Semantic Technologies for the Web of Linked Data

Nick Bassiliades,
Aristotle University of Thessaloniki
Enhancing seCurity and privAcy in the Social wEb: a user-centered
approach for the protection of minors
Funded by the Horizon H2020 Framework Programme of the European Union under grant agreement no 691025.
Semantic Technologies for the Web of Linked Data
ENCASE Summer School
Limassol, 17th July 2017

H2020 – Grant Agreement no. 691025
• Nick Bassiliades (Νικόλαος Βασιλειάδης)
‒ http://intelligence.csd.auth.gr/people/bassiliades
• Associate Professor, Department of Informatics, Aristotle University of
Thessaloniki, Greece
• Scientific specialization: Knowledge Systems
‒ Knowledge Representation & Reasoning (Rule-based systems, Logic
programming, Defeasible Reasoning, Knowledge-based systems / expert
systems)
‒ Semantic Web (Ontologies, Linked Open Data, Semantic Web Services)
‒ Multi-agent systems
‒ Intelligent Applications on e-Learning, e-Government, e-Commerce, Electric
Vehicles
N. Bassiliades, ENCASE Summer School, Limassol, 17th July 2017 2
A few words about the speaker

• Aristotle University of Thessaloniki, Greece
‒ Largest University in Greece and South-East Europe
‒ Since 1925, 41 Departments, ~2K faculty, ~45K students
• Dept. of Informatics
‒ Since 1992, 28 faculty, 5 research labs, ~1100 undergraduate students, ~200 MSc
students, ~80 PhD students, ~120 PhD graduates, >3500 pubs
• Software Engineering, Web and Intelligent Systems Lab
‒ 7 faculty, 20 PhD students, 9 Post-doctorate affiliates
• Intelligent Systems group (http://intelligence.csd.auth.gr)
‒ 4 faculty, 7 PhD students, 17 PhD graduates
‒ Research on Artificial Intelligence, Machine Learning / Data Mining, Knowledge
Representation & Reasoning / Semantic Web, Planning, Multi-Agent Systems
‒ 430 publications, 35 projects
A few words about my institution

• Semantic Web & Linked Open Data
‒ Introductory Concepts
‒ RDF, RDF Schema, Ontologies, OWL, Reasoning
• Storing and Querying RDF
‒ SPARQL, DBpedia
• Ontologies, Logic & Rules
‒ Horn Logic, OWL 2 RL, SWRL, SPIN
• Use Case: (Semantic) Entity Identification / Linking
‒ URank application (Prolog)
A few words about the talk

Semantic Web &
Linked Open Data
RDF, RDF Schema & OWL

• A shift from today’s Web, from publishing data in human readable
HTML documents to machine readable documents.
• Today, much of the data we get from the web is delivered to us in the
form of web pages
‒ HTML documents that are linked to each other through the use of hyperlinks
• Humans or machines can read (and browse/crawl) these documents
• Machines can seek keywords in a page
• Machines have difficulty extracting any meaning from these
documents themselves
The Semantic Web

H2020 – Grant Agreement no. 691025 N. Bassiliades, ENCASE Summer School, Limassol, 17th July 2017 7
Web evolution
https://atomate.net/blog/web-technologies-now-and-tomorrow/

• An extension of the Web through standard data formats and
exchange protocols
‒ Most important: RDF, OWL, SPARQL, RIF, SWRL, SPIN, …
• SW provides a common framework that allows data to be shared and
reused across application, enterprise, and community boundaries in
the web.
“… a web of data that can be processed by machines”.
Tim Berners-Lee
The Semantic Web

From the Web of Documents to the Web of Data
https://atomate.net/blog/web-technologies-now-and-tomorrow/

Web of (Linked) Data
Current Web Semantic Web
Web of Documents Web of Data
Hypertext Documents (HTML) Semi-structured Data representation in graphs
(RDF)
Interconnected documents through
URL links
Linked data through URIs
Human Consumption Machine (and human) Consumption
Use search engines and browsing to
explore
Use search engines, web (RDF) databases to query
and URI links between datasets to explore
Don't just link the documents, link the things

• Some data should be freely available to everyone to use and
republish, without restrictions from copyright, patents or other
mechanisms of control.
‒ Similar to open source, hardware, content and access.
‒ Benefits: Transparency and democratic control, Improved or new private
products and services, Improved government services, New knowledge from
combined data sources and patterns in large data volumes
• Open data must be available in a convenient and modifiable form.
• Interoperability: the ability to combine different datasets together
and to develop more and better products and services
Open Data

XLS
DOC
PDF

• But, how exactly should open data be published (technological side)?
‒ In order to be easily re-used by other people and be interoperable?
• Tim Berners-Lee, the inventor of the Web and Linked Data initiator,
suggested a 5-star deployment scheme for Open Data.
Open Data Formats & Levels

5 ★ OPEN DATA

• 3 ★ OPEN DATA (CSV):
‒ The data is available via the Web; everyone can use the data easily, with no
proprietary software
‒ But, it’s still data on the Web and not data in the Web.
• 4 ★ OPEN DATA (RDF):
‒ Data items have a URI and can be shared / bookmarked on the Web.
‒ You can reuse parts of the data.
‒ Data can be stored in RDF databases (triple stores) and can be queried via public
endpoints on the Web through remote clients (SPARQL protocol)
‒ You can combine the data safely with other data. URIs are a global scheme.
Going up the ladder

• 5 ★ OPEN DATA (LOD):
‒ Data in the Web (RDF), linked to other data in the Web (URI links via owl:sameAs).
‒ Both the consumer and the publisher benefit from the network effect.
Publisher adds value to data by linking them to other data with more information.
Consumer can discover more (related) data while consuming the data.
Consumer can directly learn about the data schema.
Publisher needs to invest resources to link data to other data in the Web.
Publisher may need to repair broken or incorrect links.
Consumer trusts the consumed data, but what about the data from external links?
Linked Open Data

