- NULL
}

This conditional expression begins by getting the length of the word vector held
in the last item of the chunks.l list and then checks to see if it is less than or
equal to (<=) one-half of chunk.size. If the condition is met (i.e. TRUE), then
the words in the last chunk
chunks.l[[length(chunks.l)]]

are added to the words in the second-to-last chunk

chunks.l[[length(chunks.l)-1]]

and the last chunk is then removed by setting it to NULL

chunks.l[[length(chunks.l)]] <- NULL

Once this is done, the word vectors in each item of the chunks.l list can be
reduced back into strings using lapply to apply the paste function to each list
item (chunks.l<-lapply(chunks.l, paste, collapse=" ")). The
resulting list can then be morphed into a data frame object using do.call and
rbind (chunks.df<-do.call(rbind, chunks.l))) exactly as you did
in Chap. 12. The entire function now looks like this:

13.3 Text Segmentation and Preparation

141

makeFlexTextChunks <- function(doc.object, chunk.size=1000, percentage=TRUE){

paras <- getNodeSet(doc.object,
"/d:TEI/d:text/d:body//d:p",
c(d = "http://www.tei-c.org/ns/1.0"))
words <- paste(sapply(paras,xmlValue), collapse=" ")
words.lower <- tolower(words)
words.lower <- gsub("[^[:alnum:][:space:]']", " ", words.lower)
words.l <- strsplit(words.lower, "\\s+")
word.v <- unlist(words.l)
x <- seq_along(word.v)
if(percentage){
max.length <- length(word.v)/chunk.size
chunks.l <- split(word.v, ceiling(x/max.length))
} else {
chunks.l <- split(word.v, ceiling(x/chunk.size))
#deal with small chunks at the end
if(length(chunks.l[[length(chunks.l)]]) <=
length(chunks.l[[length(chunks.l)]])/2){
chunks.l[[length(chunks.l)-1]] <c(chunks.l[[length(chunks.l)-1]],
chunks.l[[length(chunks.l)]])
chunks.l[[length(chunks.l)]] <- NULL
}
}
chunks.l <- lapply(chunks.l, paste, collapse=" ")
chunks.df <- do.call(rbind, chunks.l)
return(chunks.df)
}

With the function built, all you need is a simple loop that will send all of the files
in the files.v variable to this function and bind the results into a master data
frame that you can call topic.df. Naturally, however, there is one complicating
factor. In addition to keeping track of the file names, you also need to keep track of
the segment numbers. Eventually you want to be able to move from the topic model
back to the original texts and the text segments. So you need to retain this metadata
in some form or another.
You did something similar in Chap. 12 with the custom my.mapply function.
Things are easier here. You can capture the original file names from the files.v
and massage them a bit using gsub to remove the file extensions9:
textname <- gsub("\\..*","", files.v[i])

With the unique file names in hand, you can label the segments numerically by
generating a sequence of numbers from 1 to the total number of chunks, which in
this case is the same as the number of rows (found using the nrow function) in
the chunks.df variable: i.e. 1:nrow(chunk.m). All of this data can then be
bound, column-wise, into a new matrix object called segments.m which will then
be row-bound to a matrix object (topic.m) that you instantiate prior to entering
the loop, as follows10:

You did this in Chap. 12 as well.

Paul Johnson pointed out that there is a more computationally efficient method for achieving this
same result. Instead of building a matrix object inside the loop, build a list and then use do.call
to rbind the list elements after the loop is completed.
10

142

13 Topic Modeling

topic.m <- NULL

for(i in 1:length(files.v)){
doc.object <- xmlTreeParse(file.path(inputDir, files.v[i]),
useInternalNodes=TRUE)
chunk.m <- makeFlexTextChunks(doc.object, chunk.size,
percentage=FALSE)
textname <- gsub("\\..*","", files.v[i])
segments.m <- cbind(paste(textname,
segment=1:nrow(chunk.m), sep="_"), chunk.m)
topic.m <- rbind(topic.m, segments.m)
}

The result is a matrix object (topic.m) with two columns: the first column
contains the unique file-segment identifiers and the second column contains the
strings of text from each segment. To prepare this matrix for ingestion into the
mallet topic modeling package, you must now convert it to a data frame and
rename the column headers:
documents <- as.data.frame(topic.m, stringsAsFactors=F)
colnames(documents) <- c("id", "text")

Mission accomplished; now lets do some topic modeling!

13.4 The R mallet Package

The first and most important thing to know about the mallet package is that it is
not a complete wrapper for the entire MALLET toolkit. As the documentation for the
package notes: Mallet has many functions, this wrapper focuses on the topic modeling sub-package written by David Mimno. So, do not look to this R wrapper if you
want to access any of MALLETs other functions, such as document classification
or hidden Markov models for sequence tagging. This package is strictly for topic
modeling. The mallet package is installed like any other package in R (see 10.4
for instructions). Once installed it is invoked using the library(mallet) expression.
library(mallet)

## Loading required package: rJava

You will notice that mallet relies on the rjava package, which basically allows
R to create Java objects and call Java methods.11 Recall that MALLET is written
in Java, not R.

13.5 Simple Topic Modeling with a Standard Stop List

In the stylistic analysis that was covered in prior chapters, high-frequency words
were retained and used as markers of individual authorial style. In topic modeling
you will typically want to remove, or stop-out, high frequency words such as the,
11

Methods in Java are more or less synonymous with functions in R.

13.5 Simple Topic Modeling with a Standard Stop List

143

of, and, a, an, etc. because these words carry little weight in terms of thematic or
topical value. If you do not remove these common function words, the topic model
will generate topics (weighted word clusters) that are less about shared semantic
sense, that is, less about topics or themes, and more about syntactical conventions.
Here is an example showing the top seven words from 20 topics that I derived
from the exercise corpus without using a stop list:
0 0.04674
1
0.06203
2
0.1085
3
0.14428
4
0.29792
5
0.16571
6
0.45213
7
0.24112
8
0.35832
9
0.55731
10 0.59945
11 0.13994
12 0.03846
13 0.10359
14 0.50041
15 0.07111
16 0.18016
17 0.39585
18 0.12091
19 0.58441

the to a of and don in

the to and of a bryce in
of the and in a is to
the of to and they their in
the and of a in to was
the a to of and was i
you i to it a and that
the of and a in by with
i you to my and is me
the of to and in that which
a to of he was his had
the to you an a that it
hycy ye bryan an o a the
the and a to of in susan
the of and a in was which
and the a of in with
the a and of his sir is
her she and the to was a
and i a the in my
he the his and was to him

As you can see, these are more or less meaningless.

In the data directory of the exercise corpus, I have included a stop list (stoplist.csv) of 606 high frequency words. In your own work you may want to add
to or cut this list to suit your research objectives, but this list will be sufficient for
our purposes here.
The first step in generating a topic model with the mallet package is to invoke
the mallet.import function. This function takes five arguments:
1.
2.
3.
4.
5.

id.array
text.array
stoplist.file
preserve.case
token.regexp

The first argument (id.array) is an array, or vector, of document ids. You have
this information stored in the first column of the documents data frame object that
you created above. The second argument (text.array) is a vector of text strings,
and you have this data in the second column of the documents data frame object.
Next is the stop list file which will be referenced using a relative path to its location
on your computer; in this case, the path will be data/stoplist.csv. The next
argument, preserve.case, is irrelevant in this example because you have already elected to lowercase all of the words as part of the makeFlexTextChunks
function. Had you not already done this, then mallet would allow you to choose
to do so or not at this point.12 The final argument, token.regexp allows you
12

mallets default is to convert to lowercase.

144

13 Topic Modeling

to define a specific regular expression for tokenizing the text strings. The default
expression is one that keeps any sequence of one or more Unicode characters. Because you have gone to a lot of trouble to retain apostrophes, youll need to give
mallet.import a new value to replace the default token.regexp argument.
To keep those apostrophes, the default expression ([\\p{L}]+) should be replaced
with [\\p{L}]+. The complete expression, with the slightly modified regular
expression, is as follows:
mallet.instances <- mallet.import(documents$id,
documents$text,
"data/stoplist.csv",
FALSE,
token.regexp="[\\p{L}']+")

The mallet.instances object created here is a Java object that is called a

Mallet instance list. This is not an R object and must be accessed using other Java
methods. The mallet package provides other functions as a gateway, or bridge, to
those methods.
The next step is to create a topic model trainer object, which, for the moment,
can be thought of as a kind of place holder object that you will fill with data in the
next few steps. Notice that it is at this stage that the number of topics that the model
will contain is set.13 For the sake of this tutorial, I am setting the number of topics
equal to the number of novels in the corpus. The reasons for this choice are purely
pedagogical and will make more sense as we work through the rest of this chapter.
## Create a topic trainer object.
topic.model <- MalletLDA(num.topics=43)

Because the mallet package is simply providing a bridge to the Java application, this might feel a bit obtuse, and it can be a bit disconcerting when you are
unable to employ R functions, such as class and str, to explore the makeup of
these objects. If you try, youll see references to the rJava package
class(topic.model)
## [1] "jobjRef"
## attr(,"package")
## [1] "rJava"

13 How to set the number of topics is a matter of serious discussion in the topic modeling literature, and there is no obvious way of knowing in advance exactly where this number should be set. In the documentation for the MALLET program, Mimno writes: The best
number depends on what you are looking for in the model. The default (10) will provide a
broad overview of the contents of the corpus. The number of topics should depend to some
degree on the size of the collection, but 200 to 400 will produce reasonably fine-grained results. Readers interested in more nuanced solutions may wish to consult Chap. 8 of Jockers,
Matthew L. Macroanalysis: Digital Methods and Literary History. University of Illinois Press,
2013, or visit http://www.matthewjockers.net/2013/04/12/secret-recipefor-topic-modeling-themes/ for my Secret Recipe for Topic Modeling Themes.

13.5 Simple Topic Modeling with a Standard Stop List

145

In this case, jobjRef is a reference (or pointer) to a Java object that has been
created and masked behind the scenes. Unless you are willing to dig into the actual
source, that is, leave the world of R and go study the MALLET Java application, then
you will have to accept a bit of obscurity.14
With the trainer object (topic.model) instantiated, you must now fill it with
the textual data. For this you will call the loadDocuments method with the
mallet.instances object that was created a moment ago as an argument.
topic.model$loadDocuments(mallet.instances)

When invoked, some initial processing of the documents occurs as mallet prepares the data for modeling. Among other things, mallet will output to the R
console some information about the number of tokens found in the entire corpus
after stop word removal (total tokens) and about the length of the longest individual
document after stop word removal (max tokens). At this point, if you wish to access
a list of the entire vocabulary of the corpus, you can call the getVocabulary
method to return a character vector containing all the words:
vocabulary <- topic.model$getVocabulary()

You can then inspect this character vector using typical R functions:
class(vocabulary)
## [1] "character"
length(vocabulary)
## [1] 55444
head(vocabulary)
## [1] "summer"
"topsail" "schooner" "slipped"
## [5] "cove"
"trinidad"
vocabulary[1:50]
## [1] "summer"
"topsail"
"schooner"
## [4] "slipped"
"cove"
"trinidad"
## [7] "head"
"dropped"
"anchor"
## [10] "edge"
"kelp"
"fields"
## [13] "fifteen"
"minutes"
"small"
## [16] "boat"
"deposited"
"beach"
## [19] "man"
"armed"
"long"
## [22] "squirrel"
"rifle"
"axe"
## [25] "carrying"
"food"
"clothing"
## [28] "brown"
"canvas"
"pack"
## [31] "watched"
"return"
"weigh"
## [34] "stand"
"sea"
"northwest"
## [37] "trades"
"disappeared" "ken"
## [40] "swung"
"broad"
"powerful"
## [43] "back"
"strode"
"resolutely"
## [46] "timber"
"mouth"
"river"
## [49] "john"
"cardigan"
# etc. . .

