=Paper=
{{Paper
|id=Vol-2866/ceur_322_330_litvin_velichko
|storemode=property
|title=Method of Information Obtaining from Ontology on the Basis of a Natural Language Phrase Analysis
|pdfUrl=https://ceur-ws.org/Vol-2866/ceur_322_330_litvin_velichko.pdf
|volume=Vol-2866
|authors=Anna Litvin,Vitalii Velychko,Vladyslav Kaverynskyi
|dblpUrl=https://dblp.org/rec/conf/ukrprog/LitvinVK20
}}
==Method of Information Obtaining from Ontology on the Basis of a Natural Language Phrase Analysis ==
https://ceur-ws.org/Vol-2866/ceur_322_330_litvin_velichko.pdf
UДК 681.3.06
METHOD OF INFORMATION OBTAINING FROM ONTOLOGY ON THE
BASIS OF A NATURAL LANGUAGE PHRASE ANALYSIS
Anna Litvina[0000-0002-5648-9074], Vitalii Velychkoa[0000-0002-7155-9202], Vladyslav Kaverynskyib[0000-0002-6940-
579X]
a
V.M. Glushkov Institute of Cybernetics of National Academy of Sciences of Ukraine, Kyiv, Ukraine
b
I.N. Frantsevich Institute of Problems of Materials Science of National Academy of Sciences of Ukraine
Разработан метод анализа фраз на естественных языках флективного типа (украинский и русский), позволяющий выделить в
предложениях основные идеи и группы слов, при помощи которых они излагаются. Сформированные таким образом
семантические деревья высказываний, каждое из которых выражает одну конкретную идею, являются удобным исходным
материалом для построения запросов к онтологии на языке SPARQL. Метод анализа предложений включает следующую
последовательность основных этапов: разбиение на слова, выделение маркерных слов и словосочетаний, определение типа
высказывания, выделение именных групп, составление синтаксического графа предложения, построение семантических деревьев
высказываний, основанных на имеющихся типах высказываний, подстановка параметров из семантических деревьев высказываний
в соответствующие шаблоны SPARQL запросов. В статье приведена UML-диаграмма деятельности системы, которая реализует
описанную модель, и примеры шаблонов SPARQL запросов. Выбор соответствующего шаблона запроса зависит от типа
высказывания, выраженного данным семантическим деревом высказывания. Понятия, полученные в качестве ответа на запрос,
связываются с соответствующим семантическим деревом высказывания. В случае неполучения информации из онтологии,
производится редукция именных групп для выражения более общих понятий и построение запросов с их использованием. Это
позволяет всегда получить некоторый ответ, хотя и не столь точный, как при использовании полной именной группы.
Использование шаблонов SPARQL запросов требует априорно заданной структуры онтологии, которая также предлагается в
данной работе. Такая система применима для ведения диалога с помощью чат-бота, для автоматического получения ответов на
вопросы к тексту. Дальнейшими направлениями исследований являются расширение классификации высказываний пользователя и
моделирование построения уточняющих вопросов при возникновении неразрешимых неоднозначностей.
Ключевые слова: онтология, SPARQL, анализ текста, именная группа, синтаксический граф, семантическое дерево высказывания,
NLP, NLU.
Розроблено метод аналізу природно-мовних речень для мов флективного типу (українська та російська), який дозволяє виділити в
реченні основні висловлені ідеї та групи слів, за допомогою яких вони викладаються. Сформовані таким чином семантичні дерева
висловлювань, кожне з яких виражає одну конкретну ідею, є зручним вихідним матеріалом для побудови запитів до онтології
мовою SPARQL. Метод аналізу речень включає наступну послідовність основних етапів: розбиття на слова, виділення маркерних
слів і словосполучень, визначення типу висловлювання, виділення іменних груп, побудова синтаксичного графа речення, побудова
семантичних дерев висловлювань, заснованих на наявних типах висловлювань, підстановка параметрів з семантичних дерев
висловлювань до відповідних шаблонів SPARQL запитів. У статті наведено UML-діаграма діяльності системи, яка реалізує описану
модель, і приклади шаблонів SPARQL запитів. Вибір відповідного шаблону запиту залежить від типу висловлювання, яке виражено
семантичним деревом висловлення. Отримані в якості відповіді на запит набори понять зв’язуються з відповідним семантичним
деревом висловлювання. У разі неотримання інформації з онтології, проводиться редукція іменних груп для вираження більш
загальних понять і побудови запитів з їх використанням. Це дозволяє завжди отримати деяку відповідь, хоч і не настільки точну, як
при використанні повної іменної групи. Використання шаблонів SPARQL запитів вимагає апріорно заданої структури онтології,
яка також пропонується в даній роботі. Така система може бути застосована для організації діалогу з чат-ботом або для
автоматичного отримання відповідей на питання до тексту. Подальшими напрямками досліджень є розширення класифікації
висловлювань користувача і моделювання побудови уточнюючих питань при виникненні нерозв'язуваних неоднозначностей.
