Authorship Identification of Romanian Texts with Controversial Paternity
Liviu P. Dinu1,2, Marius Popescu1, Anca Dinu2,3
University of Bucharest
1Faculty of Mathematics and Computer Science,
2Centre for Computational Linguistics
3Faculty of Foreign Languages and Literature
Academiei 14, 010014 Bucharest, Romania
email@example.com, firstname.lastname@example.org, anca d email@example.com
In this work we propose a new strategy for the authorship identification problem and we test it on an example from Romanian literature:
did Radu Albala found the continuation of Mateiu Caragiale’s novel ”Sub pecetea tainei”, or did he write himself the respective contin-
uation? The proposed strategy is based on the similarity of rankings of function words; we compare the obtained results with the results
obtained by a learning method (namely Support Vector Machines -SVM- with a string kernel).
The authorship identification problem is a ancient chal-
lenge, and almost in every culture there are a lot of dis-
puted works. The problem of authorship identification is
based on the assumption that there are stylistic features that
help distinguish the real author from any other possibility.
Literary-linguistic research is limited by the human capac-
ity to analyze and combine a small number of text param-
eters, to help solve the authorship problem. We can sur-
pass limitation problems using computational and discrete
methods, which allow us to explore various text parame-
ters and characteristics and their combinations. Using these
methods van Halteren et al. (van Halteren et al., 2005)
have shown that every writer has a unique fingerprint re-
garding language use. The set of language use characteris-
tics - stylistic, lexical, syntactic - form the human stylom.
Computing and analyzing these language use characteris-
tics from various texts, we can solve text authorship prob-
Marcus (Marcus, 1989) identifies four situation in which
text authorship is disputed:
• A text attributedto oneauthor
which raises the
suspicion that there may be more than one author. If
the text was originally attributed to one author, one
must establish which fragments, if any, do not belong
to him, and who are their real authors.
• A text is anonymous. If the author of a text is un-
known, then based on the location, time frame and cul-
tural context, we can conjecture who the author may
be and test this hypothesis.
• If based on certain circumstances, arising from litera-
ture history, the paternity is disputed between two pos-
sibilities, A and B, we have to decide if A is preferred
to B, or the other way around.
• Based on literary history information, a text seems to
be the result of the collaboration of two authors, an
ulterior analysis should establish, for each of the two
authors, their corresponding text fragments.
The text characteristics and parameters used to determine
text paternity need not have aesthetic relevance. They must
be objective, un-ambiguously identifiable, and quantifiable,
such that they can be easily differentiated for different au-
In this paper we used two strategies to investigate one of
the most interesting experiments from Romanian literature.
The first strategy is based on Support Vector Machines
(SVM) with a string kernel (Section 2.). The second one is
a new strategy based on the similarity of rankings of func-
The novelty of our approach (Section 3.) resides in the
way we use information given by the function words fre-
quencies. Given a fixed set of function words (usually the
most frequent ones), a ranking of these function words ac-
cording to their frequencies is built for each text; the ob-
tained ranked lists are subsequently used to compute the
distance between two texts. To calculate the distance be-
tween two rankings we used Rank distance, a metric intro-
duced in (Dinu, 2003) and which was successfully used
in various fields as computational linguistics (in investigat-
ing the similarity of Romance languages (Dinu and Dinu,
2005)), bioinformatics (the similarity of DNA strings (Dinu
and Sgarro, 2006)), or multi-criteria classification (Dinu
and Popescu, to appear). Usage of the ranking of func-
tion words in the calculation of the distance instead of the
actual values of the frequencies may seem as a loss of infor-
mation, but we consider that the process of ranking makes
the distance more robust acting as a filter, eliminating the
noise contained in the values of the frequencies.
