See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220433607

Computational Methods in Authorship Attribution

Article in Journal of the American Society for Information Science and Technology · January 2009
DOI: 10.1002/asi.20961 · Source: DBLP

CITATIONS READS
648 3,920

3 authors:

Moshe Koppel Jonathan Schler

Bar Ilan University Holon Institute of Technology
148 PUBLICATIONS 6,476 CITATIONS 47 PUBLICATIONS 4,116 CITATIONS

SEE PROFILE SEE PROFILE

Shlomo Argamon
Illinois Institute of Technology
121 PUBLICATIONS 7,636 CITATIONS

SEE PROFILE

All content following this page was uploaded by Moshe Koppel on 11 October 2017.

The user has requested enhancement of the downloaded file.

Computational Methods in Authorship Attribution

Moshe Koppel (Corresponding Author)

Dept. of Computer Science
Bar-Ilan University
Ramat-Gan, Israel
[email protected]

Jonathan Schler
Dept. of Computer Science
Bar-Ilan University

Shlomo Argamon
Dept. of Computer Science
Illinois Institute of Technology

1
 Computational Methods in Authorship Attribution

Abstract

Statistical authorship attribution has a long history, culminating in the use of modern machine learning

classification methods. Nevertheless, most of this work suffers from the limitation of assuming a small closed

set of candidate authors and essentially unlimited training text for each. Real-life authorship attribution problems,

however, typically fall short of this ideal. Thus, following detailed discussion of previous work, three scenarios

are considered here for which solutions to the basic attribution problem are inadequate. In the first variant, the

profiling problem, there is no candidate set at all; in this case, the challenge is to provide as much demographic

or psychological information as possible about the author. In the second variant, the needle-in-a-haystack

problem, there are many thousands of candidates for each of whom we might have a very limited writing sample.

In the third variant, the verification problem, there is no closed candidate set but there is one suspect; in this case,

the challenge is to determine if the suspect is or is not the author. For each variant, it is shown how machine

learning methods can be adapted to handle the special challenges of that variant.

1. Introduction

The task of determining or verifying the authorship of an anonymous text based solely on internal

evidence is a very old one, dating back at least to the medieval scholastics, for whom the reliable

attribution of a given text to a known ancient authority was essential to determining the text’s veracity.

More recently, this problem of authorship attribution has gained greater prominence due to new

applications in forensic analysis, humanities scholarship, and electronic commerce, and the

development of computational methods for addressing the problem.

In the simplest form of the problem, we are given examples of the writing of a number of

candidate authors and are asked to determine which of them authored a given anonymous text. In this

straightforward form, the authorship attribution problem fits the standard modern paradigm of a text

2
categorization problem (Lewis & Ringuette 1994, Sebastiani 2002). The components of text

categorization systems are by now fairly well-understood: documents are represented as numerical

vectors that capture statistics of potentially relevant features of the text and machine learning methods

are used to find classifiers that separate documents that belong to different classes.

However, real-life authorship attribution problems are rarely as elegant as straightforward text

categorization problems, in which we have a small closed set of candidate authors and essentially

unlimited training text for each. There are a number of varieties of attribution problems that fall short

of this ideal. For example, we may encounter scenarios such as the following:

1. There is no candidate set at all. In this case, the challenge is to provide as much demographic

or psychological information as possible about the author. This is the profiling problem.

2. There are many thousands of candidates for each of whom we might have a very limited

writing sample. This is the needle-in-a-haystack problem.

3. There is no closed candidate set but there is one suspect. In this case, the challenge is to

determine if the suspect is or is not the author. This is the verification problem.

Our goal in this paper is to survey the history of methods used for the basic authorship

attribution scenario and to discuss some recent solutions for the more complex variants mentioned

above.

In the following section, we offer a brief history of the analytical approaches to authorship

attribution, from nineteenth century work on statistical authorial invariants to recent application of

machine learning techniques. These modern techniques, together with recent advances in natural

language processing, have enabled the development of a plethora of potential markers of authorial

style, which we survey in Section 3. In Section 4, we describe the results of a systematic comparison of

3
learning algorithms and feature sets on several representative testbeds to determine the best

combinations for authorship attribution.

We then turn to consideration of variant scenarios where we do not have a small closed

candidate set. After giving an overview of the problems and approaches in Section 5, we consider the

profiling problem in Section 6, the needle-in-a-haystack problem in Section 7, and the verification

problem in Section 8. We summarize our findings in Section 9.

2. History of Methods

Over the last century and more, a great variety of methods have been applied to authorship attribution

problems of various sorts (cf. Juola 2008). For convenience, we divide them into three classes of

approach: the earliest, unitary invariant, approach, in which a single numeric function of a text is

sought to discriminate between authors, the multivariate analysis approach, in which statistical

multivariate discriminant analysis is applied to word frequencies and related numerical features, and

the most recent, the machine learning approach, in which modern machine learning methods are

applied to sets of training documents to construct classifiers that can be applied to new anonymous

documents.

2.1. Unitary Invariant Approach

A scientific approach to the authorship attribution problem was first proposed in the late nineteenth

century, in the work of Mendenhall (1887), who studied the authorship of texts attributed to Bacon,

Marlowe and Shakespeare, and of Mascol (1888a,1888b) who studied the authorship of the gospels of

the New Testament. The key idea was that the writing of each author could be characterized by a

unique curve expressing the relationship between word length and relative frequency of occurrence;

these characteristic curves thus would provide a basis for author attribution of anonymous texts. This

4
early work was put on a firmer statistical basis in the early twentieth century with the search for

invariant properties of textual statistics (Zipf 1932). The existence of such invariants suggested the

possibility that some related feature might be found that was at least invariant for any given author,

though possibly varying among different authors. Thus, for example, Yule (1944) considered sentence

length as a potential method for authorship discrimination, though he determined that this method was

not reliable. A number of other measures have been proposed as authorial markers (see Section 3.1

below), but for the most part this approach has not proved stable (Sichel 1986; Burrows 1992; Grieve

2007) and has given way to multivariate methods.

2.2. Multivariate Analysis Approach

Mosteller and Wallace’s work (1964) on the authorship of the Federalist Papers augured in a new set

of methods for stylometric authorship attribution, based on combining information from multiple

textual clues. Mosteller and Wallace (1964) applied a then-novel method of Bayesian classification to

the papers (essentially what is now called “Naïve Bayes” classification), using as features the

frequencies of a set of a few dozen function words (i.e., words with primarily grammatical functions

such as the, of, and about). The fundamental insight was that a rigorous Bayesian methodology, applied

to the frequencies of a set of topic-independent words, could yield a measurably reliable method for

attributing authorship. This opened up the field to the exploration of new types of textual features and

new modeling techniques.

A basic intuition behind these methods is that finding the most probable attribution can be

viewed as taking documents as points in some space, and assigning a questioned document to the

author whose documents are 'closest' to it, according to an appropriate distance measure. This simple

notion is quite powerful, so such distance measures continue to be used in recent studies examining the

efficacy of different metrics and feature sets. One such method is Burrows's (2002a) Delta, which has

5
been extended and used for a variety of attribution problems (Burrows 2002b; Hoover 2004a, 2004b),

and is equivalent to an approximate probabilistic ranking based on a multidimensional Laplacian

distribution over frequently appearing words (Stein & Argamon 2006; Argamon 2008). A number of

other similarity functions computed as distance measures for authorship attribution have been applied

to different feature sets as well (Craig 1999; Chaski 2001; Stamatatos et al. 2001; Keselj et al. 2003;

van Halteren et al. 2005; Burrows 2007). Recently, Grieve (2007) has run an exhaustive battery of tests

using this type of method.

A related class of techniques was developed earlier by Burrows (1987; 1989), who applied

principal components analysis (PCA) on word frequencies to analyze authorship. The idea is to

visualize the differences between texts written by different authors by projecting high-dimensional

word-frequency vectors computed for those text onto the 2-dimensional subspace spanned by the two

principal components; if good separation is seen between documents known to be written by different

authors, then new texts may be attributed by seeing which authors' comparison documents are closest

to them in this space. This method was elaborated on by Binongo and Smith (1999), and has been used

to resolve several outstanding authorship problems (Burrows 1992; Binongo 2003; Holmes 2003). A

related method is ANOVA, as applied, for example, by Holmes and Forsyth (1995) to the Federalist.

From the probabilistic standpoint, these methods take into account, to some extent, the statistical

dependence of different words' frequencies.

Another form of dependence between words is taken into account by methods that model the

sequencing of words in a document. This may be accounted for by using a probabilistic distance

measure such as K-L divergence between Markov model probability distributions of the texts (Juola

1998; Khmelev 2001; Khmelev and Tweedie 2002; Juola & Baayen 2003; Sanderson and Guenter

2006), possibly implicitly in the context of compression methods (Kukushkina et al. 2001; Benedetto et

al. 2002; Khmelev and Teahan 2003; Marton et al. 2005).

6
2.3. Machine Learning Approach

The emergence of text categorization techniques rooted in machine learning marked an important

turning point in authorship attribution studies. The application of such methods is straightforward:

training texts are represented as labeled numerical vectors and learning methods are used to find

boundaries between classes (authors) that minimize some classification loss function. The nature of the

learned boundaries depends on the learning method used but in any case these methods facilitate the

use of classes of boundaries that extend well beyond those implicit in methods that minimize distance.

Among the earliest methods to be applied were various types of neural networks, typically using

small sets of function word as features (Matthews & Merriam 1993; Merriam & Matthews 1994; Kjell

1994a; Lowe & Matthews 1995; Tweedie et al. 1996; Hoorn 1999; Waugh et al. 2000). More recently,

Graham et al. (2005) and Zheng et al. (2006) used neural networks on a wide variety of features. Other

studies used k-nearest neighbor (Kjell et al 1995; Hoorn et al. 1999; Zhao & Zobel 2005), Naive Bayes

(Kjell 1994a; Hoorn et al. 1999; Peng et al 2004), rule learners (Holmes & Forsyth 1995; Holmes

1998; Argamon et al. 1998; Koppel & Schler 2003; Abbasi & Chen 2005; Zheng et al. 2006), support

vector machines (De Vel et al. 2001; Diederich et al. 2003; Koppel & Schler 2003, Abbasi & Chen

2005; Koppel et al. 2005; Zheng et al 2006), Winnow (Koppel et al. 2002; Argamon et al. 2003;

Koppel et al. 2006a), and Bayesian regression (Genkin et al. 2006; Madigan et al. 2006; Argamon et al.

2008). Further details regarding these studies can be found in the Appendix.

Comparative studies on machine learning methods for topic-based text categorization problems

(Dumais et al. 1998; Yang 1999) have shown that in general, support vector machine (SVM) learning is

at least as good for text categorization as any other learning method and the same has been found for

authorship attribution (Abbasi & Chen 2005; Zheng et al. 2006). Some recent studies (Koppel et al.

2003; Genkin et al. 2006) have shown that some variations of Winnow and Bayesian regression are

7
also very promising. Below, we compare the performance of several representative learning methods

for authorship attribution. As we will see below, however, the choice of the learning algorithm is no

more important than the choice of the features by which the texts are to be represented. We discuss this

issue in the next section.

