Automatic generation of semantic knowledge networks from an unstructured text
A method and an algorithm for the semantic knowledge network automated construction created from the most informative concepts in the electronic texts are proposed. Аn analysis and comparison of existing methods with their software implementations for information research in electronic texts are pre...
Gespeichert in:
| Datum: | 2018 |
|---|---|
| Hauptverfasser: | , |
| Format: | Artikel |
| Sprache: | English |
| Veröffentlicht: |
Міжнародний науково-навчальний центр інформаційних технологій і систем НАН та МОН України
2018
|
| Schriftenreihe: | Управляющие системы и машины |
| Schlagworte: | |
| Online Zugang: | http://dspace.nbuv.gov.ua/handle/123456789/142078 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Назва журналу: | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| Zitieren: | Automatic generation of semantic knowledge networks from an unstructured text / M.N. Savchenko, A.S. Kriachok // Управляющие системы и машины. — 2018. — № 1. — С. ХХ-Х71-78Х . — Бібліогр.: 10 назв. — англ. |
Institution
Digital Library of Periodicals of National Academy of Sciences of Ukraine| id |
irk-123456789-142078 |
|---|---|
| record_format |
dspace |
| spelling |
irk-123456789-1420782018-09-25T01:23:02Z Automatic generation of semantic knowledge networks from an unstructured text Savchenko, M.N. Kriachok, A.S. Методы и средства обработки данных и знаний A method and an algorithm for the semantic knowledge network automated construction created from the most informative concepts in the electronic texts are proposed. Аn analysis and comparison of existing methods with their software implementations for information research in electronic texts are presented. The results of BBC news article analysis using the proposed method are given. Цель статьи: создание алгоритмической и программной базы для построения семантических сетей знаний из самой релевантной по отношению к контексту документов информации. Методы: предложены комплексная методика, алгоритм и его реализация для построения семантической сети знаний из самой значимой информации в заданных текстах. Предложенный комплексный алгоритм сочетает в себе работу нескольких алгоритмов на основе нейронных сетей и статистического анализа. Комбинация данных алгоритмов позволяет распознавать концепты в тексте, находить между ними связи и определять, какие из концептов должны быть включены в результирующую семантическую сеть с помощью оценки их веса в заданном контексте. Результат: проведён анализ большого текстового корпуса, общей численностью около миллиона слов. На основе собранной информации с помощью разработанного алгоритма и рекурсивной грамматики естественного языка построено семантическую сеть знаний для нескольких текстов и отдельную совмещённую семантическую сеть знаний. Проведено сравнение недостатков и преимуществ разработанного алгоритма по отношению к нескольким уже существующих подходам извлечения знаний из текстов. Продемонстрированы полученные результаты. Мета статті – створення алгоритмічної і програмної бази для побудови семантичних мереж знань із найбільш релевантної інформації відносно контексту документів. Методи: Запропоновано комплексну методику, алгоритм та його реалізацію для побудови семантичної мережі знань із найбільш значимої інформації у заданих текстах. Запропонований комплексний алгоритм поєднує в собі роботу кількох алгоритмів на основі нейронних мереж та статистичного аналізу. Комбінація даних алгоритмів дозволяє розпізнавати концепти в тексті, знаходити між ними зв’язки та визначати, які із концептів мають бути включені до результуючої семантичної мережі за допомогою оцінки їх ваги. Результат: Проведено аналіз великого текстового корпусу, загальною чисельністю близько мільйону слів. На основі зібраної інформації за допомогою розробленого алгоритму і рекурсивної граматики природної мови побудовано семантичну мережу знань для декількох текстів і окрему поєднану семантичну мережу знань. Проведено порівняння недоліків і переваг розробленого алгоритму по відношенню до кількох вже існуючих підходів вилучення знань із текстів. Продемонстровано отримані результати. 2018 Article Automatic generation of semantic knowledge networks from an unstructured text / M.N. Savchenko, A.S. Kriachok // Управляющие системы и машины. — 2018. — № 1. — С. ХХ-Х71-78Х . — Бібліогр.: 10 назв. — англ. 0130-5395 DOI: https://doi.org/10.15407/usim.2018.01.071 http://dspace.nbuv.gov.ua/handle/123456789/142078 004.89 en Управляющие системы и машины Міжнародний науково-навчальний центр інформаційних технологій і систем НАН та МОН України |
| institution |
Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| collection |
DSpace DC |
| language |
English |
| topic |
