The multifactor method has been developed to enhance test validity of authorship attribution. The method is style based. An author of a text can be identified by three major factors: a style based factor, a topic based factor and an authorial style based factor. For each factor certain statistical parameters are determined. The statistical parameters are actual distributions of frequencies of occurrence of the researched language units. As the research is done on the phonological level, the language units are phonemes. To differentiate texts by different authors, the powerful statistical tests have been applied (the KolmogorovSmirnov's test, the chi-square test, the Student's t-test).
Anonymous information in the Internet has always been an important problem for
researchers. The necessity to solve this problem is growing as the number of people using
this network is increasing. The anonymous texts occur in different areas of
communication. In certain cases the information doesn’t bother anyone and can be negligible.
But it does disturb when it threatens, harasses and is inappropriate. If the reader
considers the information personal and sensitive, the anonymous or given under a
pseudonym information must be studied. It should be noted that plagiarism detection is closely
connected with disputed authorship. It is particularly important to determine the real
author of a text of the same or similar content, but having several authors. In certain
cases it is necessary to establish the author of the text written a long time ago, not clear
when. The task of determining an author involves some text categorization and
classification. The choice of most optimal classifiers and feature sets depends on various
factors. The size of the text under study is of importance. The method efficient for the
short texts may be inefficient for long texts. Each language level and feature set on this
level have their specificity. The methods selected for a certain language level and
feature sets are sure to behave differently. The number of features in a set may be increased
or decreased, depending on the applied method. The problem of authorship attribution
implies inferring a certain style, a certain topic and an authorial style. A text by an
author can be of different complexity. The simplest case is when two studied texts are
of the same style, genre and topic. The most complex case is when the texts are of
different style, genre and topic. There also intermediate cases with greater or less
similarity or difference. The three mentioned factors lie in the basis of the developed
multifactor method. Each factor effect is determined by three efficient statistical methods
– the Kolmogorov-Smirnov’s test, the chi-square test and the Student’s t-test. The
purpose of the research is to enhance test validity of authorship attribution with the help of
the multifactor method. According to the results of recent research, different authorship
attribution approaches have been used. Thus, in the field of digital text forensics,
informal chat conversations have been researched. The algorithmic solutions have been
obtained with 72,7%, 75% accuracy [
The Method Developed
The problem of authorship attribution is aimed at determining if two compared pieces
of text were written by a single author. Texts from poetry (G. Byron, T. Moore) and the
publicist style (B. Оbama, D. Trump, D. Webster, S. Logan) have been selected for
experiments. An author of a text can be revealed by three major factors: a style based
factor, a topic based factor and an authorial style based factor. The style based factor
consists of showing the difference between the two styles, the topic based factor – the
difference between the two texts on different topic, the authorial style based factor –
the difference between the two texts by different authors. The scheme is style – topic –
author. The average value of the three factor based values is considered general style
markedness of a text. The steps of the multifactor method algorithm are given below.
More detailed information regarding the steps of the Student’s t-test, the
KolmogorovSmirnov’s test and the chi-square test was presented in the previous research [
1) the Student’s t-test is performed for the texts from different styles, on different
topics and by different authors [
nm n m
Dn,m nm
sup Fn (z) Fm (z) .
n m z a a t x1 x2
S n1 n2 , n1 n2 Dn,m sup Fn (z) Fm (z) ,
z a a where x1 x 2 is a difference of average frequencies for two samples for the fixed group of consonants a, S is a variance, n is a sample size.
2) the Kolmogorov-Smirnov’s test is performed for the texts from different styles,
on different topics and by different authors [
s ˆ n2 i1 j1
k i, j
2 n j j
k n , j ij j1 where i, j is a realization number in j-th series, s is a number of consonant groups, k is a number of samples, n j is a number of sample portions, n is a number of portions for two samples. The texts can be differentiated if ˆ n2 12 ,(s1)(k1) . a a 4) determining the style based factor value 1 for a phoneme group a: x1s x 2 s (s is a style); a a 5) determining the topic based factor value 2 : x1t x 2 t ( is a topic); a a 6) determining the authorial style based factor value 3 ∶ x1а x 2 а (а is an author); 7) determining the general style markedness: 8) the authorship attribution is calculated by the difference of values of the general style markedness for two authors. To develop the software for authorial differentiation, the Java programming language has been used. The programming language is cross-platform and this is an advantage of the chosen programming language. The developed program system realizes the following algorithm (Fig. 1):
The structure of the developed software has the following tabs: “Text”, “Transcription Symbols”, “Consonant Phoneme Sample”, “Portion Division”, “Group Division”, “Calculating Phonemes in Portions”, “Calculating Phonemes in Groups”, “Statistical Test”, “Style Based Factor Value”, “Topic Based Factor Value”, “Authorial Style Based Factor Value”, “General Style Markedness Values”, “Difference by General Style Markedness Values”.
The software classes are shown in the diagram s in Fig. 2.
One of the advantages of this program system is its relative independence of the transcription site on which an English text is transcribed. The transcription site is used when the first experiments are made. The bag of words gets larger every time new texts are processed. Therefore, it is advisable to process large samples. These are some short documents in the forensic field. However, large samples are of interest when it is necessary to characterize literary legacy of some author. Such problems are usually researched in corpus linguistics. In this investigation, both short and long texts are analyzed. The sample size is 50 000 phonemes and more. 3
For the first experiment of author identification, two pieces of poetry have been selected: one by G. Byron and another by T. Moore. According to the proposed scheme, the two pieces of poetry must be analyzed in a comparison with some style having most common bag of words. Evidently, this may be the conversational style in its literary version which has few colloquial elements. This is particularly relevant in the comparison with poems by romanticists who tried to use conversational elements. On the other hand, it is necessary to compare the poems with another genre of the style of fiction. It may be Byron’s emotive prose. These two samples are sure to have common language units both being of fiction style. In the third stage of the study, the poems by the two poets are compared. The multifactor method makes it possible to calculate the average value of the values calculated for each mentioned above comparisons. The average value is general style markedness (Table 1).