14

The MALLET program is not terribly difficult to run outside of R and there are now many
good tutorials available online. A few of these are specifically written with humanities applications of topic modeling in mind. Perhaps the best place to start is with Shawn Graham, Scott Weingart, and Ian Milligans online tutorial titled Getting Started with Topic Modeling and MALLET. See http://programminghistorian.org/lessons/topicmodeling-and-mallet.

146

13 Topic Modeling

At this point, you can also access some basic information about the frequency of
words in the corpus and in the various documents of the corpus using the R mallet
method mallet.word.freqs.
word.freqs <- mallet.word.freqs(topic.model)

Calling this function will return a data frame containing a row for each unique
word type in the corpus. The data frame will have three columns:
1. words
2. term.freq
3. doc.freq
The word types are in the words column; term.freq provides a count of the total
number of tokens of that given word type in the corpus; and, finally, doc.freq
provides a count of the total number of documents that contain that word at least
once. You can look at the first few rows in the data frame using Rs head function:
head(word.freqs)
##
words term.freq doc.freq
## 1
summer
62
49
## 2 topsail
1
1
## 3 schooner
16
10
## 4 slipped
130
120
## 5
cove
7
7
## 6 trinidad
17
9

Calling head reveals that the word type summer occurs 62 times in the corpus in 49 different documents. topsail, on the other hand occurs just once, in one
document.15
With the documents all pre-processed, you are now ready to run the actual training process. Before that, however, you have the opportunity to tweak the optimization hyperparameters! Though this step is not required (if you skip it, the default
values of 200 burn-in iterations and 50 iterations between optimization will be implemented), it is worth knowing that you can control the optimization interval and
the burn-in using the following expression16:
topic.model$setAlphaOptimization(40, 80)

Because hyperparameter optimization is on by default, you can skip this step and
go directly to the training of the model. The key argument that must now be set is
the number of iterations to use in training. This argument determines the number of
sampling iterations. In theory, as you increase the number of iterations the quality
of the model will improve, but model quality is a rather subjective measure based
on human evaluation of the resulting topic word clusters. In my own tests, I have
15 Do not forget that prior to modeling you have chunked each novel from the example corpus into
1000 word segments.
16 The ramifications of resetting these values is beyond the scope of this chapter, but interested
readers may wish to consult Hanna Wallach, David Mimno and Andrew McCallum. Rethinking
LDA: Why Priors Matter. In proceedings of Advances in Neural Information Processing Systems
(NIPS), Vancouver, BC, Canada, 2009.

13.6 Unpacking the Model

147

observed that as one increases the number of iterations, topic quality increases only
to a certain point and then levels off. That is, after you reach a certain number
of iterations, the composition and quality of the resulting topics does not change
much.17 For now, set the number of iterations to 400.
topic.model$train(400)

In your own work, you may wish to experiment with different values and examine
how topic composition changes with different values.
When you run this command, a great deal of output (which I have not shown
here) will be sent to your R console. Every 50 iterations, for example, R will spit
out a set of the seven top words in each topic. Here is a small snippet of that output:
. .
8
9
10
11
12
13
14
. .

.
0.06017
0.21164
0.32898
0.02035
0.10557
0.02116
0.01771
.

school city editor story job news paper

mrs mother woman heart room life eyes
sir mr good replied man make friend
cloth mo vo gilt irish post tale
love heart father jane charles thou papa
don parker farrel pablo mike miguel kay
bryce cardigan shirley colonel pennington timber sequoia

R will also provide probabilistic information about how likely the data are given
the model at as it exists at a specific moment in the process. This figure is represented
as a log-likelihood and appears as INFO: <190> LL/token: -9.3141 in the
output. Although the meaning of the log likelihood number is beyond the scope of
this book, numbers closer to zero generally indicate better fitting models.18

13.6 Unpacking the Model

With the model now run, you can inspect the results and begin to see what is revealed
about the corpus in terms of its thematic content. You can start by exploring the
composition and coherence of the 43 topics you instructed mallet to identify. For
extracting this information from the model, mallet provides two functions that
return R objects: mallet.topic.words and mallet.top.words. Use the
first of these to generate a matrix in which each row is a topic and each column a

17 My anecdotal experience seems consistent with more scientific studies, and interested readers
may wish to consult Griffiths, T. L., & Steyvers, M. (2004). Finding scientific topics. Proceedings
of the National Academy of Science, 101, 52285235.
18 David Mimnos Topic Modeling Bibliography provides a comprehensive list of resources
for those wishing to go beyond this text. See http://www.cs.princeton.edu/~mimno/
topics.html.

148

13 Topic Modeling

unique word type in the corpus. Once run, you can determine the size of the resulting
matrix using dim:
topic.words.m <- mallet.topic.words(topic.model,
smoothed=TRUE,
normalized=TRUE)
dim(topic.words.m)
## [1]
43 55444

The values that appear in the cells of this matrix vary depending upon how you
set the normalized and smoothed arguments. In this example I have set both
normalized and smoothed to TRUE. When normalization is set to TRUE the
values in each topic (row) are converted to percentages that sum to one. This can be
easily checked with the rowSums function:
rowSums(topic.words.m)
## [1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## [24] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

When set to FALSE, the value in any given cell will be an integer representing
the count of the occurrences of that word type that were assigned to a particular
topic (row) during processing.
If you wanted to explore this matrix further, you could use bracketed sub-setting
to access the values, for example:
topic.words.m[1:3, 1:3]
##
[,1]
[,2]
[,3]
## [1,] 1.931496e-06 1.931496e-06 1.931496e-06
## [2,] 1.304767e-06 1.304767e-06 1.304767e-06
## [3,] 6.737231e-07 5.285249e-05 6.737231e-07

These results are not terribly informative because there is no column header to show
which word types are associated with each column of values. You can, however,
retrieve that information from the model and then add the column headers yourself
using the colnames function.
vocabulary <- topic.model$getVocabulary()
colnames(topic.words.m) <- vocabulary
topic.words.m[1:3, 1:3]
##
summer
topsail
schooner
## [1,] 1.931496e-06 1.931496e-06 1.931496e-06
## [2,] 1.304767e-06 1.304767e-06 1.304767e-06
## [3,] 6.737231e-07 5.285249e-05 6.737231e-07

Having set the column values, you can compare the relative weight of specific
word types (as a percentage of each topic). In the next example, I use Rs c function
to create a vector of keywords and then use that vector as a way to select named
columns from the matrix:

13.6 Unpacking the Model

149

keywords <- c("california", "ireland")

topic.words.m[, keywords]
##
california
ireland
## [1,] 1.931496e-06 1.931496e-06
## [2,] 1.213930e-03 3.133921e-03
## [3,] 6.737231e-07 1.572100e-04
## [4,] 6.167907e-07 6.167907e-07
## [5,] 4.183203e-07 1.134356e-03
## [6,] 1.047878e-06 1.047878e-06
## [7,] 3.416922e-07 3.416922e-07
## [8,] 3.400991e-07 3.400991e-07
## [9,] 5.591015e-07 1.689318e-03
## [10,] 4.720872e-03 7.170096e-07
## [11,] 3.482728e-07 3.482728e-07
## [12,] 2.372332e-05 3.516161e-04
## [13,] 3.969341e-07 3.969341e-07
## [14,] 9.542818e-03 5.372452e-03
## [15,] 1.991739e-07 1.991739e-07
## [16,] 4.264159e-07 4.264159e-07
## [17,] 3.811172e-07 3.811172e-07
## [18,] 3.377152e-07 3.377152e-07
## [19,] 3.737371e-07 3.737371e-07
## [20,] 1.755352e-07 1.755352e-07
## [21,] 1.661136e-07 1.661136e-07
## [22,] 7.287042e-07 6.779723e-04
## [23,] 1.912470e-07 1.912470e-07
## [24,] 2.671437e-07 2.671437e-07
## [25,] 2.472256e-07 9.959030e-04
## [26,] 2.002365e-07 2.002365e-07
## [27,] 1.156833e-03 8.292288e-07
## [28,] 1.924327e-07 1.924327e-07
## [29,] 7.164526e-07 7.164526e-07
## [30,] 3.467017e-07 3.467017e-07
## [31,] 1.318087e-06 1.318087e-06
## [32,] 6.406301e-07 6.406301e-07
## [33,] 5.636098e-04 2.020728e-07
## [34,] 4.986438e-07 1.390788e-03
## [35,] 3.454886e-03 8.576503e-07
## [36,] 1.901067e-06 1.901067e-06
## [37,] 4.528100e-07 4.528100e-07
## [38,] 9.729439e-07 1.371521e-02
## [39,] 1.139022e-06 1.139022e-06
## [40,] 3.667936e-07 3.667936e-07
## [41,] 7.580528e-07 7.580528e-07
## [42,] 9.611387e-07 9.611387e-07
## [43,] 3.867103e-07 3.867103e-07

You can calculate which of the topic rows has the highest concentration of these
key terms using Rs rowSums and max functions inside a call to which. Save
that row number in a new variable called imp.row.19
imp.row <- which(rowSums(topic.words.m[, keywords]) ==
max(rowSums(topic.words.m[, keywords])))

Examining these results shows that the topic in row 14 has the highest incidence
of these keywords.20 While exploring the topic.words.m object in this manner
19 Note: if you are copying and executing this code as you read along, your row values and weights
are likely to be different because the topic model employs a process that begins with a random distribution of words across topics. Though the topics you generate from this corpus will be generally
similar, they may not be exactly the same as those that appear in this text.
20 It must be noted here that in the MALLET Java program, topics are indexed starting at zero. Java,
like many programming languages begins indexing with 0. R, however, begins with 1. Were we to

150

13 Topic Modeling

can be fruitful, we are usually less interested in specific words and more interested
in examining the top or most heavily weighted words in each topic.
For this ranked sorting of topic words, mallet offers another function:
mallet.top.words. This function takes three arguments:
1. topic.model
2. word.weights
3. num.top.words
The first of these is the model itself, the second is a row from the matrix of word
weights that you have already created and stored in the topic.words.m object,
and finally a third argument stipulating a user-defined number of top words to
display. Assuming you wish to see the top 10 words from topic 14 (the row number
currently saved in the imp.row variable), you would enter:
mallet.top.words(topic.model, topic.words.m[imp.row,], 10)
##
words
weights
## irish
irish 0.014738103
## men
men 0.014632077
## san
san 0.010991843
## california california 0.009542818
## land
land 0.009048029
## state
state 0.008199819
## mr
mr 0.007881740
## people
people 0.007775714
## years
years 0.007599003
## country
country 0.007280925

The most heavily weighted word in this topic is the word irish. Were you to assign
a label to this topic, you might, after examining all these top words, choose Irish
California or Irish-American West as a general descriptor.21

13.7 Topic Visualization

Looking only at the top ten words in a topic can be a bit misleading. Bear in mind
that each topic in this model consists of values for 55,443 word types! Generally you
will want to examine more than just the top ten words when making a decision about
how to label/interpret the topical or thematic essence of a topic. It can, therefore,
be useful to visualize a larger number of the top words in the topic using a word
cloud visualization. Thanks to Ian Fellows, R has a package for generating word
cloud images from exactly the type of data returned by the mallet.top.words
function.

run this same topic modeling exercise in the Java application, the topics would be labeled with the
numbers 0 - 42. In R they are 1- 43.
21 For those who may not have intuited as much, the corpus of texts used in this book is composed
of novels written entirely by Irish and Irish-American authors.

