Semantic Web technologies:

Introduction and Survey

Lee Feigenbaum
Cambridge Semantics Inc.
W3C SPARQL Working Group Co-chair.

Adapted from a presentation by Lee Feigenbaum and Eric Prud'hommeaux at C-SHALS 2008. This work is licensed under a Creative Commons Attribution 3.0 License, with attribution to W3C.

Program

A motivating example
Introduction to RDF (data model)
Introduction to SPARQL (query language)
Introduction to GRDDL and RDFa (RDF in an XML and HTML world)
Introduction to RDFS and OWL (schemas and ontologies)

A Motivating Example: Drug discovery

The W3C HCLS interest group set out to use Semantic Web technologies to receive precise answers to a complex question:

Find me genes involved in signal transduction that are related to pyramidal neurons.

General search: 223,000 hits, 0 results

Domain-limited search: 2,580 results

Specific database: Too many silos!

A Semantic Web Approach

Integrate disparate database...

Mesh
Pubmed
Entrez Gene
Gene Ontology
...

A Semantic Web Approach (cont'd)

... so that one query ...

A Semantic Web Approach (cont'd)

... (trivially) spans several DBs ...

SenseLab Database headings and representation graph

A Semantic Web Approach (cont'd)

... to yield cross-specialty information

What was the trick?

Good-enough modeling.
- RDFS - simple subclass/subproperty relationships
- OWL - inference and more expressive modeling
Query interface tailored to data model.
- SPARQL - RDF query language
- Re-uses the familiar Web architecture
Agreement on common terms and relations.
- RDF - flexible underlying data model
- URI - unambiguous naming facilitates re-use when possible and invention when necessary

The Resource Description Framework (RDF)

Name resources and relationships with URIs
Model statements as subject, predicate, object triples
Standard serialization in XML, known as RDF/XML
Developer-friendly serializations: N3 and turtle

Patient data

<?xml version="1.0"?>
<ClinicalDocument transformation="hl7-rim-to-pomr.xslt">
  <recordTarget>
    <patientRole>
      <patientPatient>
	<name>
	  <given>Henry</given>
	  <family>Levin</family>
	</name>
	<administrativeGenderCode code="M"/>
	<birthTime value="19320924"/>
      </patientPatient>
    </patientRole>
  </recordTarget>
  <component>
    <StructuredBody>
      <Observation>
	<code displayName="Cuff blood pressure"/>
	<effectiveTime value="200004071430"/>
	<targetSiteCode displayName="Left arm"/>
	<entryRelationship typeCode="COMP">
	  <Observation>
	    <effectiveTime value="200004071530"/>
	    <value value="132" unit="mm[Hg]"/>
	  </Observation>
	</entryRelationship>
      </Observation>
      <Observation>
	<code displayName="Cuff blood pressure"/>
	<effectiveTime value="200004071530"/>
	<targetSiteCode displayName="Left arm"/>
	<entryRelationship typeCode="COMP">
	  <Observation>
	    <code displayName="Systolic BP"/>
	    <effectiveTime value="200004071530"/>
	    <value value="135" unit="mm[Hg]"/>
	  </Observation>
	</entryRelationship>
	<entryRelationship typeCode="COMP">
	  <Observation>
	    <code displayName="Diastolic BP"/>
	    <effectiveTime value="200004071530"/>
	    <value value="88" unit="mm[Hg]"/>
	  </Observation>
	</entryRelationship>
      </Observation>
    </StructuredBody>
  </component>
</ClinicalDocument>

Patient identifier
Medical history
Family medical history
Health-related behavior

RDF is good for modeling all this data...

...regardless of its source.

Simple statement

_:p   r:type galen:Patient .

Add a property ...

_:p   r:type galen:Patient ;
      foaf:familyName "Levin" .

... and another.

_:p   r:type galen:Patient ;
      foaf:familyName "Levin" ;
      foaf:givenName "Henry" .

Statements about ...

_:p   r:type galen:Patient ;
      foaf:familyName "Levin" ;
      foaf:givenName "Henry" .
_:scr dc:date "2006-03-18T18:23"^^xsd:dateTime .

... other objects ...

_:p   r:type galen:Patient ;
      foaf:familyName "Levin" ;
      foaf:givenName "Henry" .
_:scr dc:date "2006-03-18T18:23"^^xsd:dateTime ;
      edns:systolic "132"^^edns:mmHg .

... connect in the fabric.

_:p   r:type galen:Patient ;
      foaf:familyName "Levin" ;
      foaf:givenName "Henry" .
_:c   edns:patient _:p ;
      edns:screeningBP _:scr .
_:scr dc:date "2006-03-18T18:23"^^xsd:dateTime ;
      edns:systolic "132"^^edns:mmHg ;
      edns:diastolic "86"^^edns:mmHg ;
      edns:posture snomed:_163035008 .