3. Features for Authorship Attribution

One of the advantages of modern machine learning methods is that they permit us to consider a wide

variety of potentially relevant features without suffering great degradation in accuracy if most of these

features prove to be irrelevant. In this section, we consider a number of feature types that have been, or

might be, used for the attribution problem. A number of earlier works that have surveyed and/or

compared various types of feature sets include Forsyth & Holmes (1996), Holmes (1998), McEnery &

Oakes (2000), Love (2002), Zheng et al. (2006), Abbasi & Chen (2008) and Juola (2008). We note that

in addition to the above-cited work dealing with attribution of texts in a variety of genres, there has

also been a fair amount of work on attribution of programming code, music, art and other media; we

regard that work as beyond the scope of this paper.

3.1. Complexity measures

As noted in Section 2.1, early work on authorship focused on the search for a single feature that

remained invariant for a given author but varied among different authors. The search for such

invariants centered on measures of text complexity. These measures included average word length (or

more generally, word length distribution) in terms of syllables (Fucks 1952) or letters (Mendenhall

1887; Brinegar 1963) and average number of words in sentence (Yule 1944; Morton 1965). When these

measures proved inadequate, more sophisticated measures were invented, involving type-token ratio

and the number of words appearing with given frequency in a text (such as hapax legomena). Among

8
the better known of these are Yule's K-measure (1944), Sichel's S-measure (1975), and Honore's R-

measure (1979). Ultimately, none of these measures has proved especially useful on its own (Burrows

1992; Grieve 2007), though it may be that these features have marginal value as additional inputs

together with the features that we consider below (De Vel et al. 2001; Corney et al. 2002; Abbasi &

Chen 2005; Zheng et al. 2006; Li et al. 2006; Abbasi & Chen 2008).

3.2. Function words

The search for a single invariant measure of textual style was natural in the early stages of stylometric

analysis, but with the development of more sophisticated multivariate analysis techniques, larger sets of

features could be considered. Among the earliest studies to use multivariate approaches was that of

Mosteller and Wallace (1964) noted above, who considered distributions of function words. The

reason for using function words in preference to others is that we do not expect their frequencies to

vary greatly with the topic of the text, and hence we may hope to recognize texts by the same author on

different topics. It is also unlikely that the frequency of function word use can be consciously

controlled, so one may hope that use of function words for attribution will minimize the risk of being

deceived (Chung & Pennebaker 200?).

Many studies since that of Mosteller and Wallace have shown the efficacy of function words for

authorship attribution in different scenarios (Morton 1978; Burrows 1987; Karlgren & Cutting 1994;

Merriam & Matthews 1994; Kessler et al. 1997; Argamon et al. 1998; Holmes 1998; de Vel et al. 2001;

Holmes et al. 2001a, 2001b; Baayen et al. 2002; Binongo 2003; Juola & Baayen 2003; Zhao & Zobel

2005; Argamon & Levitan 2005; Koppel et al. 2005, 2006a), confirming the hypothesis that different

authors tend to have different characteristic patterns of function word use.

Typical modern studies using function words in English use lists of a few hundred words,

including pronouns, prepositions, auxiliary and modal verbs, conjunctions, and determiners. Numbers

9
and interjections are usually included as well since they are essentially independent of topic, though

they are not, strictly speaking, function words. Results of different studies using somewhat different

lists of function words have been similar, indicating that the precise choice of function words is not

crucial. Discriminators built from function word frequencies often perform at levels competitive with

those constructed from more complex features.

3.3. Syntax and parts-of-speech

A different type of feature set is based on relative frequencies of different syntactic constructions,

made possible by development of fast and reliable statistical natural language processing techniques. A

number of studies used the output of syntactic text chunkers and parsers to create features, and found

that they could considerably improve results based on traditional word based analysis alone (Baayen et

al. 1996; Stamatatos et al. 2000, 2001; Gamon 2004; van Halteren 2004; Chaski 2005; Uzuner and

Katz 2005; Hirst & Feiguina 2007). Many studies have used the frequencies of short sequences of

parts-of-speech (or combinations of parts-of-speech and other classes of words) as a simple method for

approximating syntactic features for this purpose (Argamon-Engelson et al. 1998; Kukushkina et al.

2001; De Vel et al. 2001; Koppel et al. 2002; Koppel & Schler 2003; Chaski 2005; Koppel et al. 2005,

2006a; van Halteren et al 2005; Zhao et al. 2006; Zheng et al. 2006).

3.4. Functional lexical taxonomies

Function words and some parts-of-speech features can be subsumed by considering taxonomies, based

on Systemic Functional Linguistics (Halliday & Matthiessen 2003), which represent grammatical and

semantic distinctions between classes of function words at different levels of abstraction. Such

taxonomies are represented as trees whose roots are labeled by sets of parts-of-speech (articles,

auxiliary verbs, conjunctions, prepositions, pronouns). Each node’s childrens are labeled by meaningful

10
subclasses of the parent node (such as the various sorts of personal pronouns). This bottoms out at the

leaves, which are labeled by sets of individual words. These taxonomies can be used to construct

features for stylistic text classification as has been done for authorship attribution on texts in English

(Whitelaw et al. 2004; Argamon et al. 2007, 2008) and Portuguese (Pavelec 2007).

Such feature sets might include most function words and some POS unigrams, as well as

features at intermediate levels of abstraction. It is important to note that the features so constructed are

all closed sets of words so that no part-of-speech tagging is required for identifying such features in a

text.

3.5. Content words

There are aspects of authorial identity that are not easily captured by the sorts of stylistic features

described above. For example, one author may prefer to use the words start and large, where another

may prefer begin and big (Mosteller and Wallace 1964, Koppel et al. 2006a). Such patterns of lexical

choice can be represented by modeling the relative frequencies of content words (Martindale and

McKenzie 1995; Craig 1999; Waugh et al. 2000; Diederich et al. 2003; Hoover 2004a, 2004b;

Argamon et al. 2008). Typically very rare words and those with near-uniform distribution over the

corpus of interest can be omitted (Forman 2003), so that a reasonable set of perhaps several thousand

words may used. Sequences and collocations of content words can also be useful (Hoover 2002,

2003a, 2003b).

As noted earlier, the use of content-based features for authorship studies can be problematic.

Content markers might just be artifacts of a particular writing situation or experimental setup and might

thus produce overly optimistic results that will not be borne out in real-life applications. Thus, if one

author’s training documents are all on a particular topic, the trained classifier may do very poorly at

identifying documents by that author on a different topic. We are therefore careful in this paper to

11
distinguish results that exploit content-based features from those that do not.

3.6. Character n-grams

Several authors have proposed that the frequencies of various character n-grams might be useful for

capturing lexical preferences -- and even grammatical and orthographic preferences -- without the need

for linguistic background knowledge (making application to different languages trivial). Thus, Kjell

(1994a,1994b) and Kjell et al. (1995) used relative frequencies of character n-grams for attribution of

the Federalist papers and others have used character n-grams for authorship attribution of texts in

English (Ledger & Merriam 1994; Clement and Sharp 2003; Houvardas and Stamatatos 2006;

Stamatatos 2008), Dutch (Hoorn et al. 1999), Russian (Kukushkina 2001), Italian (Benedetto et al.

2002) and Greek (Keselj et al. 2003; Peng et al. 2004). Grieve (2007) has found that character bigrams

work surprisingly well for attribution of newspaper opinion columns. Chaski (2005, 2007) found

character n-grams to work well for attribution in a forensic context. Character n-grams have also been

shown useful for related stylistic classification tasks such as document similarity (Damashek 1995) or

determining the native language of the writer (Zigdon 2005), though Graham et al. (2005) found that

character n-grams did not work as well as syntax-based features for stylistic text segmentation. Zhang

and Lee (2006) find clusters of character n-grams that prove useful for a variety of text categorization

problems.

The caveats regarding content words apply also to the use of character n-grams, as many will

be closely associated to particular content words and roots.

3.7. Other specialized features

Some other features have been found useful for authorship and stylistic classification in particular

cases. Morphological analysis has been shown to be useful for authorship attribution in languages with

12
richer morphology than English, such as Greek (Stamatatos et al. 2001) and Hebrew (Koppel et al.

2006b), where many function words are represented by prefixes and suffixes.

For unedited texts, we might identify an author according to frequency of distinctive

punctuation habits (O’Donnell 1966; Chaski 2001) or orthographic/syntactic errors and idiosyncrasies

(de Vel et al. 2001, Koppel and Schler 2003). Thus, Koppel and Schler (2002) analyzed email texts by

running them through the MS-Word spelling and grammar checker, automatically assigning each error

found an “error type” such as repeated letter (e.g. remmit instead of remit), letter substitution (e.g. firsd

instead of first), letter inversion (e.g. fisrt instead of first), or conflated words (e.g stucktogether). This

approximates methods used in manual analyses of authorship whose goal is to identify idiosyncratic

characteristics of the author that can be recognized in a questioned text (Foster 2000).

Finally, for documents such as email, blogs and other online content, formatting and other

structural features can also be profitably exploited for authorship attribution (De Vel et al. 2001;

Corney et al. 2002; Abbasi & Chen 2008).

3.8. Summary

Above, we have described a wide variety of feature sets and analysis methods that have been applied to

various authorship attribution problems over the years. In principle, any feature set can be used with

nearly any classification method, provided proper methodology is followed in study design (cf.

Rudman 1997). In practice, however, certain combinations have been more often applied and studied.

As a reference, Appendix 1 contains a summary of methods and features sets that have been used in

different authorship studies.

13
4. Comparison Studies

In this section, we consider and compare methods and features applied to three authorship attribution

problems representative of the range of classical attribution problems. The corpora are as follows:

1. A large set of emails between two of the authors of this paper (Koppel and Schler),

covering the year 2005. The set consisted of 246 emails from Koppel and 242 emails

from Schler, each stripped of headers, named greetings, signatures and quotes from

previous posts in the thread. Some of the texts were as short as a single word. The

messages prior to July 1 were used for training and the second half for testing.

2. Two books by each of nine late 19th and early 20th century authors of American and

English literature (Hawthorne, Melville, Cooper, Shaw, Wilde, C. Bronte, A. Bronte,

Thoreau, Emerson). One book of each was used for training and the other for testing.

Each 500-word chunk in the test books was tested separately.

3. The full set of posts of twenty prolific bloggers, harvested in August 2004. The number

of posts of the individual bloggers ranged from 217 to 745 with an average of just over

250 words per post. The last 30 posts of each blogger were used as a test corpus.

As can be seen, these corpora differ along a variety of dimensions, including – most prominently – the

size of the candidate sets (2, 9, 20) and the nature of the material (emails, novels, blogs).

For each corpus, we run experiments comparing the effectiveness of various combinations of

feature types and machine learning methods. The feature types and machine learning methods that we

use are given in Table 1. Note that for each feature type that we consider, there are parameters that

need to be chosen. It is beyond the scope of this paper to determine the optimal parameter settings in

each case. We show results for plausible settings that earlier work, or our own preliminary tests,

suggest work reasonably well. Thus, for POS, we use all unigrams, sufficiently frequent bigrams and

no trigrams (Koppel & Schler 2003). We show results for SFL alone, but not in combination with FW

14
and/or POS, since the overlap of SFL with each of the other types is very large. We consider only

character trigrams, since these are long enough to capture morphology without mapping too obviously

to specific words. For both content words and character n-grams, we choose the 1000 features with

highest infogain from among those that are among the 10,000 most frequent in the corpus.