• Use URIs as names for “things”, so that names are globally unique (IDs)
‒ Real inanimate or animate things, abstract concepts, …
• Use HTTP URIs, so that “things” can be looked up
‒ E.g. using a browser
• When someone looks up a URI, provide useful information, using the
standards (RDF, SPARQL)
• Include links to other URIs so that more things can be discovered
‒ Links are actually RDF properties interpreted as hyperlinks
Linked Data Principles

• URIs – A mechanism to identify things (IDs)
• HTTP – A mechanism to access things
• RDF – A mechanism to describe things and their relationships
• RDFS/OWL – A mechanism to describe vocabularies of properties and
relationships of things
Linked Data Technology Stack

• A data model (framework) for representing information (describing
resources) in the (Semantic) Web
• Compared to other data models, RDF is based on graphs
‒ Graphs have nodes and edges
In RDF graphs:
• Nodes are resources:
‒ Webized entities, i.e. “things” / objects that we want to talk about in the web
‒ Literals, i.e. constant atomic values of various data types
• Edges are properties (also called predicates):
‒ Relationships between entities
‒ Attributes of an entity, linking it with attribute values
Resource Description Framework (RDF)

RDF graph example
http://lpis.csd.auth.gr http://www.csd.auth.gr/bassiliades
http://acm.org/SemanticWeb97913
has-admin
has-phone works-on
…
…
…

• The RDF graph consists of many simple node  edge  node parts
‒ Called “triples”
• The triple can be “read” as a natural language “statement”
‒ “Nick” “has phone” “97913”
• The three parts of the triple have different names
‒ Syntactical terms
Triples / Statements
http://.../bassiliades 97913
has-phone

• Tabular data: (SQL Databases, Excel, CSV, etc.)
‒ Information arranged in a strict grid.
‒ Adding / removing data is easy
‒ Changing the shape of table is a much higher cost.
• Tree data: (JSON, XML)
‒ Loose structure / can represent easier semi- and un- structured information
‒ Tricky to modify the structure and merge data from multiple sources, especially if
those sources were not designed with the merger in mind.
• Graph data: (RDF)
‒ A set of relationships (triples) between things – it can be any shape - more flexible.
‒ Merging 2 RDF documents is trivial. (Union of two sets!)
Why graphs?

• RDF uses globally unique identifiers (URIs) for everything
‒ Things we're talking about (resources)
‒ Relationships (properties, predicates)
‒ Datatypes (literals)
• RDF is a list of unambiguous relationships
‒ Merging / combination of two graphs is trivial
Global Identifiers

• URI: Universal Resource Indicator - identifies something uniquely.
‒ A URI represents a single concept or thing
‒ Many URIs can represent the same thing
‒ If you resolve a URI it's considered good practice to return some useful triples about
the concept the URI represents (optional)
• URL: Universal Resource Location - not only identifies something, but also
describes where it is located.
• All URLs are URIs. Not all URIs are URLs.
• Example
‒ <http://dbpedia.org/resource/Julius_Caesar>: URI for Julius Caesar.
‒ <http://en.wikipedia.org/wiki/Julius_Caesar>: URL for a web page about Julius Caesar
URI vs URL

• There are several ways of writing RDF triples into a file.
• RDF+XML
• Turtle (and N3)
‒ N-triples
• RDFa
‒ Embed triples into an HTML document
‒ Triples can be extracted from the web page by tools (e.g. pyRdfa)
RDF Documents

• For URIs it's common to define related concepts in the same namespace
• A namespace is like a directory on a filesystem
‒ All the contained “things” share the same “global” address
‒ The “full address” ends with either "/" and "#"
‒ The local names (IDs) of “things” don't have "/" or "#“
‒ The global names (IDs) of “things” are constructed by namespace:local_name
• In RDF it's common to use a namespace prefix to make things readable
Namespaces

• In usual graphs (Graph Theory), all nodes and edges are of the same type
(e.g. cities and roads)
• In the RDF model, nodes can belong to different types and all edges are
labelled, so they have different meaning
• Semantic Networks
(semantic nets) (known from AI)
• How do we know which types
of nodes we can use?
‒ What’s in a type?
• What edges' labels (properties) can we use?
‒ Are they related to node types?
Semantic Networks
Web page
Researcher
Research
Topic
integer

• The term ontology originates from philosophy
‒ The study of the nature of existence
• Different meaning from computer science
‒ An ontology is an explicit and formal specification of a conceptualization
• An ontology is an artifact (1-2-3 ontologies)
• An ontology in the Semantic Web is a formally defined vocabulary that
dictates what node types are there and what properties relate which node
types
‒ And other things, as well
Ontologies

• Terms denote important concepts
(classes of objects) of the domain
‒ e.g. professors, staff, students, courses,
departments
• Relationships between these terms:
typically class hierarchies
‒ a class C is a subclass of another class B
if every object in C is also included in B
‒ e.g. all professors are staff members
Typical Components of Ontologies

• Properties / attributes:
‒ X’s phone number is 98713
• (direct) Relationships
‒ X teaches Y
• Relationship restrictions
‒ Faculty members can teach courses
• Disjointness statements
‒ faculty and general staff are disjoint
• Other restrictions on relationships between objects
‒ every department must include at least 10 faculty members
Further Components of Ontologies

• Ontologies provide a shared understanding of a domain:
‒ semantic interoperability
‒ overcome differences in terminology
‒ mappings between ontologies
• Ontologies are useful for the organization and navigation of Web sites
• Ontologies are useful for improving the accuracy of Web searches
‒ Search engines can look for pages that refer to a precise concept in an ontology
‒ If a query fails to find relevant documents, the search engine may suggest a more
general query
‒ If too many answers are retrieved, the search engine may suggest specializations
The Role of Ontologies on the Web

