Methods and software for significant indicators determination of the natural language texts author profile
Methods for the formation and optimization of author profiles are presented. The author profile is an image – a vector in a multidimensional space, which components are author’s texts measurements by a number of methods based on 4-grams, stemming, recurrence analysis and formal stochastic grammar. T...
Збережено в:
| Дата: | 2023 |
|---|---|
| Автори: | , |
| Формат: | Стаття |
| Мова: | English |
| Опубліковано: |
Інститут програмних систем НАН України
2023
|
| Теми: | |
| Онлайн доступ: | https://pp.isofts.kiev.ua/index.php/ojs1/article/view/577 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | Problems in programming |
| Завантажити файл: |  |
Репозитарії
Problems in programming| id |
pp_isofts_kiev_ua-article-577 |
|---|---|
| record_format |
ojs |
| resource_txt_mv |
ppisoftskievua/3c/5c85b2aa7fa6858811d3ff627885803c.pdf |
| spelling |
pp_isofts_kiev_ua-article-5772024-04-28T11:55:00Z Methods and software for significant indicators determination of the natural language texts author profile Методи та засоби визначення значимих показників профілю автора природно-мовних текстів Shynkarenko, V.I. Demydovych, I.M. natural language texts; authorship determination; genetic algorithm; recurrent analysis; statistical analysis; text classification; pattern recognition; formal grammars UDK 004.91 природомовні тексти; визначення авторства; генетичний алгоритм; рекурентний аналіз; статистичний аналіз; класифікація текстів; розпізнавання образів; формальні граматики УДК 004.91 Methods for the formation and optimization of author profiles are presented. The author profile is an image – a vector in a multidimensional space, which components are author’s texts measurements by a number of methods based on 4-grams, stemming, recurrence analysis and formal stochastic grammar. The author’s profile is a model of his language, including vocabulary, sentence syntax features. A comparative analysis of each of the methods effectiveness is carried out. By means of the genetic algorithm, a reduced profile of the author is formed. Insignificant indicators are excluded, which allows to reduce their number by 20%. The reduced author’s profile contains attributes that are significant for this author and is an effective attribution of a particular author.Prombles in programming 2023; 3: 22-29 Наведено методи формування та оптимізації профілів авторів. Профіль автора це образ – вектор у багатовимірному просторі, компоненти якого є вимірами текстів автора рядом методів на основі 4-грам, стемування, рекурентного аналізу та формальної стохастичної граматики. Профіль автора є моделлю його мови, включаючи словниковий запас, особливості синтаксису речень. Здійснюється порівняльний аналіз ефективності кожного із методів. Засобами генетичного алгоритму формується усічений профіль автора. Виключаються незначні показники, що дозволяє скоротити їхню кількість на 20%. Усічений профіль автора містить значущу для даного автора атрибутику і є ефективною атрибуцією конкретного автора. Дослідження виконані на україномовних текстах (мовою з низькою ресурсоємкістю). Наведені результати експериментів, виконаних на основі розроблених програмних засобів.Prombles in programming 2023; 3: 22-29 Інститут програмних систем НАН України 2023-10-06 Article Article application/pdf https://pp.isofts.kiev.ua/index.php/ojs1/article/view/577 10.15407/pp2023.03.022 PROBLEMS IN PROGRAMMING; No 3 (2023); 22-29 ПРОБЛЕМЫ ПРОГРАММИРОВАНИЯ; No 3 (2023); 22-29 ПРОБЛЕМИ ПРОГРАМУВАННЯ; No 3 (2023); 22-29 1727-4907 10.15407/pp2023.03 en https://pp.isofts.kiev.ua/index.php/ojs1/article/view/577/627 Copyright (c) 2023 PROBLEMS IN PROGRAMMING |
| institution |
Problems in programming |
| baseUrl_str |
https://pp.isofts.kiev.ua/index.php/ojs1/oai |
| datestamp_date |
2024-04-28T11:55:00Z |
| collection |
OJS |
| language |
English |
| topic |
natural language texts authorship determination genetic algorithm recurrent analysis statistical analysis text classification pattern recognition formal grammars UDK 004.91 |
| spellingShingle |
natural language texts authorship determination genetic algorithm recurrent analysis statistical analysis text classification pattern recognition formal grammars UDK 004.91 Shynkarenko, V.I. Demydovych, I.M. Methods and software for significant indicators determination of the natural language texts author profile |