Each document in each corpus is processed to produce a numerical vector, each of whose

elements represents the relative frequency of some feature in the selected feature set. Models learned

on the training sets are then applied to the corresponding test sets to estimate generalization accuracy.

Table 2 shows results for each combination of features and learning method for the email corpus. Table

3 shows the results for the classic authors corpus. Table 4 shows results for the blog corpus.

As can be seen, Naïve Bayes, which we use as a representative for multivariate methods,

performs very poorly for all feature sets. Moreover, SVM and Bayesian regression are far superior to

the other learning algorithms, for all feature sets. Moreover, for these learners, SFL features are

approximately as good as function words and parts of speech together.

In the corpora we consider here, content words prove to be very useful; in no case, do they lead

us astray. Especially surprising is the effectiveness of the character n-gram feature set. Note that

character n-grams perform almost identically to content words for the first two corpora and

significantly outperform content words for the blog corpus. Consideration of some examples of useful

character n-grams suggests that character n-grams serve as proxies for content words (e.g., dsh for

spreadsheet), as well as for function words, parts-of-speech and even formatting (e.g., the string

colon—newline—1 suggests a numbered list). In the case of the blog corpus, character n-grams have

the additional benefit of capturing acronyms and abbreviations characteristic of blog writing.

We tentatively conclude, therefore, that when the context indicates that purely stylistic features

are appropriate, the combination of parts-of-speech and function words constitutes a reasonable choice

of feature set and that SFL features can be used as an efficiently proxy for this combination. When

15
content features are appropriate, properly chosen unigrams are a good choice, with similarly chosen

character tri-grams an efficient and language-independent proxy. Using these features, as appropriate,

in conjunction with either Bayesian regression or SVM, constitutes a convenient and effective method

for ordinary authorship attribution.

TABLES 1-4 ABOUT HERE

5. Variations on the Basic Attribution Problem

The attribution problem we have considered thus far is the standard one in which we are given a

relatively small closed set of candidate authors and are asked to determine which of them is the author

of a given document. In the sections that follow we consider three variations in which no small closed

candidate set is available.

First, we consider the case in which no candidate set is available at all so that the best we can

hope to do is to profile the anonymous author. We will see in Section 6 that essentially the same

methods that we used above for distinguishing individual authors can be used to distinguish between

classes of authors, such as males and females or writers of different ages. The discussion in Section 6 is

drawn from that in Argamon et al. (2008).

Next, we consider the case in which the candidate set consists of many thousands of authors so

that learning a classifier to distinguish them is infeasible. We will see in Section 7 that this problem can

be solved if we are willing to accept Don't Know as an answer for those cases where the document to

be attributed is not sufficiently distinct to permit attribution. We use meta-learning to identify such

cases and find that in the remaining cases, where the system believes attribution is reliable, we are able

to provide highly accurate results. The discussion is Section 7 is an expansion of that given in Koppel

et al. (2006c).

16
 Finally, we consider the case where there is a single candidate author and our task is to

determine if the anonymous document was written by that author. In Section 8, we show that this

problem is solvable if the anonymous text is sufficiently long. The method used entails measuring the

"depth" of the differences between the known texts of the candidate author and the anonymous text. In

particular, we check how accurately we can distinguish between the two as the best features for doing

so are iteratively eliminated. The discussion in Section 8 is drawn from Koppel et al. (2007).

6. Profiling

As noted above, even in cases where we have an anonymous text and no candidate authors, we would

like to say something about the anonymous author. That is, we wish to exploit the sociolinguistic

observation that different groups of people speaking or writing in a particular genre and in a particular

language use that language differently (cf. Chambers et al. 2004). More specifically, we wish to use the

features and methods used above to distinguish between individual authors in order to distinguish

between classes of authors.

As in Argamon et al. (2008), we consider the following profile dimensions: author gender

(Koppel et al. 2002; Argamon et al. 2003), age (Burger and Henderson 2006; Schler et al. 2006), native

language (Koppel et al. 2005) and neuroticism level (Pennebaker & King 1999; Pennebaker, Mehl, &

Niederhoffer, 2003). For each of these, we assemble an appropriately labeled corpus and proceed

exactly as described above. Thus, for example, we learn a classifier to distinguish between male and

female writers using the same procedure we used above to distinguish between individual authors.

Other authors have considered dimensions we do not consider here, such as education level (Corney et

al. 2002).

Following our observations in Section 4, we use here SFL as our stylistic feature set. For

comparison, we also consider content features alone and stylistic features and content features together.

17
The content features are chosen exactly as described in Section 4. We use Bayesian regression (BMR)

as our learning algorithm. For each of the three feature sets, we run ten-fold cross-validation tests to

test the extent to which each profiling problem is solvable. We also present the most discriminating

features for each category within each of the four problems.

6.1 Gender

Our corpus for both gender and age, first described by Schler et al. (2006), was assembled by taking as

an initial set all 47,000 blogs in blogger.com (as of August 2004) that self-reported both age and

gender, and included at least 200 occurrences of common English words. After dividing the set into age

intervals, we selected equal numbers of male and female bloggers in each age interval by randomly

eliminating surplus. The final corpus consists of the full set of postings of 19,320 blog authors (each

text is the full set of posts by a given author) ranging in length from several hundred to tens of

thousands of words, with a mean length of 7250 words per author.

Classification results for gender are shown in the first line of Table 5. As is evident, all feature

sets give effective classification, while the content features are slightly better than style features.

In the first line of Table 6, we show the most discriminating style and content features,

respectively, for gender. As can be seen, the style features most useful for gender discrimination are

determiners and prepositions (markers of male writing) and pronouns (markers of female writing). The

content features most useful for gender discrimination are words related to technology (male) and

words related to personal life and relationships (female). Earlier studies (Argamon et al. 2003) on

author gender in both fiction and non-fiction have shown that the style features found here to be useful

for blogs are strong discriminators in other types of text as well.

6.2 Age

18
Based on each blogger’s reported age, we label each blog in our corpus as belonging to one of three age

groups: 13-17 (42.7%), 23-27 (41.9%) and 33-47 (15.5%). Intermediate age groups were removed to

avoid ambiguity, since many of the blogs were written over a period of several years. Our objective is

to identify to which of these three age intervals an anonymous author belongs.

Accuracy results for age classification are shown in the second line of Table 5. Both style and

content features give us over 76% accuracy for this three-way classification problem, while the

baseline majority-class classifier would give an accuracy of just 42.7%.

The style features most useful for age classification (Table 6) are contractions without

apostrophes (younger writing), and determiners and prepositions (older writing). Note that the strongest

style features for 20s and 30s are identical; they are those that distinguish both of these classes from

teenagers. The content features that prove to be most useful for discrimination are words related to

school and mood for teens, to work and social life for 20s, and to family life for 30s.

6.3 Native Language

For the problem of determining an author’s native language, we use a portion of the International

Corpus of Learner English (Granger et al. 2002). All the writers in the corpus are university students

(mostly in their third or fourth year) studying English as a second language and assigned to the same

proficiency level in English. We consider 1290 texts in five sub-corpora, comprising 258 writers from

Russia, the Czech Republic, Bulgaria, France, and Spain, respectively. All texts in the resulting corpus

are between 579 and 846 words long. Our objective is to determine which of the five languages is the

native tongue of an anonymous author writing in English.

Accuracy results are shown in the third line of Table 5. Both style and content features give

results above 65%, well above the baseline accuracy of 20%.

In Table 6, we can see some consistent patterns of usage in the style features. For example, as

19
might be expected, native speakers of Slavic languages (Russian, Bulgarian, Czech) tend to omit the

definite article the which does not exist in these languages. (Since we list only features that are over-

represented in a given class, this feature is seen by examining the list of features for Spanish. Indeed,

many of the most discriminating features are those that are under-represented for particular languages.)

Furthermore, those words with commonly used analogs in a given language are used with greater

frequency by native speakers of that language, such as indeed (French), over (Russian), and however

(Bulgarian).

Elsewhere (Koppel et al. 2005), we have shown that for determining native language, features

that measure stylistic idiosyncrasies and errors are particularly useful. Using such features together

with the style features considered in this section yields classification accuracy of over 80% for this

task.

Regarding content words, it should be noted that, unlike the text collections used in the other

experiments described in this paper, writers in the learner corpus did not necessarily freely choose their

writing topics, so that differences in content word usage here are plausibly artifacts of the experimental

setup.

6.4 Personality

To examine the extent to which personality type can be determined from writing style, we use a corpus

of essays written by psychology undergraduates at the University of Texas. Students were instructed to

write a short “stream of consciousness” essay wherein they tracked their thoughts and feelings over a

20-minute free-writing period. The essays range in length from 251 to 1951 words. Each writer also

filled out a questionnaire testing for the “Big Five” personality dimensions: neuroticism, extraversion,

openness, conscientiousness, and agreeableness (John et al. 1991). To illustrate personality profiling,

we consider just the dimension of neuroticism; methods and results for other personality factors are

20
qualitatively similar. To formulate this as a classification problem, we define ‘positive’ examples to be

the participants with neuroticism scores in the upper third, and ‘negative’ examples to be those with

scores in the lowest third. The rest of the data are ignored; the final corpus consists of 198 writing

samples.

Accuracy results are shown in the fourth line of Table 5. Notably, style features give a great

deal of information about personality. An accuracy rate of 65.7% in detecting neuroticism is

surprisingly high; independent studies of individuals who attempted to guess others’ neuroticism levels

have given an average accuracy of 69% -- even among people who have known each other for several

years (Vazire, 2006).

As shown in Table 6, the most discriminating style features for this task suggest that neurotics

tend more to refer to themselves, to use pronouns for subjects rather than as objects in a clause, to use

reflexive pronouns, and to consider explicitly who benefits from some action (through prepositional

phrases involving, e.g., "for" and "in order to"); non-neurotics, on the other hand, tend to be less

concrete and use less precise specification of objects or events (determiners and adjectives such as "a"

or "little"), and to show more concern with how things are or should be done (via prepositions such as

"by" or "with" and modals such as "ought to" or "should").

In fact, classifiers learned using only the ten style features shown in Table 6 give classification

accuracy of 63.6%. More surprisingly, although the results in Table 5 indicate that content words

overall are useless for classifying texts by neuroticism, using as features the ten most informative

content features (those in Table 5) gives an accuracy of 68.2%. Apparently, the vast majority of content

is irrelevant to this classification problem and masks a small number of features involving worry about

personal problems (neurotics) and relaxation activities (non-neurotics) that are quite useful for this task.

TABLES 4,5 ABOUT HERE

21
7. Finding a needle-in-a-haystack

Consider now the scenario where we seek to determine the specific identity of a document's author, but

there are many thousands of potential candidates. We call this the needle-in-a-haystack attribution

problem. In this case, standard text-classification techniques are unlikely to give reasonable accuracy,

and may require excessive computation time to learn classification models. But we will show in this

section that if we are willing to tolerate our system telling us it doesn't know the answer, we can

achieve high accuracy for the cases where the system does give us an attribution it considers reliable.