13.7 Topic Visualization

151

territory
numerous
jose irishman gentleman americans

descent chiefs santa celtic found

gold settled ancient names
great represented

countrymen son irishmen o'brien born

leading century english american young
north clare francisco
james coast
william late
dr st language
dublin large
father day
king origin
city
states street
date
mining thomas
called centuries

race

californiasan

irish
ireland

men
governor
' sic esq america
peter
birth state county
mr john historyclara

erin

senator
educated early
family nativeua york kings
represents
distinguished united famous
work
education murphy successful
national cork
lords fine
martin
written celebrated account
connaught book present
branch success

Fig. 13.1 Word cloud of topic 14

Begin by installing and then loading the wordcloud package22:

library(wordcloud)
## Loading required package: Rcpp
## Loading required package: RColorBrewer

Now employ the mallet.top.words function again to grab 100 of the top
words and their associated weights from the model. Instead of simply printing the
results to the R console, save the output into a new variable called topic.top.words.
topic.top.words <- mallet.top.words(topic.model,
topic.words.m[imp.row,], 100)

You can now call the wordcloud function providing a vector of words and
a vector of word weights from the topic.top.words object as the first two
arguments. To these I have added three more arguments that control the aesthetic
look of the final word cloud (Fig. 13.1).23
wordcloud(topic.top.words$words,
topic.top.words$weights,
c(4,.8), rot.per=0, random.order=F)

22

See Chap. 10, Sect 10.4 for package installation instructions.

To see how to control the look of the visualization, consult the help documentation for the
wordcloud function using ?wordcloud.

23

152

13 Topic Modeling

13.8 Topic Coherence and Topic Probability

Because I am familiar with this corpus, I knew that choosing the words california
and irish would prove useful in identifying a topic that deals with the Irish presence
in California and San Francisco, a topic found prominently in several books in this
corpus. Often, however, you will be dealing with larger corpora, and you will have to
inspect the makeup of each topic in order to determine if the topics are coherent. You
need to inspect them to see if they are topical or thematic in nature. If you complete
Exercise 13.1 right now, you will be able to examine 43 different word clouds.
During that inspection, you will inevitably notice a high number of character names
in many of the topics. Depending on your research goals, this presence of character
names could be a significant problem.
Lets assume that you are hoping to track thematic change throughout a corpus.
If that is the case, then the presence of character names is going to skew your results
rather dramatically. There is a topic, for example, where the words nell, tim, and
sheila are prominent. Without even doing the calculations, I can tell you that this
topic is going to be dominant in one book in the corpus (Josephine Donovans novel
Black Soil). This same topic will be comparatively absent from the other novels. I
can also predict, based on my knowledge of the corpus, that there will be another
topic featuring gerald from a book titled Gerald Ffrenchs Friends and still another
topic with john and big and flurry. These will be dominant in the book Kansas Irish
where Big Flurry is the nickname of the main character, Florence Driscoll.
As noted previously, I intentionally picked 43 topics in order to highlight this
problem (as you will recall, there are 43 books in the corpus). Even if I had not
rigged the system, we would have had a way of exploring the extent to which certain topics were more probable or present in certain documents. mallet provides
a function (mallet.doc.topics) for inspecting the probability of each topic
appearing in each document. Or in more simple terms, mallet provides a function
for assessing the proportion of a document that is about each topic.
doc.topics.m <- mallet.doc.topics(topic.model,
smoothed=T,
normalized=T)

Calling this function returns a matrix object in which each column is a topic and
each row is a document from the corpus. The values in the cells of the matrix are the
corresponding probabilities of a given topic (column) in a given document (row).
When the normalized argument is set to TRUE (as it is here) then the values
in each row will sum to one. In other words, summing the 43 topic probability
measurements for each document will return 1. This makes it easy to think about the
values as percentages or proportions of the document. In topic modeling, we assume
that documents are composed of topics in different proportions. In truth, though,
that is a bit of an oversimplification because this is a closed system and the model
is only able to assign proportions for the 43 topics in this particular configuration.
So, what we are really assuming is that documents are composed of these 43 topics
in differing proportions. It would not be entirely fair to say that the book Gerald

13.8 Topic Coherence and Topic Probability

153

Ffrenchs Friends is 36 % about topic 8 (even though 0.36 is the mean proportion of
this topic across all the segments from this book). A slightly better way to express
this proportion might be to explicitly say: of the 43 topics in this particular model
that could be assigned to Gerald Ffrenchs Friends, topic 8 is assigned with the
highest probability, a probability of .36.
Let us now write some code to explore the proportions of each topic in each
document and see if there are documents in the corpus that are dominated by specific
topics. If you discover that a particular topic is more or less unique to one particular
text, then you might have grounds to suspect a problem. Of course it is perfectly
reasonable to imagine a situation in which there is one outlier book in the corpus,
perhaps one book about vampires in a corpus of books about faeries. Here, however,
you will find that there is a problem and that it is most certainly associated with
character names.
Recall that every book in this corpus was split into segments before modeling.
You now want to look at the books as a whole again and calculate the mean topical
values across the segments as a way of assessing the general saturation of topics in
books. You begin with the doc.topics.m object, a matrix of dimension 3523,
43. You know that these 3523 rows correspond to the segments from all of the
novels, and you still have a data frame object called documents instantiated in the
workspace where these documents ids are stored. You can put those values into a
new vector called file.ids.v.
file.ids.v <- documents[,1]
head(file.ids.v)
## [1] "anonymous_1" "anonymous_2" "anonymous_3"
## [4] "anonymous_4" "anonymous_5" "anonymous_6"

Now you must massage the names in this vector so that the chunk identifier is split
off into a separate vector from the main file name. You can use the strsplit
function to break these character strings on the underscore character and return a
list object. You can then use lappy and do.call, as you have done before, to
convert these values into a two column matrix.
file.id.l <- strsplit(file.ids.v, "_")
file.chunk.id.l <- lapply(file.id.l, rbind)
file.chunk.id.m <- do.call(rbind, file.chunk.id.l)
head(file.chunk.id.m)
##
[,1]
[,2]
## [1,] "anonymous" "1"
## [2,] "anonymous" "2"
## [3,] "anonymous" "3"
## [4,] "anonymous" "4"
## [5,] "anonymous" "5"
## [6,] "anonymous" "6"

The first column provides a way of identifying which rows in the doc.
topics.m object correspond to which text files. With that information, you can
then use Rs aggregate function to calculate the topical mean for each topic in
each document. First save a copy of doc.topics.m as a data frame because you
will need an object that allows both character data and numerical values.

13 Topic Modeling

0.00

0.05

0.10

0.15

0.20

0.25

154

14

19

24

29

34

39

Fig. 13.2 Bar plot of topic means in 43 documents

doc.topics.df <- as.data.frame(doc.topics.m)

Now use cbind to bind the character data values in the first column of
file.chunk.id.m to the topical values in doc.topics.df:
doc.topics.df <- cbind(file.chunk.id.m[,1], doc.topics.df)

Rs aggregate function can then be used to calculate the mean across the segments of each document.
doc.topic.means.df <- aggregate(doc.topics.df[, 2:ncol(doc.topics.df)],
list(doc.topics.df[,1]),
mean)

The aggregate function returns a new data frame of 43 rows by 44 columns.

There is now one row for each text, a column (the first) with the file name, or what
the aggregate function titles a group and then 43 more columns, one for each of the
topics you modeled. With this data in one place, you have several options for how
to assess the mean values. Since you only have 43 documents in the corpus, you
can visualize the document means using a simple bar plot (Fig. 13.2).
barplot(doc.topic.means.df[, "V6"], names.arg=c(1:43))

A couple of things to note here: first, notice that the aggregate function renamed
the columns by prefixing each with a V. Thus when you call barplot here, you
use bracketed sub-setting to send data for all the rows and only the column labeled
V6.24 More important, notice the one outlier document, document 25. This is the
24

Your plots will not look the same since each run of the model is slightly different.

13.8 Topic Coherence and Topic Probability

155

announced
ancient
declared
artelan
nodded war
miss presently
thousand
dominic county
automobile
loustalot
mike's horse
realized father's dam bet

seat white

japanese
san girl

conway

replied

pablo's
national
mrs land porch

se

pablo palomar
mike don father

hee's
thees back

saddle
queried
japs nicol

john
valley

mother la
marcos
power

gregorio
sheep
potato jap

started

assured patio

afternoon

baron
check

toro

parker el
farrel kay

bill

miguel

mortgage

don'

ranch

panchito

race

parker's

car

riding
son brown

farrel's

okada
rancho
mission

california

south brother

informed
murmured gate
demanded carolina smile
hacienda
spanish suggested bank
continued
murray
glance
judgment
caliente anthony

Fig. 13.3 Word cloud of main characters in The Pride of Palomar

document in row 25 of the data frame. You can get its file name by retrieving the
value held in the group.1 column of row 25, like this:
filename <- as.character(doc.topic.means.df[25, "Group.1"])

filename
## [1] "Kyne1"

If you are following along, then you should be asking why I added the
as.character function. Try it for yourself without the call to as.character
and you will discover that the column called Group.1 is of the class factor.25
So while the expression will return the same answer, it will also return information
about the factor levels that you do not need.
The barplot reveals that there is one topic dominating the file called Kyne1, and,
not surprisingly, the word cloud (Fig. 13.3) for that topic indicates that it is clearly a
character-driven topic. The top words in this topic include the names parker, farrel,
don, pablo, mike, miguel, kay, and panchito.
A simple way to deal with the character name problem is to add these names to
the stop list. You will have the opportunity to do this and to compare your results in
Exercise 13.2. A more complicated way of dealing with these and related problems
of topic coherence involves pre-processing the corpus with a part-of-speech (POS)
tagger and then culling out words of different classes.
25 A full discussion of factors is beyond the scope of this chapter, but for simplicity think about
factors as a type of variable that can hold some limited set of values. These are often referred to as
categorical variables. See also: Chap. 12 Sect. 12.9, footnote 4.