Ключові слова: онтологія, SPARQL, аналіз тексту, іменна група, синтаксичний граф, семантичне дерево висловлення, NLP, NLU.
A method for analyzing sentences in natural languages of inflected type (Ukrainian and Russian) has been developed. The method allows one
to outline main expressed ideas and groups of words in the text by which they are stated. The semantic trees of sentences formed in this way,
each of which expresses one specific idea, are the convenient source material for constructing queries to the ontology in the SPARQL
language. The analysis algorithm is based on the following sequence of basic steps: word tokenize; determining of marker words and
phrases; identifying the available type of sentence; identifying nouns groups; building a syntactic graph of a sentence; building semantic trees
of sentences based on existing types of sentences; substituting parameters from semantic trees of sentences in the corresponding SPARQL
query templates. The article presents the UML-activity diagram of the system that implements the described model and examples of the
SPARQL query templates. The choice of an appropriate template depends on the type of sentence expressed by a given semantic tree of a
sentence. The sets of concepts received as an answer are tied as corresponding answers to the previously defined semantic tree of the
sentence. In the case of non-receipt of information from the ontology, the reduction of noun groups is carried out to express more general
concepts and the building queries using them. This allows us to get some answers, although not as accurate as when we use the full noun
group. The use of SPARQL query templates requires an a priori known ontology structure, which is also proposed in this paper. Such a
system is applicable for dialogue using chat-bots or for automatically receiving answers to questions from the text. Further research
directions are the expansion of the classification of the user speech and modeling of building clarifying questions when an unresolvable
ambiguity.
Keywords: ontology, SPARQL, text analysis, noun group, syntactic graph, semantic tree of a proposition, NLP, NLU.
Copyright © 2020 for this paper by its authors. Use permitted under Creative 322
Commons License Attribution 4.0 International (CC BY 4.0).
Introduction
This paper considers the application of methods for analyzing natural language texts written in languages of
inflectional type, especially, Ukrainian and Russian. Language analysis is considered in the aspect of identifying the
main ideas and intentions in the text, which is necessary for building queries to the knowledge base that provide
relevant information which is using for responses to a user's request construction. Thus, the work is devoted to the
application of NLP (natural language processing) and NLU (natural language understanding) methods to solve the
problem of organizing a dialogue between a user and a knowledge base in a natural language. In addition, attention is
paid to working with knowledge bases of ontological type.
The knowledge base is the central component of an intelligent expert system. Such systems are designed to
search for solutions to problems in a certain area, based on a subject model created by an expert-user. Knowledge bases
work in conjunction with systems for searching, extracting and analyzing information. A full-fledged knowledge base,
in addition to structured information on a certain subject, should also contain a system of semantic (meaningful)
information processing and inference rules that allow making automatic inferences. For such tasks, a hierarchical way
of concepts and their relationships representing is suitable, which is one of the types of ontology [1].
The main input source for an expert system with a knowledge base is the data provided by a user. The most
convenient form of data presentation in this case is usually natural language text. In this way, it is possible to provide a
wide range of information for various subject areas. However, for unified work with data warehouses, computer and
software systems use formalized query languages. For ontologies, SPARQL seems to be a well-known and proven
query language [2]. Its usage is recommended by the W3C consortium, and it is one of the semantic web technologies
[3]. Thus, processing of text in natural language for structured data obtaining that are necessary for SPARQL query
building, becomes an urgent task.