• The most important feature of ontologies is the reasoning support
‒ Reasoning: applying knowledge to arrive at solutions
‒ Inferencing: deriving a conclusion based on statements that only imply that
conclusion
• Reasoning/inferencing support is important for:
‒ Checking consistency of ontology and knowledge
‒ Checking for unintended relationships between classes
‒ Automatically classifying instances in classes
‒ Deriving information/knowledge not known before
Ontologies and Reasoning

• RDF is a generic data model for describing objects, their properties and
relations between them
• RDF Schema is a simple vocabulary (ontology) description language
• In RDF Schema, we can define:
‒ “Allowed” Classes (node types) and Properties (edges’ labels)
‒ Which properties go with which classes
‒ Which values a property can take
‒ Class Hierarchies and Inheritance
‒ Property Hierarchies!
Semantic Web Ontology Languages
RDF Schema

• A richer ontology language
• (More) relations between classes and properties
‒ e.g., equivalent classes/properties, disjoint classes/properties,
inverse properties, Boolean combinations of classes
• Cardinality constraints
‒ e.g. People have exactly one mother, Students attend at most five courses
• Richer typing / restrictions on properties
‒ E.g. Postgraduate students attend only Postgraduate courses
‒ E.g. Faculty members must teach at least one Undergraduate course
• Characteristics of properties
‒ Object properties vs. Datatype properties
‒ Symmetrical, transitive, functional properties
Semantic Web Ontology Languages
OWL (Web Ontology Language)

• RDF statement:
ex:index.html dc:creator exstaff:85740
35
RDF Schema example
ex:index.html exstaff:85740
dc:creator
ex:WebPage
rdf:type
ex:Employee
rdf:type
ex:Person
rdfs: subClassOf
rdfs:Class
rdf:type rdf:type
rdf:type
rdf:Property
rdfs:domain rdfs:range
rdf:type
N. Bassiliades, ENCASE Summer School, Limassol, 17th July 2017

ex:WebPage rdf:type rdfs:Class
ex:Employee rdf:type rdfs:Class
ex:Person rdf:type rdfs:Class
ex:Employee rdf:subClassOf ex:Person
dc:creator rdf:type rdf:Property
dc:creator rdfs:domain ex:WebPage
dc:creator rdfs:range ex:Person
36
RDF Schema definitions
• Resource ex:WebPage is a class
‒ An RDFS class, there also OWL classes
• Resource ex:Employee is a class
• Class ex:Employee is a subclass of
class ex:Person
• Resource dc:creator is a property
(instance of rdf:Property class)
• Property dc:creator is attached to
class ex:WebPage and takes values
instances of class ex:Person

• Resource ex:index.html is an instance of class ex:WebPage
ex:index.html rdf:type ex: WebPage
• Resource exstaff:85740 is an instance of class ex:Employee
exstaff:85740 rdf:type ex:Employee
37
Connecting RDF instances to classes

• Reasoning in RDF Schema is based on entailment rules
• Rules are logical implications:
IF a condition is true THEN the conclusion is also true
• Entailment rules:
IF such and such triples exist THEN add also these triples
(conclusions made true)
• E.g. rule for the subClassOf property:
IF ?x rdf:type ?u . AND ?u rdfs:subClassOf ?v .
THEN ?x rdf:type ?v .
Reasoning in RDF Schema

RDF Schema Inference / Query Process
Original
(explicit) set
of triples
RDF
document
translation
RDF/S
Inference rules
Inferred
(implicit) set
of triples
Set of all triples
Query

Reasoning Example in RDF Schema
Due to subClassOf
dc:creator
ex:WebPage
rdf:type
ex:Employee
rdf:type
ex:Person
rdfs: subClassOf
rdfs:Class
rdf:type rdf:type
rdf:type
rdf:Property
rdfs:domain rdfs:range rdf:type

Due to property domain
dc:creator
ex:WebPage
rdf:type
ex:Employee
rdf:type
ex:Person
rdfs: subClassOf
rdfs:Class
rdf:type rdf:type
rdf:type
rdf:Property

Due to property range
dc:creator
ex:WebPage
rdf:type
ex:Employee
rdf:type
ex:Person
rdfs: subClassOf
rdfs:Class
rdf:type rdf:type
rdf:type
rdf:Property
rdf:type

dc:creator
ex:WebPage
rdf:type
ex:Employee
rdf:type
ex:Person
rdfs: subClassOf
rdfs:Class
rdf:type rdf:type
rdf:type
rdf:Property

• OWL is mapped on a Description Logic
‒ A subset of Predicate Logic (aka First-Order Logic / FOL)
‒ Then Description Logic reasoners are used (FaCT++, RACER, Pellet, Hermit,
etc.)
• Why a subset of Predicate Logic?
‒ Reasoning in Predicate Logic is undecidable
‒ Reasoning in Description logics is (usually) decidable
‒ Efficient decision procedures have been designed and implemented
Reasoning in OWL

• Equivalence of classes
‒ If class A is equivalent to class B, and class B is equivalent to class C, then A is
equivalent to C, too
‒ If class A is subclass of B, B subclass of C and C subclass of A, then A, B, C are
equivalent to each other
• Class membership
‒ If x is an instance of a class C, and C is a subclass of D, then we can infer that x
is an instance of D
‒ If C and D are equivalent classes, then if x is an instance C, then it is also an
instance of D, and vice versa
Reasoning Tasks in OWL

• Consistency
‒ X instance of both classes A and B, but A and B are disjoint
‒ X is an instance of both A and complement of A
‒ This is an indication of an error in the ontology
• Instance Classification
‒ Certain property-value pairs are a sufficient condition for membership in a
class A
→FirstYearCourses are Courses with Year=1
‒ If an individual x satisfies such conditions, we can conclude that x must be an
instance of A

• Class Classification
‒ A driver is a person that drives a vehicle.
‒ A bus driver is a person that drives a bus.
‒ A bus is a vehicle.
‒ A bus driver drives a vehicle, …
‒ … so he/she must be a driver.
• Instance equality
‒ Every person has a unique mother
‒ X has two mothers A and B
‒ Thus, (in order to restore consistency) A = B.
‒ If we already know that A  B, then we have an inconsistency.