The blogosphere forms a convenient testbed for this problem, as it provides us with text written

by an essentially unlimited number of authors. For this study we use the blog corpus described above

in Section 6, choosing the 20,000 longest blogs in our initial set. We took 10,000 blogs randomly to

create a test set of “snippets”, each snippet comprising enough of the most recent posts of a blog to

total at least 500 words; the remainder of that blog is termed the author’s “known work”. The other

10,000 blogs are held out for training purposes as is described below.

The goal is then to determine, for each snippet, to which of the 10,000 blogs it belongs, by

comparison to the various known works. We address the problem independently for each of the

snippets in the test set, i.e., we do not make use of the fact that there is a one-to-one correspondence

between the blogs and the snippets.

7.1. A simple attribution method

Learning classification models for a 10,000-class problem with thousands of features is impractical.

So, as a first approximation, we apply the standard information retrieval technique in which we define

some distance measure over meaningful textual features and attribute each snippet to the closest blog in

that feature space. Related approaches to authorship problems have been considered by Novak et al.

22
(2004) and Abbasi & Chen (2008).

We represent each text in four different ways: three varieties of tf-idf representations based on

the 1000 most frequent content features in the text and another based on a tf-idf representation based on

style features. For each of these representation methods, we use the standard cosine measure (Salton &

Buckley 1988) to quantify the similarity of each author's known work with a given snippet. The

various authors can then be ranked according to the similarity between their known works and the

snippet under consideration, with the hope being that the highest-ranked author is the author of the

snippet. The idea is that some distinctive feature or features might render the snippet particularly

similar to just one of the candidate authors.

This simple approach to the problem actually works surprisingly well. The three content

representations assign the snippet to the actual author between 52% and 56% of the time, while the

style representation lags behind with only 6% of snippets assigned to the actual author. 64% of the

snippets are most similar to their actual authors known works in at least one of the four representation

methods. There is thus a great deal of useful information here.

7.2 Meta-Learning

While 56% may seem to be quite a high level of accuracy, given the large number of candidates and the

simplicity of the method, it is also quite useless in the sense that we are still unable to confidently assert

that a given snippet was written by a given author; after all, the system is still wrong almost half of the

time. Thus we would like to automatically determine which attributions by which representation

schemes have a high likelihood of being correct; when none of them do, the system will report that

results are inconclusive. The goal is to return specific attributions as often as possible, while ensuring a

tolerably high level of accuracy for those cases.

To accomplish this, we apply a meta-learning scheme, using the holdout set of 10,000 blogs

23
(those not included in the test set), set aside for this purpose. We consider each pair consisting of a

snippet and an author ranked most similar to that snippet for at least one representation method, in a

given blog set (holdout or test). We call the pair a successful pair if the candidate author is in fact the

actual author. The pairs over the holdout set are used as training to learn a model that distinguishes

successful pairs from unsuccessful pairs. Each example (pair) is represented in terms of a set of meta-

features reflecting, for each representation, the similarity of the author to the snippet, both absolutely

and relative to other authors, and the author’s rank in similarity relative to other authors.

A linear SVM is used for each representation method to learn a “meta-model” that decides

whether a given pair is reliable or not. To do this, we use the meta-model to compute a reliability

score, which is a monotonic function with range [0,1] of the distance of the pair's representation from

the SVM margin.

Given reliability scores for each of the representations, the system chooses the attribution of the

highest-scoring representation, provided its reliability score is above a predetermined threshold.

Otherwise, the output is Don't Know. Varying this threshold will change the number of attributions

made and the accuracy of those attributions. This enables us to plot recall/precision curves (Figure 1,

upper curve), where recall is defined as the fraction of possible attributions (number of authors

represented by snippets and by known works, in this case 10,000) that were correctly attributed, and

precision is defined as the fraction of attempted attributions that were in fact correct. Note that, for

example, we can achieve recall of 40% with precision of 87%, but if we can settle for recall of 30%, we

can get precision of 94%.

FIGURE 1 ABOUT HERE

To test the sensitivity of these results for snippet length, we ran the experiment for snippets

24
limited to 200 words. In this case (see Figure 1, lower curve), at a recall level of 30% we achieve

precision of 86% and at recall of 40% we get precision of 73%.

Thus, provided we are willing to live with the response Don't Know in a number cases, we can

achieve reasonably reliable authorship attribution even where the number of candidate authors numbers

in the many thousands and the texts are rather short.

7.3 Unattributable texts

In the real world, however, we cannot assume that the author of a questioned text will in fact be

contained in our candidate set, even if that set is very large. To evaluate the performance of our

method in such a scenario, we randomly discarded 5,000 of the known works from the candidate set,

and evaluated performance on the original 10,000 snippets. Now, half of our test cases ought to result

in an output of Don’t Know in the best case, since their actual authors are not in the candidate set. The

precision/recall curve for this case is the lower curve in Figure 2 (note that recall here is defined as the

fraction of the 5,000 possible attributions that are correctly made) shown along with the original 600-

word curve for comparison. In this case, at a recall level of 30% we achieve precision of 81% and at

recall of 40% we get precision of 72%. Clearly, performance is noticeably degraded relative to the case

where all snippets have authors in the candidate set, though useful accuracy levels are still attainable.

We should note that as the number of alternative candidates becomes much smaller, the

problem might, somewhat counter-intuitively, become more difficult. This is because our method

implicitly leverages the fact that if a document is much more similar to one author’s writing than to

those of all others, it is very likely the document was written by that author. As the number of

alternative authors decreases, the reliability of such a conclusion will similarly decrease. Thus, in the

extreme case of authorship verification, where we are faced with a single candidate author, we need an

entirely different method. It is to this problem that we now turn.

25
 FIGURE 2 ABOUT HERE

8. Authorship Verification

Consider the case in which we are given examples of the writing of a single author and are asked to

verify that a given target text was or was not written by this author. As a categorization problem,

verification is significantly more difficult than basic attribution and virtually no work has been done on

it (but see van Halteren (2004)), outside the framework of plagiarism detection (Clough 2000; Meyer

zu Eissen et al. 2007). If, for example, all we wished to do is to determine if a text had been written by

Shakespeare or by Marlowe, it would be sufficient to use their respective known writings, to construct

a model distinguishing them, and to test the unknown text against the model. If, on the other hand, we

need to determine if a text was written by Shakespeare or not, it is difficult to assemble a representative

sample of non-Shakespeare texts.

The situation in which we suspect that a given author may have written some text but do not

have an exhaustive list of alternative candidates is a common one. The problem is complicated by the

fact that a single author may vary his or her style from text to text or may unconsciously drift

stylistically over time, not to mention the possibility of conscious deception. Thus we must learn to

somehow distinguish between relatively shallow differences that reflect conscious or unconscious

changes in an author’s style and deeper differences that reflect styles of different authors.

Verification can be thought of as a one-class classification problem (Manevitz & Yousef 2001,

Scholkopf et al. 2001, Tax 2001). But, perhaps, a better way to think about authorship verification is

that we are given two example sets and are asked whether these sets were generated by the same

process (author) or by two different processes. This section, drawn from Koppel et al. (2007), describes

a method for adducing the depth of difference between two example sets, which method may have far-

26
reaching consequences for determining the reliability of classification models. The idea is to test the

extent to which the accuracy of learned models degrades as the most distinguishing features are

iteratively removed from the learning process.

This method provides a robust solution to the authorship verification problem that is

independent of language, period and genre and has already been used to settle at least one outstanding

literary attribution problem (Koppel and Schler 2004; Koppel et al. 2007).

8.1 Authorship Verification: Naive Approaches

Let us begin by considering two naive approaches to the problem. Although neither of them will prove

satisfactory, each will contribute to our understanding of the problem.

One possibility that suggests itself is what we will call the “impostors” method: assemble a

representative collection of works by other authors and to use a two-class learner, such as SVM, to

learn a model for A vs. not-A. Then chunk the mystery work X and run the chunks through the learned

model. If the preponderance of chunks of X are classed as A, then X is deemed to have been written by

A; otherwise it is deemed to have been written by someone else.

This method is straightforward but it suffers from a conceptual flaw. While it is indeed

reasonable to conclude that A is not the author if most chunks are attributed to not-A, the converse is

not true. Any author who is neither A nor represented in the sample not-A, but who happens to have a

style more similar to A than to not-A, will be falsely determined by this method to be A. Despite this

flaw, we will see later that this approach can be used to augment other methods.

Another approach, which does not depend on negative examples, is to learn a model for A vs. X

and assess the extent of the difference between A and X by evaluating generalization accuracy by

cross-validation. If cross-validation accuracy is high, then conclude that A did not write X; if cross-

validation accuracy is low, i.e., we fail to correctly classify test examples better than chance, conclude

27
that A did write X. This intuitive method does not actually work well at all.

Let us consider exactly why the last method fails, by examining a real-world example. Suppose

we are given known works by three of the authors considered in Section 4, Herman Melville, James

Fenimore Cooper and Nathaniel Hawthorne. For each of the three authors, we are asked if that author

was or was not also the author of The House of Seven Gables (henceforth: Gables). Using the method

just described and using a feature set consisting of the 250 most frequently used words in A and X, we

find that we can distinguish Gables from the works of each author with cross-validation accuracy of

above 98%. If we were to conclude, therefore, that none of these authors wrote Gables, we would be

wrong: Hawthorne in fact wrote it.

8.2 A New Approach: Unmasking

If we look closely at the models that successfully distinguish Gables from Hawthorne’s other work (in

this case, The Scarlet Letter), we find that only a small number of features are doing all the work of

distinguishing between them. These features include he (more frequent in The Scarlet Letter) and she

(more frequent in Gables). The situation in which an author will use a small number of features in a

consistently different way between works is typical. These differences might result from thematic

differences between the works, from differences in genre or purpose, from chronological stylistic drift,

or from deliberate attempts by the author to mask his or her identity.

Our main point is to show how this problem can be overcome by determining not only if A is

distinguishable from X but also how great the depth of difference between A and X is. To do this we

use a technique we call “unmasking”. The idea is to remove, by stages, 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. Our main hypothesis is that if 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

28
features, despite possible differences in theme, genre and the like.

In Figure 3, we show the result of unmasking when comparing Gables to known works of

Melville, Cooper and Hawthorne. This graph illustrates our hypothesis: when comparing Gables to

works by other authors the degradation as we remove distinguishing features from consideration is

slow and smooth but when comparing it to another work by Hawthorne, the degradation is sudden and

dramatic. Once a relatively small number of distinguishing markers are removed, the two works by

Hawthorne become nearly indistinguishable.

This phenomenon is actually quite general, as we will show below. As we will also see, the

suddenness of the degradation can be quantified in a fashion optimal for this task. Thus by taking into

account the depth of difference between two works, we can determine if they were authored by the

same person or two different people.

FIGURE 3 ABOUT HERE

8.3 Test Corpus

We use as our corpus the collection of classic nineteenth and early twentieth century books considered

in Section 4 above. To break up the two-books-per-author pattern in the corpus, we add to the corpus,

one additional work by Melville and one by Hawthorne, as well as a work by Emily Bronte, who has

no other work in the corpus.