In addition, for effective work with a knowledge base, it is required a certain structure of the ontology, which
includes some restrictions on the initial information about the subject area presentation form. This allows one to unify
SPARQL query templates for various situations, based on the types of statements and intentions from the source text.
Methods for constructing a domain ontology are a separate broad topic. This article presents some aspects of the
ontology structure adopted in the system we are developing. Practical aspects of the considered method are intelligent
chatbots, expert systems, programs for text mining and automatic generation of conclusions basing on it. It should be
noted that the system proposed in this work is designed to work primarily with grammatically and spelling correct text
of scientific and technical style.
Analysis of modern achievements in natural language processing for dealing with ontological
knowledge bases
Various approaches to building queries using as an initial data natural language text have been developed for
several decades since communication in such way is convenient for users. New ideas for natural language interfaces are
constantly being proposed. Not all of them turn out to be successful and really make a breakthrough. However, any new
research in this area is valuable because it brings new ideas and a better understanding of what really works and what
does not [4].
An important problem of expert systems with a natural language interface is their dependence on semantic
grammars adapted to a specific database. At the same time, query languages are currently quite standardized (for
example, SQL – for relational databases, SPARQL – for working with ontologies). This allows one to move away from
binding to the internal structure of the database [4].
There are several approaches to natural language interface building. For example, in [5] it is proposed to focus
on a limited set of semantically interpreted queries with an unambiguous mapping of relations, attributes and values,
and to use statistical semantic analysis. Authors of [6, 7] proposed a model in which a program basing on a natural
language text returns the problem description and forms the query code. In [8] a method is proposed for sentences
generalization applied to syntactic parse trees. Authors of [9] proposed a conversion system that consists of three base
components: 1 – conversion of a query in natural language to a query tree; 2 – interactive check of the conversion (by
asking the user); 3 – SQL query formation.
In recent years appear methods based on machine learning and neural networks. For instance, a technique adopts
reinforcement learning [10]. A neural network is used for translation of questions in natural language to the
corresponding SQL queries. However, despite the huge data set in the training set, the model's accuracy is not very
high: the execution accuracy is 59.4%, the logical form accuracy is 48.3%. The learning technique really depends on the
subject. For such system implementation, a wide collection of the training set is needed, which is often difficult to
provide.
Another possible difficultness of formal queries construction basing on natural language is cases of homonymy
and ambiguity in user’s phrases. Moreover, a user could use slang and specific terms of some subject area, words and
phrases that have a specific meaning in the context which could be different from the common one. Some approach to
this problem solving was proposed even in [11, 12]. It consists of clarifying questions addressed to the user with a
proposition to clarify which of the options is correct for this case.
Works for the creation of natural language questions to SPARQL conversion methods are also performed. For
example, LODQA system (Linked Open Data Question Answering) [13] – a program that receives a query in a natural
language as an input and returns SPARQL queries as a performance result. The system consists of several modules. The
Copyright © 2020 for this paper by its authors. Use permitted under Creative 323
Commons License Attribution 4.0 International (CC BY 4.0).
first module is responsible for syntactic analysis and for the creation of a graphic representation of the query, which is
called a pseudo graphic template. Typically, template nodes correspond to noun groups, and links are dependencies
between them. Also the pseudo graphic template points which of the nodes is the focus of the query i.e. what the user is
to receive as an answer to the query. As the first module has generated a pseudo graphic template from the natural
language query the second module turns on, which is responsible for URI and meanings search for the nodes of the
pseudo graphic template. From one pseudo graphic template, it is possible several bind templates to obtain through
normalization. The third module performs the search in the goal dataset for the corresponding parts taking into account
the possible changes that could appear in the dataset. This module tries to generate SPARQL queries for all possible
structure variations. Then the SPARQL queries are sending to the aim point, where the answers for them are to be
obtained, and then these answers are returned to the user. The arguments of a query could be a primitive type. For
making easier the identification of RDF-triplets the words of a sentence are lemmatized and the sufficient grammatical
characteristics are matched to them. The considered LODQA system is oriented to deal with the English language only.
The detailed features of its functioning are not given in [13], limiting to a brief description and examples analysis.