• The more expressive a logic is, the more computationally expensive it
becomes to draw conclusions
‒ Drawing certain conclusions may become impossible if non-computability
barriers are encountered
• Compromise:
‒ A language supported by reasonably efficient reasoners
‒ A language that can express large classes of ontologies and knowledge
Tradeoff between Expressive Power and Computational
Complexity

• OWL (OWL 2) comes with flavours / profiles of decreasing complexity,
suitable for different types of applications
‒ OWL Full (full compatibility with RDF Schema)
‒ OWL DL (all constructs allowed – not combined with RDF Schema)
‒ OWL 2 EL: useful in applications where ontologies contain very large numbers
of properties / classes, but not so many instances
‒ OWL 2 QL: aimed at applications that with very large volumes of instances;
query answering is the most important reasoning task (data resides in
relational databases)
‒ OWL 2 RL: aimed at applications that require scalable reasoning without
sacrificing too much expressive power (data resides in triplestores)
OWL flavours / profiles

Storing and Querying RDF data
with SPARQL
… and an introduction to DBpedia
52N. Bassiliades, ENCASE Summer School, Limassol, 17th July 2017

• RDF data (triples) are stored in special NoSQL databases,
called Triple stores or Graph stores.
‒ Databases optimized to import, store, and query a huge number of triples.
‒ Sesame (70 million), OpenLink Virtuoso (>15.4 billion), GraphDB (>12 b), Apache
Jena TDB (200 m), AllegroGraph (1 trillion), IBM DB2, Oracle, …
• Triple stores are queried using SPARQL
‒ Sending SPARQL queries using the SPARQL protocol
‒ Triple stores provide a (public) endpoint, where SPARQL queries can be submitted
• Clients can send queries to an endpoint using the HTTP protocol.
‒ You can issue a SPARQL query to an endpoint by entering it into the browser
‒ It’s preferable to have a client designed specifically for SPARQL.
Triple Stores

DBpedia SPARQL endpoint (http://dbpedia.org/sparql)

• A project aiming to extract structured content from the information
created as part of the Wikipedia.
‒ This structured information is then made available on the Web
• DBpedia allows users to query relationships and properties associated
with Wikipedia resources, including links to other related datasets.
• One of the more famous parts of the Linked Data project, according
to TBL
DBpedia (http://dbpedia.org)

• Wikipedia articles include structured information embedded in the
articles (mostly free text)
‒ E.g. "infobox" tables, categorization information, images, geo-coordinates and
links to external pages.
‒ This structured information is extracted and put in a uniform dataset which can
be queried.
• The DBpedia project uses RDF to represent the extracted information.
‒ 8.8 billion RDF triples
→1.1 billion from the English edition
→4.4 billion from other language editions (125)
→3.2 billion from DBpedia Commons and Wikidata
DBpedia dataset

Wikipedia extraction to DBpedia
Infobox
Title

Wikipedia extraction to DBpedia
Infobox
Categories

• The same concepts can be expressed using different properties in
templates
‒ E.g. birthplace and placeofbirth
‒ Queries about where people were born must search for both properties to get
complete results.
• The DBpedia Mapping Language helps mapping properties to an ontology
‒ Reduces the number of synonyms
• The development of the ontology and the mappings are open to public
‒ Due to large diversity of infoboxes and properties
DBpedia challenges

SPARQL examples
• Find all cities in Cyprus with a University
select DISTINCT ?c where {
?u rdf:type dbo:University .
?u dbo:country dbr:Cyprus .
?u dbo:city ?c .
} ORDER BY ?c

SPARQL examples
• Find cities in Cyprus without Universities
select ?c where {
?c rdf:type dbo:City .
?c dbo:country dbr:Cyprus .
FILTER NOT EXISTS {
?u dbo:city ?c .
}
}
Empty
result!

• Find all cities in Cyprus
select ?c where {
?c rdf:type dbo:City .
FILTER NOT EXISTS {
?u dbo:city ?c .
}
}
SPARQL examples
What happened to
the rest of the cities?

DBpedia entry for Limassol
There is no dbr:Limassol rdf:type dbo:City . triple in DBpedia

DBpedia: the Truth!
• DBpedia has a lot of wrong or missing information
• Not all Wikipedia properties (infobox) have been correctly
mapped to the corresponding DBpedia ontology property
• Thus, in order to retrieve the correct information sometimes
you have to become a detective!

Query that returns more results
{ ?c rdf:type dbo:City . }
UNION
{ ?x dbo:city ?c . }
} order by ?c
?c is a city, or
?c is the city of something ?x

Filtering results
UNION
?c dbo:populationTotal ?p .
FILTER ( ?p >= 50000 )
} order by ?c

Return cities and their mayors
select DISTINCT ?c ?n where {
UNION
?c dbo:leaderName ?n .
} order by ?c
What happened to
the rest of the cities?
They have not their mayor
mentioned in DBpedia

Return cities and their mayors (if mentioned)
select DISTINCT ?c ?n where {
UNION
OPTIONAL { ?c dbo:leaderName ?n . }
} order by ?c
The OPTIONAL keyword allows
for flexibility in pattern matching
• Needed for “null values”

What about smaller towns?
{ ?c rdf:type dbo:Settlement . }
UNION
FILTER ( ?p > 1000 )
} order by ?c
dbo:Settlement is
appropriate for villages,
towns, even neighborhoods
Too many results!