In the practical side of this project, we tested the upper
strategies to address the following situation from Romanian
literature. Mateiu Caragiale, one of the most important Ro-
manian novelists, died on 1936, at age of 51. In 1929 he be-
gun to works to the novel ”Sub pecetea tainei”, but unfor-
tunately he died before finishing this novel. Some decades
later, in the 70’s, a rumor has agitated the Romanian liter-
ary world: it seemed that it was founded the last part of
the novel ”Sub pecetea tainei”. Few human experts agreed
that the founded text is in concordance with Mateiu’s style,
and in the next months almost everybody talked about the
huge finding. We have to say that the one who claimed that
he has found the last part of the novel was an author (Radu
Albala) who’s literary style was the closest to Mateiu Cara-
giale, regarding all the successors of Mateiu. When Albala
sees that the claimed last part of novel passed the human
experts judgement, he stopped the discussions and said that
he is the real author of respective text, text which appears
latter with the name ”ˆIn deal, pe Militari” . In fact, this was
his challenge: to continue the unfinished novel of Mateiu.
In the following we will show that our methods distinguish
between the texts of Albala and the texts of Mateiu, and we
also show that Albala was closest to Mateiu’s style in the
first part of its continuation novel.
The goal of the classification experiments was to see if
the style of Albala can be distinguished from the style
of Mateiu in a supervised machine learning scenario. In
our experiments we followed the usual setting, treating the
problem as a binary classification problem. Each one of
the two alleged texts, ”Sub pecetea tainei”, and ”ˆIn deal,
pe Militari” had to be classified as being written by Mateiu
(class −1) or by Albala (class +1). For training were used
all the others works of the two authors. In order to have
a balanced training set, in terms of the number of exam-
ples for each author and the length of each example (text),
we treated each chapter of the Mateiu’s novel ”Craii de
Curtea-Veche” as a separate text. Thus, 5 negative exam-
ples (texts written by Mateiu: the 4 chapters of ”Craii de
Curtea-Veche” and the novella ”Remember”) and 5 posi-
tive examples (texts written by Albala: all the novellas pub-
lished by Albala excepting ”ˆIn deal, pe Militari”) resulted.
In Table 1 the title of each text, its author and its length (in
characters) are listed.
As learning method we used Support Vector Machines
(SVM) with a string kernel. String Kernels proved to be
effective in authorship attribution (Sanderson and Guenter,
2006; Popescu and Dinu, 2007) and because they treat text
as characters string they are language independent.
SVM learning algorithm works by embedding the data into
a feature space (a Hilbert space), and searching for linear
itly, that is by specifying the inner product between each
pair of points rather than by giving their coordinates ex-
plicitly. Details about SVM can be found in (Taylor and
The kernel function offers to the SVM the power to natu-
rally handle input data that are not in the form of numerical
vectors, such for example strings. The kernel function cap-
tures the intuitive notion of similarity between objects in a
specific domain and can be any function defined on the re-
spective domain that is symmetric and positive definite. For
strings, a lot of such kernel functions exist with many ap-
plications in computational biology and computational lin-
guistics (Taylor and Cristianini, 2004).
One of the most natural ways to measure the similarity of
two strings is to count how many substrings of length p
the two strings have in common. This give rise to the p-
spectrum kernel. Formally, for two strings over an alphabet
Σ, s,t ∈ Σ∗, the p-spectrum kernel is defined as:
where numv(s) is the number of occurrences of string v as
a substring in s1The feature map defined by this kernel as-
sociate to each string a vector of dimension |Σ|pcontaining
the histogram of frequencies of all its substrings of length
p. Taking into account all substrings of length less than p it
will be obtained a kernel that is called the blended spectrum
As in (Popescu and Dinu, 2007), in our experiments we
used the blended spectrum kernel. More precisely we used
a normalized version of the kernel to allow a fair compari-
son of strings of different length:
The reason for using this kernel is the fact that, in our opin-
ion, similarity of two strings as it is measured by string ker-
nels reflect the similarity of the two texts as it is given by
the short words (2-5 characters) which usually are function
suffixes (”ing” for example) which also can be good indi-
cators of author’s style.
Because the string kernels work at the character level, we
didn’t need to split the texts in words or to do any prepro-
cessing. The only editing done to the texts was the replac-
new line, etc.) with only one space character. This nor-
malization was needed in order to not increase or decrease
artificially the similarity between texts because of different
In all the experiments we used a normalized blended spec-
trum kernel of 5 characters,ˆk5
because it proved to be good in previous attribution tests
(Popescu and Dinu, 2007), but also because the most im-
portant style indicators in a text are function words which
usually are short (2-5 characters).
