Working prototypes of the scalable semantic web portals, which are deployed on cloud platforms and intended for use in universities educational activity, are discussed. The first project is related to teaching in the field of nuclear physics and nuclear power engineering. The second project is related to training in computer science and programming. The possibility of using the DLLearner software in conjunction with the Apache Jena Reasoners in order to refine the ontologies that are designed on the basis of the SROIQ(D) description logic is shown. A software agent for the context-sensitive searching for new knowledge in the WWW has been developed as a toolkit for ontologies enrichment. The binary Pareto relation and Levenshtein metrics are used in order to evaluate the measure of compliance of the found content concerning a specific domain. It allows the knowledge engineer to calculate the measure of the proximity of an arbitrary network resource about classes and objects of specific knowledge graphs. The suggested software solutions are based on cloud computing using DBaaS and PaaS service models to ensure the scalability of data warehouses and network services. Examples of applying the software and technologies under discuss are given.
The world of data is a place where computers rule. Supercomputers have wondrous capabilities, but they often find it difficult when it comes to acquiring new knowledge and experience or existing knowledge categorization. While it's natural for a human to decide whether two or more things are related based on cognitive associations, a computer often fails to do it. The endowment of machines with common sense, as well as domain–specific knowledge in order to give them an understanding of certain problem domains, has been and remains the main goal of research in the field of artificial intelligence. While the amount of data on the WWW, as well as in corporative intranets headily grows, knowledge databases engineering still remains a challenge. This paper discusses, how the semi–automatic methods work for knowledge graphs refinement and enrichment.
A recent authoritative review of the latest achievements and current issues in the
designated field of the Semantic Web is given in [
Ontologies are often considered as special knowledge repositories that can be read
and recognised both by people and computers, separated from the developer and
reused. Ontology in the context of information technology is a formal specification with
a hierarchical structure, which is created to represent knowledge. Typically, an
ontology holds descriptions of classes of entities (concepts) and their properties (roles)
with respect to a certain subject domain of knowledge, as well as associations
between entities and constraints on how these associations can be used. Further, we
adhere to the formal definition of ontology, which is given in [
The initial ontology design is performed by a knowledge engineer with the involvement of domain experts. In the process of creating a quality ontology, clearly articulated databases normalization principles should be taken into account that reflects best practice. Web ontologies at the design stage already define a hierarchical structure of knowledge, which can be expressed explicitly or it may be detected indirectly, based on the available axioms of categorization. During the subsequent practical use of the ontology, it may turn out that some designed classes are sparsely populated since the created hierarchy reflected the subjective point of view of the knowledge engineer (ontology designer), which does not correspond to the actual filling of the ontology with specific resources.
In this case, the reengineering the structure of the created ontology is inevitable.
Adequate allocation of resources into appropriate classes is a fundamental intellectual
service of the semantic web. Existing clustering methods offer an effective solution to
support a variety of complex related actions, such as building ontologies, taking into
consideration the inherent incompleteness underlying the representation of the facts.
Among the large number of algorithms proposed in the machine learning literature,
the concept cluster approach [
There are two features that should be taken into account during the ontology design. The first feature is the Open World Assumption. OWL by default considers the world open, that is, everything that is not explicitly specified in the ontology is not considered as false, but considered as possible. Facts and statements that are false or impossible should be clearly stated as so in the ontology.
The second feature concerns the use of the disjointness axioms in the ontology. Why is class disjointness important? This is due to the reliability of the logical entailments that can be obtained from the ontology. The point is that OWL does not imply class disjointness until it is explicitly declared. That is, by default, classes may overlap unless otherwise specified. If anyone simply declares two classes A and B in the ontology and say nothing more about them, than any ontology model can interpret these classes as it pleases, for example, as nested classes, as intersecting classes or as disjoint classes. However, if two entities in the domain really belong to different classes, this fact should be reflected in the ontology. One of the purposes of ontologies is to make domain knowledge explicit and reliable, which is why the axioms of disjointness are so important.
As an illustration of the ontology engineering process in the project [
The particularity of our case study was that when working with the DL-Learner,
we did not use the built-in Pellet, FaCT ++, HermiT or OWLlink reasoners, since they
are mainly focused on the use of ALC-level description logics. We investigated the
ontology, which is designed on the basis of the more rich SROIQ(D) description
logic, therefore the more advanced Apache Jena Reasoner to the DL-Learner [
Nowadays the DL-Learner software product is perhaps the most popular of the
available solutions for semi–automated ontology engineering. To reveal dependencies
hidden in the ontology, the refinement operators, heuristic measures and training
algorithms named OCEL and CELOE are used, see [
As a toolkit that provides data for ontologies refinement and enrichment, a software agent (which is actually a specialized meta–search engine) for the survey context– sensitive search for new knowledge in the WWW is implemented. To begin with it, should be noted several essential features of public search engines that are well known to most maximum users.