Just count them!
select count(DISTINCT ?c) as ?Num
where {
UNION
FILTER ( ?p > 1000 )
} order by ?c
The results contain also cities,
towns, …, from the pseudo-
state of “North Cyprus”

Exclude results that are (directly) part of
where {
UNION
FILTER NOT EXISTS {
?c dbo:isPartOf dbr:Northern_Cyprus .
}
FILTER ( ?p > 1000 )
}
The results still contain places
that are indirectly related to
the pseudo-state of “Northern
Cyprus”
(e.g. dbr:Yeni_Jami,_Nicosia)

Exclude results that are directly or indirectly part of
where {
UNION
FILTER NOT EXISTS {
?c dbo:isPartOf+ dbr:Northern_Cyprus .
}
FILTER ( ?p > 1000 )
}
The + sign indicates a
path of 1 or more
dbo:isPartOf edges in
the RDF graph
dbo:isPartOf
dbo:isPartOf/dbo:isPartOf
dbo:isPartOf/dbo:isPartOf/dbo:isPartOf
…

• dbr:Nicosia expands to full URI: http://dbpedia.org/resource/Nicosia
‒ Unique ID that represents the resource in the Web of Data
• However, if you type the above URI at a browser, you will be re-directed
at: http://dbpedia.org/page/Nicosia
‒ Manifestation (or visualization) of the resource in the Web of Documents
• If you type at the browser: http://dbpedia.org/data/Nicosia
‒ The browser will retrieve an RDF/XML file that contains all information for Nicosia
‒ All triples with http://dbpedia.org/resource/Nicosia as a subject
BTW: LOD principles in action

Ontologies, Logic & Rules
Broadening reasoning capabilities

• Can ontologies express every piece of knowledge needed in the
Semantic Web?
• They can express static information (e.g. knowledge about the world)
• They have some reasoning capabilities, but…
• They cannot combine information from several different individuals in
order to come to a complex conclusion about the world
• They cannot reason about which actions should be performed at each
situation
Ontology shortcomings

• Consists of logical implications (rules): A1, . . ., An  B
‒ Ai and B are atomic formulas
• There are 2 ways of reading such a rule:
‒ Deductive rules: If A1,..., An are known to be true, then B is also true
‒ Reactive rules: If the conditions A1,..., An are true, then carry out the action B
• Horn logic is tractable and is supported by efficient reasoning tools
Horn Logic: a(nother) Predicate Logic subset

• Neither of them is a subset of the other
‒ Both are needed in the Semantic Web
• Horn logic example:
‒ Persons who study and live in the same city are “home students”
studies(X,U), lives(X,A), loc(U,C), loc(A,C)  homeStudent(X)
‒ It is impossible to state that in OWL
• OWL example:
‒ A person is either a man or a woman
‒ Easily expressed in OWL using disjoint union
‒ It is impossible to state that in Horn logic
Horn Logic vs. Description Logics (aka ontologies)

• Horn logic (rules) and description logics (ontologies) are orthogonal
‒ Both are subsets of first-order logic (FOL)
‒ Neither is the subset of each other
Horn logic vs. Description logics
FOL
Horn LogicDescription
logics
OWL 2 RL
SWRL



• The simplest integration approach is the intersection of both logics
• OWL 2 RL is an interesting sublanguage of OWL 2 DL
‒ Inherits open-world assumption and non-unique-name assumption
‒ These assumptions do not make a difference
OWL2 RL

• OWA: We cannot conclude some statement x to be false simply because
we cannot show x to be true.
‒ E.g. a patient’s clinical history does not include a particular allergy
‒ It would be incorrect to assume that the patient does not suffer from that allergy
‒ It is unknown, unless more information is given
• CWA: If something cannot be proved, then it is false
‒ E.g. we are looking for a direct flight between Larnaca and Madrid in a DB
application for airline reservations
‒ The flight doesn’t exist in the database
‒ Expected / correct answer: “There is no direct flight between Austin and Madrid.”
• OWL is committed to OWA, Horn logic to CWA
Open-World Assumption (OWA) vs.
Closed-World Assumption (CWA)

• Statement: “Juan is a citizen of the USA.”
• Question: “Is Juan a citizen of Colombia?”
‒ CWA answer: no
‒ OWA answer: I don’t know
• Additional statements:
‒ “A person can only be citizen of one country”
‒ “Juan is a citizen of Colombia.”
• CWA: error (we assume that Colombia and USA are different things)
• OWA: “USA and Colombia must be the same thing”
OWA vs. CWA
Examples

• When two individuals are known by different names, they are in fact
different individuals.
‒ Sometimes works well and sometimes not
‒ Example in favor: when two products in a catalog are known by different codes, they
are different
‒ Example against: two people in a social environment initially known with different
identifiers (e.g., “Prof. van Harmelen” and “Frank”) are sometimes the same person
• CWA systems (Horn logic) have UNA
• OWA systems (OWL) do not have UNA
‒ However, one could manually add the UNA
‒ Using owl:allDifferent or owl:differentFrom
Unique-Name Αssumption (UNA)

• OWL 2 RL is the largest fragment of OWL on which the choice for
CWA/OWA and UNA does not matter
‒ Weak enough so that differences between choices don’t show up.
‒ Still large enough to enable useful representation and reasoning tasks.
• Constructs of OWL that can be expressed using Horn logic rules
‒ Subclass, sub-property, class and property equivalence
‒ Equality-inequality between individuals
‒ Inverse, transitive, symmetric and functional properties
‒ Intersection of classes
• Excluded constructors
‒ Union, existential quantification, and arbitrary cardinality constraints
OWL 2 RL

