Journal of Law and Policy

Volume 21
Issue 2
SYMPOSIUM: Article 7
Authorship Attribution Workshop

2013

On the Robustness of Authorship Attribution

Based on Character N-gram Features
Efstathios Stamatatos, Ph.D.

Follow this and additional works at: https://brooklynworks.brooklaw.edu/jlp

Recommended Citation
Efstathios Stamatatos, Ph.D., On the Robustness of Authorship Attribution Based on Character N-gram Features, 21 J. L. & Pol'y (2013).
Available at: https://brooklynworks.brooklaw.edu/jlp/vol21/iss2/7

This Article is brought to you for free and open access by the Law Journals at BrooklynWorks. It has been accepted for inclusion in Journal of Law and
Policy by an authorized editor of BrooklynWorks.
 ON THE ROBUSTNESS OF AUTHORSHIP
ATTRIBUTION BASED ON CHARACTER
N-GRAM FEATURES

Efstathios Stamatatos*

ABSTRACT

A number of independent authorship attribution studies have

demonstrated the effectiveness of character n-gram features for
representing the stylistic properties of text. However, the vast
majority of these studies examined the simple case where the
training and test corpora are similar in terms of genre, topic,
and distribution of the texts. Hence, there are doubts whether
such a simple and low-level representation is equally effective in
realistic conditions where some of the above factors are not
possible to remain stable. In this study, the robustness of
authorship attribution based on character n-gram features is
tested under cross-genre and cross-topic conditions. In addition,
the distribution of texts over the candidate authors varies in
training and test corpora to imitate real cases. Comparative
results with another competitive text representation approach
based on very frequent words show that character n-grams are
better able to capture stylistic properties of text when there are
significant differences among the training and test corpora.
Moreover, a set of guidelines to tune an authorship attribution
model according to the properties of training and test corpora is
provided.

* Assistant Professor, University of the Aegean.

421
422 JOURNAL OF LAW AND POLICY

I. INTRODUCTION

Authorship attribution is the line of research dealing with the

identification of the author of a text under investigation given a
set of candidate authors (e.g., suspects) and samples of known
authorship for each one of them. Indeed, in many forensic
examinations, part of the evidence refers to texts (e.g., notes,
e-mail messages, SMS messages, written reports, etc.). The
ability to verify that a text was written by one of the suspects
could be crucial to support a case. During the last decades,
significant progress has been achieved in the automation of this
procedure by incorporating statistical and/or machine learning
techniques (i.e., algorithms that can learn from data).1 There is
strong potential for this technology to be used as evidence in a
judicial process, given that it provides effective results in well-
designed experimental tests. So far, a primitive and controversial
technique has been used in British courts.2 In addition, Chaski
discusses examples of the use of a semiautomated author
identification method in U.S. courts.3
From the machine-learning point of view, authorship
attribution can be viewed as a multiclass, single-label
classification problem (i.e., there may be multiple suspect
authors, one of whom must be selected) and can be studied

1
See Patrick Juola, Authorship Attribution, 1 FOUND. & TRENDS IN
INFO. RETRIEVAL 234, 235, 284–86 (2006); Moshe Koppel et al.,
Computational Methods in Authorship Attribution, 60 J. AM. SOC’Y FOR
INFO. SCI. & TECH. 9, 10–13 (2009); Efstathios Stamatatos, A Survey of
Modern Authorship Attribution Methods, 60 J. AM. SOC’Y FOR INFO. SCI. &
TECH. 538, 538 (2009).
2
R.A. Hardcastle, CUSUM: A Credible Method for the Determination of
Authorship?, 37 J. FORENSIC SCI. SOC’Y 129, 137–38 (1997).
3
See Carol E. Chaski, Who’s at the Keyboard? Authorship Attribution in
Digital Evidence Investigations?, INT’L J. DIGITAL EVIDENCE, Spring 2005,
at 9, 10–11 (providing examples of cases in which the syntactic analysis
method of authorship identification has been used in U.S. courts); Carol E.
Chaski, Empirical Evaluations of Language-Based Author Identification
Techniques, 8 FORENSIC LINGUISTICS 1, 1–2 (2001) (discussing the
admissibility of FBI forensic stylistics methods in a federal district court
case).
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 423