Методы и средства обработки данных и знаний Методы и средства обработки данных и знаний |
| spellingShingle |
Методы и средства обработки данных и знаний Методы и средства обработки данных и знаний Savchenko, M.N. Kriachok, A.S. Automatic generation of semantic knowledge networks from an unstructured text Управляющие системы и машины |
| description |
A method and an algorithm for the semantic knowledge network automated construction created from the most informative concepts in the electronic texts are proposed. Аn analysis and comparison of existing methods with their software implementations for information research in electronic texts are presented. The results of BBC news article analysis using the proposed method are given. |
| format |
Article |
| author |
Savchenko, M.N. Kriachok, A.S. |
| author_facet |
Savchenko, M.N. Kriachok, A.S. |
| author_sort |
Savchenko, M.N. |
| title |
Automatic generation of semantic knowledge networks from an unstructured text |
| title_short |
Automatic generation of semantic knowledge networks from an unstructured text |
| title_full |
Automatic generation of semantic knowledge networks from an unstructured text |
| title_fullStr |
Automatic generation of semantic knowledge networks from an unstructured text |
| title_full_unstemmed |
Automatic generation of semantic knowledge networks from an unstructured text |
| title_sort |
automatic generation of semantic knowledge networks from an unstructured text |
| publisher |
Міжнародний науково-навчальний центр інформаційних технологій і систем НАН та МОН України |
| publishDate |
2018 |
| topic_facet |
Методы и средства обработки данных и знаний |
| url |
http://dspace.nbuv.gov.ua/handle/123456789/142078 |
| citation_txt |
Automatic generation of semantic knowledge networks from an unstructured text / M.N. Savchenko, A.S. Kriachok // Управляющие системы и машины. — 2018. — № 1. — С. ХХ-Х71-78Х . — Бібліогр.: 10 назв. — англ. |
| series |
Управляющие системы и машины |
| work_keys_str_mv |
AT savchenkomn automaticgenerationofsemanticknowledgenetworksfromanunstructuredtext AT kriachokas automaticgenerationofsemanticknowledgenetworksfromanunstructuredtext |
| first_indexed |
2025-07-10T14:06:08Z |
| last_indexed |
2025-07-10T14:06:08Z |
| _version_ |
1837269122059075584 |
| fulltext |
ISSN 0130-5395, УСиМ, 2018, № 1 71
DOI: https://doi.org/10.15407/usim.2018.01.071
UDK 004.89
M. SAVCHENKO, Student (Masters), zitros.lab@gmail.com
O. KRIACHOK, Candidate of Engineering’s Sciences, Associate Professor,
alexandrkriachok@gmail.com
National Technical University of Ukraine ' Igor Sikorsky Kyiv Polytechnic Institute',
Kyiv, boulevard of Victory 37, corps 5, 03056.
AUTOMATIC GENERATION OF SEMANTIC KNOWLEDGE
NETWORKS FROM AN UNSTRUCTURED TEXT
A method and an algorithm for the semantic knowledge network automated construction created from the most informative concepts
in the electronic texts are proposed. Аn analysis and comparison of existing methods with their software implementations for in-
formation research in electronic texts are presented. The results of BBC news article analysis using the proposed method are given.
Keywords: building semantic networks, knowledge extraction, knowledge models, natural language processing
Introduction
In today's information environment, through the
huge amount of unstructured text information,
there is a need in the search, seizure, formalizing
and processing of the most essential knowledge
laid down by the authors in the texts. Such knowl-
edge may be hidden in the concepts presented in
the document, and the characteristic relationship
between those concepts.
Considering the large number of texts, such as
news articles or scientific publications, one can
notice that each text has a certain, sense-unique
characteristic only for the given text. Only after
reading the text, one can briefly describe it, head-
ing to highlight the most important concepts in it
and combine their logical relations with other
known concepts. For the rest of this document we
will call every single text corpus as a document, a
single word in it as a term and a word or group of
words that represent a specific entity as a concept.
When it comes to a huge data set, Big Data or
massive text corpus, just a single person can’t read
through it quickly, taking all the important infor-
mation from there. There is a need to structure the
knowledge in texts and present texts in a struc-
tured form that can be quickly analyzed. There are
various methods of search and evaluation of the
relevant information in the document on which
one is about to make a decision using an in-depth
analysis of certain parts. Additional statistical es-
timation methods (weighing) and the identifica-
tion of the most relevant terms in the particular
document include the following: [1].