First we did cross validation in order to establish values
for parameters ν for SVM. Also the cross validation had
the role of estimating the generalization error of learning
methods used, or how reliable these methods are. The rel-
ative small number of training examples allowed us to use
leave one out cross validation which is considered an al-
most unbiased estimator of generalization error. Leave one
out technique consists of holding each example out, train-
ing on all the other examples and testing on the hold out
example. For value ν = 0.7 we obtained 0% leave one out
error for SVM.
1. The value of 5 was chosen
1Note that the notion of substring requires contiguity. See
(Taylor and Cristianini, 2004) for discussion about the ambiguity
between the terms ”substring” and ”subsequence” across different
traditions: biology, computer science.
Chapters of the novel
”Craii de Curtea-Veche”
ˆIntˆ ampinarea crailor
Cele trei hagialˆ acuri
Asfint ¸itul crailor
Sub pecetea tainei
Propyl¨ aen Kunstgeschichte
Nis ¸te cires ¸e
Femeia de la miezul nopt ¸ii
ˆIn deal, pe Militari
Length (in characters)
Table 1: Texts used in the experiments
Tested on the two texts in the test set, SVM correctly at-
tributed ”ˆIn deal, pe Militari” to Albala and ”Sub pecetea
tainei” to Mateiu, but the degrees of confidence of the two
predictions were different. ”Sub pecetea tainei” was at-
tributed to Mateiu with a probability of 62.56%, while ”ˆIn
deal, pe Militari” was attributed to Albala with a probabil-
ity of 50.56%. This very low confidence indicates that in
”ˆIn deal, pe Militari” Albala was very close (concerning
the style) to Mateiu.
We repeated the above experiment doing a different pre-
processing of the texts. Apart from normalizing spaces
(as in the previous experiment), we removed all punctua-
tion marks from the texts. All the other settings remained
exactly the same as in the previous experiment (the same
training set, the same kernelˆk5
out cross validation error was 0% for the SVM parameter
ν = 0.7.
Tested on the two texts this time, ”Sub pecetea tainei” was
again correctly attributed to Mateiu with a confidence of
66.87%, but ”ˆIn deal, pe Militari” was also attributed to
Mateiu with a confidence of 58.94%.
The role of punctuation in authorship identification prob-
lem was anticipated by Chaski (Chaski, 1996).
The punctuation (especially , and ; ) is the one who be-
trayed on Albala in his challenge to continue the novel of
Mateiu, and we can say that the Albala’s stylom is different
on the Mateiu’s stylom either on a alternative breathing of
Compared with other machine learning and statistical ap-
proaches, clustering was relatively rarely used in stylistic
investigations. However, few researchers (Holmes et al.,
2001; Labb´ e and Labb´ e, 2006; Luyckx et al., 2006) have
recently proved that clustering can be a useful tool in com-
putational stylistic studies.
An agglomerative hierarchical clustering algorithm (Duda
et al., 2001) arranges a set of objects in a family tree (den-
dogram) according to their similarity. The goal of the clus-
tering experiments was to see how the dendogram of the
works of the two authors look like (if the texts belonging
to one author are cluster together) and to see where in this
family tree are placed the two texts of interest, ”ˆIn deal, pe
Militari” and ”Sub pecetea tainei”.
1). Again, the leave one
In order to work, an agglomerative hierarchical clustering
algorithm needs to measure the similarity between objects
by a distance function defined on the set of respective ob-
In our experiments we used a new distance measure
(Popescu and Dinu, forthcoming) designed to reflect stylis-
tic similarity between texts. As style markers it use the
function word frequencies. Function words are generally
considered good indicators of style because their use is very
unlikely to be under the conscious control of the author and
because of their psychological and cognitive role (Chung
and Pennebaker, 2007). Also function words prove to be
very effective in many author attribution studies. The nov-
elty of the distance measure resides in the way it use the
information given by the function word frequencies. Given
a fixed set of function words (usually the most frequent
ones), a ranking of these function words according to their
frequencies is built for each text; the obtained ranked lists
are subsequently used to compute the distance between
two texts. To calculate the distance between two rankings
we used Rank distance (Dinu, 2003), an ordinal distance
tightly related to the so-called Spearman’s footrule (Diaco-
nis and Graham, 1977).