Thus, we can see that despite quite long existing term of the problem of natural language phrases to formal
queries conversion it still remains actual and new approaches to its solution are proceeding to develop. Most of the
works are devoted to constructing SQL queries basing on natural language (NL2SQL). There are rather few works on
the conversion of natural language phrases to queries language into queries to ontologies and most of them do not reveal
some technical details of the system realization. Most of the programs for natural language to formal query formation
are oriented to English, but inflectional languages, such as Ukrainian and Russian) have rather different structures and
need specific analysis methods.
The proposed model of natural language text analysis to obtain data for construction of
queries to ontology
Ontology is a formalization of some area of knowledge through a conceptual scheme, i.e. linked data structure.
Ontologies are stored in a way convenient for computer processing [14]. A standard storage format is OWL. For formal
queries to ontology the most widespread language is SPARQL. However, for a human more comfortable and
understandable will be information representation in a natural language. Therefore, the task of transition from a phrase
or a set of phrases in natural language to a package of SPARQL queries becomes urgent. Information obtained from
ontology for most of the cases will be reasonable to make out into natural language phrases.
Ontologies as a rule are built in a way to suit to a certain topic or subject. There could be several ontologies in a
system. For selection of the most suitable ontology, a file could be used that contains the key terms for each of the
ontologies. These lists of keywords are to be matched with words from input texts. That ontology will be selected to
work with, for which list of keywords the maximum percent of coincidences in the analyzed text will be found.
Let us describe the sequence of a phrase in the natural language of inflective type analysis. To the languages of
inflective type Ukrainian and Russian belong.
The first stage is graphematic analysis, which consists of tokenization of a text string into sentences, parts of
complex sentences and separate words. For this purpose in our system are used instruments from NLTK library [15].
The words are represented as program objects also able to contain the word’s characteristics, its place in the sentence
and relationships with the other words. The usage of such structure for words representation simplifies the realization of
morphological and syntactic analysis stages.
The second stage is the recognition of the marker words in the sentence for the determination of the sentence
type. In Ukrainian and Russian languages the main criterion in the sentence type determination (statement or
interrogative) in writing document is the presence of a question mark at the end of the sentence. Determination of a
more narrow subtype of a phrase is performed through the presence of certain marker words. The accepted here
classification of phrases is based on [1]. In our model for the moment the following simple subtypes are defined:
“neutral narration” (no marker words), “time”, “place”, “cause”, “object”, “way of doing”, “direction”, “need”. There
could be several marker words or their groups. In such case, a complex subtype is determined, for example, “place and
time”, “cause and time” etc.
In the next stage of the analysis, the noun groups are to be allocated in the sentence. A noun group is a set of
linked nouns and adjectives, which in aggregate describe a certain entity [16]. The main word in a noun group is a noun.
All the external relationships of a noun group are going to this main word. A noun from a noun group is recognized as
linked to its main word if there is a case consistency between them and is standing near it or is separated only with
adjectives that are coherent with the main word or this word. The adjectives from the noun group could also be linked
with adverbs.
Then the building of a syntactic graph is to be performed. The noun groups are nodes of this graph as the
separate words do which are not in groups. The essence of the syntactic graph is the representation of the relationships
between the words of the sentence. Fore inflective languages a syntactic graph could be constructed basing on a certain
combination of parts of speech in the corresponding word forms. In our model, the types of links between words are
based on the parts of a sentence. The name of a link type is related to the dependent part of a sentence to which the link
is aimed. There is some connection of these types to the syntactic relationships in a phrase [17]: attributed, objected,
circumstantial, appositive and compliant ones.
Copyright © 2020 for this paper by its authors. Use permitted under Creative 324
Commons License Attribution 4.0 International (CC BY 4.0).