| topic_facet |
natural language texts authorship determination genetic algorithm recurrent analysis statistical analysis text classification pattern recognition formal grammars UDK 004.91 природомовні тексти визначення авторства генетичний алгоритм рекурентний аналіз статистичний аналіз класифікація текстів розпізнавання образів формальні граматики УДК 004.91 |
| format |
Article |
| author |
Shynkarenko, V.I. Demydovych, I.M. |
| author_facet |
Shynkarenko, V.I. Demydovych, I.M. |
| author_sort |
Shynkarenko, V.I. |
| title |
Methods and software for significant indicators determination of the natural language texts author profile |
| title_short |
Methods and software for significant indicators determination of the natural language texts author profile |
| title_full |
Methods and software for significant indicators determination of the natural language texts author profile |
| title_fullStr |
Methods and software for significant indicators determination of the natural language texts author profile |
| title_full_unstemmed |
Methods and software for significant indicators determination of the natural language texts author profile |
| title_sort |
methods and software for significant indicators determination of the natural language texts author profile |
| title_alt |
Методи та засоби визначення значимих показників профілю автора природно-мовних текстів |
| description |
Methods for the formation and optimization of author profiles are presented. The author profile is an image – a vector in a multidimensional space, which components are author’s texts measurements by a number of methods based on 4-grams, stemming, recurrence analysis and formal stochastic grammar. The author’s profile is a model of his language, including vocabulary, sentence syntax features. A comparative analysis of each of the methods effectiveness is carried out. By means of the genetic algorithm, a reduced profile of the author is formed. Insignificant indicators are excluded, which allows to reduce their number by 20%. The reduced author’s profile contains attributes that are significant for this author and is an effective attribution of a particular author.Prombles in programming 2023; 3: 22-29 |
| publisher |
Інститут програмних систем НАН України |
| publishDate |
2023 |
| url |
https://pp.isofts.kiev.ua/index.php/ojs1/article/view/577 |
| work_keys_str_mv |
AT shynkarenkovi methodsandsoftwareforsignificantindicatorsdeterminationofthenaturallanguagetextsauthorprofile AT demydovychim methodsandsoftwareforsignificantindicatorsdeterminationofthenaturallanguagetextsauthorprofile AT shynkarenkovi metoditazasobiviznačennâznačimihpokaznikívprofílûavtoraprirodnomovnihtekstív AT demydovychim metoditazasobiviznačennâznačimihpokaznikívprofílûavtoraprirodnomovnihtekstív |
| first_indexed |
2024-09-16T04:08:05Z |
| last_indexed |
2024-09-16T04:08:05Z |
| _version_ |
1818568353789247488 |
| fulltext |
Прикладне програмне забезпечення
22
UDK 004.91 http://doi.org/10.15407/pp2023.03.22
V.I. Shynkarenko, I.M. Demydovych
METHODS AND SOFTWARE FOR SIGNIFICANT INDICATORS
DETERMINATION OF THE NATURAL LANGUAGE TEXTS
AUTHOR PROFILE
Methods for the formation and optimization of author profiles are presented. The author profile is an image – a
vector in a multidimensional space, which components are author’s texts measurements by a number of methods
based on 4-grams, stemming, recurrence analysis and formal stochastic grammar. The author’s profile is a
model of his language, including vocabulary, sentence syntax features. A comparative analysis of each of the
methods effectiveness is carried out. By means of the genetic algorithm, a reduced profile of the author is formed.
Insignificant indicators are excluded, which allows to reduce their number by 20%. The reduced author’s profile
contains attributes that are significant for this author and is an effective attribution of a particular author.
Keywords: natural language texts, authorship determination, genetic algorithm, recurrent analysis, statistical
analysis, text classification, pattern recognition, formal grammars
Introduction
Attribution of authorship is the prob-
lem of identifying an anonymous text author or
a text whose authorship is in doubt [1]. There
are many examples in the literature of differ-
ent countries, when doubts arose in the work
authorship and authorship was not reliably es-
tablished.