Having calculated the value of general style markedness for each poet, the author identification test can be performed. The difference of the two authorial styles is calculated by the difference of values general style markedness (Table 2). It is equal to 1.8.
m getNames() 1 1 C ConsonantUtils m countConsonantTypes(String) Map<ConsonantType, Long> m countConsonants(String) Map<String, Long> m countConsonantTypes(Map<String, Long>) Map<ConsonantType, Long>
m processConsonants(String, String) MultiValueMap<ConsonantType, Long> C PropertyUtils m getltem(String) m saveltem(String, String) m builder() +
create C Builder m path(String) Builder m build() PropertyUtils
void Builder ϟ NoSuchConsonantException m NoSuchConsonantException() m NoSuchConsonantException(char) Comparison with the Comparison with Comparison with Value of general 1-st style the 2-nd style another author style markedness GB-СLS – 16 GB-GBP – 17 GB-TM – 7 GB – 13.3 TM-СLS –15 TM-GBP –13 TM-GB –7 TM –11.5 In Table 1 the following designations are used: GB is Byron’s poetry, TM is Moore’s poetry, GBP is Byron’s emotive prose, СLS is the conversational style. For the second experiment four authors have been selected. They are: B. Оbama, D. Trump, D. Webster, S. Logan. The pieces of writing represent the publicist style. In this case the multifactor method involves comparison with the 1-st author, comparison with the 2-nd author and comparison with the 3-rd author. The average value calculated from the three values got in each comparison is the general style markedness by which the author can be identified. Though the samples are of the same style, the topic varies from sample to sample having common bag of words. The topic reflects international political events all over the world. Therefore, the content is relatively homogeneous. This relative homogeneity creates a problem of its own. The topic based factor affects the final result of author identification. Sometimes it is rather difficult to draw a distinct demarcation line between effect of the topic based factor and the authorial style based factor. The case is easier with documents following strict standards of conveying information. The same situation can hardly be observed in the publicist style. On the other hand, the publicist style is the style in which the individual peculiarities of an author’s manner of writing can be vividly revealed. The effect of the three factors mentioned is expressed in the value of the general style markedness given in Table 3. The value of general style markedness is the highest for Webster’s authorial style. It equals to 16. The lowest value is for Obama’s writing style (14). But, as a significant role is played here by the topic based factor, the authorial writing characteristics can be different in another verbal content. Thus the essential information about author identification can be got with the help of the multifactor method. The results of comparisons of texts by different authors are shown in Table 4. Table 4 shows that different effect of the topic and author based factors causes great difference in a comparison Оbama-Webster, less difference – Оbama-Trump, ОbamaLogan, still less difference – Trump-Webster, Webster-Logan and practically no difference – Trump-Logan. The last pair of texts shows similarity of bag of words.
Among the used statistical tests two are the most powerful. These are the Kolmogorov-Smirnov’s test and the chi-square test. With the help of the former, the authorial styles differ essentially in all eight phoneme groups. The latter is a little less powerful – the difference in six of eight groups.
The efficiency of the multifactor method may be analyzed for each of eight groups of phonemes. Reduction of the number of phoneme groups makes the whole procedure more economical. Consequently, it is necessary to analyze author-differentiating capability for every group.
The degree of author-differentiating capability of phoneme group depends on the number of times the essential differences have been established. If the essential difference has been revealed by three statistical tests, the phoneme group takes number 3, by two statistical tests – number 2, by one statistical test – number 1. In Table 5, for labial group, number 3 (Оbama-Logan) has been got once, number 1 – thrice. The group takes the second degree of differentiation power. Compared with labial group, dorsal group has higher degree of differentiating capability (Table 6). Only in two pairs of texts one statistical test has proved efficient. More differences have been obtained by two tests. The group takes the first degree of differentiation power.
The results of the conducted research have shown that the multifactor method is efficient in authorship attribution. The established general style markedness of a text has made it possible to classify each sample under study in accordance with three basic factors – style, topic and author’s manner of writing. Taking into account the three mentioned factors is particularly efficient when the compared samples represent different style and topic. In this case it is impossible to characterize the authorial specificity of writing because of the influence of style and topic factors. Having determined the style based and topic based features, the authorial features can be identified. 4
In order to single out particular features of an individual writing style, the style based features and topic based features must be separated. To solve this task, the multifactor method must be applied. In accordance with this method, the average value of the three factor based values is calculated. The three factor based values involve: comparison Оbama-Trump Оbama-Webster Оbama-Logan Trump-Webster Trump-Logan Webster-Logan Byron-Moore 1 1 3 2 2 2 2 with the text least marked by the style elements, comparison with the text of the same topic, but different author and comparison with the text of the same style and topic, but another author. The average value of these three values is general style markedness. The author identification is calculated by the difference of general style markedness values. The results show that the greatest difference is in the pair Obama – Webster (2), less difference – in the pairs Obama – Trump, Obama – Logan (1.6), still less – in the pairs Trump – Webster, Webster – Logan (0.4), the least in the pair Trump – Logan. The test validity has been enhanced up to 95%, 97%.
The developed software on the Java programming language has performed the author identification procedure in a fewer number of consonant groups making it more automated. The next step in our research will be concentrated on the other statistical methods.