By the semantic tree of a statement we mean some elementary act performed by the object or with the object or
changing of some property of the object which occurs in the certain given conditions. Structurally it is a subgraph of the
syntactic graph of the sentence, to which the belonging to one of the phrase types is determined, expressed by this
sentence [18]. In the case of a marker word presence, it becomes the root of such tree. The main term of a semantic tree
is in direct relationship with the marker word. Terms linked to the main term, and the other ones linked to them
according to the syntactic graph, form the so-called additional circumstances of the semantic tree. Let us consider an
example of the sentence (in Russian): «Где можно купить книги по программированию на Python?» (eng.: “Where
is possible to buy books on Python programming?”). The marker word here will be «где можно» (eng.: “where is
possible”), hence it a question of place with an additional predicate with the meaning of conformity (variants with the
meaning of nonconformity might be, for instance, «где нельзя» (eng.: “where is not possible”), «где невозможно»
(eng.: “where is impossible”), «где запрещено» (eng.: “where is prohibited”) etc.). The main term here will be the verb
«купить» (eng.: “to buy”). The additional circumstances here are the words «книги» (eng.: “books”),
«программирование» (eng.: “programming”) and «Python» according to the syntactic graph.
In case of the absence of marker words, the root of the semantic tree of the phrase will be an abstract zero term,
which represents the neutrality of the statement for the given tree and does not have a verbal representation in the
sentence. It is needed only for the unification of the trees of phrase in the program implementation. The main term of
the semantic tree of the phrase in such case can become the predicate or the subject. The type of phrase in this case will
be classified as a neutral narration (for a narrative sentence) or a common question (for an interrogative sentence).
Linked with it words will be treated as the additional circumstances.
The semantic trees of the phrase are used for SPARQL queries to ontology constructing. The predetermined
templates are used for this purpose, which type depends on the phrase type. For the process of the templates creation is
important the structure of the ontology accepted in this system. The program performance, which realizes the described
above parsing model, is presented as a UML-diagram of activity shown in figure 1.
Fig. 1. UML activity diagram illustrating the general scheme of the system operation
Copyright © 2020 for this paper by its authors. Use permitted under Creative 325
Commons License Attribution 4.0 International (CC BY 4.0).
A brief description of the ontology structure accepted in the developed system
The conception of ontology structure accepted in our system for the moment is following. There are classes of
the highest level of the hierarchy, which classify terms in the ontology: “action”, “cause”, “method”, “object”, “place”,
“time”. Other variants are also possible according to accepted types of phrases. These classes are devised into more
narrow subclasses, but still abstract enough. For example, class “action” could be devised into “active action” and
“passive action”. An “active action” means a really performed action, activity (“to organize”, “to develop”, “to begin”,
“to perform”). A ‘passive action” is something expressed by a verb but does not assume a real activity and rather
characterize the object's state (“stand”, “exist”, “consist”). The terms of lower hierarchy could be presented by arbitrary
nested classes.
The ontology structure used here does not assume individuals. Limitation to classes only for terms description
unifies SPARQL queries. Properties could also be present in the ontology. They introduce not hierarchical relationships
between concepts. According to the standard of OWL properties have “Domain” and “Range” fields. In the proposed
structure of ontology the “Domain” section consists of a set of independent terms, the “Range” section represents a set
of terms that are the consequence of the combination of the concepts presented in the “Domain”. Thus, the properties
here are peculiar functions where independent variables (factors) are in “Domain” and dependent (responses) in
“Range”. For example, in “Domain” there are the following terms: “begin”, “develop”, “conception of informatization”,
“Ukraine”, and in “Range” the term is “60-s years of XX century”. It should be noticed that the concept “60-s years of
XX century” is a child of the class “Time”. The considered property expresses that the conception of informatization in
Ukraine began to be developed in the 60-s years of the XX century. The belonging of terms to the "Domain" is
important for the correct interpretation of the word’s essence. In this case terms “begin” and “develop” are actions and
“Ukraine” is a place.
Building principles and examples of SPARQL queries
As it has been mentioned above, basing on the semantic trees of phrases, which are formed during the analysis of
the initial phrase, SPARQL queries are constructed. The structure of the query depends on the type of phrase that is in
the base of the semantic tree of the phrase. It is also possible more precise tuning of the variants of the SPARQL queries
in the range of each of the types. Let us consider some examples of SPARQL queries depending on the phrase type.
A neutral narration does not assume a certain answer but it is possible to comment it by finding related concepts.
In this case, if there are no additional circumstances, the SPARQL query template might look like this:
SELECT DISTINCT ?res WHERE {
?y rdfs:domain :inst_name.
