Literary detective work on the computer

293 Pages·2014·3.088 MB·English

by  Oakes, Michael P

Literary Detective Work on the Computer
Natural Language Processing (NLP)
The scope of NLP ranges from theoretical Computational Linguistics topics
to highly practical Language Technology topics. The focus of the series is on
new results in NLP and modern alternative theories and methodologies.
For an overview of all books published in this series, please see
http://benjamins.com/catalog/nlp
Editor
Ruslan Mitkov
University of Wolverhampton
Advisory Board
Sylviane Cardey Inderjeet Mani Constantin Orăsan
Institut universitaire de France, Yahoo! Labs, Sunnyvale, USA University of Wolverhampton
Université de Franche-Comté Carlos Martín-Vide Manuel Palomar
Gloria Corpas Rovira i Virgili Un., Tarragona University of Alicante
University of Malaga Rada Mihalcea Khalil Simaan
Robert Dale University of Michigan University of Amsterdam
Macquarie University, Sydney Andrei Mikheev Richard Sproat
Eduard Hovy Daxtra Technologies Google Research
University of Southern Roberto Navigli Key-Yih Su
California
Universita di Sapienza, Roma Behaviour Design Corp.
Alma Kharrat
John Nerbonne Benjamin Tsou
Microsoft Research
University of Groningen The Hong Kong Institute of
Richard Kittredge Nicolas Nicolov Education
GoGenTex Inc, Ithaca Microsoft Research Yorick Wilks
Lori Lamel Kemal Oflazer Florida Institute of Human
LIMSI, CNRS and Machine Cognition
Carnegie Mellon University,
Qatar
Editorial Assistant
Miranda Chong
University of Wolverhampton
Volume 12
Literary Detective Work on the Computer
by Michael P. Oakes
Literary Detective Work
on the Computer
Michael P. Oakes
University of Wolverhampton
John Benjamins Publishing Company
Amsterdam / Philadelphia
TM The paper used in this publication meets the minimum requirements of
8
the American National Standard for Information Sciences – Permanence
of Paper for Printed Library Materials, ansi z39.48-1984.
Library of Congress Cataloging-in-Publication Data
Oakes, Michael P.
Literary Detective Work on the Computer / Michael P. Oakes.
p. cm. (Natural Language Processing, issn 1567-8202 ; v. 12)
Includes bibliographical references and index.
1. Computational linguistics--Research. 2. Imitation in literature. 3. Plagiarism.
4.  Linguistics--Research--Methodology. 5. Authorship--Study and teaching.
I. Title.
P98.5.O25 2014
006.3’5--dc23 2014007366
isbn 978 90 272 4999 9 (Hb ; alk. paper)
isbn 978 90 272 7013 9 (Eb)
© 2014 – John Benjamins B.V.
No part of this book may be reproduced in any form, by print, photoprint, microfilm, or any
other means, without written permission from the publisher.
John Benjamins Publishing Co. · P.O. Box 36224 · 1020 me Amsterdam · The Netherlands
John Benjamins North America · P.O. Box 27519 · Philadelphia pa 19118-0519 · usa
Table of contents
Preface ix
chapter 1
Author identification 1
1. Introduction 1
2. Feature selection 5
2.1 Evaluation of feature sets for authorship attribution 8
3. Inter-textual distances 11
3.1 Manhattan distance and Euclidean distance 12
3.2 Labbé and Labbé’s measure 14
3.3 Chi-squared distance 15
3.4 The cosine similarity measure 16
3.5 Kullback-Leibler Divergence (KLD) 18
3.6 Burrows’ Delta 18
3.7 Evaluation of feature-based measures for inter-textual distance 23
3.8 Inter-textual distance by semantic similarity 26
3.9 Stemmatology as a measure of inter-textual distance 28
4. Clustering techniques 30
4.1 Introduction to factor analysis 31
4.2 Matrix algebra 35
4.3 Use of matrix algebra for PCA 38
4.4 PCA case studies 44
4.5 Correspondence analysis 45
5. Comparisons of classifiers 47
6. Other tasks related to authorship 50
6.1 Stylochronometry 50
6.2 Affect dictionaries and psychological profiling 53
6.3 Evaluation of author profiling 58
7. Conclusion 58
vi Literary Detective Work on the Computer
chapter 2
Plagiarism and spam filtering 59
1. Introduction 59
2. Plagiarism detection software 62
2.1 Collusion and plagiarism, external and intrinsic 63
2.2 Preprocessing of corpora and feature extraction 63
2.3 Sequence comparison and exact match 64
2.4 Source-suspicious document similarity measures 65
2.5 Fingerprinting 66