along the lines of other text categorization tasks.4 However,

there are some properties of authorship attribution that
differentiate it from other text categorization tasks.5 First, and
perhaps most important, the stylistic choices of an author are far
more difficult to capture and quantify in comparison to topic-
related information. Stylistic information is usually based on
very frequent patterns that are encountered in texts by the same
author. On the other hand, it is preferable to focus on stylistic
choices that are unconsciously made by the author and remain
stable over the text length. To this end, a very large number of
such features have been proposed, including measures about the
length of words or sentences, vocabulary richness measures,
function word frequencies, character n-gram6 frequencies, and
syntactic-related or even semantic-related measures.7 In several
independent studies, it has been demonstrated that function
words (defined as the set of the most frequent words of the
training set) and character n-grams are among the most effective
stylometric features, though the combination of several feature
types usually improves the performance of an attribution model.8
Practical applications of authorship attribution usually
provide a limited number of samples of known authorship
unevenly distributed over the candidate authors. Therefore, it is
essential for the attribution model to be able to handle limited
9
and imbalanced training sets. Moreover, the availability of
many samples for one candidate author does not necessarily
increase the probability that the author is the true author of

4
See Fabrizio Sebastiani, Machine Learning in Automated Text
Categorization, ACM COMPUTING SURVEYS, Mar. 2002, at 5 (listing “author
identification for literary texts of unknown or disputed authorship” as an
application of text categorization).
5
See Stamatatos, supra note 1, at 553.
6
For example, the character 3-grams of the beginning of this footnote
would be “For”, “or ”, “r e”, “ ex”, etc.
7
See Stamatatos, supra note 1, at 539–44.
8
KIM LUYCKX, SCALABILITY ISSUES IN AUTHORSHIP ATTRIBUTION 124–
26 (2010); Jack Grieve, Quantitative Authorship Attribution: An Evaluation of
Techniques, 22 LITERARY & LINGUISTIC COMPUTING 251, 266–67 (2007).
9
See Efstathios Stamatatos, Author Identification Using Imbalanced and
Limited Training Tests, PROC. EIGHTEENTH INT’L WORKSHOP ON DATABASE
& EXPERT SYS. APPLICATIONS: DEXA 2007, at 237, 237–41.
424 JOURNAL OF LAW AND POLICY

another text. This is in contrast to other text categorization tasks

(e.g., thematic classification of texts) where well-represented
classes have high prior probability.10 In addition, in authorship
attribution applications it is probable to have samples of known
authorship on a certain thematic area (e.g., politics) while the
unknown texts are on another thematic area (e.g., sports). The
same can be said about the genre (e.g., known samples are
scientific papers while the unknown texts are e-mail messages).
In other words, in authorship attribution it is very likely to have
heterogeneous training and test sets in terms of distribution of
samples over the training authors, topic of texts, and genre of
texts. Note that in text categorization research, it is usually
assumed that the test set follows the properties of the training
set.11
Most of the authorship attribution studies examine the simple
case where the topic and genre are controlled in both the
training and the test corpus.12 While this differs from most
practical applications, it aims at ensuring that the authorial style
will be the crucial factor responsible for the differences among
texts. In some cases, a variety of topics are covered but the