• A triple s p o is expressed as a fact p(s, o)
• An instance declaration rdf:type(a,C)
‒ a is an instance of class C
‒ expressed as C(a)
• C is a subclass of D: C(X) → D(X)
• P is a sub-property of Q: P(X,Y) → Q(X,Y)
• Domain and Range Restrictions
‒ D is the domain of property P: P(X, Y) → D(X)
‒ R is the range of property P: P(X, Y) → R(Y)
RDF constructs

• equivalentClass(C,D): C(X) → D(X), D(X) → C(X)
• equivalentProperty(P,Q): P(X,Y) → Q(X,Y), Q(X,Y) → P(X,Y)
• Transitive Properties: P(X,Y), P(Y,Z) → P(X,Z)
• allValuesFrom(P,D): C(X), P(X,Y) → D(Y)
‒ Necessary restriction (e.g. all undergraduate students must attend only
undergraduate courses)
‒ UGStudent(X), attends(X,Y) → UGCourse(Y)
• someValuesFrom(P,D): P(X,Y), D(Y) → C(X)
‒ Sufficient restriction / Instance classification rule (e.g. if someone attends a
postgraduate course, then he/she is a postgraduate student)
‒ attends(X,Y), PGCourse(Y) → PGStudent(X)
OWL Constructs

• A proposed Semantic Web language combining OWL 2 DL with Horn logic
‒ Syntax: Datalog RuleML
• Allows the definition of Horn-logic rules on top of OWL 2 DL ontologies
‒ Rule conclusions are “stored” back in the ontology
• SWRL unites the expressivities of DL and Horn-logic
‒ OWL 2 RL combines the advantages of both languages in their common
sublanguage
• SWRL is intractable
‒ DL-safe rules: tractable subset of SWRL
→Every variable must appear in a non-DL atom in the rule body
Semantic Web Rules Language (SWRL)

Man(?m) → Person(?m)
‒ Possible in OWL - subclassOf relation
‒ Some rules are OWL syntactic sugar
Person(?m)  hasSex(?m,male) → Man(?m)
‒ Possible in OWL – hasValue (sufficient) restriction
‒ Not all such reclassifications are possible in OWL
Person(?m)  hasSpouse(?m,?w)  works_at(?w,?j)  publicOrg(?j) →
MarriedToPublicServantPerson(?m)
‒ Not possible in OWL
Example SWRL Rules: Reclassification

hasParent(?x, ?y)  hasBrother(?y, ?z) → hasUncle(?x, ?z)
‒ Property chaining
‒ Possible in OWL 2 - Not possible in OWL 1
Person(?p)  hasSibling(?p,?s)  Man(?s) → hasBrother(?p,?s)
‒ Not possible in OWL
Publication(?p)  hasAuthor(?p,?y)  hasAuthor(?p,?z)  differentFrom(?y,?z)
→ cooperatedWith(?y, ?z)
‒ SWRL does not adopt the UNA
‒ Individuals must also be explicitly stated to be different (using owl:allDifferent
restriction) in the OWL ontology
Example SWRL Rules: Property Value Assignment

• Built-ins dramatically increase expressivity
‒ Most rules are not expressible in OWL 1
‒ Some built-ins can be expressed in OWL 2
Person(?p)  hasAge(?p,?age)  swrlb:greaterThan(?age,17) → Adult(?p)
Person(?p)  hasNumber(?p, ?number)  swrlb:startsWith(?number, "+") →
hasInternationalNumber(?p, true)
Person(?p)  hasSalaryInPounds(?p, ?gbp)  swrlb:multiply(?d, ?gbp, 2) →
hasSalaryInDollars(?p, ?dollars)
Example SWRL Rules: Built-ins

Person(?p)  not hasCar(?p, ?c) → CarlessPerson(?p)
• Not possible – rule language does not support negation
• Potential invalidation - what if a person later gets a car?
N. Bassiliades, ENCASE Summer School, Limassol, 17th July 2017 5-90
SWRL is Monotonic: does not Support Negation

• De-facto industry standard to represent SPARQL rules and constraints
on Semantic Web models.
• Provides meta-modeling capabilities that allow users to define their
own SPARQL functions and query templates.
• Includes a ready to use library of common functions.
• SPIN follows the CWA
‒ A special kind on negation exists (negation-as-failure)
SPIN (SPARQL Inferencing Notation)

• Rules can be expressed in SPARQL using the CONSTRUCT feature
parent(X,Y), parent(Y,Z)  grandparent(X,Z)
CONSTRUCT {
?X grandParent ?Z .
}
WHERE {
?X rdf:type Person .
?X parent ?Y .
?Y parent ?Z .
}
Rules in SPARQL: SPIN
optional

• Rules can be associated to classes
‒ Rules can represent behavior of the instances of that class (an OOP feature!)
‒ Even constructors can be define!
‒ Global rules can also be defined
• SPIN variants
‒ Inference / Entailment Rules: CONSTRUCT, INSERT
‒ Production Rules: DELETE
‒ Integrity constraints: ASK
SPIN

• The grandParent rule is stored at the Person class
‒ It will be executed only for instances of this class (and subclasses)
‒ Increases accuracy and efficiency
CONSTRUCT {
?this grandParent ?Z .
}
WHERE {
?this parent ?Y .
?Y parent ?Z .
}
Rules in SPARQL: SPIN
?this means an instance of the
class, where the rule is stored