To resolve such controversial issues, an
analysis of the other authors works is carried
out, during which it is required to determine
the significant characteristics of the text and
the author’s style as a whole. Subsequently, the
belonging of the text to one or another author’s
pen will be determined by the closeness of the
text under study writing style to one of them.
In most cases, such a task of determining the
text authorship refers to classification tasks.
There are various subtasks in text clas-
sification, and they can be divided into themat-
ic and non-thematic. The traditional classifica-
tion of texts is based on their subject matter.
However, over the past 20 years, areas
of non-thematic classification have also been
actively used, for example, in such subtasks as
genre classification [2,5], sentiment classifica-
tion, spam identification, language identifica-
tion, authorship identification, and plagiarism
detection [3].
Many algorithms have been developed
to evaluate text authorship. These algorithms
rely on the fact that the authors are character-
ized by the linguistic features of their own lan-
guage at all levels – semantic, syntactic, lexico-
graphic, spelling and morphological [4], which
manifest themselves in the writing of texts.
As a rule, these features appear uncon-
sciously in the authors works and thus provide
a useful basis for determining authorship. The
most common approach to determining au-
thorship is to use stylistic analysis, which takes
place in two stages: first, certain style markers
are extracted, then, some classification proce-
dures are applied to the resulting model.
These methods are usually based on the
calculation of lexical measures representing
the author’s vocabulary richness and the com-
monly used words appear frequency [5].
The extraction of style markers is usual-
ly done using some form NLP analysis, such as
tagging, parsing, and morphological analysis.
However, this standard approach has
several drawbacks. First, the methods used to
extract style markers are language specific. For
example, the English parser is not applicable
to texts in German, Ukrainian, or Chinese.
Second, feature selection is not a trivial
process and usually involves setting thresholds
to exclude non-informative features [6].
These decisions can be extremely sub-
tle because although rare features contribute
less signal than common features, they can still
have an important cumulative effect [7].
©V.I.Shynkarenko, I.M.Demydovych, 2023
ISSN 1727-4907. Проблеми програмування. 2023. №3
Прикладне програмне забезпечення
23
Thirdly, modern authorship attribu-
tion systems – determining the author of a
text – invariably analyze by words. However,
although word-level analysis seems intuitive,
it ignores the fact that morphological features
can also play an important role, and in addi-
tion, many Asian languages such as Chinese
and Japanese do not have well-defined word
boundaries in text.
When working with a small number of
authors and their works, the number of mea-
sures for comparison will also be small. How-
ever, if the number of authors or classes is
much larger, it is necessary to set a limit on the
amount of information about the author, i.e.
create an author profile that will include only
the most informative indicators from a large
list of them.
At present, approaches starting with
the theory of pattern recognition, mathemat-
ical statistics and probability theory, algo-
rithms of neural networks and cluster analy-
sis, and many others are used for text attribu-
tion.
This article solves the problem of deter-
mining the text authorship various attributions
effectiveness – from the set of text attributes
obtained by different methods, their subset is
distinguished, which is sufficient to identify
a specific author of the text. We will consider
these subsets as effective attribution of a par-
ticular author.
The work is carried out on Ukrainian lit-
erary texts and explores the features of speech
constructions and sentence construction that
are specific to the Ukrainian language.
The allocation of effective attribution
of the author is carried out on the basis of ex-
periments with texts of different Ukrainian au-
thors by means of a genetic algorithm.
Methods
Several methods are used to analyze the
texts of different authors, form their profiles,
highlight the most significant indicators, and
then reduce the data of each profile to reduce
the time and computational resources required
during the experiment.
Below is a general scheme for high-
lighting the effective attribution of authors
(Fig. 1).
Figure 1 – General experiment scheme
In the selecting weights process for
each of the indicators using a genetic algorithm,
the following is performed: the initial weight
vector Wk of the first generation is randomly
formed, the fitness function is determined, and
the best ones are selected with a crossover and
mutation to form a new generation Wk.
Fitness function where
– is the profile of the k-th work author, –
are the measurement weights corresponding to
this author, ρ – is a function that experimen-
tally determines whether the authorship of the
k-th work is established correctly.
The last two steps are repeated until the
improvement of the function result stops, af-
Прикладне програмне забезпечення
24
ter which the process is considered completed,
and the weights are determined.