10
See Stamatatos, supra note 1, at 540, 553.
11
See Sebastiani, supra note 4, at 19.
12
See Stamatatos, supra note 9 (addressing the problem of author
identification); Moshe Koppel et al., Authorship Attribution in the Wild, 45
LANGUAGE RESOURCES & EVALUATION 83, 83–94 (2011) (explaining how
similarity-based methods can be used with “high precision” to attribute
authorship to a “set of known candidates [that is] extremely large (possibly
many thousands) and might not even include the actual author”); Moshe
Koppel et al., Measuring Differentiability: Unmasking Pseudonymous
Authors, 8 J. MACHINE LEARNING RES. 1261, 1261–76 (2007) (presenting “a
new learning-based method for adducing the ‘depth of difference’ between
two example sets and offer[ing] evidence that this method solves the
authorship verification problem with very high accuracy”); Efstathios
Stamatatos et al., Automatic Text Categorization in Terms of Genre and
Author, 26 COMPUTATIONAL LINGUISTICS 471, 471–95 (2000) (presenting “an
approach to text categorization in terms of genre and author for Modern
Greek”); Hans van Halteren et al., New Machine Learning Methods
Demonstrate the Existence of a Human Stylome, 12 J. QUANTITATIVE
LINGUISTICS 65, 65–77 (2005) (explaining how the ability to distinguish
between writings of less experienced authors “implies that a stylome exists
even in the general population”).
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 425

same topics may be found in both the training and test set.13
Although this setting makes sense in laboratory experiments, it
is rarely the case in practical applications where usually the
available texts of known authorship and the texts under
investigation are completely different with respect to thematic
area and genre. The control for topic and genre in training and
test sets provide results that may overestimate the effectiveness
of the examined models in more difficult (but realistic) cases. In
a recent study,14 the authors present a cross-genre authorship
verification experiment where the well-known unmasking
15
method is applied on pairs of documents that belong to two
different genres (e.g., prose works and theatrical plays) and the
performance is considerably decreased in comparison to
intragenre document pairs. In order for authorship attribution
technology to be used as evidence in courts, more complicated
tests should be performed to verify the robustness of this
technology under realistic scenarios.
In this paper, an experimental authorship attribution study is
presented where authorship attribution models based on
character n-gram and word features are stress-tested under cross-
topic and cross-genre conditions. In contrast to the vast majority
of the published studies, the performed experiments better match
the requirements of a realistic scenario of forensic applications
where the available texts by the candidate authors (e.g.,
suspects) may belong to certain genres and discuss specific
topics while the texts under investigation belong to other genres
and are about completely different topics. We examine the case
where the training set contains texts on a certain thematic area

13
LUYCKX, supra note 8, at 96–99.
14
Mike Kestemont et al., Cross-Genre Authorship Verification Using
Unmasking, 93 ENG. STUD. 340, 340 (2012).
15
See generally Koppel et al., Measuring Differentiability, supra note
12, at 1264 (“The intuitive idea of unmasking is to iteratively remove those
features that are most useful for distinguishing between A and X and to gauge
the speed with which cross-validation accuracy degrades as more features are
removed. . . . [I]f A and X are by the same author, then whatever
differences there are between them will be reflected in only a relatively small
number of features, despite possible differences in theme, genre and the
like.”).
426 JOURNAL OF LAW AND POLICY

or genre while the test set includes texts on another thematic

area or genre. Moreover, we make sure that the distribution of
texts over the candidate authors differs in training and test sets,
again to imitate realistic conditions. Two of the most successful
stylometric features are tested: frequent words and character
n-grams. Moreover, it is demonstrated that, when training and
test corpora have significant differences, the most crucial
decision concerns the appropriate selection of the representation
dimensionality (i.e., number of features). Based on the
experimental results, a set of general guidelines is provided to
tune an attribution model according to specific properties of
training and test corpora.
The next section compares the stylometric features we
examine. Section III describes the corpus used in this study
while Section IV includes the performed experiments. Finally,
Section V summarizes the main conclusions and proposes future
work directions.