156

13 Topic Modeling

13.9 Pre-processing with a POS Tagger

In my research for Chap. 8 of Macroanalysis, I discovered that by modeling the
nouns in my documents, I could generate what I considered to be highly coherent
and highly thematic topics.26 In order to do this kind of modeling, I had to first
pre-process the corpus with a POS tagger and identify which words were nouns and
which were verbs, adjectives, etc. In that work, I used the Stanford Log-linear Partof-Speech Tagger outside of the R environment and then post-processed the results
in R. Now, however, there is an R package for POS tagging and the entire process can
be done in one place. Having said that, POS tagging is a time-consuming process,
so for this book I have provided another directory (taggedCorpus) containing
pre-tagged versions of all of the files. Since you may want to tag your own files,
here is the code you would need:
library(openNLP)
library(NLP)
for(i in 1:length(files.v)){
doc.object <- xmlTreeParse(file.path(inputDir, files.v[i]),
useInternalNodes=TRUE)
paras <- getNodeSet(doc.object,
"/d:TEI/d:text/d:body//d:p",
c(d = "http://www.tei-c.org/ns/1.0"))
words <- as.String(paste(sapply(paras,xmlValue),
collapse=" "))
# Need sentence and word tokens first
sent_token_annotator <- Maxent_Sent_Token_Annotator()
word_token_annotator <- Maxent_Word_Token_Annotator()
a2 <- annotate(words, list(sent_token_annotator,
word_token_annotator))
# now pos tags
pos_tag_annotator <- Maxent_POS_Tag_Annotator()
a3 <- annotate(words, pos_tag_annotator, a2)
a3w <- subset(a3, type == "word")
tags <- sapply(a3w$features, `[[`, "POS")
tagged_text <- paste(sprintf("%s/%s", words[a3w], tags),
collapse=" ")
write(tagged_text, paste("data/taggedCorpus/",
files.v[i], ".txt", sep=""))
}

This code will iteratively load each XML file in the corpus, run it through the
openNLP POS tagger, and then write the result to the directory called
taggedCorpus. You do not need to run this code now since I have already tagged
the files and put them into the directory for you.27
With the files tagged, you can now write a modified version of the text chunking
script that you used for the untagged files. In this new script, you will add a few lines
of R code to remove all the words that are not nouns prior to segmentation. Begin
just as you did with the XML files, but this time reference the taggedCorpus
directory. You should decrease the chunk.size to 500 bearing in mind that in

26

All 500 of them can be viewed at http://www.matthewjockers.net/

macroanalysisbook/macro-themes/.
27 Note that both the NLP and openNLP packages must be installed first.

13.9 Pre-processing with a POS Tagger

157

this new process you will be segmenting based on nouns only, thus each segment
will have 500 nouns.
inputDir <- "data/taggedCorpus"
files.v <- dir(path=inputDir, pattern=".*xml")
chunk.size <- 500

Before you get into the main script, youll need to write a couple of new
functions for handling the POS tagged text. If you open one of the files in the
taggedCorpus directory, you will see that each word token is followed by a
forward slash and then an abbreviated POS marker.28 Here, for example, is the first
line of the file titled anonymous.xml.txt:
In/IN the/DT summer/NN of/IN 1850/CD a/DT topsail/NN schooner/NN slipped/VBD
into/IN the/DT cove/NN under/IN Trinidad/NNP Head/NNP and/CC dropped/VBD anchor/NN at/IN the/DT edge/NN of/IN the/DT kelp-fields./NN.

So begin by writing a function to split the tagged file into a vector in which each
value is a single word/POS pair:
splitText <- function(text) {
unlist(strsplit(text," "))
}

The elements of this function should be familiar to you: it uses strsplit and
unlist and returns the result.29
You must now write another function that will walk through this vector and pick
out only those values that contain certain target POS markers. Name this function
selectTaggedWords and set it up to take two arguments: a vector of word/POS
pairs and a target.tag to search for in the vector.
selectTaggedWords <- function(tagged.words, target.tag) {
tagged.words[grep(target.tag, tagged.words)]
}

In this function, I have embedded a call to grep inside the brackets of the
tagged word vector. When grep finds a match for the target.tag inside the
tagged.words vector it returns TRUE. The output from grep will be a vector
of TRUE/FALSE values that then gets used to mark the location of the matching
strings. In this way, grep is similar to which, returning only those positions that
meet a certain condition.
You now need a function to strip off the POS markers. You already know that
each word token is followed by a forward slash and then a POS marker, so you can
identify this pattern using a regular expression. In order to write the right regex,
however, you also need to know a bit about the variety of POS tags that are used
28 The openNLP tagger implements the Penn Treebank tag set. See http://www.ling.
upenn.edu/courses/Fall_2003/ling001/penn_treebank_pos.html.
29 In previous functions you have written, you have explicitly called the return function to send
the results of the function back to the main script. This is not always necessary as Rs default
behavior is to return the last object generated by the function. This function is simple enough that
I have chosen to leave off an explicit call to return.

158

13 Topic Modeling

by the tagger. As it happens, these tags are always composed of two or three capital
letters. So you can write a regular expression to find a pattern that begins with a
forward slash and is followed by two or three capital letters. Such an expression
looks like this: /[A-Z]{2,3}. You can use Rs sub function (a kin of grep)
to remove matches found for this pattern by replacing (subbing) them with nothing.
The whole function, therefore, is as simple as this:
removeTags <- function(word.pos) {
sub("/[A-Z]{2,3}", "", word.pos)
}

You can now wrap calls to all of these functions inside a larger function that
will be almost identical to the makeFlexTextChunks function written at the
beginning of this chapter.
makeFlexTextChunksFromTagged <- function(tagged.text,
chunk.size=500, percentage=TRUE){
tagged.words <- splitText(tagged.text)
tagged.words.keep <- c(selectTaggedWords(tagged.words,"/NN$"))
words <- removeTags(tagged.words.keep)
words.lower <- tolower(words)
word.v <- gsub("[^[:alnum:][:space:]']", "", words.lower)
x <- seq_along(word.v)
if(percentage){
max.length <- length(word.v)/chunk.size
chunks.l <- split(word.v, ceiling(x/max.length))
} else {
chunks.l <- split(word.v, ceiling(x/chunk.size))
if(length(chunks.l[[length(chunks.l)]]) <=
length(chunks.l[[length(chunks.l)]])/2){
chunks.l[[length(chunks.l)-1]] <c(chunks.l[[length(chunks.l)-1]], chunks.l[[length(chunks.l)]])
chunks.l[[length(chunks.l)]] <- NULL
}
}
chunks.l <- lapply(chunks.l, paste, collapse=" ")
chunks.df <- do.call(rbind, chunks.l)
return(chunks.df)
}

Save this new function into your corpusFunctions.R file as

makeFlexTextChunksFromTagged along with the others you have just written and then load them into your active workspace using a call to source:
source("code/corpusFunctions.R")

With all of the new functions loaded, you run a loop that is almost identical to
the one you ran earlier in the chapter.
topic.m <- NULL
for(i in 1:length(files.v)){
tagged.text <- scan(file.path(inputDir, files.v[i]),
what="character", sep="\n")
chunk.m <- makeFlexTextChunksFromTagged(tagged.text,
chunk.size, percentage=FALSE)
textname <- gsub("\\..*","", files.v[i])
segments.m <- cbind(paste(textname,
segment=1:nrow(chunk.m), sep="_"),
chunk.m)
topic.m <- rbind(topic.m, segments.m)
}

13.9 Pre-processing with a POS Tagger

159

Finally, with the chunk matrix now prepared, you can run the topic modeling code
again on the refined data:
documents <- as.data.frame(topic.m, stringsAsFactors=F)
colnames(documents) <- c("id", "text")
library(mallet)
mallet.instances <- mallet.import(documents$id,
documents$text,
"data/stoplist.csv",
FALSE,
token.regexp="[\\p{L}']+")
topic.model <- MalletLDA(num.topics=43)
topic.model$loadDocuments(mallet.instances)
vocabulary <- topic.model$getVocabulary()
word.freqs <- mallet.word.freqs(topic.model)
topic.model$train(400)
topic.words.m <- mallet.topic.words(topic.model,
smoothed=TRUE,
normalized=TRUE)
colnames(topic.words.m) <- vocabulary

Based on what you have learned here and what you will learn by completing the
exercises, you should now be able to inspect these new topics and visualize them as
word clouds. When you do so, you will see that they have improved considerably.30

Practice
13.1. Write a script that uses a for loop to iterate over all of the topics data in order
to produce a word cloud for each.
13.2. In the data directory you will find a fairly comprehensive stop list of common
names and high frequency words: stoplist-exp.csv. Replace the reference to
stoplist.csv in the example code of this chapter with stoplist-exp.csv
and generate a new model with a new set of topics and document proportions. Plot
the means as you did in this chapter, and then assess the extent to which these
new topics are distributed across the corpus. Make word clouds for each topic and
consider how the new ones compare to those that included character names.

30

In the topicClouds sub-directory of the data directory, you will find two .pdf files showing 43 word clouds each. The file titled fromUnTagged.pdf contains the clouds produced without POS-based pre-processing, fromTagged.pdf includes the clouds produced only using words
tagged as nouns.

Appendix A

Variable Scope Example

This is an example of scope within functions. First create a variable outside of a

function.
my.var.to.process <- 10

Now create a function that uses the same name for an argument as an existing
variable.
my.func <- function(my.var.to.process){
# overwrite the value in my.var.to.process
# with a new value that adds ten
my.var.to.process <- my.var.to.process + 10 # add ten
# return the new value
return(my.var.to.process)
}

The value returned by calling my.func is 20.

my.func(my.var.to.process)
## [1] 20

But the value in the original variable is still 10 even though the same name was
used inside the function.
my.var.to.process
## [1] 10

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4,
Springer International Publishing Switzerland 2014

161

Appendix B

The LDA Buffet

A version of what follows was originally posted to

http://www.matthewjockers.net/macroanalysisbook/lda/
August 12, 2012.

on

. . . imagine a quaint town, somewhere in New England perhaps. The town is a writers
retreat, a place they come in the summer months to seek inspiration. Melville is there,
Hemingway, Joyce, and Jane Austen just fresh from across the pond. In this mythical town
there is spot popular among the inhabitants; it is a little place called the LDA Buffet.
Sooner or later all the writers go there to find themes for their novels. . . .
One afternoon Herman Melville bumps into Jane Austen at the bocce ball court, and
they get to talking.
You know, says Austen, I have not written a thing in weeks.
Arrrrgh, Melville replies, me neither.
So hand in hand they stroll down Gibbs Lane to the LDA Buffet. Now, down at the
LDA Buffet no one gets fat. The buffet only serves light (leit?) motifs, themes, topics, and
tropes (seasonal). Melville hands a plate to Austen, grabs another for himself, and they
begin walking down the buffet line. Austen is finicky; she spoons a dainty helping of words
out of the bucket marked dancing. A slightly larger spoonful of words, she takes from the
gossip bucket and then a good ladles worth of courtship.
Melville makes a bee line for the whaling trough, and after piling on an Ahabsized handful of whaling words, he takes a smaller spoonful of seafaring and then just
a smidgen of cetological jargon.
The two companions find a table where they sit and begin putting all the words from
their plates into sentences, paragraphs, and chapters.
At one point, Austen interrupts this business: Oh Herman, you must try a bit of this
courtship.
He takes a couple of words but is not really fond of the topic. Then Austen, to her credit,
asks permission before reaching across the table and sticking her fork in Melvilles pile of
seafaring words, just a taste, she says. This work goes on for a little while; they order a
few drinks and after a few hours, voila! Moby Dick and Persuasion are written. . . .
[Now, dear reader, our story thus far provides an approximation of the first assumption
made in LDA. We assume that documents are constructed out of some finite set of available
topics. It is in the next part that things become a little complicated, but fear not, for you
shall sample themes both grand and beautiful.]