• the content found are ordered by the public search engine in accordance with its intrinsic algorithm, which does not always meet the interests of a special user; • users are not always convenient to manage the context of the search query, clarify and focus the search;
• links to the business sites usually have a higher rank than other search results. Such effect is gained through the use of so-called search engine optimization (SEO) to artificially inflate the positions of commercial network content on pages of public search engines, in order to boost the flow of potential consumers for the subsequent monetization of traffic.
It appears that the above points and inclinations make known search engines an increasingly incompetent tool for extracting knowledge in the WWW for educative purposes. The context-sensitive search is based on a simple idea: to create such a mediator (a software agent) between the knowledge engineer and public search engines that help to arrange search results in accordance with his professional wants, by effectively sifting inappropriate content and trash. The aim is to include the power of the modern search engines in the maximum level, including built-in query languages and other search handles.
When the «Context-sensitive search» software agent is operating, the global content search, as well as the search on the specific web resources, is originally conducted by the conventional search engines (Google Ajax Search, Yandex, Yahoo, Mail.ru), the communication with which occurs asynchronously via the vibrant pool of the proxy servers, each of which is hosted on the Google Cloud Platform. The results of the activity of the conventional search engines are a kind of «raw material» for extra processing. Especially designed proxy servers on the cloud platform parse these results and generate the feeds, which are then forwarded to the client computer, where from the feeds the snippets are formed. These snippets, which metadata, before they arrive on the monitor of the client computer, undergo supplementary processing, screening and sorting, as outlined below. In particular, for each snippet, its relevance, persistence and a number of other indexes are computed, which are then used to organise and clustering search results retrieved. 3.1
The query language of some search engines may involve the so-called «search context». It is about using immediately in the text of the search query of particular operators, which allow the user to designate the presence and relative locating of specific tokens in the content found. In this paper, a «search context» is understood a slightly another way, namely, as a certain limited on length text that designates the domain that is currently of matter to the knowledge engineer.
When setting the search context, the subsequent data sources are available:
taxonomies, thesauri, keywords, ontologies, textual files from the user computer, arbitrary
content from the WWW. Any mixture of the above methods for setting the search
context is enabled. The resulting context is the union of the chosen options. The
context defined in this way allows us to choose, sort and classify information that
originates from the search engines through the proxy servers. Fig. 3 below exposes the
possible options for setting the search context.
For the goals of this paper, the relevance of the snippet is the measure of the similarity
between the snippet and the text of the search query. Under the pertinence of the
snippet is intended the measure of the similarity between the snippet and search
context, that was determined earlier. These and other measures are estimated by means a
fuzzy matching of the corresponding texts. To quantify these measures, «Context–
sensitive search» software agent uses the Levenshtein metrics [
Various options for classifying search results in the final output of the
«Contextsensitive search» software agent is permitted. Deserves a particular mention the
sorting by aspect named «dominance index», which supplies a joint account of the values
of many metrics that describe the adequacy of the snippets. For example, the
dominance index, in addition to the relevance and pertinence of the snippets, can also take
into account the measure of the similarity between the snippet and the keywords,
categories and attributes of the educational portal in total. For the practical calculation
of the values of the dominance index, it seems reasonable to use the formalism of
Pareto dominance relation [
Let given the original set of snippets, from which one should choose some optimal subset. The choice should be made on the basis of specific ideas about the adequacy of snippets (the principle of optimality). The choice task is a simple one if there is only a single aspect by which it is reasonable to compare any two snippets and directly indicate which one is more adequate. The solution to the simple choice tasks is naive. In real circumstances, it is not possible to single out any one aspect. Furthermore, it is often commonly hard to single out aspects. The selection and ranking of aspects that are quintessential for subsequent choice, in turn, is the task of choice. If some of the aspects are more significant (priority) of other aspects, this circumstance should be taken into account in the mathematical model of choice.
The selection task is the algebra , where is a set of alternatives (in our case, a set of snippets), and is the optimality principle. The task makes sense if the set of alternatives is known. Usually the principle of optimality is unknown.
For further discussion, suppose that each snippet x is characterized by a finite set of aspects x (x1, x2, ..., xm ) . Let {1, ... , m } be the set of aspect numbers to consider when choosing; {} is the set of all subsets .