Ranging the terms ti in the document Dj by the
number of a particular concept occurrences ti Dj
— TF, which stands for Term Frequency. It is sta-
tistically investigated that if the document de-
scribes the specific area of knowledge, the most
typical concepts in a particular document are re-
peated relatively a lot of times.
Ranging the terms ti in the document Dj by the
number of occurrences of a specific concept and
inversely related to the total number of all other
documents '
jD j i , ti Dj –TF-IDF, which
stands for Term Frequency — Inverse Document
M. Savchenko, O. Kriachok
72 ISSN 0130-5395, Control systems and computers, 2018, № 1
Frequency. In contrast to the TF, TF-IDF allows
us to dramatically reduce the rank of terms found
in almost all the documents, such as, for example,
articles, prepositions, and other insignificant
commonly used words.
Other methods, such as, for example, Okapi
BM25, sigma method, etc.
The above methods allow us to highlight the
most relevant concepts for a specific text. Using
this information, when the relations between con-
cepts are established, it can be dropped in case of
these concepts a little weight hence giving priority
to the concepts with the highest score.
To examine the documents and obtain the for-
malized representation of unstructured informa-
tion that they contain, it is proposed to construct a
semantic network of concepts and relationships
between them. There are many approaches to the
construction of semantic networks, and in this
article only the most common ones are described.
Matrix method. To identify the most expressed
relationship between two terms {ti, tji j }, we
can use the matrix method. In this method, at
first, all terms are weighted by one of weighting
methods (for example, TF-IDF), and each term ti
is assigned its unique corresponding weight pi. Af-
ter selecting n the most highly expressed set of
terms, the adjacency matrix A is constructed and
filled with zeros. Next, the document is divided
into groups gk . Each group may be represented,
for example, as one individual sentence.
The adjacency matrix for the matrix method
completes as follows. For each group gk there is a
pair of terms {ti, tji j, ti gk , tj gk}. The number
of groups into which a pair of these terms are car-
ried, {ti , tji j } is added to the adjacency matrix A.
In this case, we can also consider the weight of such
pair of terms {ti , tji j } and use it as a coefficient
for the resulting value in the adjacency matrix A.
To demonstrate the results, this article will use
the text corpus of approximately 2500 BBC News
articles in English, presented in a separate text
document containing the title and full text of arti-
cles. Articles are divided into 5 main categories:
business, politics, sport, entertainment and tech-
nology. Around 500 documents match each of the
categories. For in-depth analysis we will use the
article «Ink helps drive democracy in Asia», pub-
lished February 19, 2005 on the website BBC
News [2].
An algorithm for constructing semantic net-
work using the matrix method [6]:
Take an evaluation (weighting) of each term in
a given text in relation to all other terms in the text
body by using one of the above methods (we will
use TF-IDF).
Create a list of unique terms that are graded in
descending order of their weight in the given text.
30 the most weighted terms are selected. In auto-
matic mode, the analysis can be selected, for ex-
ample, 3% of the most highly ranked terms.
Adjacency matrix is constructed by the algo-
rithm described above.
The result of building is listed in the file with
the extension *.csv and visualizing via Gephi visu-
alization tool. The resulting graph is shown in
Figure 1.
The above graph visualizes terms which have a
connection with other terms in the text. The more
text has the direct links between the two terms, the
thicker is the connection between these terms on a
graph. The size of a concept on a graph is also pro-
portional to its weight of the weighting method.
It may be noted that concepts are mainly or-
ganized in separate clusters. For example, among
other clusters, it can be noted that the most ex-
pressed terms are «voter enter uv station polling».
Undoubtedly, this cluster expresses the most rele-
vant part of the knowledge of the text, as com-
pared to all other texts. Among the other most
expressed clusters, such as «upcoming presidential
parliamentary elections» or «Kyrgyz elections use
Fig. 1. Analysis of entities in the article using the matrix
method
Automatic generation of semantic knowledge networks from an unstructured text
ISSN 0130-5395, УСиМ, 2018, № 1 73
ultraviolet ink», you can get a clear idea of the na-
ture and content of the text.