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4,
Springer International Publishing Switzerland 2014

163

164

B The LDA Buffet

. . . Filled with a sense of deep satisfaction, the two begin walking back to the lodging
house. Along the way, they bump into a blurry-eyed Hemingway, who is just then stumbling
out of the Rising Sun Saloon.
Having taken on a bit too much cargo, Hemingway stops on the sidewalk in front of
the two literati. Holding out a shaky pointer finger, and then feigning an English accent,
Hemingway says: Stand and Deliver!
To this, Austen replies, Oh come now, Mr. Hemingway, must we do this every season?
More gentlemanly then, Hemingway replies, My dear Jane, isnt it pretty to think so.
Now if you could please be so kind as to tell me whats in the offing down at the LDA
Buffet.
Austen turns to Melville and the two writers frown at each other. Hemingway was recently banned from the LDA Buffet. Then Austen turns toward Hemingway and holds up
six fingers, the sixth in front of her now pursed lips.
Six topics! Hemingway says with surprise, but what are todays themes?
Now wouldnt you like to know that you old sot. Says Melville.
The thousand injuries of Melville, Hemingway had borne as best he could, but when
Melville ventured upon insult he vowed revenge. Grabbing their recently
completed manuscripts, Hemingway turned and ran toward the South. Just before disappearing down an alleyway, he calls back to the dumbfounded writers: All my life Ive looked
at words as though I were seeing them for the first time. . . tonight I will do so again!. . .
[Hemingway has thus overcome the first challenge of topic modeling. He has a corpus
and a set number of topics to extract from it. In reality determining the number of topics to
extract from a corpus is a bit trickier. If only we could ask the authors, as Hemingway has
done here, things would be so much easier.]
. . . Armed with the manuscripts and the knowledge that there were six topics on the
buffet, Hemingway goes to work.
After making backup copies of the manuscripts, he then pours all the words from the
originals into a giant Italian-leather attache. He shakes the bag vigorously and then begins
dividing its contents into six smaller ceramic bowls, one for each topic. When each of the
six bowls is full, Hemingway gets a first glimpse of the topics that the authors might have
found at the LDA Buffet. Regrettably, these topics are not very good at all; in fact, they
are terrible, a jumble of random unrelated words. . . .
[And now for the magic that is Gibbs Sampling.]
. . . Hemingway knows that the two manuscripts were written based on some mixture of
topics available at the LDA Buffet. So to improve on this random assignment of words to
topic bowls, he goes through the copied manuscripts that he kept as back ups. One at a time,
he picks a manuscript and pulls out a word. He examines the word in the context of the
other words that are distributed throughout each of the six bowls and in the context of the
manuscript from which it was taken. The first word he selects is heaven, and at this word
he pauses, and asks himself two questions:
How much of Topic A, as it is presently represented in bowl A, is present in the
current document? Which topic, of all of the topics, has the most heaven in it?. . .
[Here again dear reader, you must take with me a small leap of faith and engage in a bit
of further make believe. There are some occult statistics here accessible only to the initiated.
Nevertheless, the assumptions of Hemingway and of the topic model are not so far-fetched
or hard to understand. A writer goes to his or her imaginary buffet of themes and pulls them
out in different proportions. The writer then blends these themes together into a work of art.
That we might now be able to discover the original themes by reading the book is not at
all amazing. In fact we do it all the timeevery time we say that such and such a book is
about whaling or courtship. The manner in which the computer (or dear Hemingway)

B The LDA Buffet

165

does this is perhaps less elegant and involves a good degree of mathematical magic. Like all
magic tricks, however, the explanation for the surprise at the end is actually quite simple: in
this case our magician simply repeats the process 10 billion times!]
. . . As Hemingway examines each word in its turn, he decides based on the calculated
probabilities whether that word would be more appropriately moved into one of the other
topic bowls. So, if he were examining the word whale at a particular moment, he would
assume that all of the words in the six bowls except for whale were correctly distributed.
Hed now consider the words in each of those bowls and in the original manuscripts, and he
would choose to move a certain number of occurrences of whale to one bowl or another.
Fortunately, Hemingway has by now bumped into James Joyce who arrives bearing a
cup of coffee on which a spoon and napkin lay crossed. Joyce, no stranger to bags-of-words,
asks with compassion: Is this going to be a long night.
Yes, Hemingway said, yes it will, yes.
Hemingway must now run through this whole process over and over again many times.
Ultimately, his topic bowls reach a steady state where words are no longer needing to be
being reassigned to other bowls; the words have found their proper context.
After pausing for a well-deserved smoke, Hemingway dumps out the contents of the
first bowl and finds that it contains the following words:
whale sea men ship whales penfon air side life bounty night oil natives shark seas beard
sailors hands harpoon mast top feet arms teeth length voyage eye heart leviathan islanders
flask soul ships fishery sailor sharks company. . . .
He peers into another bowl that looks more like this:
marriage happiness daughter union fortune heart wife consent affection wishes life
attachment lover family promise choice proposal hopes duty alliance affections feelings
engagement conduct sacrifice passion parents bride misery reason fate letter mind resolution
rank suit event object time wealth ceremony opposition age refusal result determination
proposals. . . .
After consulting the contents of each bowl, Hemingway immediately knows what topics
were on the menu at the LDA Buffet. And, not only this, Hemingway knows exactly what
Melville and Austen selected from the Buffet and in what quantities. He discovers that Moby
Dick is composed of 40 % whaling, 18 % seafaring and 2 % gossip (from that little taste he
got from Jane) and so on. . . .
[Thus ends the fable.]

My thanks to David Mimno for reviewing this little story. See Mimnos Topic
Modeling Bibliography at http://www.cs.princeton.edu/mimno/
topics.html for a less fictional perspective.

Appendix C

Start up Code

All of the code in this appendix can be found in the start.up.code directory of the
downloadable supporting materials for this book. See http://www.
matthewjockers.net/text-analysis-with-r-for-studentsof-literature

C.1 Chapter 3
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")
start.metadata.v <- text.v[1:start.v -1]
end.metadata.v <- text.v[(end.v+1):length(text.v)]
metadata.v <- c(start.metadata.v, end.metadata.v)
novel.lines.v <- text.v[start.v:end.v]
novel.v <- paste(novel.lines.v, collapse=" ")
novel.lower.v <- tolower(novel.v)
moby.words.l <- strsplit(novel.lower.v, "\\W")
moby.word.v <- unlist(moby.words.l)
not.blanks.v <- which(moby.word.v!="")
moby.word.v <- moby.word.v[not.blanks.v]
whale.hits.v <- length(moby.word.v[which(moby.word.v=="whale")])
total.words.v <- length(moby.word.v)
moby.freqs.t <- table(moby.word.v)
sorted.moby.freqs.t <- sort(moby.freqs.t , decreasing=T)

C.2 Chapter 4
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")
novel.lines.v <- text.v[start.v:end.v]
novel.v <- paste(novel.lines.v, collapse=" ")
novel.lower.v <- tolower(novel.v)
moby.words.l <- strsplit(novel.lower.v, "\\W")
moby.word.v <- unlist(moby.words.l)
not.blanks.v <- which(moby.word.v!="")
moby.word.v <- moby.word.v[not.blanks.v]

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4,
Springer International Publishing Switzerland 2014

167

168

C Start up Code

C.3 Chapter 5
# Chapter 4 Start up code
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")
novel.lines.v <- text.v[start.v:end.v]
novel.lines.v <- unlist(novel.lines.v)
chap.positions.v <- grep("^CHAPTER \\d", novel.lines.v)
last.position.v <- length(novel.lines.v)
chap.positions.v <- c(chap.positions.v , last.position.v)
chapter.freqs.l <- list()
chapter.raws.l <- list()
for(i in 1:length(chap.positions.v)){
if(i != length(chap.positions.v)){
chapter.title <- novel.lines.v[chap.positions.v[i]]
start <- chap.positions.v[i]+1
end <- chap.positions.v[i+1]-1
chapter.lines.v <- novel.lines.v[start:end]
chapter.words.v <- tolower(paste(chapter.lines.v, collapse=" "))
chapter.words.l <- strsplit(chapter.words.v, "\\W")
chapter.word.v <- unlist(chapter.words.l)
chapter.word.v <- chapter.word.v[which(chapter.word.v!="")]
chapter.freqs.t <- table(chapter.word.v)
chapter.raws.l[[chapter.title]] <- chapter.freqs.t
chapter.freqs.t.rel <- 100*(chapter.freqs.t/sum(chapter.freqs.t))
chapter.freqs.l[[chapter.title]] <- chapter.freqs.t.rel
}
}
whale.l <- lapply(chapter.freqs.l, '[', 'whale')
whales.m <- do.call(rbind, whale.l)
ahab.l <- lapply(chapter.freqs.l, '[', 'ahab')
ahabs.m <- do.call(rbind, ahab.l)
whales.v <- as.vector(whales.m[,1])
ahabs.v <- as.vector(ahabs.m[,1])
whales.ahabs.m <- cbind(whales.v, ahabs.v)
colnames(whales.ahabs.m) <- c("whale", "ahab")

C.4 Chapter 6
# Chapter 6 Start up code
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")
novel.lines.v <- text.v[start.v:end.v]
novel.lines.v <- unlist(novel.lines.v)
chap.positions.v <- grep("^CHAPTER \\d", novel.lines.v)
last.position.v <- length(novel.lines.v)
chap.positions.v <- c(chap.positions.v , last.position.v)
chapter.freqs.l <- list()
chapter.raws.l <- list()
for(i in 1:length(chap.positions.v)){
if(i != length(chap.positions.v)){
chapter.title <- novel.lines.v[chap.positions.v[i]]
start <- chap.positions.v[i]+1
end <- chap.positions.v[i+1]-1

C.5 Chapter 7
chapter.lines.v <- novel.lines.v[start:end]
chapter.words.v <- tolower(paste(chapter.lines.v, collapse=" "))
chapter.words.l <- strsplit(chapter.words.v, "\\W")
chapter.word.v <- unlist(chapter.words.l)
chapter.word.v <- chapter.word.v[which(chapter.word.v!="")]
chapter.freqs.t <- table(chapter.word.v)
chapter.raws.l[[chapter.title]] <- chapter.freqs.t
chapter.freqs.t.rel <- 100*(chapter.freqs.t/sum(chapter.freqs.t))
chapter.freqs.l[[chapter.title]] <- chapter.freqs.t.rel
}
}

C.5 Chapter 7
# Chapter 7 Start up code
text.v <- scan("data/plainText/melville.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")
novel.lines.v <- text.v[start.v:end.v]
novel.lines.v <- unlist(novel.lines.v)
chap.positions.v <- grep("^CHAPTER \\d", novel.lines.v)
last.position.v <- length(novel.lines.v)
chap.positions.v <- c(chap.positions.v , last.position.v)
chapter.freqs.l <- list()
chapter.raws.l <- list()
for(i in 1:length(chap.positions.v)){
if(i != length(chap.positions.v)){
chapter.title <- novel.lines.v[chap.positions.v[i]]
start <- chap.positions.v[i]+1
end <- chap.positions.v[i+1]-1
chapter.lines.v <- novel.lines.v[start:end]
chapter.words.v <- tolower(paste(chapter.lines.v, collapse=" "))
chapter.words.l <- strsplit(chapter.words.v, "\\W")
chapter.word.v <- unlist(chapter.words.l)
chapter.word.v <- chapter.word.v[which(chapter.word.v!="")]
chapter.freqs.t <- table(chapter.word.v)
chapter.raws.l[[chapter.title]] <- chapter.freqs.t
chapter.freqs.t.rel <- 100*(chapter.freqs.t/sum(chapter.freqs.t))
chapter.freqs.l[[chapter.title]] <- chapter.freqs.t.rel
}
}
chapter.lengths.m <- do.call(rbind, lapply(chapter.raws.l,sum))

169

Appendix D

R Resources for Further Reading

Books
Baayen, R. Harald (2008). Analyzing linguistic data: a practical introduction to
statistics using R. English. Cambridge, UK; New York: Cambridge University
Press. ISBN: 9780521882590 0521882591 9780521709187 0521709180.
Dalgaard, Peter (2008). Introductory statistics with R. English. New York:
Springer. ISBN: 9780387790534 0387790535 9780387790541 0387790543.
Gries, Stefan Thomas (2009). Quantitative corpus linguistics with R: a practical introduction. English. New York, NY: Routledge. ISBN: 9780415962711
0415962714 9780415962704 0415962706 9780415997805 0415997801
9780203880920 0203880927.
Teetor, Paul (2011). R cookbook. English. Beijing: OReilly. ISBN:
9780596809157 0596809158.
Wickham, Hadley (2009). ggplot2 elegant graphics for data analysis. English. Dordrecht; New York: Springer. ISBN: 9780387981413 0387981411.