• Constraint are expressed via the ASK SPARQL construct
• For each instance that the ASK query is true, then the constraint is
violated
‒ A constraint violation warning is issued
ASK WHERE {
?this hasAge ?age .
FILTER (?age < 18) .
}
SPIN Constraints
The SPIN constraint is
stored at class Student

• A more radical solution to constraint violation would be to delete the
instance that violates the constraint
‒ This can be done via the DELETE construct
‒ The SPIN rules should be declared as a constructor
‒ Constructor rules run each time a new instance is created
DELETE { ?this rdf:type Student . }
WHERE {
?this hasAge ?age .
FILTER (?age < 18) .
}
SPIN Constructors
The SPIN constructor is
stored at class Student

• Unlike SWRL, in SPIN there is a special negation (negation-as-failure)
‒ When something is not found to be true, it is false, thus its negation is true
Person(?p)  not hasCar(?p, ?c) → CarlessPerson(?p)
CONSTRUCT {
?this rdf:type CarlessPerson .
}
WHERE {
?this rdf:type Person .
FILTER NOT EXISTS { ?this hasCar ?x . }
}
Using negation in SPIN
If we store the SPIN rule at class
Person we do not need this line

C1(X), equivalentClass(C1,C2)  C2(X)
CONSTRUCT {
?X a ?C2 .
}
WHERE {
?X a ?C1.
?C1 equivalentClass ?C2.
}
OWL 2 RL rules in SPIN
Actually, in some Semantic Web
systems, the OWL 2 RL semantics
are implemented via SPIN rules
(TopBraid Composer)

A Semantic Entity Linking
Use Case
The URank system
URank
Entity
Extractor
Site-specific
transformations
Extraction
Rules
Extracted
Data
Ranking
sites
Entity
Linker
Ranking
datasets
Merged
dataset
Entity
Merger
Domain-
specific
filtering
Ranking
ontology

• University rankings
‒ Means of advertisement
‒ There are so many of them! (>20 global rankings)
‒ Do we need all of them? Are they similar? Are they robust?
‒ Comparative Statistical Analysis is needed
• Collecting the data from multiple web sites
‒ Web data extraction
‒ Each ranking site produces a ranking table
‒ A single table needs to be constructed to feed the Statisticians
‒ The ranking tables need to be merged into a single all-rankings table
Motivating example

Name Rank …
… … …
Aristotle University of Thessaloniki 491-500
… … …
The easy case…
QS
Name Rank …
… … …
Aristotle University of Thessaloniki 401-500
… … …
THE
…
Name QS THE …
… … … …
Aristotle University of Thessaloniki 491-500 401-500 …
… … … …
Merged table

Name Rank …
… … …
The Imperial College of Science,
Technology and Medicine
22
… … …
The difficult case…
ARWU
Name Rank …
… … …
Imperial College London 8
… … …
THE
…
Name ARWU THE …
… … … …
The Imperial College of Science,
Technology and Medicine
22 - …
Imperial College London - 8 …
… … … …
Merged table
Levenshtein Distance = 38
Substring similarity* = 0.605
*A string metric for ontology alignment by Giorgos Stoilos, 2005.

• Use string matching and semantic search to find a unique ID for each
university in each ranking table
• Where to search? DBpedia
‒ DBpedia / Wikipedia have entries for almost all World Universities
‒ Each DBpedia entity has an ID
• Hopefully different variations of the University name will retrieve the
same University entity in DBpedia
• Then merging is straight forward
Merging tables: Solution

URank[1] System
URank
Entity
Extractor
Site-specific
transformations
Extraction
Rules
Extracted
Data
Ranking
sites
Entity
Linker
Ranking
datasets
Merged
dataset
Entity
Merger
Domain-
specific
filtering
Ranking
ontology
[1] N. Bassiliades: “Collecting University Rankings for
Comparison Using Web Extraction and Entity Linking
Techniques”, Springer CCIS Vol.469, pp.23-46, 2014.
A Prolog application that:
1. Extracts data from
ranking sites
2. Links Universities to
DBpedia entities
through semantic
search and string
similarity
3. Generates merged table

• Entity linking: the task of determining the identity of entities mentioned
in text
• Different from Named Entity Recognition, which identifies the
occurrence or mention of a named entity in text but it does not
identify which specific entity it is.
• Entity linking requires a knowledge base containing the entities to
which entity mentions can be linked.
• In URank:
‒ Named Entity Recognition is not needed - All University names are entities
‒ Entity Linking uses DBpedia as a knowledge base
Entity Linking

• DBpedia University entities do not always belong to the correct class
‒ dbo:University or dbo:EducationalInstitution
‒ Need to loose that criterion carefully
• University mergers /splits
‒ University of Paris split in 1970 into 13 Universities named with very similar
names: University of Paris I, II, …
‒ University of Montpellier split into three universities (I, II, II) in 1970; I and II
merged back in 2015
‒ Need to check if the University is currently operating
• Newcastle University, UK vs University of Newcastle, Australia
‒ Need to check the country
Entity Linking challenges

• US vs USA
‒ Need to have a common representation for countries
• Imperial College London vs Imperial College of Science, Technology and
Medicine
‒ Need to have access to alternative University names
• University of Montpellier II vs University of Montpellier 2
‒ Need to convert between Arabic and roman literals
• Universität vs University
‒ Need to translate between different languages
Entity Linking challenges

• DBpedia Spotlight: annotate mentions of DBpedia resources in NL text
‒ http://demo.dbpedia-spotlight.org/
‒ Have tried it for the University Ranking use case with ~86% F-measure
• Silk: integrate heterogeneous data sources
‒ http://silkframework.org/
‒ Generates links between related data items within different Linked Data sources.
‒ Linked Data publishers can use Silk to set RDF links from their data sources to other
data sources on the Web.
‒ Experimentation for the use case has even lower F-measure
• Domain-specific knowledge must be used to face all challenges
General-purpose Entity Linking Tools