The last step is to reduce the number
of indicators. хj and wj are successively elim-
inated such that . If the result
remains the same or slightly deteriorates, the
profile reduction continues. As soon as the re-
sult begins to deteriorate significantly, the con-
traction stops and is considered complete.
Frequency analysis in creating
an author profile
Frequency analysis is one of the most
common text analysis methods. For many lan-
guages and a large number of authors, linguists
compiled an author’s language frequency dic-
tionary or for the individual author’s texts [8,
9]. The basis of such text processing is the cal-
culation of a single character occurrence fre-
quency for a particular text. Based on the data
obtained, it can be concluded that each text
will be characterized by its own individual fre-
quency structure.
This method is based on the fact that
there is a non-standard statistical distribution
of characters within the text.
Practical application of this approach
can be very different. A large number of works
have been devoted to this problem. Also, the
problems of frequency analysis occur when
the process of decoding is necessary, the nec-
essary set of data selection in large arrays, the
analysis of texts that were written in ancient
languages, and the conduct of categorization
processes. The implementation of frequency
analysis can be used in expert systems. At the
same time, the frequency component under-
lines the measure of texts proximity.
The method of text analysis using
N-grams is a relatively new method and in most
cases is used to search for plagiarism in various
text sources [10, 12]. This method also shows
the best results in determining the authorship of
texts using frequency analysis [12, 13].
In the current work, 4-grams are used
due to their greatest efficiency in determining
authorship in previous works [12, 13].
Based on the obtained frequencies of
4-grams, a recurrent analysis adapted for work-
ing with texts is carried out – a time series is
built based on the frequency of occurrence of
each 4-gram in order (advance to the next cor-
responding element is taken as a unit of time),
on the basis of which a recursive diagram is
formed. According to the resulting diagram,
the following indicators are calculated: for re-
peating statistically similar symbols, 𝐷𝐼𝑉 – is
a value, reverse maximum length of diagonal
structures; 𝐸𝑁𝑇 – indicate the frequency dis-
tribution of the statistically similar characters
repetition, 𝐿𝐴𝑀 – indicates the repetition of
statistically similar characters, 𝑇𝑇 – indicates
the average frequency of statistically similar
characters repetition. [12, 13].
An example of 4-grams from the work
“Доля” by T. Shevchenko:
Ти не лукавила зо мною,
Ти другом, братом і сестрою…
Obtained 4-grams: тине, инел, нелу,
елук, лука, укав, кави, авил, вила, илаз,
лазо, азом, зомн, омно, мною, ноют, оюти,
ютид,…
Using stems to form an author
profile
Stemming is the process of shortening
a word to its base by cutting off parts, such
as an ending or a suffix. The basic concept of
stemming is words with the same stem or root
that refer to the same concept.
The results of stemming are some-
times very similar to determining the root of
a word, but its algorithms are based on oth-
er principles. Therefore, the word after pro-
cessing by the stemming algorithm may dif-
fer from the morphological root of the word.
Stemming is used in linguistic morphology
and information retrieval [16]. Many search
systems use stemming to establish synon-
ymous relationships if they have the same
forms after stemming.
Martin Porter’s stemming algorithm
has become widespread and has become the
de facto standard stemming algorithm for the
English language.
In this work, Porter’s stemmer adapt-
ed to the Ukrainian language is also used and
studied from its effectiveness point of view for
determining authorship [14, 15]. It is used to
work directly with the texts of various authors
and also to build a various stems frequency
profile, specific to each author.
Прикладне програмне забезпечення
25
An example of the same passage from
the work “Доля” by T. Shevchenko after stem-
ming: т, лукав, мн, друг, брат, сестр.
Using dictionaries to create author
profile
To conduct an experiment in this paper,
we studied the effectiveness of using a dictio-
nary. In general, the dictionary was developed
on the basis of two approaches. The first, dic-
tionary was the public dictionary the Large
Electronic Dictionary of Ukrainian (VESUM)
[17]. And the second, one was formed on the
basis of Ukrainian text bank, including literacy
texts, messages, posts, etc.
Based on it, a complex dictionary was
built containing unique word stems, their end-
ings and prefixes. To reduce its size, a prelimi-
nary selection of unique endings lists was car-
ried out and only an index from it was assigned
to the stem of the word. Maintaining a list of
vowel alternations in words is also supported.