Our objective is to run 209 independent authorship verification experiments representing all

possible author/book pairs (21 books * 10 authors but excluding just the pair Emily Bronte/Wuthering

Heights which can’t be tested since it is the author’s only work).

As above, we partitioned each book into approximately equal-length sections of at least 500

29
words without breaking up paragraphs. For each author A and each book X, let AX consist of all the

works by A in the corpus unless X is in fact written by A, in which case AX consists of all works by A

except X. Our objective is to assign to each pair <AX,X> the value same-author if X is by A and the

value different-author otherwise.

8.4 Unmasking Applied

Now let us introduce the details of our new method based on our observations above regarding iterative

elimination of features. We choose as an initial feature set the n words with highest average frequency

in AX and X (that is, the average of the frequency in AX and the frequency in X, giving equal weight

to AX and X). Note that our objective here is not to maximize accuracy, but rather to measure the

degradation of accuracy; thus, it is enough to choose a simple feature set, rather the best possible one.

Using an SVM with linear kernel we run the following unmasking scheme:

1. Determine the accuracy results of a ten-fold cross-validation experiment for AX

against X. (If one of the sets, AX or X, includes more chunks than the other, we

randomly discard the surplus. Accuracy results are the average of five runs of ten-fold

cross-validation in which we discard randomly for each run.)

2. For the model obtained in each fold, eliminate the k most strongly weighted positive

features and the k most strongly weighted negative features.

3. Go to step 1.

In this way, we construct degradation curves for each pair <AX,X>. In Figure 4, we show such

curves (using n=250 and k=3) for An Ideal Husband against each of ten authors, including Oscar

Wilde.

30
 FIGURE 4 ABOUT HERE

8.5 Meta-learning: Identifying Same-Author Curves

We wish now to quantify the difference between same-author curves and different-author curves. To

do so, we first represent each curve as a numerical vector in terms of its essential features. These

features include, for i = 0,…,m:

• accuracy after i elimination rounds

• accuracy difference between round i and i+1

• accuracy difference between round i and i+2

• ith highest accuracy drop in one iteration

• ith highest accuracy drop in two iterations

We sort these vectors into two subsets: those in which AX and X are the by same author and those in

which AX and X are by different authors. We then apply a meta-learning scheme in which we use

learners to determine what role to assign to various features of the curves. (Note that although we have

20 same-author pairs, we really only have 13 distinct same-author curves, since for authors with

exactly two works in our corpus, the comparison of AX with X is identical for each of the two books.)

In order to assess the accuracy of the method, we use the following cross-validation

methodology. For each book B in our corpus, we run a trial in which B is completely eliminated from

consideration. We use unmasking to construct curves for all author/book pairs <AX,X> (where B does

not appear in AX and is not X) and then we use a linear SVM to meta-learn to distinguish same-author

curves from different-author curves. Then, for each author A in the corpus, we use unmasking to

31
construct a curve for the pair <AB,B> and use the meta-learned model to determine if the curve is a

same-author curve or a different-author curve.

Using this testing protocol, we obtain the following results: All but one (Pygmalion by Shaw) of

the twenty same-author pairs are correctly classified. In addition, 181 of 189 different-author pairs are

correctly classified. Among the exceptions are the attributions of The Professor by Charlotte Bronte to

each of her sisters. Thus, we obtain overall accuracy of 95.7% with errors almost identically distributed

between false positives and false negatives. (It should be noted that some of the 8 misclassified

different-author pairs result in a single book being attributed to two authors, which is obviously

impossible. Nevertheless, since each of our author/book pairs is regarded as an independent

experiment, we do not leverage this information.)

Note that the algorithm includes three parameters: n, the size of the initial feature set; k, the

number of eliminated features from each extreme in each iteration; m, the number of iterations we

consider. The results reported above are based on experiments using n=250, k=3, and m=10. We chose

n=250 because experimentation indicated that this was a reasonable rough boundary between common

words and words tightly tied to a particular work. In Koppel et al. (2007), it is shown that results are

somewhat robust with regard to choice of k and m (in fact, some parameter choices turn out to be better

than those shown here), but the recall results for same-author degrades considerably as the size of the

initial feature set increases. Moreover, parameter settings that proved successful on the English

literature corpus considered here also proved successful on a corpus of Hebrew legal writings, thus

demonstrating some degree of robustness over variation in language and genre.

It is further shown in Koppel et al. (2007) that unmasking can be augmented by exploiting

known negative examples, using the “impostors” method described above; the augmented method

classes all 189 different-author pairs and 18 of 20 same-author pairs correctly. Finally, one limitation

32
of unmasking that should be noted is that it requires a large amount of training text (Sanderson and

Guenter 2006) ; preliminary tests suggest that the minimum would be in the area of 5000 to 10,000

words..

In Figure 5, we summarize the entire algorithm (including the optional augmentation using

negative examples).

FIGURE 5 ABOUT HERE

9. Conclusions

We have surveyed the variety of feature types and categorization methods that have been proposed in

the past for authorship attribution. These methods range from early attempts to find individual

statistical markers that could serve as authorial fingerprints, through multivariate methods of varying

degrees of sophistication, and ultimately to text categorization methods rooted in machine learning. We

conclude that two of the most sophisticated machine learning methods, SVM and Bayesian regression,

used in conjunction with word classes derived from systemic functional linguistics or with character n-

grams, offer easily scalable, efficient and effective solutions to the ordinary authorship attribution

problem, assuming proper methodological controls for text genre and the like.

Since many realistic authorship problems don’t fit the standard attribution paradigm, we

consider also three variations that are likely to arise in practice. For the profiling problem, where no

individual candidates are known, we find that we can identify, with varying degrees of accuracy, an

author’s gender, age, native language and personality type. For the needle-in-a-haystack problem,

where there are possibly many thousands of candidate authors, we find that information retrieval

methods can be used to identify the correct author – of even very short texts – with high accuracy for

some considerable fraction of cases. These cases can be isolated using meta-learning methods that take

33
into account the degree to which a single author is more likely than any of the other candidates to be

the actual author. Finally, for the verification problem, where we need to determine of a given author

wrote a given text, we find that our unmasking technique is highly effective at identifying actual

authors, though it is limited to cases in which the attested text is sufficiently long.

34
References

Abbasi, A., and Chen, H. (2005), Applying authorship analysis to extremist-group Web forum messages, IEEE
Intelligent Systems,

Abbasi, A. and Chen, H. 2008. Writeprints: A stylometric approach to identity-level identification and similarity
detection. ACM Transactions on Information Systems (26:2), no. 7.

Argamon, S. (2008), Interpreting Burrows’s Delta: Geometric and probabilistic foundations, Literary and
Linguistic Computing, in press.

Argamon, S. and S. Levitan (2005), Measuring the usefulness of function words for authorship attribution. In
the Proceedings of the ACH/ALLC Conference, Victoria, BC, Canada, June 2005.

Argamon, S., Whitelaw, C., Chase, P., Hota, S., Garg, N., Levitan, S. (2007), Stylistic text classification using
functional lexical features, Journal of the American Society for Information Science and Technology 58(6),
802-821.

Argamon, S., Koppel , M., Fine, J. and Shimoni, A. (2003), Gender, Genre, and Writing Style in Formal Written
Texts, Text 23(3), August 2003.

Argamon, S., Koppel, M., Pennebaker, J. and Schler, J. (2008), Automatically Profiling the Author of an
Anonymous Text, Communications of the ACM , in press

Argamon-Engelson, S., Koppel, M., Avneri, G. (1998), Style-based text categorization: What newspaper am I
reading?, in Proc. of AAAI Workshop on Learning for Text Categorization, 1998, pp. 1-4

Baayen, H., van Halteran, H., Neijt, A., Tweedie, F. (2002), An Experiment in Authorship Attribution, Journees
internationales d'Analyse statistique des Donnees Textuelles 6.

Baayen, H., Van Halteren, H. and Tweedie, F.J. (1996), Outside the Cave of Shadows: Using Syntactic
Annotation to Enhance Authorship Attribution, Literary and Linguistic Computing, 11, 121-131.

Benedetto, D., Caglioti, E. and Loreto, V. (2002), Language Trees and Zipping, Phys. Rev. Lett. 88(4), 487-490

Binongo, J.N.G. and Smith, M.W.A. (1999), The application of principal component analysis to stylometry, Lit
Linguist Computing 14: 445-466

Binongo, J. N. G. (2003), Who wrote the 15th Book of Oz? An application of multivariate analysis to
authorship attribution. Chance 16(2), pp. 9-17.

Brill, E. (1992), A simple rule-based part-of-speech tagger, Proceedings of 3rd Conference on Applied Natural
Language Processing, pp. 152-155

35
Brinegar, C. S. (1963), "Mark Twain and the Quintus Curtius Snodgrass Letters: A Statistical Test of
Authorship," Journal of the American Statistical Association 58, pp. 85–96.

Burger J. and Henderson, J. (2006), An exploration of features for predicting blogger age. In AAAI Spring
Symposium on Computational Approaches to Analyzing Weblogs.

Burrows, J. F. (1992), Computers and the study of literature. In C. Butler, editor, Computers and Written Text,
Applied Language Studies, pages 167-204. Blackwell, Oxford.

Burrows, J.F. (1987), "Word Patterns and Story Shapes: The Statistical Analysis of Narrative Style", Literary
and Linguistic Computing, 2, 61-70.

Burrows, J.F. (1989), ‘An ocean where each kind..’: Statistical analysis and some major determinants of literary
style, Computers and the Humanities 23(4), 309-321.

Burrows, J.F. (1992), Not Unles You Ask Nicely: The Interpretative Nexus Between Analysis and Information,
Literary and Linguistic Computing 1992 7(2):91-109

Burrows, J.F. (2002a), Delta: a measure of stylistic difference and a guide to likely authorship, Liter-
ary and Linguistic Computing 17, pp. 267–287

Burrows, J.F. (2002b), The Englishing of Juvenal: Computational stylistics and translated texts,” Style 36, pp.
677–699.

Burrows, J. (2007), All the Way Through: Testing for Authorship in Different Frequency Strata , Literary and
Linguistic Computing 21, pp. 27-47

Chambers, J. K., P. Trudgill, and N. Schilling-Estes (2004), The Handbook of Language Variation and Change.
Blackwell, Oxford.

Chang, C.C. and Lin, C. (2001), LIBSVM: a Library for Support Vector Machines (Version 2.3)

Chaski, C. (2001), Empirical Evaluations of Language-Based Author Identification Techniques, Forensic

Linguistics 81: 1-65.

Chaski, C. (2005), “Who’s at the keyboard: Authorship attribution in digital evidence investigations,”
International Journal of Digital Evidence, vol. 4, no. 1, 2005

Chaski, C. (2007), "Multilingual Forensic Author Identification through N-Gram Analysis", Presented at the 8th
Biennial Conference on Forensic Linguistics/Language and Law, July 2007, Seattle, WA.

Chung, C.K. & Pennebaker, J.W. (2007). The psychological function of function words. In K. Fiedler (Ed.),
Social communication: Frontiers of social psychology (pp 343-359). New York: Psychology PressClement, R.
and Sharp, D. (2003). Ngram and Bayesian classification of documents. Literary and Linguistic Computing, 18:

36
423–47.