Online
Cookbook for R >> Cookbook for R. URL: http://www.cookbook-r.com/
(visited on 12/05/2013).
Fridolin, Wild. CRAN Task View: Natural Language Processing. URL: http://
c ran.r-project.org/web/views/
NaturalLanguageProcessing.html (visited on 12/05/2013).
Newest r Questions - Stack Overflow. URL: http://stackoverflow.com/
questions/tagged/r (visited on 12/05/2013).
Quick-R: Home Page. URL: http://www.statmethods.net/ (visited on
12/05/2013).

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4,
Springer International Publishing Switzerland 2014

171

172

D R Resources for Further Reading

R 1.1 - Initial Setup and Navigation - YouTube. URL: http://www.youtube.

com/watch?v=iffR3fWv4xw (visited on 12/05/2013).
R: Mailing Lists. URL: http://www.r-project.org/mail.html
(visited on 12/05/2013).
R Programming - Wikibooks, open books for an open world. URL: http://en.
wikibooks.org/wiki/R_Programming (visited on 12/05/2013).
Revolutions. URL: http://blog.revolutionanalytics.com/
atom.xml (visited on 12/05/2013).
RSeek.org R-project Search Engine. URL: http://www.rseek.org/
(visited on 12/05/2013).
Rydberg-Cox, Jeff. Statistical Methods for Studying Literature Using R. URL:
http://www.chlt.org/StatisticalMethods/index.html
(visited on 12/05/2013).
Try R. URL: http://tryr.codeschool.com (visited on 12/05/2013).
Videos from Courseras four week course in R. URL: http://blog.
revolutionanalytics.com/2012/12/coursera-videos.html
(visited on 12/05/2013).

Practice Exercise Solutions

Practice Exercises for Chapter 1

1.1
10+5
## [1] 15
10-5
## [1] 5

1.2
10*1576
## [1] 15760

1.3
15760/10
## [1] 1576

1.4
10+pi
## [1] 13.14159
10/pi
## [1] 3.183099

1.5
10^2
## [1] 100

1.6
x <- 10
x - 3
## [1] 7

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4,
Springer International Publishing Switzerland 2014

173

174

1.7
x <- 10
x - 3 + 10 / 2
## [1] 12

1.8
x <- 10

1.9
sqrt(12)
## [1] 3.464102
abs(-23)
## [1] 23
round(pi)
## [1] 3

1.10
1:10
## [1] 1 2 3 4 5 6 7 8 9 10
12:37
## [1] 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
## [16] 27 28 29 30 31 32 33 34 35 36 37

Practice Exercise Solutions

Practice Exercise Solutions

175

Practice Exercises for Chapter 2

2.1 The top ten most frequent words are found in the first through tenth position in
the sorted vector. To see them, just enter:
sorted.moby.freqs.t[1:10]
## moby.word.v
##
the
of
and
a
## 14175 6469 6325 4636
##
his
i
## 2495 2114

to
4539

in
4077

that
3045

it
2497

Visualizing the results is really as simple as using

plot(sorted.moby.freqs.t[1:10])

But adding a few more arguments to the plot function gives you a more informative graph.

10000 14000
6000
2000

Word Count

plot(sorted.moby.freqs.t[1:10], type="b",
xlab="Top Ten Words", ylab="Word Count",xaxt = "n")
axis(1,1:10, labels=names(sorted.moby.freqs.t[1:10]))

the

of

and

to

in

Top Ten Words

that

it

his

176

Practice Exercise Solutions

Practice Exercises for Chapter 3

3.1 First load Moby Dick
text.v <- scan("data/plainText/melville.txt",
what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1. Loomings.")
end.v <- which(text.v == "orphan.")

Now remove the boilerplate

start.metadata.v <- text.v[1:start.v -1]
end.metadata.v <- text.v[(end.v+1):length(text.v)]
metadata.v <- c(start.metadata.v, end.metadata.v)
novel.lines.v <- text.v[start.v:end.v]
novel.v <- paste(novel.lines.v, collapse=" ")
novel.lower.v <- tolower(novel.v)
moby.words.l <- strsplit(novel.lower.v, "\\W")
moby.word.v <- unlist(moby.words.l)
not.blanks.v <- which(moby.word.v!="")
moby.word.v <- moby.word.v[not.blanks.v]
moby.freqs.t <- table(moby.word.v)
sorted.moby.freqs.t <- sort(moby.freqs.t , decreasing=T)

This next line was not in Exercise 2

sorted.moby.rel.freqs.t <- 100*(sorted.moby.freqs.t/sum(sorted.moby.freqs.t))
plot(sorted.moby.rel.freqs.t[1:10],
main="Moby Dick",
type="b",
xlab="Top Ten Words",
ylab="Percentage",
xaxt = "n")
axis(1,1:10, labels=names(sorted.moby.rel.freqs.t[1:10]))

Percentage
3 4 5

Moby Dick

the

of

and

to

in

that

it

his

Top Ten Words

Now do something similar by loading Sense and Sensibility. Since well only
need the final variables
(sorted.moby.rel.freqs.t and sorted.sense.rel.freqs.t), we
can reuse many of the same variable names.

Practice Exercise Solutions

177

text.v <- scan("data/plainText/austen.txt",

what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1")
end.v <- which(text.v == "THE END")
novel.lines.v <- text.v[start.v:end.v]
novel.v <- paste(novel.lines.v, collapse=" ")
novel.lower.v <- tolower(novel.v)
sense.words.l <- strsplit(novel.lower.v, "\\W")
sense.word.v <- unlist(sense.words.l)
not.blanks.v <- which(sense.word.v!="")
sense.word.v <- sense.word.v[not.blanks.v]
sense.freqs.t <- table(sense.word.v)
sorted.sense.freqs.t <- sort(sense.freqs.t , decreasing=T)
sorted.sense.rel.freqs.t <- 100*
(sorted.sense.freqs.t/sum(sorted.sense.freqs.t))
plot(sorted.sense.rel.freqs.t[1:10],
main="Sense and Sensibility", type="b",
xlab="Top Ten Words", ylab="Percentage",xaxt = "n")
axis(1,1:10, labels=names(sorted.sense.rel.freqs.t[1:10]))

2.5
2.0
1.5

Percentage

3.0

Sense and Sensibility

to

the

of

and

her

in

was

Top Ten Words

3.2
unique(c(names(sorted.moby.rel.freqs.t[1:10]),
names(sorted.sense.rel.freqs.t[1:10])))
## [1] "the" "of"
"and" "a"
"to"
"in"
## [7] "that" "it"
"his" "i"
"her" "was"

3.3
names(sorted.sense.rel.freqs.t[
which(names(sorted.sense.rel.freqs.t[1:10])
%in% names(sorted.moby.rel.freqs.t[1:10]))])
## [1] "to" "the" "of" "and" "a"
"i"
"in"
## [8] "it"

3.4
presentSense <- which(names(sorted.sense.rel.freqs.t[1:10])
%in% names(sorted.moby.rel.freqs.t[1:10]))
names(sorted.sense.rel.freqs.t[1:10])[-presentSense]
## [1] "her" "was"
presentMoby <- which(names(sorted.moby.rel.freqs.t[1:10])
%in% names(sorted.sense.rel.freqs.t[1:10]))
names(sorted.moby.rel.freqs.t[1:10])[-presentMoby]
## [1] "that" "his"

it

178

Practice Exercise Solutions

Practice Exercises for Chapter 4

4.1
whales.l <- lapply(chapter.freqs.l, '[', 'whale')
whales.m <- do.call(rbind, whales.l)
whales.v <- as.vector(whales.m[,1])
ahabs.l <- lapply(chapter.freqs.l, '[', 'ahab')
ahabs.m <- do.call(rbind, ahabs.l)
ahabs.v <- as.vector(ahabs.m[,1])
queequeg.l <- lapply(chapter.freqs.l, '[', 'queequeg')
queequeg.m <- do.call(rbind, queequeg.l)
queequeg.v <- as.vector(queequeg.m[,1])

0.0

0.5

1.0

1.5

2.0

whales.ahabs.queequeg.m <- cbind(whales.v, ahabs.v,queequeg.v)

barplot(whales.ahabs.queequeg.m, beside=T, col="grey")

whales.v

ahabs.v

queequeg.v

4.2
whale.raw.l <- lapply(chapter.raws.l, '[', 'whale')
whale.raw.m <- do.call(rbind, whale.raw.l)
whale.raw.v <- as.vector(whale.raw.m[,1])
ahab.raw.l <- lapply(chapter.raws.l, '[', 'ahab')
ahab.raw.m <- do.call(rbind, ahab.raw.l)
ahab.raw.v <- as.vector(ahab.raw.m[,1])
whales.ahabs.raw.m <- cbind(whale.raw.v, ahab.raw.v)
barplot(whales.ahabs.raw.m, beside=T, col="grey")

179

20

40

60

80

100

Practice Exercise Solutions

whale.raw.v

ahab.raw.v

180

Practice Exercise Solutions

Practice Exercises for Chapter 5

5.1
my.l <- lapply(chapter.freqs.l, "[", "my")
my.m <- do.call(rbind, my.l)
my.v <- as.vector(my.m[,1])
i.l <- lapply(chapter.freqs.l, "[", "i")
i.m <- do.call(rbind, i.l)
i.v <- as.vector(i.m[,1])
whales.ahabs.my.i.m <- cbind(whales.v, ahabs.v, my.v, i.v)
whales.ahabs.my.i.m[which(is.na(whales.ahabs.my.i.m))] <- 0
cor(whales.ahabs.my.i.m)
##
whales.v
ahabs.v
my.v
## whales.v 1.0000000 -0.24701018 -0.26124495
## ahabs.v -0.2470102 1.00000000 0.09353653
## my.v
-0.2612449 0.09353653 1.00000000
## i.v
-0.2825659 0.07104268 0.77746408
##
i.v
## whales.v -0.28256590
## ahabs.v
0.07104268
## my.v
0.77746408
## i.v
1.00000000

5.2
my.i.m <- cbind(my.v, i.v)
my.i.m[which(is.na(my.i.m))] <- 0
my.i.cor.data.df <- as.data.frame(my.i.m)
cor(my.i.cor.data.df$i, my.i.cor.data.df$my)
## [1] 0.7774641
i.my.cors.v <- NULL
for(i in 1:10000){
i.my.cors.v <- c(i.my.cors.v,
cor(sample(my.i.cor.data.df$i), my.i.cor.data.df$my))
}
min(i.my.cors.v)
## [1] -0.2688589
max(i.my.cors.v)
## [1] 0.3833696
range(i.my.cors.v)
## [1] -0.2688589 0.3833696
mean(i.my.cors.v)
## [1] 4.293087e-05
sd(i.my.cors.v)
## [1] 0.08678589

Practice Exercise Solutions

181

Practice Exercises for Chapter 6

6.1
ttr.v <- as.vector(ttr.m)
chapter.lengths.m <- do.call(rbind, lapply(chapter.raws.l,sum))
chap.len.v <- as.vector(chapter.lengths.m)
cor(ttr.v, chap.len.v)
## [1] -0.7971883

A Correlation coefficient of -0.7971883 indicates strong negative correlation.