Matrix method includes the most well-inter-
connected terms, but one of the major disadvan-
tages of this method can be regarded as the relative
complexity of the allocating clusters which are the
most expressed in the graph. It is not hard to do,
looking at the graph visually with a small number of
terms, but as the number of connections grows, it
becomes very difficult to analyze and find such
clusters, even using the software. In addition, for
the resulting graph it is almost impossible to include
such terms that have relatively low weight accord-
ing to the algorithm of entities evaluation. This
plays a key role for the data crawler.
Horizontal visibility graph application. Another
interesting method for constructing semantic net-
work of terms is the method proposed by Lan-
de D.V., which combines features of the graph
with the horizontal visibility evaluation methods
in terms of a single text [3]. This method can be
applied not only to build a network of basic con-
cepts of terminology in the text, but also to build a
semantic network as a whole.
An algorithm for constructing a semantic net-
work for this method is as follows:
Text entities, similarly to previous method are
evaluated (weighed) against relation to all other
terms in the text body with TF-IDF method.
The algorithm for constructing the horizontal
visibility graph applies to the values of the term
weights. The horizontal axis is taken the term po-
sition in the text, and the vertical as a term weight.
Before constructing the actual result, some proce-
dures like stemming and rejection of the terms
listed in the list of stop words are performed.
Figure 2 shows the principle of the visibility
graph construction with the horizontal normalized
TF-IDF values of evaluation. After term weights
are put on the horizontal axis, for each term ti in
the document D a «horizontal search» applies to
the corresponding document using a horizontal
search algorithm [4]. Thus, two terms ti and tj are
compared and, a bond is formed there between.
The weight of this connection in the graph may be
proportional to a predetermined «farsightedness»
(horizon). In other words, terms that are adjacent
in the document form a strong bond, thereby form
a strong relation between concepts.
One of the possible semantic network option
construction for a given text using horizontal visi-
bility graph is shown in Figure 3. This graph has
been built with a threshold value of 0.25, which
means that only those terms whose relative weight
is greater than 25% get to the final result. The visi-
bility horizon was set to 20 words, and the weight
of a definite connection between the two periods
was measured by the total amount of a linear dis-
tance between two terms.
Building a semantic network based on the ap-
proach of the horizontal visibility graph construc-
tion has an advantage in the formation of stable
relations between concepts, as it allows us to ex-
plicitly highlight concepts and existing links be-
tween them through other concepts. But in this
method, without modifications, the same disad-
vantages apply as in the previous one, matrix
method: is difficult to determine the logical order
of relationships between entities and, in addition,
as in the matrix method, some medium or low
valued words can be missed.
InterSystems iKnow technology. InterSystems
Corporation is developing its own proprietary al-
gorithms for in-depth analysis of the texts that also
Fig. 2. The principle of the horizontal visibility graph con-
struction
Fig. 3. The article analysis using the horizontal visibility
graph algorithm
M. Savchenko, O. Kriachok
74 ISSN 0130-5395, Control systems and computers, 2018, № 1
can be used for the construction of semantic net-
works of words [7]. A set of tools for application
analysis of texts, included in the InterSystems
iKnow corporate product, allows to identify con-
cepts in the text, the relation of similarity and
connecting relation between these concepts. Any
solutions can be built based on these tools for
«structuring» unstructured information, such as,
for example, a solution for semantic analysis of
sentences, revealing modern trends through
analysis of the news and so on.
The algorithm by which InterSystems iKnow
finds the relationship between the concepts in the
text, as well as similar concepts is mainly based on
the use of the stable structures and words from
natural language [8]. For example, some of the
concepts in the text are resilient — they are chang-
ing rapidly, and there are too many to remember
to find out the role of each concept in the text. But
there are concepts which can be considered as
relatively stable between different ages, for exam-
ple: «no», «replace», «performing», «it is», «use»,
«used by», «stored in the» and so on. InterSystems
iKnow recognizes such patterns for the specified
language in the text and distinguishes separate
concepts ratio there between from other minor
parts of speech. Additional metrics are also com-
puted for concepts, like the total numbers of con-
cepts, their spread (the average distance between
same concepts in the text), relevance and the total
score, which is combined from the previous met-
rics by the formula.
To demonstrate the algorithm in action, as an
example, let’s take the sentence «clever cat eats
cheese and breathes on a mouse burrows». iKnow
technology will first consider the text as a set of
sustainable language constructs, i.e. «__ __ eats
and breathes on a __». After that, instead of «__»
the other concepts are considered, whose presence
there is almost guaranteed. The concepts also dif-
fer in terms of similarity, such that «smart cat» is
similar to the concept of «cat» and «mouse holes»
through «holes».