II. FREQUENT WORDS VERSUS CHARACTER N-GRAMS

An intuitive way to quantify a text is based on frequencies of

occurrence of words. For authorship attribution, as well as any
style-based text categorization task, the most frequent words
16
have proved to be the most useful features. Interestingly, in
topic-related text categorization, very frequent words (e.g.,
articles, prepositions, conjunctions, etc.) are usually excluded
since they carry no semantic information. Hence, they are
frequently called “stopwords” or function words. There are two
main methods to define a set of such words to be used in an
authorship attribution model: 1) using a predefined list of words
belonging to specific closed-class parts of speech, such as
17
articles, prepositions, etc., or 2) using the most frequent words

16
Stamatatos, supra note 1, at 540.
17
Shlomo Argamon et al., Stylistic Text Classification Using Functional
Lexical Features, 58 J. AM. SOC’Y INFO. SCI. & TECH. 802, 803 (2007); see
also Ahmed Abbasi & Hsinchun Chen, Applying Authorship Analysis to
Extremist Group Web Forum Messages, IEEE INTELLIGENT SYS., Sept. 2005,
at 67, 68 (focusing on the use of lexical, syntactic, structural, and content-
specific features).
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 427

of the training corpus.18 In the latter case, the top words with
respect to their frequency correspond to function words. As we
descend the ranked list, we encounter more and more nouns,
verbs, and adjectives (possibly related with thematic choices).
One disadvantage of lexical features is that they fail to capture
any similarity in cases of noisy word forms (probably the result
of errors in language use). For example, “stylometric” and
“stilometric” are considered two different words. Another
shortcoming is that in some languages, mostly East Asian ones,
it is not easy to define what a word is.
Nowadays, character n-grams provide a standard approach to
represent texts. Each text is considered as a mere sequence of
characters. Then, all the overlapping sequences of n consecutive
characters are extracted. For example, the character 3-grams of
the beginning of this sentence would be “For,” “or,” “r e,”
“ex,” etc. Character n-gram features have several important
advantages: simplicity of measurement; language independence;
tolerance to noise (“stylometric” and “stilometric” have many
Figure 1: An example of an online article and the extracted main text.

18
J.F. Burrows, Not Unless You Ask Nicely: The Interpretative Nexus
Between Analysis and Information, 7 LITERARY & LINGUISTIC COMPUTING
91, 91–109 (1992).
428 JOURNAL OF LAW AND POLICY

common character 3-grams); effectiveness in authorship

attribution tasks, as has been proven in several studies and
competitions;19 and they require a high-dimensional
representation based on information difficult to understand by
humans, so deception attempts are less likely to be successful.
On the other hand, the high dimensional representation
requirement means that they can only be used in combination
with certain classification algorithms able to support thousands
of features. Furthermore, they capture small pieces of stylistic
information, making the interpretation of the stylistic property of
text very difficult if not impossible. Such an interpretation is
crucial in case the authorship attribution technology is used as
evidence in a judicial process.
Another common intuition is that character n-grams
unavoidably capture thematic information in addition to the
stylistic information. Under the assumption that all the available
texts are on the same thematic area, this property of character n-
grams can be viewed as an advantage since they provide a richer
representation including preference of the authors on specific
thematic-related choices of words or expressions (e.g., vehicle
vs. automobile). However, when the available texts are not on
the same thematic area, a topic-independent approach to
represent texts, like the use of a few dozen function words,
sounds more promising. In this paper we examine this
assumption and show that, contrary to intuition, character n-
grams are more robust features than frequent words when the
thematic area or the genre of the texts is not controlled.

III. THE GUARDIAN CORPUS

The corpus used in this study is composed of texts published

in The Guardian daily newspaper. The texts were downloaded
using the publicly available API20 and preprocessed to keep the
21
unformatted main text. An example is depicted in Table 1.

19
See Grieve, supra note 8, at 259; Vlado Keselj et al., N-Gram-Based
Author Profiles for Authorship Attribution, PROC. PAC. ASS’N FOR
COMPUTATIONAL LINGUISTICS, 2003, at 255, 255–64; Stamatatos, supra note
1, at 538–56; Stamatatos, supra note 9, at 237–41.
20
Open Platform, GUARDIAN, http://explorer.content.guardianapis.com/
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 429

Table 1: The Guardian corpus.