As the length of the chapter increases the TTR scores decrease.
6.2
mean.word.use.v <- as.vector(mean.word.use.m)
cor(mean.word.use.v, chap.len.v)
## [1] 0.8924301

A Correlation coefficient of 0.8924301 indicates a strong positive correlation.

As the length of a chapter increases, the overall mean word frequency increases as
well. More words in the chapter means more repeated words.
6.3
cor(ttr.v, chap.len.v)
## [1] -0.7971883
my.cors.v <- NULL
for(i in 1:10000){
my.cors.v <- c(my.cors.v, cor(sample(ttr.v), chap.len.v))
}
min(my.cors.v)
## [1] -0.3193845
max(my.cors.v)
## [1] 0.307482
range(my.cors.v)
## [1] -0.3193845 0.3074820
mean(my.cors.v)
## [1] 0.0009244654
sd(my.cors.v)
## [1] 0.08625174

The permutation test reveals that the observed correlation is highly unlikely to be
seen by mere chance alone. In 10,000 iterations the highest positive correlation
was 0.307482 and the lowest negative -0.3193845. The mean hovered near
zero indicating that the observed correlation was far outside the norm expected by
chance.

182

Practice Exercise Solutions

Practice Exercises for Chapter 7

7.1
lengths.v.hapax <- cbind(chapter.hapax.v, chapter.lengths.m)
cor(lengths.v.hapax [,1], lengths.v.hapax [,2])
## [1] 0.9677132

In this case the correlation is extremely strong with an R-value of 97. As the chapters of Moby Dick get longer, not only do we observe the same words repeated more
often, but we also see an increase in the number of new words being introduced.
7.2
ranks <- order(hapax.percentage, decreasing=TRUE)

In this case the correlation is extremely strong with an R-value of 97. As the chapters of Moby Dick get longer, not only do we observe the same words repeated more
often, but we also see an increase in the number of new words being introduced.
7.3
text.v <- scan("data/plainText/austen.txt", what="character", sep="\n")
start.v <- which(text.v == "CHAPTER 1")
end.v <- which(text.v == "THE END")
novel.lines.v <- text.v[start.v:end.v]
novel.lines.v <- unlist(novel.lines.v)
chap.positions.v <- grep("^CHAPTER \\d", novel.lines.v)
last.position.v <- length(novel.lines.v)
chap.positions.v <- c(chap.positions.v , last.position.v)
sense.raws.l <- list()
for(i in 1:length(chap.positions.v)){
if(i != length(chap.positions.v)){
chapter.title <- novel.lines.v[chap.positions.v[i]]
start <- chap.positions.v[i]+1
end <- chap.positions.v[i+1]-1
chapter.lines.v <- novel.lines.v[start:end]
chapter.words.v <- tolower(paste(chapter.lines.v, collapse=" "))
chapter.words.l <- strsplit(chapter.words.v, "\\W")
chapter.word.v <- unlist(chapter.words.l)
chapter.word.v <- chapter.word.v[which(chapter.word.v!="")]
chapter.freqs.t <- table(chapter.word.v)
sense.raws.l[[chapter.title]] <- chapter.freqs.t
}
}
sense.chapter.hapax <- sapply(sense.raws.l, function(x) sum(x == 1))
sense.chapter.lengths <- do.call(rbind, lapply(sense.raws.l,sum))
sense.hapax.percentage <- sense.chapter.hapax / sense.chapter.lengths

A correlation coefficient of 0.8673559 indicates that Jane Austen is less consistent than Melville when it comes to the introduction of new words into her novel
even while she increases the length of her chapters. It turns out, in fact, that in terms
of vocabulary size and richness, Austen is very consistent. Her working vocabulary in Sense and Sensibility contains 6,325 unique word types and from one of
her novels to the next she rarely deviates very far from a base vocabulary of about
6300 words types. For comparison, recall that Melvilles vocabulary in Moby Dick
contains 16,872 unique word types spread over 214,889 tokens. Austen uses
6,325 types over 120,766 tokens. Even though Austens Sense and Sensibility

Practice Exercise Solutions

183

is much shorter that Moby Dick, Austen has a smaller vocabulary and repeats words
much more often. Austen uses each word an average of 19 times whereas Melville
uses each word in his vocabulary only about 13 times on average.
7.4
# Test the Moby Dick Result
cor(lengths.v.hapax[,1], lengths.v.hapax[,2])
## [1] 0.9677132
moby.cors.v <- NULL
for(i in 1:10000){
moby.cors.v <- c(moby.cors.v,
cor(sample(lengths.v.hapax[,1]),
lengths.v.hapax[,2])
)
}
min(moby.cors.v)
## [1] -0.2922685
max(moby.cors.v)
## [1] 0.3938905
range(moby.cors.v)
## [1] -0.2922685 0.3938905
mean(moby.cors.v)
## [1] -0.001119979
sd(moby.cors.v)
## [1] 0.08593544
# Test the Sense and Sensibility Result
cor(sense.chapter.hapax, sense.chapter.lengths)
##
[,1]
## [1,] 0.8673559
sense.cors.v <- NULL
for(i in 1:10000){
sense.cors.v <- c(sense.cors.v,
cor(sample(sense.chapter.hapax),
sense.chapter.lengths)
)
}
min(sense.cors.v)
## [1] -0.4899
max(sense.cors.v)
## [1] 0.4802502
range(sense.cors.v)
## [1] -0.4899000 0.4802502
mean(sense.cors.v)
## [1] -0.001468671
sd(sense.cors.v)
## [1] 0.143424

184

Practice Exercise Solutions

Practice Exercises for Chapter 8

8.1
context <- 5
dog.positions.sense <- which(my.corpus.l[[1]][]=="dog")
dog.positions.moby <- which(my.corpus.l[[2]][]=="dog")
# Answer for Sense and Sensibility
for(i in 1:length(dog.positions.sense)){
start <- dog.positions.sense[i]-context
end <- dog.positions.sense[i]+context
cat(my.corpus.l[[1]][start:end], "\n")
}
## a fellow such a deceitful dog it was only the last
# Answer for Moby Dick
for(i in 1:length(dog.positions.moby)){
start <- dog.positions.moby[i]-context
end <- dog.positions.moby[i]+context
cat(my.corpus.l[[2]][start:end], "\n")
}
## all over like a newfoundland dog just from the water and
## a fellow that in the dog days will mow his two
## was seen swimming like a dog throwing his long arms straight
## filling one at last down dog and kennel starting at the
## not tamely be called a dog sir then be called ten
## t he call me a dog blazes he called me ten
## sacrifice of the sacred white dog was by far the holiest
## life that lives in a dog or a horse indeed in
## the sagacious kindness of the dog the accursed shark alone can
## boats the ungracious and ungrateful dog cried starbuck he mocks and
## intense whisper give way greyhounds dog to it i tell ye
## to the whale that a dog does to the elephant nevertheless
## aries or the ram lecherous dog he begets us then taurus
## is dr bunger bunger you dog laugh out why don t
## to die in pickle you dog you should be preserved to
## round ahab and like a dog strangely snuffing this man s
## lad five feet high hang dog look and cowardly jumped from
## as a sagacious ship s dog will in drawing nigh to
## the compass and then the dog vane and then ascertaining the

8.2
for(i in 1:length(dog.positions.moby)){
start <- dog.positions.moby[i]-context
end <- dog.positions.moby[i]+context
before <- my.corpus.l[[2]][start:(start+context-1)]
after <- my.corpus.l[[2]][(start+context+1):end]
keyword <- my.corpus.l[[2]][start+context]
cat("----------------------", i, "----------------------", "\n")
cat(before,"[",keyword, "]", after, "\n")
}
## ---------------------- 1 ---------------------## all over like a newfoundland [ dog ] just from the water and
## ---------------------- 2 ---------------------## a fellow that in the [ dog ] days will mow his two
## ---------------------- 3 ---------------------## was seen swimming like a [ dog ] throwing his long arms straight
## ---------------------- 4 ---------------------## filling one at last down [ dog ] and kennel starting at the
## ---------------------- 5 ----------------------

Practice Exercise Solutions

##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##

not tamely be called a [ dog ] sir then be called ten

---------------------- 6 ---------------------t he call me a [ dog ] blazes he called me ten
---------------------- 7 ---------------------sacrifice of the sacred white [ dog ] was by far the holiest
---------------------- 8 ---------------------life that lives in a [ dog ] or a horse indeed in
---------------------- 9 ---------------------the sagacious kindness of the [ dog ] the accursed shark alone can
---------------------- 10 ---------------------boats the ungracious and ungrateful [ dog ] cried starbuck he mocks and
---------------------- 11 ---------------------intense whisper give way greyhounds [ dog ] to it i tell ye
---------------------- 12 ---------------------to the whale that a [ dog ] does to the elephant nevertheless
---------------------- 13 ---------------------aries or the ram lecherous [ dog ] he begets us then taurus
---------------------- 14 ---------------------is dr bunger bunger you [ dog ] laugh out why don t
---------------------- 15 ---------------------to die in pickle you [ dog ] you should be preserved to
---------------------- 16 ---------------------round ahab and like a [ dog ] strangely snuffing this man s
---------------------- 17 ---------------------lad five feet high hang [ dog ] look and cowardly jumped from
---------------------- 18 ---------------------as a sagacious ship s [ dog ] will in drawing nigh to
---------------------- 19 ---------------------the compass and then the [ dog ] vane and then ascertaining the

185

186

Practice Exercise Solutions

Practice Exercises for Chapter 9

9.1
doitKwicBetter <- function(named.text.word.vector.l){
show.files(names(named.text.word.vector.list))
# ask the user for three bits of information
file.id <- as.numeric(
readline("Which file would you like to examine? Enter a number: \n"))
context <- as.numeric(
readline("How much context do you want to see? Enter a number: \n"))
keyword <- tolower((readline("Enter a keyword: \n")))
hits.v <- which(named.text.word.vector.l[[file.id]] == keyword)
if(length(hits.v)>0){
result <- NULL
for(h in 1:length(hits.v)){
start <- hits.v[h]-context
if(start < 1){
start <- 1
}
end <- hits.v[h]+context
cat(named.text.word.vector.l[[file.id]][start:end], "\n")
myrow <- cbind(hits.v[h],
paste(
named.text.word.vector.l[[file.id]][start:(hits.v[h]-1)],
collapse=" "),
paste(
named.text.word.vector.l[[file.id]][hits.v[h]],
collapse=" "),
paste(
named.text.word.vector.l[[file.id]][(hits.v[h]+1):end],
collapse=" "))
result <- rbind(result,myrow)
}
}
colnames(result) <- c("position", "left", "keyword", "right")
return(result)
}

9.2
doitKwicBest <- function(named.text.word.vector.l){
show.files(names(named.text.word.vector.l))
# ask the user for three bits of information
file.id <- as.numeric(
readline("Which file would you like to examine? Enter a number: \n"))
context <- as.numeric(
readline("How much context do you want to see? Enter a number: \n"))
keyword <- tolower((readline("Enter a keyword: \n")))
hits.v <- which(named.text.word.vector.l[[file.id]] == keyword)
if(length(hits.v)>0){
result <- NULL
for(h in 1:length(hits.v)){
start <- hits.v[h]-context
if(start < 1){
start <- 1
}
end <- hits.v[h]+context
cat("\n-----------------------", h, "-------------------------\n")
cat(named.text.word.vector.l[[file.id]][start:(hits.v[h]-1)], sep=" ")
cat(" [", named.text.word.vector.l[[file.id]][hits.v[h]],"] ", sep="")
cat(named.text.word.vector.l[[file.id]][(hits.v[h]+1):end], sep=" ")