Usage of the ranking of function words in the calculation of
the distance instead of the actual values of the frequencies
may seem as a loss of information, but we consider that
the process of ranking makes the distance measure more
robust acting as a filter, eliminating the noise contained in
the values of the frequencies. The fact that a specific func-
tion word has the rank 2 (is the second most frequent word)
in one text and has the rank 4 (is the fourth most frequent
word) in another text can be more relevant than the fact that
the respective word appears 349 times in the first text and
only 299 times in the second.
Rank distance (Dinu, 2003) is an ordinal metric able to
compare different rankings of a set of objects.
A ranking of a set of n objects can be represented as a per-
mutation of the integers 1,2,...,n, σ ∈ Sn. σ(i) will rep-
resent the place (rank) of the object i in the ranking. The
Rank distance in this case is simply the distance induced by
Figure 1: Dendogram of the works of Mateiu and Albala
Figure 2: Dendogram of the works of Mateiu and Albala
|σ1(i) − σ2(i)|
This is a distance between what is called full rankings.
However, in real situations, the problem of tying arises,
when two or more objects claim the same rank (are ranked
equally). For example, two or more function words can
have the same frequency in a text and any ordering of them
would be arbitrary.
The Rank distance allocates to tied objects a number which
is the average of the ranks the tied objects share. For in-
stance, if two objects claim the rank 2, then they will share
the ranks 2 and 3 and both will receive the rank num-
ber (2 + 3)/2 = 2.5. In general, if k objects will claim
the same rank and the first x ranks are already used by
other objects, then they will share the ranks x + 1,x +
2,...,x+k and all of them will receive as rank the number:
= x +k+1
will be no longer a permutation (σ(i) can be a non integer
value), but the formula (1) will remain a distance (Dinu,
Rank distance can be used as a stylistic distance between
texts in the following way:
First a set of function word must be fixed. The most fre-
quent function words may be selected or other criteria may
be used for selection. In all our experiments we used a set
of 120 most frequent Romanian function words.
Once the set of function words is established, for each text
a ranking of these function word is computed. The ranking
is done according to the function word frequencies in the
text. Rank 1 will be assigned to the most frequent function
word, rank 2 will be assigned to the second most frequent
function word, and so on. The ties are resolved as we dis-
cussed above. If some function words from the set don’t
appear in the text, they will share the last places (ranks) of
The distance between two texts will be the Rank distance
between the two rankings of the function words corre-
sponding to the respective texts.
Having the distance measure, the clustering algorithm ini-
tially assigns each object to its own cluster and then re-
peatedly merges pairs of clusters until the whole tree is
formed. At each step the pair of nearest clusters is selected
for merging. Various agglomerative hierarchical clustering
algorithms differ in the way in which they measure the dis-
tance between clusters. Note that although a distance func-
tion between objects exists, the distance measure between
clusters (set of objects) remains to be defined. In our ex-
periments we used the complete linkage distance between
clusters, the maximum of the distances between all pairs of
objects drawn from the two clusters (one object from the
first cluster, the other from the second).
the same texts that we used in classification experiments
(Table 1). The resulted dendrogram is shown in Figure 1. It
is easy to see that Mateiu’s workss and Albala’s works are
clustered into two distinct groups, and, that the two investi-
gated texts are placed in their corresponding branch.
2. In this case, a ranking
To see if indeed Albala wanted to write in the matein style,
we made an ad-hoc experiment: we concatenated the last
part of the novel ”Sub pecetea tainei” with the first part of
the ”ˆIn deal, pe Militari” and used this artificial text in the
same experiment as the previous one. The result was (see
Figure 2) that the new ad-hoc text is placed in the Mateiu’s
branch. Conclusion is that Albala wrote in the beginning of
the ”ˆIn deal, pe Militari” as Mateiu, but his concentration
decreased towards the end of the novel and eventually his
stylom was detectable.