2.6 Language models 67
2.7 Natural language processing 68
2.8 Intrinsic plagiarism detection 70
2.9 Plagiarism of program code 73
2.10 Distance between translated and original text 74
2.11 Direction of plagiarism 76
2.12 The search engine-based approach used at PAN-13 78
2.13 Case study 1: Hidden influences from printed sources
in the Gaelic tales of Duncan and Neil MacDonald 81
2.14 Case study 2: General George Pickett and related writings 83
2.15 Evaluation methods 84
2.16 Conclusion 85
3. Spam filters 86
3.1 Content-based techniques 87
3.2 Building a labeled corpus for training 87
3.3 Exact matching techniques 88
3.4 Rule-based methods 89
3.5 Machine learning 90
3.6 Unsupervised machine learning approaches 92
3.7 Other spam-filtering problems 93
3.8 Evaluation of spam filters 94
3.9 Non-linguistic techniques 94
3.10 Conclusion 97
4. Recommendations for further reading 98
chapter 3
Computer studies of Shakespearean authorship 99
1. Introduction 99
2. Shakespeare, Wilkins and “Pericles” 101
2.1 Correspondence analysis for “Pericles” and related texts 105
3. Shakespeare, Fletcher and “The Two Noble Kinsmen” 108
4. “King John” 110
Table of contents vii
5. “The Raigne of King Edward III” 111
5.1 Neural networks in stylometry 111
5.2 Cusum charts in stylometry 113
5.3 Burrows’ Zeta and Iota 116
6. Hand D in “Sir Thomas More” 118
6.1 Elliott, Valenza and the Earl of Oxford 118
6.2 Elliott and Valenza: Hand D 121
6.3 Bayesian approach to questions of Shakespearian authorship 122
6.4 Bayesian analysis of Shakespeare’s second person pronouns 127
6.5 Vocabulary differences, LDA and the authorship of Hand D 130
6.6 Hand D: Conclusions 131
7. The three parts of “Henry VI” 132
8. “Timon of Athens” 132
9. “The Puritan” and “A Yorkshire Tragedy” 133
10. “Arden of Faversham” 134
11. Estimation of the extent of Shakespeare’s vocabulary
and the authorship of the “Taylor” poem 136
12. The chronology of Shakespeare 141
13. Conclusion 147
chapter 4
Stylometric analysis of religious texts 149
1. Introduction 149
1.1 Overview of the New Testament by correspondence analysis 151
1.2 Q 153
1.3 Luke and Acts 169
1.4 Recent approaches to New Testament stylometry 171
1.5 The Pauline Epistles 175
1.6 Hebrews 188
1.7 The Signs Gospel 188
2. Stylometric analysis of the Book of Mormon 190
3. Stylometric studies of the Qu’ran 198
4. Conclusion 206
chapter 5
Computers and decipherment 207
1. Introduction 207
1.1 Differences between cryptography and decipherment 208
1.2 Cryptological techniques for automatic language recognition 209
1.3 Dictionary approaches to language recognition 212
1.4 Sinkov’s test 212
viii Literary Detective Work on the Computer
1.5 Index of coincidence 213
1.6 The log-likelihood ratio 214
1.7 The chi-squared test statistic 215
1.8 Entropy of language 215
1.9 Zipf’s Law and Heaps’ Law coefficients 218
1.10 Modal token length 219
1.11 Autocorrelation analysis 220
1.12 Vowel identification 221
2. Rongorongo 224
2.1 History of Rongorongo 224
2.2 Characteristics of Rongorongo 226
2.3 Obstacles to decipherment 227
2.4 Encoding of Rongorongo symbols 227
2.5 The “Mamari” lunar calendar 228
2.6 Basic statistics of the Rongorongo corpus 228
2.7 Alignment of the Rongorongo corpus 229
2.8 A concordance for Rongorongo 231
2.9 Collocations and collostructions 233
2.10 Classification by genre 234
2.11 Vocabulary richness 237
2.12 Podzniakov’s approach to matching frequency curves 241
3. The Indus Valley texts 243
3.1 Why decipherment of the Indus texts is difficult 243
3.2 Are the Indus texts writing? 244
3.3 Other evidence for the Indus Script being writing 248
3.4 Determining the order of the Markov model 248
3.5 Missing symbols 249
3.6 Text segmentation and the log-likelihood measure 249
3.7 Network analysis of the Indus Signs 251
4. Linear A 252
5. The Phaistos disk 255
6. Iron Age Pictish symbols 256
7. Mayan glyphs 256
8. Conclusion 257
References 259
Index 281
Preface
Computer stylometry is the computer analysis of writing style. This enables infer-
ences to be made, especially about the sometimes disputed provenance of texts,
but also about the dating of texts and also how texts reveal broad personality types.