Practice Exercise Solutions

187

myrow <- cbind(hits.v[h],

paste(named.text.word.vector.l[[file.id]][start:(hits.v[h]-1)],
collapse=" "),
paste(named.text.word.vector.l[[file.id]][hits.v[h]],
collapse=" "),
paste(named.text.word.vector.l[[file.id]][(hits.v[h]+1):end],
collapse=" "))
result <- rbind(result,myrow)
}
colnames(result) <- c("position", "left", "keyword", "right")
toprint <- as.numeric((
readline("Would you like to save this result to a file:
enter 1=yes or 0=no \n")))
if(toprint==1){
write.csv(result,
paste("results/", keyword,"_In_",
context, names(named.text.word.vector.l)[file.id], ".csv", sep=""))
}
} else {
cat("YOUR KEYWORD WAS NOT FOUND\n")
}
}

188

Practice Exercise Solutions

Practice Exercises for Chapter 10

10.1
(respStmt <- xpathApply(doc,
"/tei:TEI//tei:fileDesc//tei:titleStmt//tei:respStmt",
namespaces=c(tei = "http://www.tei-c.org/ns/1.0"))[[1]])
## <respStmt>
##
<resp>converted into TEI-conformant markup by</resp>
##
<name type="contributor">Eric Lease Morgan</name>
##
<resp>Modified for use as an exercise by</resp>
##
<name type="contributor">Matthew Jockers</name>
## </respStmt>

Practice Exercise Solutions

189

Practice Exercises for Chapter 11

11.1
smaller.m <- final.m[,apply(final.m,2,mean)>=2.5]
dim(smaller.m)
## [1] 43 4
dm <- dist(smaller.m)
cluster <- hclust(dm)
cluster$labels <- names(book.freqs.l)
plot(cluster)

LeFanu3.xml
Quigley1.xml
Quigley2.xml

Donovan2.xml
Donovan1.xml
Jessop2.xml
Jessop3.xml
Jessop1.xml
Driscoll1.xml
Driscoll2.xml
Driscoll3.xml
Carleton2.xml
LeFanu2.xml
Carleton5.xml
Lewis.xml
Carleton12.xml
Carleton14.xml
Polidori1.xml
Carleton6.xml
Carleton8.xml
Carleton1.xml
Carleton11.xml
anonymous.xml
Kyne1.xml
Kyne2.xml
McHenry1.xml
McHenry2.xml
Carleton4.xml
Carleton10.xml
Carleton13.xml
Carleton9.xml
Edgeworth1.xml
Carleton3.xml
Carleton7.xml
Norris2.xml
Norris3.xml
Norris4.xml
LeFanu6.xml
Norris1.xml
LeFanu5.xml
LeFanu7.xml
LeFanu1.xml
LeFanu4.xml

3
0

Height

Cluster Dendrogram

dm
hclust (*, "complete")

Using a mean threshold of 2.5 reduces the feature set to just 4 words, and the
anonymous text is still found closest to Kyne. However, a mean threshold of 2.75
reduces the feature set to just 3 words, and the anonymous text is then found closest
to Donovan.
11.2
random.m <- final.m[,sample(colnames(final.m), 100)]
dim(random.m)
## [1] 43 100
dm <- dist(random.m)
cluster <- hclust(dm)
cluster$labels <- names(book.freqs.l)
plot(cluster)

Norris3.xml
Donovan1.xml
Jessop3.xml
Carleton5.xml
Lewis.xml
Edgeworth1.xml
Polidori1.xml
LeFanu2.xml
LeFanu3.xml
Quigley2.xml
Quigley1.xml
anonymous.xml
Driscoll3.xml
Driscoll1.xml
Driscoll2.xml
Carleton8.xml
LeFanu1.xml
Jessop2.xml
Carleton10.xml
Norris4.xml
Carleton6.xml
Donovan2.xml
Carleton4.xml
Jessop1.xml
Kyne1.xml
Kyne2.xml
Carleton1.xml
Carleton11.xml
Carleton3.xml
Carleton7.xml
Carleton9.xml
Norris2.xml
McHenry2.xml
Carleton12.xml
LeFanu7.xml
McHenry1.xml
Norris1.xml
Carleton13.xml
Carleton2.xml
LeFanu5.xml
LeFanu4.xml
LeFanu6.xml

Carleton14.xml

0.00 0.02 0.04 0.06 0.08 0.10 0.12 0.14

Height

190
Practice Exercise Solutions

Cluster Dendrogram

dm
hclust (*, "complete")

Practice Exercise Solutions

Practice Exercises for Chapter 12

12.1
keepers.v <- which(freq.means.v >=.00014)
length(keepers.v)
## [1] 681
smaller.df <- authorship.df[, c(names(authorship.df)[1:3],
names(keepers.v))]
anon.v <- which(smaller.df$author.v == "anonymous")
train <- smaller.df[-anon.v,4:ncol(smaller.df)]
class.f <- smaller.df[-anon.v,"author.v"]
model.svm <- svm(train, class.f)
pred.svm <- predict(model.svm, train)
testdata <- smaller.df[anon.v,4:ncol(smaller.df)]
final.result <- predict (model.svm, testdata)
as.data.frame(final.result)
##
final.result
## anonymous_1
Carleton
## anonymous_10
Carleton
## anonymous_2
Carleton
## anonymous_3
Carleton
## anonymous_4
Carleton
## anonymous_5
Carleton
## anonymous_6
Carleton
## anonymous_7
Carleton
## anonymous_8
Carleton
## anonymous_9
Carleton

12.2
keepers.v <- which(freq.means.v >=.00014)
smaller.df <- authorship.df[, c(names(authorship.df)[1:3],
names(keepers.v))]
author.sums <- aggregate(smaller.df[,
4:ncol(smaller.df)],
list(smaller.df[,1]), sum)
reduced.author.sums <- author.sums[,
colSums(author.sums==0) == 0]
keepers.v <- colnames(
reduced.author.sums)[2:ncol(reduced.author.sums)]
new.smaller.df <- smaller.df[,
c("author.v","sampletext","samplechunk", keepers.v)]
anon.v <- which(new.smaller.df$author.v == "anonymous")
train <- new.smaller.df[-anon.v,4:ncol(new.smaller.df)]
class.f <- new.smaller.df[-anon.v,"author.v"]
model.svm <- svm(train, class.f)
pred.svm <- predict(model.svm, train)
testdata <- new.smaller.df[anon.v,4:ncol(new.smaller.df)]
final.result <- predict (model.svm, testdata)
as.data.frame(final.result)
##
final.result
## anonymous_1
Kyne
## anonymous_10
Kyne
## anonymous_2
Kyne
## anonymous_3
Kyne
## anonymous_4
Kyne
## anonymous_5
Kyne
## anonymous_6
Kyne
## anonymous_7
Kyne
## anonymous_8
Kyne
## anonymous_9
Kyne

191

192

Practice Exercise Solutions

Practice Exercises for Chapter 13

13.1
for(i in 1:43){
topic.top.words <- mallet.top.words(topic.model,
topic.words.m[i,], 100)
print(wordcloud(topic.top.words$words,
topic.top.words$weights,
c(4,.8), rot.per=0,
random.order=F))
}

13.2
mallet.instances <- mallet.import(documents$id,
documents$text,
"data/stoplist-exp.csv",
FALSE,
token.regexp="[\\p{L}']+")
topic.model <- MalletLDA(num.topics=43)
topic.model$loadDocuments(mallet.instances)
vocabulary <- topic.model$getVocabulary()
word.freqs <- mallet.word.freqs(topic.model)
topic.model$train(400)
topic.words.m <- mallet.topic.words(topic.model,
smoothed=TRUE,
normalized=TRUE)
vocabulary <- topic.model$getVocabulary()
colnames(topic.words.m) <- vocabulary
for(i in 1:43){
topic.top.words <- mallet.top.words(topic.model,
topic.words.m[i,], 100)
print(wordcloud(topic.top.words$words,
topic.top.words$weights,
c(4,.8), rot.per=0,
random.order=F))
}

Index

A
Accessing word data, 2527
apply, 4143
B
barplot, 46
C
c, 15
cbind, 4446
class, 17
Classification
colMeans, 129
cross tabulation, 126127
data organization, 125126
mapping the data, 127129
R list conversion, matrix, 124125
with SVM, 130132
text segmentation, 120124
Clustering
corpus ingestion, 103106
data, 116117
data preparation, 115116
and Euclidean metric, 109111
getTEIWordTableList, 107
oddities in R, 102103
R list conversion, data matrix, 112115
cor, 47
Correlation
analysis, 4750
data frames, 5052
testing, 5256
cor.test, 54
D
do.call, 4344
Document object model (DOM), 91

F
for, 3539
G
grep, 3135
H
Hapax
legomena, 69
mini-conditional function, 7071
sapply, 6970
I
if, 3539
K
Keyword in context (KWIC)
custom functions, 7476
fixation problem, 8586
readline, 83
reuse, 82
user interaction, 8283
L
lapply, 41
LDA buffet, 163165
length, 14
Lexical variety
calculating TTR inside lapply, 6566
mean word usage, 6164
mean word frequency, 6061
ranking, 6465
and type-token ratio, 59
ls, 33
M
Mallet instance list, 144
Mallet package, 142
max, 54

M.L. Jockers, Text Analysis with R for Students of Literature, Quantitative Methods
in the Humanities and Social Sciences, DOI 10.1007/978-3-319-03164-4,
Springer International Publishing Switzerland 2014

193

194
mean, 41, 54
min, 54
N
names, 27
O
Operators
assignment, 10, 13
division, 26
multiplication, 27
Optical character recognition (OCR), 8990
P
Parsing XML
with R, 91
paste
collapse, 15
Pearson product-moment correlation
coefficient, 49
plot, 22, 27
axis, 27
POS tagger, 156159
Pythagorean metric, 109
R
R
console, 7
download, 45
install, 4
packages, 9192
prompt, 7
resources, 171172
and RStudio, 4
text analysis, 1123
range, 54
rbind, 39
R data types
vector, 12
Recycling, 2728
Regular expressions, 12, 33
Relative frequencies, 26
rep, 30
R help, 17
rm, 32
R packages
installing, 9192
RStudio
download, 5
install, 5
layout, 6

Index
S
sample, 52
sapply, 6970
scan, 11
sd, 54
seq, 29
sort, 22
Start up code, 167169
str, 16
strsplit, 16
SVM, 130132
T
table, 22
Terminal/shell, 6
Text encoding initiative (TEI), 9091
Token distribution analysis
cbind, 4446
dispersion plots, 2931
do.call, 4344
lapply, 4143
for loop and if conditional, 3539
rbind, 40
recycling, 4041
searching with grep, 3135
tolower, 16
Topic coherence and probability, 152155
Topic modeling
coherence and probability, 152155
mallet package, 142
pre-processing with POS tagger, 156159
and R, 136137
with standard stop list, 142147
text segmentation and preparation, 137142
unpacking, 147150
and visualization, 150151
Type-token ratio (TTR), 59
U
unique, 21
unlist, 18
V
Variable scope, 161
Vector
indexing, 13, 20
recycling, 4041
W
which, 13
X
XML package, 9296