To create lists of prefixes for the bas-
es, the formed dictionary was analyzed for the
presence of bases that differ only in the pres-
ence of a prefix by simple enumeration. As a
result, the original dictionary of bases has de-
creased – all key bases have been assigned the
corresponding index from the list of prefixes,
and the extra bases with prefixes have been re-
moved.
The advantage of the resulting dictio-
nary is its support for taking into account all
word forms for stems, each of them will be as-
signed a unique index. Thus, all cases, differ-
ent forms of words, as well as words obtained
by adding a prefix, will unmistakably lead to a
single stem.
The process of dictionary formation
and its form is described in more detail in the
previous works of the authors [18].
Using formal stochastic grammar
to model sentence structure
Stochastic grammar is used to create
rules that describe the structure of sentences in
a text. For each rule, the probability of appli-
cation in a particular work is determined. The
probability of inferring the entire sentence is
defined as the probabilities of the speech parts
sequences product used in it. The resulting
rules will generate a language characteristic of
the processed and structurally similar a certain
author works [19].
To describe the structure of the text
under study, speech parts are used as a char-
acteristic of the word. Thus, each word in the
sentence is replaced by the part of speech that
it is. For more information about the structure
of sentences and the rules for their construc-
tion, characteristic of a particular author, read-
ing not only parts of speech, but also forms,
numbers, gender, etc. for the word under study
[19].
For each speech part, its occurrence
probability in a certain place of the sentence in
the given text is calculated. The certain speech
part appearance probability in the studied se-
quence will more accurately capture the each
of the authors under study individual writing
style characteristic. After receiving the text in
the form a speech parts sequences set in sen-
tences with the probability of their occurrence
in a particular place, rules are formed. The pro-
cess is described in more detail in the previous
work of the authors [19].
An example of the same passage from
the work “Доля” by T. Shevchenko in terms
of rules:
where σ – is the initial nonterminal,
the -th nonterminal in the rule of the -th
level, – is the probability of applying the
corresponding rule when parsing this work.
More details are given in the work of
the authors. [12].
Forming an author profile
To obtain the profile of a specific au-
thor, calculations are carried out to determine
each of the studied indicators groups for all the
works of the author in the training sample. Fur-
ther, they are all collected in one vector X – the
profile of the author.
For example, when working with
4-grams, based on the obtained indicators, a
vector is formed that contains the frequency
of each such 4-gram occurrence in the text. To
compile the author’s profile, such vectors are
taken into account for each texts in the train-
Прикладне програмне забезпечення
26
ing sample and the average value for each of
them is found. A similar procedure is repeated
to form vectors based on the remaining groups
of indicators.
An example of a vector image of
T. Shevchenko based on 4-grams::
= [АБАЗ, АБАЙ, АБАР, АБАС,
АБАТ, АБАУ, АБЕР, АБІК, АБІЛ, АБЛА,
АБОГ, АБОТ, АБОЮ, АБОЯ, АБУД, АБУД,
АБУЛ, АБУС, АБУТ, АВАБ, АВАВ, АВАЛ,
…].
In total, there are 8748 4-grams used in
the text in the vector. And their frequencies:
[0.0001249, 0.0001565, 0.0001249,
0.0001565, 0.0001249, 0.0001249, 0.0001565,
0.0001565, 0.0001249, 0.0001249, 0.0004998,
0.0001249, 0.0001249, 0.0001249, 0.0004381,
0.0001249, 0.0001249, 0.0001249, 0.0001565,
0.0002499, 0.0001565, 0.0004696, …].
As can be seen, there are a large num-
ber of obtained 4-grams and their frequencies,
which is time-consuming and computationally
expensive to work with. However, since each
author has his own style of writing, different
4-grams may be most informative for different
authors. In addition, often the least common
letter combinations can be of the greatest im-
portance, as they will be a characteristic fea-
ture of the author’s language. Thus, the list of
received frequencies requires additional anal-
ysis of their informativeness and subsequent
data reduction to work with only the most sig-
nificant indicators.
To optimize performance and obtain
best result, when working with different indi-
cators in the vectors, a genetic algorithm was
applied to determine the weights of each of
them in each group.