Clough, P. (2000), Plagiarism in natural and programming languages: an overview of current tools and
technologies, Research Memoranda: CS-00-05, Department of Computer Science, University of Sheffield, UK.

Corney, M., de Vel, O., Anderson, A. and Mohay, G. (2002), "Gender-Preferential Text Mining of E-mail
Discourse", in Proc. of 18th Annual Computer Security Applications Conference

Craig , H. (1999), Authorial attribution and computational stylistics: if you can tell authors apart, have you
learned anything about them?, Lit Linguist Computing 14: 103-113

Damashek, M. (1995), Gauging similarity with n-grams: language independent categorization of text. Science,
267(5199), 843--848.

De Vel, O., Anderson, A., Corney, M., Mohay, G. M. (2001), Mining e-mail content for author identification
forensics. SIGMOD Record 30(4), pp. 55-64

De Vel, O., M. Corney, A. Anderson and G. Mohay (2002), E-mail Authorship Attribution for Computer
Forensics, in Applications of Data Mining in Computer Security, Barbará, D. and Jajodia, S. (eds.), Kluwer.

Diederich, J., Kindermann, J., Leopold, E. and Paass, G. (2003), Authorship Attribution with Support Vector
Machines, Applied Intelligence 19(1), pp. 109-123

Dumais, S., J. Platt, J., Heckerman, D. and Sahami, M. (1998). Inductive learning algorithms and
representations for text categorization, Proceedings of ACM-CIKM98, 148-155

Forman, G. (2003), An extensive empirical study of feature selection metrics for text classification. Journal of
Machine Learning Research 3(1), pages 1289-1305.

Forsyth, R.S. and Holmes, D.I. "Feature-Finding for Text Classification", Literary and Linguistic Computing,
11, 4, (1996).

Foster, D. (2000), Author Unknown: On the Trail of Anonymous, New York: Henry Holt, 2000.

Fucks W. (1952). On the mathematical analysis of style. Biometrica 39, pp. 122-129.

Gamon, M. (2004), Linguistic correlates of style: authorship classification with deep linguistic analysis features.
In Proc. 20th Int. Conf. Computational Linguistics (COLING), pages 611–617, Geneva.

Genkin, A., Lewis, D. and Madigan, D. (2006), Large-scale Bayesian logistic regression for text categorization,
Technometrics

Graham, N., Hirst, G. and Marthi, B. (2005), ``Segmenting documents by stylistic character.'' Natural Language
Engineering, 11(4), December 2005, 397—415

37
Granger, S., Dagneaux, E., Meunier, F. (2002), The International Corpus of Learner English. Handbook and
CD-ROM. Louvain-la-Neuve: Presses Universitaires de Louvain

Grieve, J. (2007), Quantitative Authorship Attribution: An Evaluation of Techniques, Literary and Linguistic
Computing 22(3):251-270.

Halliday and Matthiessen (2003), An Introduction to Functional Grammar. Hodder Arnold.

Hirst, G. and Feiguina, O. (2007), Bigrams of syntactic labels for authorship discrimination of short texts,
Literary and Linguistic Computing, 22(4), pp. 405-417.

Holmes, D. (1998), The evolution of stylometry in humanities scholarship, Literary and Linguistic Computing,
13, 3, 1998, pp. 111-117.

Holmes, D. (2003), Stylometry and the Civil War, Chance 16(2)

Holmes, D. I., Gordon, L., Wilson, C. (2001a), A Widow and her Soldier: Stylometry and the American Civil
War, Literary and Linguistic Computing 16(4), pp. 403-420

Holmes, D. I., Robertson, M., Paez, R. (2001b), Stephen Crane and the New-York Tribune: A case study in
traditional and non-traditional authorship attribution. Computers and the Humanities 35(3) pp. 315-331.

Holmes, D. and Forsyth, R. (1995), The Federalist revisited: New directions in authorship attribution, Literary
and Linguistic Computing, pp. 111--127.

Honore (1979), Some Simple Measures of Richness of Vocabulary, Association for Literary and Linguistic
Computing Bulletin 7(2), pp. 172-177.

Hoorn, J., Frank, S., Kowalczyk, W., van der Ham, F. (1999), Neural network identification of poets using letter
sequences. Literary and Linguistic Computing, 14(3) pp. 311-338.

Hoover, D. L. (2002), Frequent Word Sequences and Statistical Stylistics, Literary and Linguistic Computing
17: 157–180

Hoover, D. L. (2003a), Frequent Collocations and Authorial Style, Literary and Linguistic Computing 18: 261–
286.

Hoover, D. L. (2003b), Multivariate Analysis and the Study of Style Variation, Literary and Linguistic
Computing 18: 341–360.

Hoover, D. L. (2003c), Another Perspective on Vocabulary Richness, Computers and the Humanities 37: 151–
178.

38
Hoover, D. (2004a), Testing Burrows’s Delta, Literary and Linguistic Computing 19(4), pp. 453–475,

Hoover, D. (2004b), Delta prime?, Literary and Linguistic Computing 19(4), pp. 477–495.

Houvardas, J. and E. Stamatatos (2006), N-gram feature selection for authorship identification, in Proc. of the
12th Int. Conf. on Artificial Intelligence: Methodology, Systems, Applications, pp. 77-86

Joachims, T. (1998), Text categorization with support vector machines: learning with many relevant features. In
Proc. 10th European Conference on Machine Learning ECML-98, pp. 137-142

John, O. P., (1990) , The “Big Five” factor taxonomy: Dimensions of personality in the natural language and in
questionnaires, in John, O. P. and L. A. Pervin, eds., Handbook of Personality: Theory and Research, Guilford
Press, pp. 66-100.

Juola, P. (1998). Cross-entropy and linguistic typology. In Proceedings of New Methods in Language
Processing 3. Sydney, Australia.

Juola, P. (2008), Author Attribution, Foundations and Trends in Information Retrieval, in press

Juola, P. & Baayen, H. (2005), A Controlled-corpus Experiment in Authorship Identification by Cross-Entropy,

Literary and Linguistic Computing 20(Suppl 1), pp. 59-67

Karlgren, J. and Cutting, D. (1994), Recognizing text genres with simple metrics using discriminant analysis. In
Proceedings of the 15th Conference on Computational Linguistics, Kyoto, Japan, August 1994, pp. 1071-1075.

Keselj, V., Peng, F., Cercone, N., Thomas, C. (2003), N- Gram-Based Author Profiles for Authorship
Attribution. In proceeding of PACLING'03, Halifax, Canada pp. 255-264.

Kessler, B., G. Nunberg, and H. Schütze (1997), Automatic detection of genre. In Proc. 35th Annual Meeting
of the Association for Computational Linguistics and the 8th Meeting of the European Chapter of the
Association for Computational Linguistics, pp. 32-38.

Khmelev D.V. (2001) Disputed Authorship Resolution through Using Relative Empirical Entropy for Markov
Chains of Letters in Human Language Text, Journal of Quantitative Linguistics, 7( 3), 201-207

Khmelev D. V., Teahan W. J. (2003), A repetition based measure for verification of text collections and for text
categorization, Proceedings of the 26th SIGIR conference, pp. 104-110

Khmelev, D. V., Tweedie, F. J. (2002), Using Markov chains for identification of writers. Literary and
Linguistic Computing, 16(4) pp. 299-307.

Kjell, B. (1994a), Authorship attribution of text samples using neural networks and Bayesian classifiers. In
IEEE International Conference on Systems, Man and Cybernetics, San Antonio, TX.

39
Kjell, B. (1994b), Authorship determination using letter pair frequencies with neural network classifiers.
Literary and Linguistic Computing, 9(2) pp. 119-124.

Kjell, B., Woods, W. A., Frieder, O. (1995), Information retrieval using letter tuples with neural network and
nearest neighbor classifiers. In IEEE International Conference on Systems, Man and Cybernetics, volume 2, pp.
1222-1225, Vancouver, BC.

Koppel, M., Akiva, N. and Dagan, I. (2006a), Feature Instability as a Criterion for Selecting Potential Style
Markers, Journal of the American Society for Information Science and Technology 57(11), pp. 1519-1525.

Koppel, M., Argamon, S. Shimoni, A. (2002), Automatically categorizing written texts by author gender,
Literary and Linguistic Computing 17(4), pp. 401-412

Koppel, M., Mughaz, D. and Akiva, N. (2006b), New Methods for Attribution of Rabbinic Literature, Hebrew
Linguistics: A Journal for Hebrew Descriptive, Computational and Applied Linguistics 57, Jan. 2006, pp. 5-18.

Koppel, M. and Schler, J. (2003), Exploiting Stylistic Idiosyncrasies for Authorship Attribution, in Proceedings
of IJCAI'03 Workshop on Computational Approaches to Style Analysis and Synthesis, pp. 69-72.

Koppel, M. and Schler, J. (2004), Authorship Verification as a One Class Classification Problem, in
Proceedings of ECML, Banff, Canada

Koppel M., Schler J. and Zigdon K. (2005), Determining an Author’s Native Language by Mining a Text for
Errors, Proceedings of KDD ’05, Chicago IL.

Koppel, M., Schler, J., Argamon, S. and Messeri,E. (2006c). Authorship Attribution with Thousands of
Candidate Authors, in Proc. 29th ACM SIGIR Conference on Research & Development on Information
Retrieval.

Koppel, M., Schler, J. and Bonchek-Dokow, E. (2007), Measuring Differentiability: Unmasking Pseudonymous
Authors, JMLR 8, pp. 1261-1276

Kukushkina, O. V., Polikarpov, A. A., and Khmelev, D. V. (2001), Using Literal and Grammatical Statistics for
Authorship Attribution, Probl. Inf. Transm. 37, 2 (Apr. 2001), 172-184

Ledger, G. and Merriam, T. (1994), Shakespeare, Fletcher, and the Two Noble Kinsmen, Lit Linguist
Computing 9: 235-248

Lewis, D.D. and Ringuette, M. (1994), Comparison of two learning algorithms for text categorization. In
Proceedings of the Third Annual Symposium on Document Analysis and Information Retrieval (SDAIR 94).

Li, J., Zheng, R., and Chen, H. 2006. From fingerprint to writeprint. Communications of the ACM (49:4), pp.
76-82.

40
Littlestone, N. (1988), Learning quickly when irrelevant attributes abound: A new linear threshold algorithm.
Machine Learning 2(4), pp. 285-318.

Love, H. (2002). Attributing Authorship: An Introduction. Cambridge University Press

Lowe, D. and Matthews, R.(1995), Shakespeare vs. Fletcher: A stylometric analysis by Radial Basis Functions.
Computers and the Humanities, 29 pp. 449-461.

Madigan, D., Genkin, A., Lewis, D.D., Argamon, S., Fradkin, D. & Ye, L. (2006), Author Identification on the
Large Scale, Proc. of Classification Society of N. America, 2005

Manevitz, L. M., Yousef, M. (2001), One-class svms for document classification,. Journal of Machine Learning
Research 2 pp. 139-154

Martindale, C. and McKenzie, D. "On the Utility of Content Analysis in Author Attribution: The 'Federalist'",
Computers and the Humanities, 29, 259-270, (1995).