Thus, the article is analyzed and such concepts as,
for example, «readers» and similar «ultraviolet
readers», «effort» and «general effort» are identi-
fied. Figure 4 depicts a graph — word semantic
network constructed on the basis of the article for
analysis by InterSystems iKnow technology and
visualized using the visualization tool iKnow En-
tity Browser. Note that the main concept of the
graph is the «ink» – «Ink», from which arrows
show which concepts it is related in the ink. The
concept of the circle corresponds to the size of its
assessment, which holds iKnow. The arrows com-
ing out of these concepts reflect the «similar» con-
cepts.
Thus, using iKnow toolkit can not only identify
the most relevant concepts for this text, but also
relationships between them and the nature of that
relationships (ratio for denying the similarities).
Importantly, InterSystems iKnow concept is
built mainly for working with entities, in fact, that
basically requires a modern market. After gather-
ing the necessary data about a particular concept
using iKnow tools further expert should independ-
ently find all relevant entities in the text and to
analyze individual sentences to update the basic
knowledge that has been assigned to him. In addi-
tion, the technology is closed and is supplied with
the main product of InterSystems' — DBMS
Cachй (since 2018 — IRIS platform).
The complex method of constructing a seman-
tic network. Taking into account all the advan-
tages of the above-described methods for building
semantic web (knowledge network) for a single
text, the complex approach has been developed
Fig . 4. The article analysis using iKnow technology with
iKnow Entity Browser
Automatic generation of semantic knowledge networks from an unstructured text
ISSN 0130-5395, УСиМ, 2018, № 1 75
and investigated that includes all the benefits of
the methods described above and adds its own.
This method combines algorithms for evalua-
tion of words with algorithms Part of Speech Tag-
ging (POS Tagging, identification of parts of
speech) that allows you to find the concepts with
their characteristics in the text and to build rela-
tionships between them based solely on informa-
tion obtained from the text and a small number of
basic rules for a single language.
Also, this method is based on constructing a full
semantic network (knowledge network) that con-
tains the logical connection between the called
and the concepts, as opposed to the simple binary
«yes-no» relationships.
A simplified algorithm for constructing a se-
mantic network using the complex method looks
as follows:
Without a change in the original text, the iden-
tification of parts of speech (POS Tagging) is ap-
plied. It is important not to spend Stemming or
normalization of terms at this stage, as for the
identification of parts of speech the semantic
meaning of the original text is important, as well
as the register of symbols, punctuation marks, etc.
As a result, each period is recorded in the pair
identified with his part of the language.
Each term in the text gets its weight assigned,
reduced to the normal form (lowercase Stemming
for complex languages).
On the basis of the information received, concepts
are determined in this text and, based on a language
rules, relationships between concepts are added.
A semantic network is built
Steps 1–2 are presented in Figure 5. The timing
marks are shown in a normalized form. The blue
columns in the picture identify the concepts,
namely, nouns and their corresponding adjectives
(ink, Kyrgyz republic). Red bars represent verb
(helps, is using). For convenience, let's call them
functional terms. Green represents the auxiliary
parts of speech (the, of, in, to).
Figure 5 shows that functional terms are almost
always between two non-functional terms, i.e.
concepts. Thanks to the definition of the parts of
speech, it is possible to build the relationships be-
tween concepts from the given information. For
example, on Figure 5 it is clearly observed two
such relationships: «ink helps drive democracy»
and «former soviet republic is using invisible ink».
Thus, the algorithm picks three concepts: «(invisi-
ble) ink», «drive democracy», «(former soviet)
republic», as well as the relationship between the
two functional concepts of «helps» and «is using».
This idea of functional terms allocation is a few
similar to the idea used in the InterSystems iKnow
technology, but it is not identical. In contrast with
InterSystems iKnow, complex method uses POS
Tagging for identification of such terms but not
pre-defined dictionaries. In turn, POS Tagging
algorithm implementation may vary, including
specially modified for a particular task, or one in
which the neural network applied, which can
greatly improve the accuracy of output results,
even in the case of adding new words to the lan-
guage or in case of mistakes in the text.
Algorithms for identifying parts of speech can
now correctly recognize part of speeches with
more than 97% accuracy [5], which is an accept-
able value for constructing semantic network
based on them.