Opinion articles Book
Author
Politics Society World UK reviews
CB 12 4 11 14 16
GM 6 3 41 3 0
HY 8 6 35 5 3
JF 9 1 100 16 2
MK 7 0 36 3 2
MR 8 12 23 24 4
NC 30 2 9 7 5
PP 14 1 66 10 72
PT 17 36 12 5 4
RH 22 4 3 15 39
SH 100 5 5 6 2
WH 17 6 22 5 7
ZW 4 14 14 6 4
Total: 254 94 377 119 160
The majority of the corpus comprises opinion articles
(comments). The newspaper describes the opinion articles using
a set of tags indicating its subject. There are eight top-level tags
(World, U.S., U.K., Belief, Culture, Life&Style, Politics,
Society), each one of them having multiple subtags. It is
possible (and very common) for an article to be described by
multiple tags belonging to different main categories (e.g., a
specific article may simultaneously belong to U.K., Politics, and
Society). In order to have a clearer picture of the thematic area
of the collected texts, we only used articles that belong to a
single main category. Therefore, each article can be described
by multiple tags, all of them belonging to a single main
category. Moreover, articles coauthored by multiple authors
were discarded.
In addition to opinion articles on several thematic areas, the
presented corpus comprises a second text genre—book reviews.
The book reviews are also described by a set of tags similar to
the opinion articles. However, no thematic tag restriction was
taken into account when collecting book reviews, since our main
concern was to find texts of a specific genre that cover multiple

(last visited Mar. 2, 2013).

21
Titles, names of authors, dates, tags, images, etc. were removed.
430 JOURNAL OF LAW AND POLICY

thematic areas. Note that since all texts come from the same
newspaper, they are expected to have been edited according to
the same rules, so any significant difference among the texts is
not likely to be attributed to the editing process.
Table 1 shows details about The Guardian Corpus (“TGC”).
It comprises texts from thirteen authors selected on the basis of
having published texts in multiple thematic areas (Politics,
Society, World, U.K.) and different genres (opinion articles and
book reviews). At most 100 texts per author and category have
been collected—all of them published within a decade (from
1999 to 2009). Note that the opinion article thematic areas can
be divided into two pairs of low similarity, namely Politics-
Society and World-U.K. In other words, the Politics texts are
more likely to have some thematic similarities with World or
U.K. texts than with the Society texts.
TGC provides texts on two different genres from the same
set of authors. Moreover, one genre is divided into four
thematic areas. Therefore, it can be used to examine authorship
attribution models under cross-genre and cross-topic conditions.

IV. EXPERIMENTS

Two types of text representation features are examined—

namely, words and character 3-grams. In both cases, the
features are selected according to their total frequency of
occurrence in the training corpus, a method proven to be
suitable for authorship attribution tasks.22 Let V be the
vocabulary of the training corpus (the set of different words or
character 3-grams) and F = {f1, f2,..., fi,..., fv} be the set of
features ordered in decreased frequency of occurrence in the
training corpus. Given a predefined threshold t, the feature set Ft
includes all the features with fi ≥ t. The higher the t, the lower
the dimensionality of the representation and vice versa.
Therefore, it is possible to examine different sizes of the feature

22
John Houvardas & Efstathios Stamatatos, N-Gram Feature Selection
for Authorship Identification, in ARTIFICIAL INTELLIGENCE: METHODOLOGY,
SYSTEMS, AND APPLICATIONS 77, 82–84 (Jérôme Euzenat & John Domingue
eds., 2006).
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 431