It can be assumed that choosing between any two snippets x and y with only one of any aspect taken into account is a simple task. If this is not the case, the corresponding aspect can be decomposed and presented as a group of simpler aspects. For each pair of snippets (x, y) we define a family of functions j (x, y) as follows: 1, if x exceed y in aspect j j (x, y) where j ; x, y ; 0, if y exceed x in aspect j
If x and y are equal or not comparable in some aspect with the number j , then for such number j the function j (x, y) is not defined. Let's form a set J of numbers of such aspects that x and y differ in these aspects
J { j : j ; j (x, y) is defined }, J {};
Next, we construct a metric that takes into account the number of aspects by which a particular snippet is inferior to all other snippets. Let there be two snippets x, y . Denote d ( y, x) j ( y, x)
jJ the number of aspects in which y is better than x . Then the value
D (x) max d ( y, x) y (1) is called the dominance index of x when presenting the set. This value characterizes the number of aspects of the snippet x that are not the best in comparison with all other snippets available in the set.
Let us define the function C D () for selecting the best snippets as follows: C D () { x : D (x) min D (z)} z Here, the value D min D (x) is called the index of dominance of the whole x set. Snippets with a minimum value of the dominance index form the Pareto set. The Pareto set includes snippets that are the best with respect to all the considered aspects, including relevance and pertinence.
In the projects [
Unlike conventional lexical search where search engines look for literal matches of the query words and their modifications, semantic annotation attempts to interpret ordinary language close to how people do it. During semantic annotation, all allusions to cases related to entities in the ontology are appreciated. Semantic annotation is the adhesive that ties ontologies into document spaces, via metadata.
The workbench for executing the semantic annotation process is shown in Fig. 4 below. At the top of the workbench is a workspace for entering and editing network content addresses (URLs) to be annotated. The data in this workspace can be originated from any source, including manually. However, a more technologically high-level approach is to first obtain on the WWW those network resources that are most satisfying to a given domain using the «Context-sensitive search» software agent. The found suitable content can then be obviously loaded using the «Download resources» button and included in the list for annotation with a single mouse click. The settings sheaf for the semantic annotation process is shown in Fig. 5 below. For annotation, you can select any of the knowledge graphs that are presented in the semantic repository, as well as any aggregate of them. To calculate measures of similarity between the annotated content and entities from knowledge graphs, both text analysis methods and neural networks that are trained on existing knowledge graphs can be practised.
It is possible to annotate network resources using classes (concepts) of the ontology (TBox - terminological components), using objects (individuals) of knowledge graphs (ABox - assertion components), or using both of them.
The depth of the semantic analysis can be limited by considering textual metadata inherent in network resources and entities in knowledge graphs. It can be very expensive to carry out full-text semantic analysis and in many ways redundant. Improving the accuracy of annotation in full–text analysis often does not justify the increased consumption of computing resources.
The number of entities from the knowledge graphs can be limited by the user. The
entities that are most adequate to the annotated resource appear at the top of the
output of the «Semantic annotation» software agent. All the results can be saved in files
on the user's computer for later study.As an example of the use of a software agent
"Semantic Annotation", Fig. 6 below shows the results of the semantic annotation of
the network resource. It can be seen that have been discovered semantic annotation of
five different graphs knowledge. With one click the user can open a browser and
visualize RDF annotations found in any of the graphs of knowledge, as well as anyone
can see the environment found entities, such as classes and their neighboring objects.
This information is necessary for the knowledge engineering, which deals with
refinement and enrichment of knowledge graphs..
Communities of scientists from the University of Manchester, Stanford University,
the University of Bari, Leipzig University, Cambridge University and a number of
other universities focus on the development of the theory and technology
implementation for the Semantic Web, Description Logic and Machine Learning. Among the
publicly available working frameworks that are designed to enrich knowledge graphs
with content from the WWW, the REX project should be mentioned first [
It is worth commenting on the use of the Levenshtein distance to calculate a
measure of similarity between two pieces of text. Today there are more than three dozen
algorithms that solve similar problems in various ways [
As for the use of the binary Pareto relation for multi-criteria ranking and clustering
of the found network content, the use of this fruitful idea is not fundamentally
innovative. For example, the Skyline software [
It is necessary to develop and improve tools for the intuitive perception of linked
data for non–professionals. VOWL [
Unlike to the above decisions, the projects [
The reported study was funded by the Russian Foundation for Basic Research and Government of the Kaluga Region according to the research project 19–47–400002.