To review and search for «neighboring» concepts
and relationships, like in method with horizontal
visibility graph, only those terms are taken which
relative score is not less than a certain value. To
construct the graph shown in Figure 6, the value of
0.25 or 25% was used. Increasing this value will add
to the graph less relevant concepts, but also, in
turn, increase the number of explicit relations in it
and build logic circuits, which may lead to the dis-
covery of new logical sequences in the text.
Unlike horizontal search using horizontal visi-
bility graph, the complex method provides a com-
Fig. 5. Evaluated words and identified parts of speech
M. Savchenko, O. Kriachok
76 ISSN 0130-5395, Control systems and computers, 2018, № 1
prehensive way of breadth-first search (or searches
extension of scope), starting from the term and
moving in a both sides. An important advantage of
this approach is that the graph counts also those
entities whose weight is relatively low, but only
those that have a certain relationship with a highly
weighted concept through the functional word.
For example, in the bar chart shown in Figure
5, the concept «drive democracy» has a rating less
than 25% but, as is evident from Figure 6, it is
present in the graph, because it is associated with
more highly regarded concept «ink».
As a result of the depth-first search of concepts,
starting from functional terms, a natural relation-
ship between the concepts is formed in the seman-
tic network, which does not distort the meaning of
the original proposal. Moreover, the identical con-
cepts are combined with many other concepts
through a variety of relationships, it creates the pos-
sibility of a new way to treat the knowledge that has
been invested in the text, without further distortion.
Figure 6 shows the results of complex text
analysis method based on a BBC article «Ink helps
drive democracy in Asia» [2]. Note that the ob-
tained knowledge graph wasn’t manually modified
(except for the visual arrangement of the elements
on the graph, which is performed using Gephi
visualization tool).That is, it was built in a fully
automatic way.
To provide a broader example of the developed
complex knowledge graph building algorithm, we
built knowledge graph of another article, «Turkey
turns on the economic charm», published by BBC
in approximately the same time frame comparing
to «Ink helps drive democracy in Asia». The sec-
ond knowledge graph is presented on Figure 7.
Moreover, to demonstrate that the retrieved
data is accumulated and properly related using the
same algorithm on multiple articles, we built a
single knowledge graph from the two articles men-
tioned above, both published in BBC. The aggre-
gated result is presented on Figure 8.
Joining multiple graphs into a single graph,
which is presented on Figure 8 can be improved in
future to consider filtering of even less relevant
entities. For example, a set of texts may have
many words in common. These common words
can be considered when weighting concepts, by
modifying TF-IDF metric with new parameters
and hence cleaning the knowledge graph.
The algorithm, its implementation and the
source code are published online [10], where one
can find detailed instructions of how to reproduce
the presented result and use this algorithm to gen-
erate semantic knowledge networks for new texts.
Conclusion. Complex method of the semantic
networks combines construction all the advan-
tages of the methods described in this article and
does not inherit their drawbacks. With the help of
Fig. 6. A semantic network constructed using the complex
method
Fig. 7. A semantic network built from another BBC article
Fig. 8. A combined semantic network built from the both
articles
Automatic generation of semantic knowledge networks from an unstructured text
ISSN 0130-5395, УСиМ, 2018, № 1 77
complex method, one can build semantic net-
works (knowledge graphs) from any number of
texts in a fully automatic mode without the need
of the system experts. In the results of data extrac-
tion from the basic text, the most relevant infor-
mation presented as a graph knowledge can be
used in the future, e.g., for the development of
automatic intelligent analysis of any text data.
Designed information extraction algorithm,
which considers only the most relevant informa-
tion in the given texts is flexible and easily extend-
able. With just a few basic rules for the given lan-
guage in the text, on average complex algorithm
covers more than 50% of all entities in the English
text. Within each rule, this percentage increases
and could theoretically reach 100%, which is also
reflected in the success of parts of speech recogni-
tion algorithms. By developing a set of rules for a
particular language, it can be widely applied to
any texts, and is not limited to only technical lit-
erature, but even to the individual texts written in
a particular style of information representation.
REFERENCES
1. Savchenko M.M., Kriachok O.S. Using models of knowledge for the analysis of unstructured text. Modern problems of
scientific support for Energy. Proc. XV Int. scientific-practical conf. of graduate students, undergraduates and stu-
dents. Kiev, Apr. 25–28, 2017, Igor Sikorsky KPI. 2. P. 121.