Marton, Y. Wu, N. and Hellerstein, L (2005), On compression-based text classification, in Proceedings of the
27th European Conference on IR Research, pp. 300--314

Mascol, C. (1888a), Curves of pauline and pseudo-pauline style i. Unitarian Review, 30:452-460,1888.

Mascol, C. (1888b), Curves of pauline and pseudo-pauline style ii. Unitarian Review, 30:539-546,1888.

Matthews, R., Merriam, T. (1993), Neural computation in stylometry : An application to the works of
Shakespeare and Fletcher. Literary and Linguistic Computing, 8(4), pp. 203-209.

Matthiessen (1992), Lexicogrammatical Cartography, International Languages Sciences Publishers, Tokyo.

McEnery, A. and Oakes, M. (2000), Authorship studies/textual statistics, in R. Dale, H. Moisl, H. Somers eds.,
Handbook of Natural Language Processing (Marcel Dekker, 2000).

Mealand, D.L. (1995), Correspondence Analysis of Luke, Lit Linguist Computing 10: 171-182

Mendenhall, T. C. (1887), The characteristic curves of composition. Science 9, pp. 237-249.

Merriam, T. (1996), Marlowe's hand in Edward III revisited. Literary and Linguistic Computing, 11(1) pp. 19-
22.

Merriam, T. and Matthews, R. (1994), Neural compuation in stylometry II: An application to the works of
Shakespeare and Marlowe. Literary and Linguistic Computing 9, pp. 1-6.

Meyer zu Eissen, S., Stein, B. and Kulig, M. (2007), Plagiarism detection without reference collections, in R.
Decker and H. J. Lenz (eds.), Advances in Data Analysis, pages 359-366

Morton, A.Q., "The Authorship of Greek Prose", Journal of the Royal Statistical Society (A), 128, 169-233,

41
(1965).

Morton, A.Q., Literary Detection, New York:Scribners, (1978).

Mosteller, F., Wallace, D. L. (1964), Inference and Disputed Authorship: The Federalist. Reading, Mass.
Addison Wesley.

Novak, J., Raghavan, P., and Tomkins, A. 2004. Anti-aliasing on the web. In Proceedings of the 13th
International World Wide Web Conference, pp. 30-39

O’Donnell, B. (1966). Stephen Crane’s The O’Ruddy: A Problem In Authorship Discrimination. In Leed (ed.),
The Computer and Literary Style. Kent, OH: Kent State University Press, pp. 107–15.

Pavelec, D., Justino, E., and Oliveira, L. S. 2007. Author identification using stylometric features. Inteligencia
Artificial (11:36), pp. 59-65.

Peng, F., Schuurmans, D., ,Wang, S. (2004), Augumenting Naive Bayes Text Classifier with Statistical
Language Models , Information Retrieval, 7 (3-4), pp. 317 - 345

Pennebaker, J. W. and King, L. A. (1999), Linguistic Styles: Language Use as an Individual Difference. Journal
of Personality and Social Psychology, 77 (6) pp. 1296-1312.

Pennebaker, J.W., Mehl, M.R., & Niederhoffer, K. (2003). Psychological aspects of natural language use: Our
words, our selves. Annual Review of Psychology, 54, pp. 547-577.

Platt, J. (1998), Sequential minimal optimization: A fast algorithm for training support vector machines. In
Technical Report MST TR 98(14) Microsoft Research.

Quinlan, J.R. (1986), Induction of Decision Trees. Machine Learning 1(1), 81-106.

Rudman, J. (1997), The State of Authorship Attribution Studies: Some Problems and Solutions, Computers and
the Humanities 31(4), pp. 351-365.

Salton, G. and Buckley, C. (1988), Term-weighting approaches in automatic text retrieval, Information
Processing and Management: an International Journal 24(5), pp. 513-523.

Sanderson, C. and Guenter, S. (2006), Short Text Authorship Attribution via Sequence Kernels, Markov Chains
and Author Unmasking: An Investigation, in Int’l Conference on Empirical Methods in Natural Language
Processing, pp. 482-491

Schler, J. (2007), Authorship attribution in the absence of a closed candidate set, Ph.D. Dissertation, Dept. of
Computer Science, Bar-Ilan University, 2007.

Schler, J., Koppel, M., Argamon, S. and Pennebaker, J. (2006), Effects of Age and Gender on Blogging. In

42
proceedings of AAAI Spring Symposium on Computational Approaches for Analyzing Weblogs

Schölkopf, B., Platt, J., Shawe-Taylor, J., Smola, A. J., Williamson, R. C. (2001), Estimating the support of a
high-dimensional distribution. Neural Computation, 13, pp. 1443-1471.

Sebastiani, F. (2002), Machine learning in automated text categorization, ACM Computing Surveys 34(1), pp.
1-47.

Sichel, H. S., "Word frequency distributions and type-token characteristics", Mathematical Scientist, 11, pp. 45-
72, (1986).

Sichel, H.S., "On a Distribution Law for Word Frequencies", Journal of the American Statistical Association,
70, 542-547, (1975).

Stamatatos, E. 2008. Author identification: Using text sampling to handle the class imbalance problem.
Information Processing and Management (44:2), pp. 790-799

Stamatatos, E., Fakotakis, N., Kokkinakis, G. (2000), Automatic text categorization in terms of genre and
author, Computational Linguistics 26(4), pp. 471-495

Stamatatos, E., Fakotakis, N., Kokkinakis, G. (2001), Computer-based authorship attribution without lexical
measures, Computers and the Humanities 35, pp. 193-214.

Stein, S. and S. Argamon (2006), A mathematical explanation of Burrows’s Delta. In the Proceedings of the
Digital Humanities Conference, Paris, 2006.

Tax, D. M. J. (2001), One-class classification. PhD thesis, TU Delft.

Tweedie, F. J., Baayen, ,R. H. (1998), How Variable May a Constant Be? Measures of Lexical Richness in
Perspective, Computers and the Humanities, 32 (1998), 323-352.

Tweedie, S Singh, and D I Holmes (1996), Neural network applications in stylometry: The Federalist Papers.
Computers and the Humanities, 30(1):1-10.

Uzuner, O. and Katz, B. (2005), A Comparative Study of Language Models for Book and Author Recognition,
in Springer Lecture Notes in Computer Science, Vol. 3651, pp. 969-980

van Halteren, H. (2004) Linguistic profiling for authorship recognition and verification, Proc. of 42nd Conf. Of
ACL, July 2004, pp. 199-206

van Halteren, H., Baayen, H., Tweedie, F., Haverkort, M. and Neijt, A., New Machine Learning Methods
Demonstrate the Existence of a Human Stylome, Journal of Quantitative Linguistics 12(1): 65-77 (2005)

Vazire , S. (2006), Informant reports: A cheap, fast, and easy method for personality assessment. Journal of

43
Research in Personality 40(5), pp. 472-481.

Waugh, S., Adams, A., and Tweedie, F. J. (2000), Computational stylistics using Artificial Neural
Networks. Literary and Linguistic Computing, 15(2) pp. 187-198.

Whitelaw, C., Herke-Couchman, M. and Patrick, J. (2004), Identifying interpersonal distance using systemic
features, AAAI Spring Symp. on Exploring Attitude and Affect in Text

Witten, I. H., Frank, E. (2000), Data Mining: Practical Machine Learning Tools with Java Implementations.
Morgan Kaufmann, San Francisco.

Yang, Y. (1999), An evaluation of statistical approaches to text categorization. Journal of Information Retrieval,
1 (1-2), pp 67--88.

Yule, G. U. (1938), On Sentence Length as a Statistical Characteristic of Style in Prose with Application to Two
Cases of Disputed Authorship, Biometrika, 30, 363-390.

Yule, G. U. (1944), The Statistical Study of Literary Vocabulary. Cambridge University Press, Cambridge.

Zhang, D. and Lee, W. S. (2006), Extracting key-substring-group features for text classification, in Proc. of the
12th ACM Int’l Conference on Knowledge Discovery and Data Mining, pp. 474-483

Zhao, Y. & Zobel, J. (2005), Effective authorship attribution using function word, in 'Proc. 2nd AIRS Asian
Information Retrieval Symposium', Springer, pp. 174-190

Zhao, Y. and Zobel, J. (2007), Searching with style: authorship attribution in classic literature, Proc.of 30th
Australasian Conference on Computer Science, Vol. 62, pp. 59-68

Zhao, Y., Zobel, J. & Vines, P. (2006), Using relative entropy for authorship attribution, in ‘Proc. 3rd AIRS
Asian Information Retrieval Symposium’, pp. 92–105

Zheng, R., Li, J., Chen, H. and Huang, Z. (2006), “A framework for authorship identification of online
messages: Writing-style features and classification techniques,” Journal of the American Society for
Information Science and Technology, vol. 57, no. 3, pp. 378–393.

Zigdon, K. (2005), Automatically determining an author’s native language, M.Sc. Thesis, Dept. of Computer
Science, Bar-Ilan University, 2005

Zipf, G. K. (1932), Selected Studies of the Principle of Relative Frequency in Language. Harvard University
Press, Cambridge, MA.

44
FW a list of 512 function words, including conjunctions, prepositions, pronouns, modal verbs,
determiners and numbers.
POS 38 part-of-speech unigrams and 1000 most common bigrams using the Brill (1992) part-of-
speech tagger
SFL all 372 nodes in SFL trees for conjunctions, prepositions, pronouns and modal verbs, based
on Matthiessen (1992)
CW the 1000 words with highest information gain (Quinlan 1986) in the training corpus among
the 10,000 most common words in the corpus
CNG the 1000 character trigrams with highest information gain in the training corpus among the
10,000 most common trigrams in the corpus (cf. Keselj 2003)

NB WEKA’s implementation (Witten and Frank 2000) of Naïve Bayes (Lewis 1998) with
Laplace smoothing
J4.8 WEKA’s implementation of the J4.8 decision tree method (Quinlan 1986) with no pruning

RMW our implementation of a version of Littlestone’s (1988) Winnow algorithm, generalized to

handle real-valued features and more than two classes (Schler 2007)
BMR Genkin et al.’s (2006) implementation of Bayesian multi-class regression

SMO Weka’s implementation of Platt’s (1998) SMO algorithm for SVM with a linear kernel and
default settings

Table1. Feature types and machine learning methods used in our experiments

45
features/learner NB J4.8 RMW BMR SMO
FW 60.2% 58.7% 66.1% 68.2% 63.8%
POS 61.0% 59.0% 66.1% 66.3% 67.1%
FW+POS 65.9% 61.6% 68.0% 67.8% 71.7%
SFL 57.2% 57.2% 65.6% 67.2% 62.7%
CW 67.1% 66.9% 74.9% 78.4% 74.7%
CNG 72.3% 65.1% 73.1% 80.1% 74.9%
CW+CNG 73.2% 68.9% 74.2% 83.6% 78.2%
Table 2: Accuracy on test set attribution for a variety of feature sets and
learning algorithms applied to authorship classification for the email corpus.

features/learner NB J4.8 RMW BMR SMO

FW 51.4% 44.0% 63.0% 73.8% 77.8%
POS 45.9% 50.3% 53.3% 69.6% 75.5%
FW+POS 56.5% 46.2% 61.7% 75.0% 79.5%
SFL 66.1% 45.7% 62.8% 76.6% 79.0%
CW 68.9% 50.3% 57.0% 80.0% 84.7%
CNG 69.1% 42.7% 49.4% 80.3% 84.2%
CW+CNG 73.9% 49.9% 57.1% 82.8% 86.3%
Table 3: Accuracy on test set attribution for a variety of feature sets and
learning algorithms applied to authorship classification for the literature corpus.

features/learner NB J4.8 RMW BMR SMO

FW 38.2% 30.3% 51.8% 63.2% 63.2%
POS 34.0% 30.3% 51.0% 63.2% 60.6%
FW+POS 47.0% 34.3% 62.3% 70.3% 72.0%
SFL 35.4% 36.3% 61.4% 69.2% 71.7%
CW 56.4% 51.0% 62.9% 72.5% 70.5%
CNG 65.0% 48.9% 67.1% 80.4% 80.9%
CW+CNG 69.9% 51.6% 75.4% 86.1% 85.7%
Table 4: Accuracy test set attribution for a variety of feature sets and learning
algorithms applied to authorship classification for the blog corpus.