• Each University entity at each ranking site is matched against a
DBpedia entry using 3 alternative methods
‒ DBpedia lookup service
‒ DBpedia SPARQL endpoint, using approximate string matching filtering
functions
‒ Keyword search engine of Wikipedia
• At each step, if a satisfactory match is found (substring matching), the
algorithm terminates
• Otherwise, all matching entries are collected and scored
‒ Top scored candidate is returned as a match
Entity Linking in URank

• String distance is measured using a metric for ontology alignment [1]
‒ Concerns substring matching
‒ More appropriate for matching names of Universities, than e.g. Levenshtein
‒ E.g. “Imperial College” vs. “Imperial College of Science, Technology and Medicine”
‒ Materialized as a built-in predicate in SWI-Prolog (isub/4)
• Satisfactory match is above a high similarity threshold
‒  0,97 depending on the algorithm step
[1] Stoilos, G., Stamou, G., Kollias, S.: A String Metric for Ontology Alignment. ISWC 2005, LNCS, vol.
3729, pp. 624-637. Springer (2005)
String matching

http://lookup.dbpedia.org/api/search.asmx/KeywordSearch?Q
ueryClass=University&QueryString=Imperial%20College%20Lo
ndon
• RESTful API parameters:
‒ QueryString: a string for which a DBpedia URI
should be found (University name)
‒ QueryClass: a DBpedia class from the Ontology
that the results should have (University)
‒ MaxHits: the maximum number of returned
results (default: 5)
• Results in XML should be parsed
‒ using native SWI-Prolog XPath built-ins
111
DBpedia lookup service
N. Bassiliades, ENCASE Summer School, Limassol, 17th July
2017

• If DBpedia lookup does not return a satisfactory match then the DBpedia
SPARQL endpoint is used (provided by OpenLink Virtuoso RDF DB engine)
SELECT ?univ, ?name WHERE {
?univ rdf:type Class .
?univ ?property ?val .
?val bif:contains University.Name.Words option (score ?score) .
?univ rdfs:label ?name .
FILTER (lang(?name) = "en") }
ORDER BY DESC (?score*0.3+sql: rnk_scale(<LONG::IRI_RANK> (?univ)))
LIMIT Top-N2
DBpedia SPARQL endpoint (http://dbpedia.org/sparql)

SPARQL query example @ DBpedia
Results are actually returned in RDF,
not HTML

• If neither native “DBpedia methods” return a satisfactory match, the
Wikipedia search engine is used
http://en.wikipedia.org/w/index.php?search=Univ.name&limit=Top-N3&go=Go
• Results are verified using equivalent DBpedia entity
‒ Wikipedia URLs are uniquely transformed to DBpedia URIs
‒ https://en.wikipedia.org/wiki/Imperial_College_London 
http://dbpedia.org/resource/Imperial_College_London
Keyword search engine of Wikipedia

• Spatiotemporal constraints are used for filtering
‒ Not easy, due to incomplete and inaccurate information from DBpedia/Wikipedia
• Retrieved university must be located in the same country with the extracted
university
‒ Using dbo:country property (not present in all entities)
‒ Using alternative properties dbo:state, dbo:city, dbo:location, and try to locate
country (geographical areas containment)
‒ Using redirection to another entity (owl:sameAs, dbo:wikiPageRedirects)
• Countries are not always represented using the same name
‒ E.g. US, U.S., USA, U.S.A., United States of America
‒ E.g. UK, United Kingdom, United Kingdom of Great Britain and Northern Ireland
‒ Synonym matrix for most problematic cases
Spatiotemporal constraints

• select ?x where { <DBPediaURL> dbo:country ?x . }
• select ?x where { <DBPediaURL> dbo:state ?x . }
• select ?x where { <DBPediaURL> dbo:city ?x . }
• select ?x where { <DBPediaURL> dbo:location ?x . }
• …
Geographical area containment

• Retrieved university must be still operating
‒ E.g. University of Paris split, University of Montpellier split and merge, etc.
• Checking if the property dbp:closed exists
‒ Not all “closed” Universities have this property
‒ In this case heuristic scrapping
from Wikipedia entries is performed
Spatiotemporal constraints

Conclusions

• Semantic technologies and techniques – not necessarily using Semantic Web
standards – have been adopted for building and using knowledge graphs
‒ Google, Facebook, Yandex, Baidu, Bing have embedded Semantic technologies into
their core businesses
‒ The presentage of Web content that uses microdata/RDFa/schema.org information to
enhance search results reached double digits
‒ IBM’s Watson (AlchemyLanguage) returns concept information as
DBpedia/yago/freemix URIs, including a Linked Data API
• Semantic technologies are rapidly finding their ways into consumer products
‒ The average developer barely realizes it, but semantics are now just about everywhere
The Future of Semantic Technologies

• Semantic Web is a branch of AI
• The market interest and justification is in solving problems
‒ AI has problem solving in its core
• Semantic technologies will become an essential enabler to the
development of true knowledge-based AI
‒ What things mean and how people understand what they mean
• AI applications currently focus on Machine Learning (ML) and NLP
‒ Knowledge graphs can help enhance the accuracy of ML/NLP
‒ NLP: top-down processing (use ontology terms to disambiguate text)
‒ ML: bottom-up processing (learn from data and match to generalized concepts)
120
Semantic Web and Artificial Intelligence

Enhancing seCurity and privAcy in the Social wEb: a user-centered
approach for the protection of minors

Semantic Technologies for the Web of Linked Data

More Related Content

What's hot

Similar to Semantic Technologies for the Web of Linked Data

Recently uploaded

Semantic Technologies for the Web of Linked Data