Following the PAN conferences, studies of disputed authorship, plagiarism and
spam (unwarranted email campaigns) are considered together, partly because they
often uncover fraudulent behaviour, but also because they are all examples of text
classification: either a text is by author A or author B, or an email message is either
spam or a legitimate message. The first two chapters will show many ways in which
these techniques overlap, particularly regarding the question of how similar one
text is to another. Chapters 3 and 4 aim to be comprehensive surveys of how com-
puter stylometry has been used to examine the work of Shakespeare and the New
Testament, both of great cultural significance. This book takes the standpoint that
all of Shakespeare’s most famous plays were indeed written by Shakespeare, so the
focus is on the so-called Shakespeare “apocrypha” – plays for which there is some
historical evidence to suggest that Shakespeare might have had a hand in their
composition. Examples we will look at are the “Two Noble Kinsmen”, written with
John Fletcher; “Pericles”, written with George Wilkins, and “Edward III”, possibly
written with Kit Marlowe. We will see how computers have been able to indicate
the extent of Shakespeare’s contribution in each case. Computer stylometry has
also considered the evidence for whether the handwritten fragment “Hand D” is
by Shakespeare, as well as the play “Arden of Faversham”, where the association
with Shakespeare is simply that it was very popular in his day, is very good, but
we simply do not know who wrote it. This book also considers evidence that the
writing style of Shakespeare is distinct from that of a recently popular “claimant”,
Edward de Vere. Like Shakespeare, the King James version of the Bible has greatly
influenced the English language. However, in the survey of computer stylometry
given in Chapter 4, most of the studies considered have used texts in the original
Greek of the New Testament. The findings largely agree with the beliefs of mod-
ern theologians, including a “cautious preference for Q”, a possible source of the
Gospels of Matthew and Luke. The vast majority of computer studies on religious
texts are concerned with the New Testament, but new work is also starting to
emerge on the Book of Mormon and the Qu’ran. In the final chapter, the aspect of
literary detective work we consider is the decipherment of lost languages. In some
x Literary Detective Work on the Computer
respects, computer techniques can only scratch the surface of this, and there are
difficulties in showing that a discovered script even constitutes language. However,
the mathematics behind these techniques, and what these techniques do show us,
are of considerable interest in themselves. The most extensive case studies in this
chapter are the Rongorongo writings of Easter Island and the Indus Valley seals.
I would like to thank the series editor, Prof. Ruslan Mitkov, for suggesting in
the first place a book centred around disputed authorship, plagiarism and spam.
I am also grateful to Harry Erwin who kindled my interest in computer studies
of religious texts, through his interest in the Signs Gospel, a possible precursor of
the Gospel of John. I also wish to thank Miranda Chong, for her valuable feed-
back on Chapters 2 and 5. Finally, the following people were kind enough to send
detailed responses to my emailed questions on their work: Raf Alvarado, Ward
Elliott, Richard Forsyth, Antonius Linmans, David Mealand, Richard Sproat and
Robert Valenza.
Michael P. Oakes
Wolverhampton, January 2014