In this work, on the basis of all the
above indicators and further determination of
their weight, profiles of the authors were com-
piled. In total, the author’s profile included four
main groups according to the methods studied.
Each of the groups includes a list of indica-
tors with individual weights for each. Thus, for
each author, a list of indicators was determined
that most accurately reflect his author’s style
and allow you to identify similar elements in
the texts of the control sample.
An example of a T. Shevchenko profile
vector based on stems, created on the basis of
the Large Electronic Dictionary of Ukrainian
(VESUM):
[…а, аа, аб, абатів, абатівськ,
абатств, абатськ, абет, абетк, аби, аби-
аби, абиде, абиколи, абикуди, аби-но,
абискільки, абись, аби-то, абич, …].
In total, there are 7239 stems used in
the text in the vector. As can be seen from the
data obtained, the number of topics for analy-
sis is as large as previous, which will also re-
quire subsequent reduction and selection the
most informative of them.
Their weights for the profile
T. Shevchenko:
= […0.91, 0.12, 0.55, 0.08, 0.18,
0.82, 0.9, 0.85, 0.99, 0.89, 0.17, 0.86, 0.38,
0.99, 0.42, 0.58, 0.98, 0.62, 0.43, 0.34, …].
And working with the rules when cre-
ating a profile, all the rules obtained in the pro-
cess of analyzing the texts in the training sam-
ple were collected in a single database, and for
each of them was also found a weight. The to-
tal number of rules was 6946, the following is
an example of a vector with weights for them:
[…0.35, 0.88, 0.25, 0.44, 0.21,
0.6, 0.41, 1, 0.08, 0.2, 0.72, 0.21, 0.86, 0.49,
0.62, 0.12, 0.54, 0.14, 0.12, 0.24, …].
The number of rules is somewhat less,
but still requires the selection of the most im-
portant and informative ones for the correct
determination of authorship with the least ex-
penditure of resources.
For a repeat experiment, the profile
of each author was reduced for each group of
indicators. The indicators with the smallest
weights for each of the groups were discard-
ed in order to reduce the time and computing
power of the computer.
During the experiment, the authorship
of natural language texts was determined by
two samples. The sample included works of
art due to the presence of the author style char-
acteristic in them and confirmed information
about their authorship, which is not subject to
doubt.
For the first experiment, 40 texts of fic-
tion by 10 Ukrainian authors were selected in
the training sample. The control sample con-
sisted of 60 texts by the same authors.
The works of the following authors are
presented: IB – I. Bahrianyi, AV – A. Vyshnia,
Прикладне програмне забезпечення
27
MV – M. Vovchok, AD – A. Dovzhenko, HK –
H. Kvitka-Osnovianenko, PM – P. Myrnyi,
VN – V. Nestaiko, VP – V. Pidmohylnyi, IF –
I. Franko, MK – M. Khvylovyi.
Attribution results
In working with a control sample, when
determining the authorship of a text based on
the author’s profile, the following results were
obtained.
Based on the data presented, working
with the author’s profile, the number of works
with correctly identified authorship in the con-
trol sample was 54 works out of 60. The meth-
od under study made it possible to determine
the authorship of most texts correctly, with
some exceptions. While when comparing the
profile of the following authors – Bahrianyi,
Vovchok, Kvitka-Osnovianenko, Franko and
Khvylovyi – one of the works was not correct-
ly identified and showed a great similarity with
the profile of another author in the sample.
During analyzing the result obtained,
some similarity of styles in the two works was
shown by Bahrianyi and Franko, and it can
also be argued that Khvylovyi’s style most
often echoes the styles of other authors: in 3
cases out of 6.
Table 1 – Authorship establishing result
with the full profiles
real defined real defined
IB IB MV MV
IB IB MV MK
IB IB MV MV
IB IB AD AD
IB IB AD AD
IB IF AD AD
AV AV AD AD
AV AV AD AD
AV AV AD AD
AV AV HK HK
AV AV HK HK
AV AV HK MK
MV MV HK HK
MV MV HK HK
MV MV HK MK
PM PM VP VP
PM PM VP VP
PM PM VP VP
PM PM IF IF
PM PM IF IF
PM PM IF IF
VN VN IF IB
VN VN IF IF
VN VN IF IF
VN VN MK MK
VN VN MK MK
VN VN MK MK
VP VP MK MK
VP VP MK MK
VP VP MK IF
When excluding from the list of indica-
tors the least significant for each author. Thus,
the number of 4-grams in the profile decreased
by 1750, stem by 1448 and rules by 1390,
which amounted to 20% in each of the classes.