2. Mikosz D. Ink helps drive democracy in Asia. David Mikosz. 2005. Resource Access:
http://news.bbc.co.uk/2/hi/technology/4276125.stm.
3. Lande D.V. Building of Networks of Natural Hierarchies of Terms Based on Analysis of Texts Corpora. E-preprint
ArXiv 1405.6068.
4. Lande D.V., Snarskii A.A., Yagunova E.V., Pronoza E.V. The Use of Horizontal Visibility Graphs to Identify the Words
that Define the Informational Structure of a Text. 12th Mexican International Conference on Artificial Intelligence,
2013. P. 209–215.
5. Toutanova K., Klein D., Manning C.D., Singer Y. Feature-rich part-of-speech tagging with a cyclic dependency net-
work. In: NAACL 3. (2003), P. 252–258.
6. Lande D.V., Snarskii A.A., Bezsudnov I.V. Internetika. Navigation in complex networks. Moscow: “LIBROKOM”
Book House, 2009. 264 p.
7. Van Hyfte, M. Bouzinier, M. Tsatsulin, S. Richards, K. Lee, C. Almond. “Mining medical texts for cancer intelligence
using iKnow”. Cancer Outcomes Conference in 2013.
8. Bronselaer G. De Tre, “Concept-Relational Text Clustering”, international journal of intelligent systems, 2012,
vol. 27, 970–993 (2012).
9. De Boe M., Bouzinier D., Van Hyfte. “Extending the PMML Text Model for Text Categorization”, PMML workshop
@ KDD13, August November 2013, Chicago, Illinois, USA.
10. Savchenko M. Auto Semantic Knowledge Network Builder. 2017, https://github.com/ZitRos/edu-semantic-
knowledge-network-auto-builder.
Received 21.02.2018
М.М. Савченêо, стóдент (маãістр), zitros.lab@gmail.com
О.С. Êрячоê, ê.т.н., доцент, alexandrkriachok@gmail.com
Національний технічний óніверситет Óêраїни «Êиївсьêий політехнічний інститóт
ім. Іãоря Сіêорсьêоãо». м. Êиїв, пр. Перемоãи, 37, êорп. 5, 03056.
АВТОМАТИЧНА ПОБÓДОВА СЕМАНТИЧНИХ МЕРЕЖ ЗНАНЬ ІЗ НЕСТРÓÊТÓРОВАНИХ ТЕÊСТІВ
Встóп. Ó сóчасномó інформаційномó просторі через величезнó êільêість нестрóêтóрованої теêстової інформації
існóє потреба пошóêó, вилóчення, формалізації та обробêи найбільш сóттєвих знань, заêладених автором ó теêст.
Таêими знаннями можóть бóти êонцепти, представлені в доêóментах, та хараêтерні відносини між ними. Êож-
ний теêст бóдь-яêоãо теêстовоãо êорпóсó несе певний óніêальний зміст, хараêтерний лише для даноãо теêстó.
Аêтóальною задачею є розробêа алãоритмічної та проãрамної бази, яêа дозволяла б обробляти лише найбільш
змістовнó частинó теêстів та вилóчати з неї знання, релевантні для даноãо êонтеêстó.
Мета статті. Створення алãоритмічної і проãрамної бази для побóдови семантичних мереж знань з най-
більш релевантної інформації відносно êонтеêстó доêóментів.
M. Savchenko, O. Kriachok
78 ISSN 0130-5395, Control systems and computers, 2018, № 1
Методи. Запропоновано êомплеêснó методиêó, алãоритм та йоãо реалізацію для побóдови семантичної ме-
режі знань з найбільш значної інформації ó заданих теêстах. Запропонований êомплеêсний алãоритм поєднóє
роботó êільêох алãоритмів на основі нейронних мереж та статистичноãо аналізó. Êомбінація даних алãоритмів
дозволяє розпізнавати êонцепти в теêсті, знаходити між ними зв’язêи та визначати, яêі з êонцептів мають бóти
вêлючені до резóльтóючої семантичної мережі шляхом оцінêи їх ваãи.
Резóльтат. Проведено аналіз велиêоãо теêстовоãо êорпóсó, заãальною чисельністю близьêо мільйона слів.
На основі зібраної інформації за виêористання розробленоãо алãоритмó і реêóрсивної ãраматиêи природної мо-
ви побóдовано семантичнó мережó знань для деêільêох теêстів і оêремó поєднанó семантичнó мережó знань.