?y rdfs:range ?z.
?z rdfs:label ?res.}
In the template inst_name is a variable that is formed from the main term of the semantic tree of the phrase in a
way to fit the classes naming assumed in the ontology, which are formed from a set of the term’s words. The query
assumes to return “label”, which is a field that contains the terms in a more human understandable manner (the class
names are formed by solid join of words stems).
If the semantic tree of the phrase also has additional circumstances, the query template is more complicated:
SELECT DISTINCT ?res WHERE {
?y rdfs:domain :inst_name.
?y rdfs:domain :cur_sup_name_1.
…
?y rdfs:domain :cur_sup_name_n.
?y rdfs:range ?z.
?z rdfs:label ?res.}
In the template inst_name, as it was in the previous case, is a variable formed from the main term of the semantic
tree of the phrase; cur_sup_name_1 … cur_sup_name_n are variables that are forming from the additional
circumstances.
The following example illustrates a SPARQL query for the case of the semantic tree of a cause phrase:
SELECT DISTINCT ?res WHERE {
?x0 rdfs:subClassOf :cause.
?x_n rdfs:subClassOf ?x0.
…
?last_x rdfs:subClassOf ?x_n.
:inst_name rdfs:subClassOf last_x.
?y rdfs:domain :inst_name.
?y rdfs:domain :sup_1.
…
?y rdfs:domain :sup_i.
?y rdfs:range ?z.
?z rdfs:label ?res.}
Copyright © 2020 for this paper by its authors. Use permitted under Creative 326
Commons License Attribution 4.0 International (CC BY 4.0).
There are the following parameters in the template: inst_name is a variable formed from the main term of the
semantic tree of the phrase; sup _1 … sup _n are variables that are forming from the additional circumstances. Such
template structure guarantees that the main concept of the semantic tree will be considered in a context of cause
regardless of the given concept hierarchy depth.
The next template of SPARQL query is suitable for a semantic tree of a time phrase:
SELECT DISTINCT ?res WHERE {
?y rdfs:domain :inst_name.
?y rdfs:domain :sup_1.
…
?y rdfs:domain :sup_i.
?x rdfs:subClassOf :time.
?y rdfs:range ?z.
?z rdfs:subClassOf ?x.
?z rdfs:label ?res. }
In the template there are the following parameters: inst_name is a variable formed from the main term of the
semantic tree of the phrase; sup _1 … sup _n are variables that are forming from the additional circumstances. Here for
simplification, it is assumed that all the classes denoting time terms are directly inherited from “time” class.
SPARQL queries templates for the semantic trees of phrases of other types are in a hole similar in their structure
to the mentioned above ones and so do not given here. On the other hand, it shows sufficient versatility of the proposed
model to deal with phrases of different types.
The result of a SPARQL query is a set of concepts that match its selection criteria. If there are no suitable objects
in the ontology, then an empty list is returned. In case of obtained data absence, the program executes new SPARQL
queries using the reduced noun groups and/or reduced terms list of additional circumstances. The answer, in this case,
may appear less relevant, but nevertheless, some answers to the request will be received more probably.
The set of concepts obtained as a result of a query is certainly essential, but this information might be not enough
for further machine processing and formation of a natural language phrase for the answer. Therefore, it is to be clarified
which class of the higher hierarchy is inherited by each of the found concepts. For this purpose for each of the obtained
concepts, a series of SPARQL queries is to be performed, aimed at finding which class is the parent of the class used in
the previous iteration. The iterations end when the “Thing” class, which is the root for all the classes, appears as the
parent of the class at the current iteration. The template for such request is shown below:
SELECT DISTINCT ?res WHERE {
?x rdfs:label current_class.
?x rdfs:subClassOf ?y.
?y rdfs:label ?res. }
In the template current_class is a value of the “label” parameter for the class from the previous iteration,
on the first iteration, it is one of the terms returned by the main query.
Thus, it is possible to get a chain of the hierarchy of higher-level concepts to which belongs the concept returned
by the main query. This information is useful for a more informative and correct construction of a natural language
answer which is based on the found by the main query concepts. So knowing whether a given concept is an object,
action, time, place, etc. it is easier to arrange words in the proper order and, if necessary, supplement them with
appropriate prepositions.