The authorship identification problem is a ancient chal-
lenge, and almost in every culture there are a lot of dis-
puted papers. In this work we proposed a new strategy for
the authorship identification problem and we have tested on
an example from Romanian literature: Radu Albala found
the continuing of Mateiu Caragiale’s novel ”Sub pecetea
tainei”, or he write himself the respective continuing? The
answer is that Albala write himself the continuing.
Research supported by MEdC-ANCS, PNII-Idei, project
228. We want to thank to S ¸tefan Agopian for fruitful dis-
cussions and to Humanitas Ed. for the electronic source of
Carole E. Chaski. 1996. Linguistic methods of determin-
ing authorship. In National Institute of Justice Research
Seminar. 49th American Academy of Forensic Sciences
Meeting, Nashville, TN.
CindyK.ChungandJamesW.Pennebaker. 2007. Thepsy-
chological function of function words. In K. Fiedler, ed-
itor, Social communication: Frontiers of social psychol-
ogy, pages 343–359. Psychology Press, New York.
P. Diaconis and R.L. Graham. 1977. Spearman’s footrule
as a measure of disarray. Journal of the Royal Statistical
Society, Series B (Methodological), 39(2):262–268.
Anca Dinu and Liviu Petrisor Dinu. 2005. On the syllabic
similarities of romance languages. In CICLing, Lecture
Notes in Computer Science, pages 785–788.
Liviu Petrisor Dinu and Marius Popescu. to appear. A
multi-criteria decision method based on rank distance.
Liviu Petrisor Dinu and Andrea Sgarro. 2006. A low-
complexity distance for dna strings. Fundamenta Infor-
Liviu Petrisor Dinu. 2003. On the classification and aggre-
gation of hierarchies with different constitutive elements.
Fundamenta Informaticae, 55(1):39–50.
R. O. Duda, P. E. Hart, and D. G. Stork. 2001. Pattern
Classification (2nd ed.). Wiley-Interscience Publication.
David I. Holmes, Lesley J. Gordon, and Christine Wilson.
2001. A widow and her soldier: Stylometry and the
american civil war. Literary and Linguistic Computing,
Cyril Labb´ e and Dominique Labb´ e. 2006. A tool for liter-
ary studies: Intertextual distance and tree classification.
Literary and Linguistic Computing, 21(3):311–326.
Kim Luyckx, Walter Daelemans, and Edward Vanhoutte. Download full-text
2006. Stylogenetics: Clustering-based stylistic analysis
of literary corpora. In Proceedings of LREC-2006, the
fifth International Language Resources and Evaluation
Conference, pages 30–35.
Solomon Marcus. 1989. Inventie si descoperire. Ed.
Cartea Romaneasca, Bucuresti.
Marius Popescu and Liviu P. Dinu. 2007. Kernel methods
and string kernels for authorship identification: The fed-
eralist papers case. In Proceedings of the International
Conference RANLP - 2007 (Recent Advances in Natural
Language Processing), Borovets, Bulgaria, September.
Marius Popescu and Liviu P. Dinu. forthcoming. Rank dis-
tance as a stylistic similarity.
Conrad Sanderson and Simon Guenter. 2006. Short text
authorship attribution via sequence kernels, markov
chains and author unmasking: An investigation. In Pro-
ceedings of the 2006 Conference on Empirical Methods
in Natural Language Processing, pages 482–491, Syd-
ney, Australia, July. Association for Computational Lin-
John S. Taylor and Nello Cristianini. 2004. Kernel Meth-
ods for Pattern Analysis. Cambridge University Press,
New York, NY, USA.
Hans van Halteren, M. Haverkort, H. Baayen, A. Neijt,
and F. Tweedie. 2005. New machine learning methods
demonstrate the existence of a human stylome. Journal
of Quantitative Linguistics, 12:65–77.