Проведено порівняння недоліêів і переваã розробленоãо алãоритмó відносно êільêох існóючих підходів вилóчен-
ня знань з теêстів. Продемонстровано резóльтати.
Висновоê. Êомплеêсний метод побóдови семантичних мереж поєднóє всі переваãи описаних в статті методів
і не наслідóє їх основних недоліêів. Êомплеêсним методом можна бóдóвати семантичні мережі (ãрафи знань) з
теêстів ó повністю автоматичномó режимі та без втрóчання еêспертів. Резóльтати вилóчення з теêстів основної,
найбільш релевантної інформації, представленої ó виãляді ãрафó знань можна виêористовóвати в подальшомó,
наприêлад, для розробêи систем автоматичноãо інтелеêтóальноãо аналізó бóдь-яêих теêстових даних.
Êлючовi слова: побóдова семантичних мереж, вилóчення знань, моделi знань, обробêа природної мови
Н.Н. Савченêо, стóдент (маãистр), zitros.lab@gmail.com
А.С. Êрячоê, ê.т.н., доцент, alexandrkriachok@gmail.com
Национальный техничесêий óниверситет Óêраины «Êиевсêий политехничесêий инститóт
им. Иãоря Сиêорсêоãо». ã. Êиев, пр. Победы, 37, êорпóс 5, 03056.
АВТОМАТИЧЕСÊОЕ ПОСТРОЕНИЕ СЕМАНТИЧЕСÊОЙ СЕТИ ЗНАНИЙ
ИЗ НЕСТРÓÊТÓРИРОВАННЫХ ТЕÊСТОВ
Введение. В связи с наличием в современном информационном пространстве оãромноãо êоличества нестрóêтó-
рированной теêстовой информации сóществóет потребность в поисêе, изъятии, формализации и обработêе наи-
более сóщественных знаний, заложенных авторами в теêсты. Таêими знаниями моãóт быть êонцепты, представ-
ленные в доêóментах, и хараêтерные отношения междó ними. Êаждый теêст любоãо теêстовоãо êорпóса несёт
определённый óниêальный смысл, хараêтерный тольêо для данноãо теêста. Аêтóальная задача — разработêа
алãоритмичесêой и проãраммной базы, позволяющей обрабатывать тольêо наиболее содержательнóю часть теê-
стов и извлеêать из нее знания, релевантные для данноãо êонтеêста.
Цель статьи. Создание алãоритмичесêой и проãраммной базы для построения семантичесêих сетей знаний
из релевантной относительно êонтеêста доêóментов информации.
Методы. Предложены êомплеêсная методиêа, алãоритм и еãо реализация для построения семантичесêой
сети знаний из самой значимой информации в заданных теêстах. Данный êомплеêсный алãоритм сочетает рабо-
тó несêольêих алãоритмов на основе нейронных сетей и статистичесêоãо анализа. Êомбинация этих алãоритмов
позволяет распознавать êонцепты в теêсте, находить междó ними связи и определять, êаêие из êонцептов долж-
ны быть вêлючены в резóльтирóющóю семантичесêóю сеть с помощью оценêи их веса в заданном êонтеêсте.
Резóльтат. Проведен анализ большоãо теêстовоãо êорпóса, общей численностью оêоло миллиона слов. На
основе собранной информации с помощью разработанноãо алãоритма и реêóрсивной ãрамматиêи естественноãо
языêа построена семантичесêая сеть знаний для несêольêих теêстов и отдельная совмещенная семантичесêая
сеть знаний. Проведено сравнение недостатêов и преимóществ разработанноãо алãоритма относительно не-
сêольêих сóществóющих подходов извлечения знаний из теêстов. Продемонстрированы резóльтаты.
Выводы. Êомплеêсный метод построения семантичесêих сетей сочетает все преимóщества описанных ме-
тодов и не наследóет их основных недостатêов. Êомплеêсным методом можно строить семантичесêие сети (ãра-
фы знаний) из теêстов в полностью автоматичесêом режиме без вмешательства эêспертов. Резóльтаты извлече-
ния из теêстов основной, наиболее релевантной информации, представленной в виде ãрафа знаний, можно ис-
пользовать в дальнейшем, например, для разработêи систем автоматичесêоãо интеллеêтóальноãо анализа любых
теêстовых данных.
Êлючевые слова: построение семантичесêих сетей, извлечение знаний, модели знаний, обработêа естественноãо языêа
|