When working with optimized vectors, the fol-
lowing results were obtained.
As a result of the experiment with a re-
duced author profile, the result was 53 works
with correctly established authorship out of 60.
Results and discussion
As a result of the experiment using a
genetic algorithm and obtaining the best solu-
tion, the following results were obtained: out
of 60 texts in the control sample, the author-
ship of 54 works was established correctly,
which amounted to a total 90%.
Table 2 – Authorship establishing result
with the reduced profiles
real elimi real elimin
IB IB MV MV
IB IB MV MHK
IB IB MV MV
IB IB AD AD
IB IB AD AD
IB IF AD AD
AV AV AD AD
AV AV AD AD
AV AV AD AD
AV AV HK HK
AV AV HK HK
AV AV HK MKH
MV MV HK HK
MV MV HK HK
Прикладне програмне забезпечення
28
MV MV HK MKH
PM PM VP VP
PM PM VP VP
PM PM VP VP
PM PM IF IF
PM PM IF IF
PM PM IF IF
VN VN IF IB
VN VN IF IF
VN VN IF IF
VN VN MKH MKH
VN VN MKH MKH
VN VN MKH MKH
VP VP MKH MKH
VP VP MKH MKH
VP PM MKH IF
For comparison in previous works and
the application of these methods separately,
the following results were obtained. The best
indicator – 91% coincidence of the texts au-
thorship – was obtained when working with
4-grams. Working with the basics of words us-
ing dictionaries and stemming gave a result of
88%.
As you can see, the combination of
different approaches and methods did not sig-
nificantly improve the result, however, it made
it possible to take into account additional fea-
tures of the text due to working with grammars.
Based on the data obtained, the most
successful methods of working with text are
4-grams – working with them is average in
terms of resources and time, relative to other
methods, and gives the best result. As well as
work with stochastic grammars, due to the dis-
play the features of the phrases and sentences
construction by the author, however, this meth-
od requires significant computational and time
resources.
The result of working with stems and
dictionaries shows that they are less informa-
tive. Taking into account the high cost of these
methods in calculations and time, the methods
are the most expensive and the least informa-
tive among all those used.
With the exception of the least signifi-
cant indicators and, as a result, a reduction in
their number, the result obtained was 52 works
with correctly established authorship, which is
a good result – 87% the accuracy of the defi-
nition.
This approach made it possible to sig-
nificantly reduce the complexity and time of
calculation, while the result did not decrease
significantly.
Conclusions
In the work, various approaches were
explored for the formation of the general au-
thor profile: work with 4-grams, stems, recur-
rent analysis and sentence structure formalized
by means of a formal stochastic grammar.
This approach made it possible to ob-
tain an effective profile of the author, taking
into account the various features of his personal
language, from the use of individual words to
the peculiarities of constructing sentences. The
results obtained demonstrate the effectiveness
of an integrated approach that provides better
results compared to approaches that take into
account individual aspects of the author’s style.
References
1. H. Love. 2002. Attributing Authorship: An Intro-
duction. Cambridge University Press.
2. Aidan Finn and Nicholas Kushmerick. 2003.
Learning to classify documents according to
genre. In IJCAI-03 Workshop on Computation-
al Approaches to Style Analysis and Synthesis.
3. D. Khmelev and W. Teahan. 2003. A repetition
based measure for verification of text collections
and for text categorization. In SIGIR’2003, To-
ronto, Canada.
4. M. Ephratt. 1997. Authorship attribution – the
case of lexical innovations. In Proc. ACH-
ALLC-97.
5. E. Stamatatos, N. Fakotakis, and G. Kokkinakis.
2001. Computer-based authorship attribution
without lexical measures. Computers and the
Humanities, 35:193–214.
6. S. Scott and S. Matwin. 1999. Feature engi-
neering for text classification. In Proceedings
ICML-99.