set by modifying t. In this study, the following frequency

threshold values were used: 500, 300, 200, 100, 50, 30, 20, 10,
5, 3, 2, 1.
The well-known Support Vector Machine (“SVM”)
classifier23 is used. It is a powerful classification model that can
handle high dimensional and sparse data, and it is considered
one of the best algorithms for text categorization tasks. The
linear kernel (which is used to produce a linear boundary
between the classes) is used since the dimensionality of the
representation is usually high, including several hundreds or
thousands of features.24 There is no attempt to optimize the
classification model by using different classification algorithms,
since our aim is to highlight the capability of text representation
features to remain robust in cross-topic and cross-genre
conditions.
In each experiment, we follow the procedure described
below:
 An attribution model is learned based on SVM and texts
from a single topic category of TGC (e.g., Politics). At
most, ten texts per author are used in the training phase.
This provides an imbalanced training corpus.
 The learned classifier is applied to the texts of a category of
TGC. Again, at most ten texts per author are used. If the
selected category is Politics, that is the same as the topic
category used in the training phase (intratopic attribution).
The first ten texts are skipped, so there is no overlapping
with the texts used in the training corpus. If the selected
category is U.K., World, Society (cross-topic attribution) or
Books (cross-genre attribution), then an imbalanced test
corpus is compiled. Note that the distribution of the training
corpus over the candidate authors is not necessarily the same
with the corresponding distribution of the test corpus. This
ensures that in case the attribution model favors the authors
with the most training texts, it will produce many errors.

23
See Corinna Cortes & Vladimir Vapnik, Support-Vector Networks, 20
MACHINE LEARNING 273, 274–75 (1995).
24
See Thorsten Joachims, Text Categorization with Support Vector
Machines: Learning with Many Relevant Features, MACHINE LEARNING:
ECML-98: 10TH EUR. CONF. ON MACHINE LEARNING, 1998, at 137.
432 JOURNAL OF LAW AND POLICY

A. Intratopic Attribution

In the first experiment, we examine the simplest (but

unrealistic) scenario that all texts included in both training and
test corpora belong to the same genre and the same thematic
area. That way, the personal style of the author is more likely to
be the most significant factor for discriminating between texts.
Using TGC, the texts of the Politics thematic category were
used for both training and test (recall, there is no overlap
between training and test texts). The distribution of test texts
over the candidate authors is unavoidably similar to the
corresponding distribution of the training texts.
The classification accuracy results are shown in Figure 2 for
models based on frequent words and character 3-grams with a
varying number of features (acquired by the different values of
the frequency threshold). As can be seen, the models based on
character 3-grams are far more effective than models based on
words and achieve perfect classification accuracy. Their
performance seems to increase with the dimensionality of the
representation. This indicates that even the most rare character
n-grams carry information that help the classifier to discriminate
between author choices. Since all the texts are on the same
thematic area, these choices also include preferences of the
authors on specific thematic-related words or phrases.

100

90

80
Accuracy (%)

70
Words
60

50

40

30
0 5000 10000 15000
Features

Figure 2: Performance of the intratopic attribution models

(training on Politics, test on Politics).
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 433

As concerns models using word features, their performance

constantly increases until about 1,500 features, then drops a little
bit and then increases again. Hence, low-frequency words,
probably associated with thematic-related choices, provide useful
information to the classifier. In conclusion, when all the texts
are controlled in terms of genre and topic, it seems that a very
high dimensionality of the representation is a reliable option for
both character n-gram and word features.

B. Cross-Topic Attribution

Next, and more interestingly, we examine the cross-topic

scenario where the classifier is trained using the Politics texts
and then applied to the other thematic categories (that is,
Society, World, and U.K.) of the same genre. Recall that the
test texts distribution over the candidate authors does not follow
the corresponding distribution of the training texts. The results
are shown in Figures 3, 4, and 5, respectively.
In all three cases, character 3-gram features are significantly
more effective than words. When the topic of the test texts is
distant with respect to training texts (i.e., Society), the
performance steadily increases until about 3,500 features and
then significantly drops. In the cases of thematic areas unrelated
with the training texts (i.e., World and U.K.), there is a similar
pattern but the performance does not drop so much when the
dimensionality increases. This indicates that low frequency
features found in the training corpus (usually associated with
thematic information) should be avoided when the thematic area
of the test corpus is distant with respect to the thematic area of
the training corpus. On the other hand, these rare features are
not so crucial when the thematic area of the test corpus is not
specifically related to that of the training corpus. The best
performance is acquired by different frequency thresholds. In the
World texts the performance peak is at about 6,000 features
while in the U.K. texts the peak is at about 2,500 features.
Therefore, it seems that one very crucial decision in cross-topic
attribution to achieve high performance is the appropriate
selection of the number of features.
434 JOURNAL OF LAW AND POLICY