46
 Baseline Style Content Style+Content
Gender (2 classes) 50.0 72.0 75.1 76.1
Age (3 classes) 42.7 66.9 75.5 77.7
Language (5 classes) 20.0 65.1 82.3 79.3
Neuroticism (2 classes) 50.0 65.7 53.0 63.1

Table 5: Classification accuracy for profiling problems using different feature sets.

Class Style Features Content Features

Female personal pronoun, I, me, him, my cute, love, boyfriend, mom, feel
Male determiner, the, of, preposition-matter, as
system, software, game, based, site

Teens im, so, thats, dont, cant haha, school, lol, wanna, bored
Twenties preposition, determiner, of, the, in apartment, office, work, job, bar
Thirties+ preposition, the, determiner, of, in years, wife, husband, daughter, children

Bulgarian conjunction-extension, pronoun-interactant, bulgaria, university, imagination, bulgarian,

however, pronoun-conscious, and theoretical
Czech personal pronoun, usually, did, not, very czech, republic, able, care, started
French indeed, conjunction-elaboration, will, identity, europe, european, nation, gap
auxverb-future, auxverb-probability
Russian can’t, i, can, over, every russia, russian, crimes, moscow, crime
Spanish determiner-specific, this, going_to, because, spain, restoration, comedy, related, hardcastle
although
Neurotic myself, subject pronoun, reflexive pronoun, put, feel, worry, says, hurt
preposition-behalf, pronoun-speaker
Non- little, auxverbs-obligation, nonspecific reading, next, cool, tired, bed
determiner, up, preposition-agent
neurotic

Table 6: Most important Style and Content features (by information gain) for each class of texts in
each profiling problem.

47
Figure 1: Precision/Recall curves for attribution, adjusting the SVM threshold for deciding whether the
highest-scoring attribution should in fact be made. Upper curve is for snippets limited to 600 words
and lower is for snippets limited to 200 words. Recall (percentage of possible attributions correctly
made) is on the x-axis, and Precision (percentage of actual attributions correctly made) on the y-axis.

48
Figure 2: Precision/recall curves (as in Figure 1) for attribution of 10,000 snippets where all 10,000 are
theoretically attributable (upper curve) and where only 5,000 are theoretically attributable.

49
 100

90

80

70

60

50
0 1 2 3 4 5 6 7 8

Figure 3. Ten-fold cross-validation accuracy of models distinguishing House of Seven Gables from each of Hawthorne, Melville and
Cooper. The x-axis represents the number of iterations of eliminating best features at previous iteration. The curve well below the others
is that of Hawthorne, the actual author.

100

90

80

70

60
0 1 2 3 4 5 6 7 8

Figure 4. Unmasking An Ideal Husband against each of the ten authors (n=250, k=3). The curve below all the authors is that of Oscar
Wilde, the actual author. (Several curves are indistinguishable.)

50
Given: anonymous book X, works of suspect author A,
(optionally) impostors {A1,…,An}

Step 1 – Impostors method(optional)

if impostors {A1,…,An} are given then
{
foreach impostor Ai
{
Build model Mi for classifying A vs. Ai
Test each chunk of X with built model Mi
}
If for some Ai number of chunks assigned to Ai > number of chunks assigned
to A
then
return different-author
}
Impostors_Method_END

Step 2 - Unmasking
Build degradation curve <A,X>
Represent degradation curve as feature vector (see text)
Test degradation curve vector (see text)
if test result positive
return same-author
else
return different-author
Unmasking_END

Method Build Degradation Curve:

Use 10 fold cross validation for learning A against X
foreach fold
{
Do m iterations
{
Build a model for A against X
Evaluate accuracy results
Add accuracy number for iteration m to degradation curve <A,X> (as
average for all iterations m on current fold)
Remove k top contributing features (in each direction) from data
}
}
Method_END

Figure 5: Overview of the authorship verification algorithm.

51
Appendix. History of studies on authorship attribution problems. For each, we identify the corpus on
which methods were tested, the feature types used and the categorization method used.

(NB=Naïve Bayes; NN=neural nets; k-NN=k nearest neighbors; MVA=multivariate analysis; PCA=principle component analysis;
LDA=linear discriminant analysis)

Paper Reference Year Corpus Feature Types Classification

Method
Mendenhall 1887 Bacon/Marlowe/Shakespeare Sentence length, word length distance
Mascol 1888 (a, b) Pauline Epistles FW(10s), punctuation distance
Yule 1938 de Gerson Sentence length distance
Yule 1944 de Gerson Vocabulary richness(K- distance
measure)
Fucks 1952 English and German authors Word length distance
Brinegar 1963 QCS letters Word length distance
Mosteller & Wallace 1964 Federalist Papers FW(10s) NB
Morton 1965 Ancient Greek Prose sentence length distance
Burrows 1987 Austen/S.Fielding/H.Fielding FW(10s) MVA+PCA
Burrows 1992 (a) Brontes FW(10s) MVA+PCA
Matthews & Merriam 1993 Shakespeare/Fletcher FW(1s) NN
Kjell 1994 (a,b) Federalist Papers Character n-grams NN, NB
Merriam & Matthews 1994 Shakespeare/Marlowe FW(1s) NN
Ledger & Merriam 1994 Shakespeare/Fletcher character n-grams MVA
Holmes & Forsyth 1995 Federalist Papers FW(10s), vocabulary richness MVA, genetic algorithm
Kjell et al. 1995 WSJ character n-grams NN, k-NN
Lowe & Matthews 1995 Fletcher/Shakespeare FW(1s) RBF-NN
Martindale & McKenzie 1995 Federalist Papers words MVA+LDA, NN
Mealand 1995 Book of Luke FW(10s), POS MVA
Baayen et al. 1996 Federalist papers syntax NN
Merriam 1996 Shakespeare FW(1s) MVA+PCA
Tweedie et al. 1996 Federalist Papers FW(1s) NN
S. Argamon et al. 1998 newspapers&magazines FW(100s), POS n-grams ID3, Ripper
Tweedie & Baayen 1998 English prose FW(10s), vocabulary richness distance, MVA+PCA
Binongo & Smith 1999 Shakespeare FW(10s) MVA+PCA
Craig 1999 Middleton words distance
Hoorn et al. 1999 Dutch poets Character n-grams NN, NB, k-NN
Stamatatos et al. 2000 Greek newspapers Syntactic chunks distance
Waugh et al. 2000 Renaissance plays, Federalist words NN
Papers
Kukushkina et al. 2001 Russian texts Character n-grams, POS n- distance (Markov)
grams
Chaski 2001 Four women Syntax, punctuation, various distance
De Vel et al. 2001 emails FW(10s), complexity, various SVM
Holmes et al. 2001 (a) Pickett letters FW(10s) MVA+PCA
Holmes et al. 2001 (b) Crane articles (purported) FW(10s) MVA+PCA
Stamatatos 2001 Greek newspapers Syntactic chunks distance (LDA)
Baayen et al. 2002 Dutch texts FW(10s), syntax MVA+PCA
Benedetto et al. 2002 Italian texts Character n-grams distance (compression)
Burrows 2002 (a, b) Restoration-era poets FW(10s) MVA+PCA
Hoover 2002 novels and articles Words, word n-grams MVA
Khmelev & Tweedie 2002 Federalist papers, various character n-grams distance (Markov)
Binongo 2003 Oz books FW(10s) MVA+PCA
Clement & Sharp 2003 Movie reviews character n-grams NB
Diederich et al. 2003 German newspapers words SVM
Hoover 2003 (a) novels and articles words, word n-grams MVA
Hoover 2003 (b) Orwell/Golding/Wilde words, word n-grams MVA
Hoover 2003 (c) Novels vocabulary richness MVA

52
 Keselj et al. 2003 English novels, Greek character n-grams MVA
newspapers
Khmelev & Teahan 2003 Russian texts character n-grams distance (Markov)
Koppel & Schler 2003 Emails FW(100s), POS n-grams, SVM, J4.8
idiosyncrasies
Argamon et al. 2003 BNC FW(100s), POS n-grams Winnow
Hoover 2004 (a) American novels words MVA+PCA
Hoover 2004 (b) novels and articles words MVA+PCA
Peng et al. 2004 Greek newspapers character n-grams, word n- NB
grams
van Halteren 2004 Dutch texts Word n-grams, syntax MVA
Abbasi & Chen 2005 Arabic forum posts Characters, words, vocabulary SVM, J4.8
richness, various
Chaski 2005 10 anonymous authors character n-grams, word n- distance (LDA)
grams, POS n-grams, various
Juola & Baayen 2005 Dutch texts FW(10s) distance (cross-entropy)
Zhao & Zobel 2005 newswire stories FW(100s) NB, J4.8, k-NN
Koppel et al. 2005 Learner English FW(100s), POS n-grams, SVM
idiosyncrasies
Koppel et al. 2006a Brontes, BNC FW(100s), POS n-grams Balanced Winnow
Zhao et al. 2006 AP stories, English novels FW(100s), POS, punctuation SVM, distance
Madigan et al. 2006 Federalist papers characters, FW(100s), words, Bayesian regression
various
Zheng et al. 2006 English and Chinese characters, FW(100s), syntax, NN, J4.8, SVM
Li et al. newsgroups vocabulary richness, various
Argamon et al. 2007 novels and articles FW(100s), syntax, SFL SVM
Burrows 2007 Restoration poets Words MVA+zeta
Hirst & Feiguina 2007 Brontes syntax SVM
Pavelec et al. 2007 Portuguese newspapers conjunction types SVM
Zhao & Zobel 2007 Shakespeare, Marlowe, various FW(100s), POS, POS n-grams distance(infogain)
Abbasi & Chen 2008 emails, online comments, chats characters, FW(100s), syntax, SVM, PCA, other
vocabulary richness, various
Argamon et al. 2008 blogs, student essays, learner words, SFL Bayesian regression
English
Stamatatos 2008 English and Arabic news character n-grams SVM

53

View publication stats