7. A. Aizawa. 2001. Linguistic techniques to im-
prove the performance of automatic text cate-
gorization. In Proceedings 6th NLP Pac. Rim
Symp. NLPRS-01.
8. Darchuk N. 2023. Automatic frequency dic-
tionary of connectivity by Lina Kostenko and
Mykola Vingranovskyi. Linguistic and concep-
Прикладне програмне забезпечення
29
tual pictures of the world, 73 (1), 10.17721/2520-
6397.2023.1.01.
9. Danyliuk, I., Zagnitko, A. and Sytar, G., 2019.
Text corpus of Yury Shevelyov: structure,
functions, navigation. APPLIED LINGUIS-
TICS. LINGUISTICS. 10.18523/1p.2522-
9281.2019.5.158-169.
10. Kuzma, K.T., 2020. Information technology
for estimating the level of simslarity of strings
based on the N-gram method. Academic notes
of TNU named after V.I. Vernadskyi. Series:
technical sciences. 31 (7), p. 96-98. 10.32838/
TNU-2663-5941/2020.6-1/16.
11.H. Gómez-Adorno, JP. Posadas-Durán, G.
Sidorov, Document embeddings learned on var-
ious types of n-grams for cross-topic authorship
attribution. Computing 100 (2018) 741–756.
doi: 10.1007/s00607-018-0587-8.
12. V.I. Shynkarenko, I.M. Demidovich Determi-
nation of the attributes of authorship of natural
texts. Artificial Intelligence 3 (2018) 27-35.
13. V.I. Shynkarenko, I.M. Demidovich Author-
ship Determination of Natural Language Texts
by Several Classes of Indicators with Custom-
izable Weights, in: Proceedings of the 5th Inter-
national Conference on Computational Linguis-
tics and Intelligent Systems (COLINS 2021).
Volume I: Main Conference. Lviv, Ukraine,
April 22-23, 2021, pp. 832-844.
14. T. V. Golub, M. Yu. Tyagunova, Method of
steaming Ukrainian-language texts for classi-
fication of documents based on Porter’s algo-
rithm. Scientific works of Donetsk National
Technical University. Series: Informatics, cy-
bernetics and computer engineering No. 1(24)
(2017) 59–63.
15. Dukhnovska KK, Strashok YaA, Shilo PV. In-
formation technology for performing lemma-
tization and steming in Ukrainian-language
texts. Applied systems and technologies in the
information society. Pp.. 119-127.
16. S. Memon, K. Memon, F. Dehraj and others.
2020. Comparative Study of Truncating and
Statistical Stemming Algorithms. Internation-
al Journal of Advanced Computer Science
and Applications.
17. Great electronic dictionary of the Ukrainian
language (VESUM). URL: https://github.
com/brown-uk/dict_uk.
18. I. Demidovich, V. Shynkarenko, O. Kuropiat-
nyk, O. Kirichenko, Processing Words Effec-
tiveness Analysis in Solving the Natural Lan-
guage Texts Authorship Determination Task,
XVI International Scientific and Technical
Conference (CSIT’2021). September 22-25,
2021, Lviv, Ukraine.
19. V. I. Shynkarenko, I. M. Demidovich Natu-
ral Language Texts Authorship Establishing
Based on the Sentences Structure, in: Pro-
ceedings of the 6th International Conference
on Computational Linguistics and Intelligent
Systems (COLINS 2022), Volume I: Main
Conference, Gliwice, Poland, May 22- 23,
2022, pp. 328-337
Received: 07.09.2023
About the authors:
Viktor Shynkarenko,
Doctor of Science, Professor,
Number of scientific publications
in Ukrainian publications – more than 200
Number of scientific publications
in foreign publications – more than 30
Index Girsh – 6
https://orcid.org/0000-0001-8738-7225
Scopus Author ID: 26635896100
Inna Demydovych,
PhD student,
Number of scientific publications
in Ukrainian publications – 4
Number of scientific publications
in foreign publications – 1
Index Girsh – 2
https://orcid.org/0000-0002-3644-184X
Scopus Author ID: 57224201949
Place of work:
Ukrainian State University
of Science and Technologies,
49010, Ukraine, Dnipro, str. Lazaryana, 2
E-mail:office@ust.edu.ua
|