Let us consider an example of the process that includes analyzing a phrase, building a request and receiving a
response from ontology. The example uses an ontology compiled according to the report of the academician
V. S. Mikhalevich on the concept of society informatization. Here is the user's question in Russian: «В чём состоит
значение информатизации для человеческого общества?» (eng.: "What is the significance of informatization for the
human society?") From this phrase, the system selects one semantic tree: the type of phrase is “question – request for a
list of objects”; marker phrase is «в чём состоит» (eng.: “what is”); the main concept is the noun group «значение
информатизации» (eng.: “the significance of informatization”); additional circumstances are represented by the noun
group «человеческое общество» (eng.: “the human society”). For this case, a SPARQL template is provided, into
which lemmatized and stemmed concepts are substituted (marked by italics and underlining):
SELECT DISTINCT ?res WHERE {
?y rdfs:domain :ЗначенИнформатизац.
?y rdfs:domain :ЧеловеческОбществ.
?y rdfs:range ?z.
?z rdfs:label ?res. }
This case is rather simple, there is only one additional circumstance, and concepts are not specifically marked to
be descendants of certain classes. The result returned by the query is a set of concepts (terms): «решение научно-
технических проблем» (eng.: “solving scientific and technical problems”), «внесение значительного
экономического вклада» (eng.: “making a significant economic contribution”), «повышение производительности
труда», (eng.: “increasing labor productivity”).
The resulting sets of concepts can be used for the following processing. For example, if the application is a chat-
bot, then the concept sets, along with the semantic data and the history of the dialogue, can be used to form a natural
Copyright © 2020 for this paper by its authors. Use permitted under Creative 327
Commons License Attribution 4.0 International (CC BY 4.0).
language phrase for the response. However, a detailed description of the SPARQL query results processing is beyond
the scope of this paper.
Conclusions and perspectives for the following investigations
A method of analysis is developed for a natural language text in Ukrainian and Russian allowing building
semantic trees of phrases, which are convenient for SPARQL queries to ontology forming. The semantic trees of
phrases are characterized by the marker words and the expression type. In the initial sentence, several semantic trees
could be determined and for each of them, an appropriate SPARQL query can be formed.
Templates of SPARQL queries that make it possible to form the queries to ontology basing on natural language
expressions of various types are presented. The type of template is determined by the type and the structure of the given
expression, which is represented through the semantic its tree. The template contains inline parameters that are obtained
from the corresponding nodes of the semantic tree of the phrase.
The model uses the approach of the noun groups reduction for the situation if the result of a SPARQL query is
unsatisfactory (no data obtained). This allows the system to request information on more general concepts, and get some
answer, although might be less relevant.
The proposed system could become an element of a user’s friendly natural language interface dealing with a
knowledge base represented as an ontology or set of ontologies. Also, it may be used for intelligent chat-bots creation.
The future perspectives of the program system development might be an extension and more precise
classification of user’s phrases, modeling of clarifying questions construction for cases of unsolvable ambiguities,
generation of grammatically correct phrases for the answers which form depends on the semantic of the question.
References
1. Gavrilova, T. A. & V. F. Khoroshevsky (2000) Knowledge Base of Intelligent Systems. St. Petersburg: Peter.
2. Antoniou, G. (2016) Semantic Web. Moscow: DMK-Press.
3. W3C (2013) SPARQL 1.1 Query Language [Online] Available from: https://www.w3.org/TR/sparql11-query/ [Accessed: 11 February 2020].
4. Galitsky, B. (2019) Developing Enterprise Chatbots. Learning Linguistic Structures. San Jose: Springer.
5. Popescu, A. M., Etzioni, O. & Kautz, H. A. (2003) Towards a theory of natural language interfaces to databases. IUI. p. 149 – 157.
6. Galitsky, B. & Usikov, D. (2015) Programming Spatial Algorithms in Natural Language. AAAI Workshop Technical Report WS-08-11. p. 16 –
24.
7. Quirk, C., Mooney, R. & Galley, M. (2015) Language to code: learning semantic parsers for if-this-then-that recipes. ACL. p. 878 – 888.