100

90

Accuracy (%) 80

70
Words
60
Char 3-
50 grams

40

30
0 5000 10000 15000
Features

Figure 3: Performance of the cross-topic attribution models

(training on Politics, test on Society).
100

90

80
Accuracy (%)

70
Words
60
Char 3-
50 grams

40

30
0 5000 10000 15000
Features

Figure 4: Performance of the cross-topic attribution models

(training on Politics, test on World).

The performance of the models based on word features has

similar characteristics. It steadily grows or remains practically
stable until about 1,500 features and then drops significantly.
The drop is much more abrupt in the case of Society texts
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 435

indicating that thematic-related words have a very negative effect

when the test texts are about a topic distant from that of the
training texts. In comparison to character n-grams, the word
features are far more vulnerable by low frequency features in
cross-topic conditions. Moreover, the models based on word
features achieve their best performance with about 1,000
features (Society), 1,500 features (World), and 250 features
(U.K.). Again, the appropriate selection of the dimensionality of
the representation seems to be crucial. In comparison to
character n-grams, word features need lower dimensionality to
achieve good results in cross-topic attribution.

C. Cross-Genre Attribution

Finally, we applied the classifier learned on opinion articles

about Politics to texts of another genre, book reviews. As with
the cross-topic experiments, the test set is imbalanced but its
distribution over the candidate authors does not follow that of
the training texts. The classification accuracy results for
attribution models based on word and character 3-gram features
are shown in Figure 6.
Again, character n-gram representation seems to be far better
than the word representation. The best achieved performance is
lower than all the best performances for the three cross-topic
experiments, indicating that cross-genre attribution is a more
difficult case. However, the average performance of the cross-
genre models is very close to the average performance of the
cross-topic models. Another interesting point is that the best
performance is achieved with considerably higher dimensionality
(about 9,000 features) with respect to the best performance of
the cross-topic attribution models. It seems that low frequency
features, probably related to thematic information, are helpful in
cross-genre conditions. Some of the book reviews included in
the test corpus may refer to books about Politics. Hence, when
text genre varies between training and test corpora, topic-related
choices may assist the attribution model.
436 100 JOURNAL OF LAW AND POLICY

90

80

Accuracy (%)
70

60 Words

50
Char 3-
grams
40

30
0 5000 10000 15000
Features

Figure 5: Performance of the cross-topic attribution models

(training on Politics, test on UK).
100

90

80
Accuracy (%)

70

Words
60

50 Char 3-
grams
40

30
0 5000 10000 15000
Features

Figure 6: Performance of the cross-genre attribution models

(training on Politics, test on Book reviews).

The models based on word features achieve their best

performance at about 400 features, far lower than the character
n-gram representation. However, the performance of the models
based on more features does not drop dramatically as happens in
cross-topic experiments. Again, this confirms the above
conclusion about the usefulness of thematic-related information
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 437

in cross-genre attribution. On the other hand, the appropriate

selection of the number of features is very important to achieve
the best possible results.