8. Galitsky, B., De La Rosa, J. L. & Dobrocsi, G. (2011) Mapping syntactic to semantic generalizations of linguistic parse trees. Proceedings of
the twenty-fourth international Florida artificial intelligence research society conference. p. 168 – 173.
9. Li, F. & Jagadish, H. V. (2016) Understanding natural language queries over relational databases. SIGMOD Record. 45. p. 6 – 13.
10. Zhong, V., Xiong, G. & Socher, R. (2017) Seq2SQL: generating structured queries from natural language using reinforcement learning.
[Online] Available from: https://arxiv.org/pdf/1709.00103.pdf [Accessed: 11 February 2020].
11. Kupper, D., Strobel, M. & Rosner, D. (1993) Nauda – a cooperative, natural language interface to relational databases. SIGMOD conference. p.
529 – 533.
12. Li, Y., Yang, H. & Jagadish, H. V. (2005) Nalix: an interactive natural language interface for querying xml. SIGMOD conference. p. 900 – 902.
13. Shaik, S., Kanakam, P., Hussain, S. M., Suryanarayana, D. (2016) Transforming Natural Language Query to SPARQL for Semantic Information
Retrieval. International Journal of Engineering Trends and Technology. 7. p. 347 – 350.
14. Lapshin, V. A. (2010) Ontologies in computer systems. Moscow: Scientific World.
15. NLTK Project (2019) Natural Language Toolkit. NLTK 3.4.5 documentation. [Online] Available from: https://www.nltk.org [Accessed: 11
February 2020].
16. Crystal D. A (2008) Dictionary of Linguistics and Phonetics Wiley-Blackwell.
17. Kurysheva, M.V. (2014) Russian language: syntactic analysis of phrases and simple sentences. Tomsk: Tomsk State Pedagogical University.
18. Shelmanov, A. O. (2015) Ph.D. Tresses: Study of methods for automatic text analysis and development of an integrated system of
semantic-syntactic analysis. Moscow.
Об авторах:
Литвин Анна Андреевна,
аспирант Института кибернетики им. В. М. Глушкова НАН Украины.
Количество научных публикаций в украинских изданиях – 2.
Количество научных публикаций в зарубежных изданиях – 1.
http://orcid.org/0000-0002-5648-9074
Величко Виталий Юрьевич,
кандидат технических наук, доцент,
старший научный сотрудник отдела микропроцессорной техники Института кибернетики им. В.М. Глушкова
НАН Украины; ведущий научный сотрудник отдела создания и использования интеллектуальных сетевых
инструментов Национального центра «Малая академия наук Украины» по совместительству;
Количество научных публикаций в украинских изданиях – 75
Количество научных публикаций в зарубежных изданиях – 27.
H-index: Google Scholar – 11
Scopus – 1,
http://orcid.org/0000-0002-7155-9202.
Copyright © 2020 for this paper by its authors. Use permitted under Creative 328
Commons License Attribution 4.0 International (CC BY 4.0).
Каверинский Владислав Владимирович,
кандидат технических наук,
старший научный сотрудник отдела износостойких и коррозионностойких порошковых конструкционных
материалов Института проблем материаловедения им. И. Н. Францевича НАН Украины.
Количество научных публикаций в украинских изданиях – 81.
Количество научных публикаций в зарубежных изданиях – 18.
H-index: Google Scholar – 4
Scopus – 2,
http://orcid.org/0000-0002-6940-579X
Место работы авторов:
Институт кибернетики им. В. М. Глушкова НАН Украины.
03187, Киев-187, проспект Академика Глушкова, 40.
Литвин Анна Андреевна, Тел.: (097) 570-99-84, E-mail: litvin_any@ukr.net
Величко Виталий Юрьевич, Тел.: (096) 139-96-28, E-mail: aduisukr@gmail.com
Института проблем материаловедения им. И. Н. Францевича НАН Украины.
03142, Киев, ул. Кржижановского, 3.
Каверинский Владислав Владимирович, Тел.: (050) 212-17-24, E-mail: insamhlaithe@gmail.com
Copyright © 2020 for this paper by its authors. Use permitted under Creative 329
Commons License Attribution 4.0 International (CC BY 4.0).