V. DISCUSSION

One main conclusion of this study is that, in addition to the

simple intratopic attribution, character n-grams produce models
more effective and robust than those based on word features in
both cross-topic and cross-genre conditions. In general, models
based on words require fewer features to achieve their best
results, but they are significantly inferior to the best models
based on character n-grams. An authorship attribution model
based on character 3-grams in combination with a SVM
classifier with linear kernel, although simple, proves to be very
effective and can be used as a baseline approach, with which
every new or advanced model should be compared.
The simple scenario of intratopic (in combination with
intragenre) attribution seems to be a relatively tractable problem
for current technology. The performance based on both
character n-grams and words is very high, and unlikely to be
matched by human experts, even when there are multiple
candidate authors and relatively short texts. However, taking
into account only such cases, the accuracy of the attribution
models may be overestimated.25 The presented cross-topic and
cross-genre experiments show that the performance is affected
sometimes considerably when topic and genre of training and
test texts are not controlled. On the other hand, in such difficult
cases, if the models are fine-tuned to the appropriate
dimensionality of the representation, then the classification
results remain surprisingly high. Hence, in the general case of
applying authorship attribution technology to real world
applications, a one-model-fits-all approach is not adequate.
According to the properties of the texts of known authorship and
the texts under investigation, one should fine-tune the attribution
models appropriately to maintain a high level of effectiveness.

25
See LUYCKX, supra note 8, at 4; Kestemont et al., supra note 14, at
343.
438 JOURNAL OF LAW AND POLICY

Several observations from the performed experiments may be

used as guidelines for tuning an attribution model:
 In intratopic attribution, a very high dimensionality of the
representation is advisable. Surely, high frequency features
are the most important. However, it seems that low
frequency features also contribute to the discrimination
ability of the model.
 In cross-topic attribution, if the topic is distant from the topic
of the training texts (e.g., Politics vs. Society, World vs.
U.K.), low frequency features should be avoided. Since they
are closely related with nuances of thematic choices, they
harm the effectiveness of the attribution models. The crucial
decision is the appropriate selection of the representation
dimensionality.
 In cross-topic attribution, if the topic is not specifically
associated to the topic of the training texts (e.g., Politics vs.
World), low frequency features are not so harmful.
However, it is better to exclude them, and again there is a
crucial decision about the appropriate selection of the
representation dimensionality.
 In cross-genre attribution, a high representation
dimensionality seems to be advisable, especially when topic
similarities are likely to be found in training and test texts.
An interesting conclusion that can be drawn from this study
is that cross-topic attribution where the topic of the training and
test texts can be regarded as highly dissimilar (e.g., Politics vs.
Society) may be more challenging than cross-genre attribution.
Additionally, in cross-genre attribution, perhaps
counterintuitively, models based on thousands of features (both
character n-grams and words) are either better than or
competitive with ones that use only a few hundreds of features.
Surely, more experiments are needed to verify all these
conclusions. An interesting direction for future work is to
explore the role of the candidate set size and how it affects the
appropriate representation dimensionality. The combination of
different feature types should also be examined since this
approach usually improves the performance of the attribution
models, as is exemplified by some of the most successful
participant methods in the recently organized competitions on
 ON THE ROBUSTNESS OF AUTHORSHIP ATTRIBUTION 439

authorship attribution.26 Finally, a missing block in the

authorship attribution research that is necessary to use this
technology as evidence in court is the ability to explain the
automatically derived decisions. In the case of attribution models
based on low-level information, like character n-grams, that
seem to be the most robust and effective approach, what is
needed is a way to associate this highly dimensional information
to some human interpretable high-level features.

26
See Shlomo Argamon & Patrick Juola, Overview of the International
Authorship Identification Competition at PAN-2011 (Sept. 19–22, 2011),
http://www.uni-weimar.de/medien/webis/research/events/pan-11/pan11-
papers-final/pan11-authorship-identification/juola11-overview-of-the-
authorship-identification-competition-at-pan.pdf; Patrick Juola, An Overview
of the Traditional Authorship Attribution Subtask Notebook for PAN at CLEF
2012 (Sept. 17–20, 2012), http://www.uni-weimar.de/medien/webis/research/
events/pan-12/pan12-papers-final/pan12-author-identification/juola12-overview-
of-the-traditional-authorship-attribution-subtask.pdf.