A First Look at Turtle

<http://thefigtrees.net/lee/id#lee> 
  <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://xmlns.com/foaf/0.1/Person> .
<http://thefigtrees.net/lee/id#lee> 
  <http://xmlns.com/foaf/0.1/name> "Lee Feigenbaum" .
<http://thefigtrees.net/lee/id#lee>
  <http://xmlns.com/foaf/0.1/homepage> <http://thefigtrees.net/lee/> .

... is more succinctly represented as:

@prefix rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> .
@prefix foaf: <http://xmlns.com/foaf/0.1/> .

<http://thefigtrees.net/lee/id#lee> rdf:type      foaf:Person ;
                                    foaf:name     "Lee Feigenbaum" ;
                                    foaf:homepage <http://thefigtrees.net/lee/> .

Patient Data in RDF

_:p1  a galen:Patient ;
      foaf:family_name "Levin" ;
      foaf:firstName "Henry" .

_:c1a edns:patient _:p1 ;
      edns:screeningBP [
      a cpr:clinical-examination ;
      dc:date "2000-04-07T15:30:00" ;
      edns:systolic [
          a galen:AbsoluteMeasurement ;
          ex:unit "mm[Hg]" ;
          r:value "132" ;
          skos:prefLabel "Systolic BP" 
      ] ;
      edns:diastolic [
          a galen:AbsoluteMeasurement ;
          ex:unit "mm[Hg]" ;
          r:value "86" ;
          skos:prefLabel "Diastolic BP"
      ] ;
      edns:location snomed:_66480008 ; # SNOMED:left arm
      edns:posture snomed:_163035008   # SNOMED:sitting      
   ] .

There is a blood-pressure examination of a patient named Henry Levin. The examination was on 7-April-2000 at 3:30pm and was conducted on the patient's left arm while he was sitting. The examination resulted in a systolic blood pressure measurement of 132 and a diastolic measurement of 86.

Introduction to SPARQL

Why SPARQL?
SELECTing variables
Triple patterns
Simple query patterns
GRAPH constraints

Why SPARQL?

SPARQL is the query language of the Semantic Web. It lets us:

Pull values from structured and semi-structured data
Explore data by querying unknown relationships
Perform complex joins of disparate databases in a single, simple query
Transform RDF data from one vocabulary to another

SELECTing variables

SPARQL variables bind to RDF terms
- Ex. ?journal, ?disease, ?price
Like SQL, we pick the variables we want from a query with a SELECT clause
- Ex. SELECT ?article ?author ?published
A SELECT query results in a table of values:

?artist	?album	?times_platinum
Michael Jackson	Thriller	27
Led Zeppelin	Led Zeppelin IV	22
Pink Floyd	The Wall	22

Triple patterns

A triple pattern is an RDF triple that can have variables in any of the subject, predicate, or object positions.

Examples:

Find countries and their capital cities:
- ?country geo:capital ?capital .
Given a FOAF URI, find the person's name:
- <http://thefigtrees.net/id#lee> foaf:name ?name .
What direct relationships exist between two employees?
- emp:8A0120 ?relationship emp:D29J10 .

Simple query pattern

We can combine more than one triple pattern to retrieve multiple values and easily traverse an RDF graph:

Find countries, their capital cities, and their populations:
- ?country geo:capital ?capital .
  ?country geo:population ?population .
Given a FOAF URI, find the person's name and friends' names:
- <http://thefigtrees.net/id#lee> foaf:name ?name .
  <http://thefigtrees.net/id#lee> foaf:knows ?friend .
  ?friend foaf:name ?friend_name .
Retrieve all third-line managers in the company:
- ?emp hr:managedBy ?first_line .
  ?first_line hr:managedBy ?second_line .
  ?second_line hr:managedBy ?third_line .

GRAPH constraints

SPARQL lets us query different RDF graphs in a single query. Consider movie reviews:

Target one authoritative data source (What does Roger Ebert say?):

GRAPH <http://example.org/reviews/rogerebert> {
  ex:atonement rev:hasReview ?review .
  ?review rev:rating ?rating .
}

Relate multiple sources (How do my reviews compare to Ebert's?):

GRAPH <http://example.org/reviews/rogerebert> {
  ?movie rev:hasReview ?rev1 .
  ?rev1 rev:rating ?ebert .
}
GRAPH <http://example.org/reviews/me> {
  ?movie rev:hasReview ?rev2 .
  ?rev2 rev:rating ?me .
}

Retrieve provenance data (Which reviewers have given out perfect ratings?):
- ```
GRAPH ?reviewer_graph {
  ?review rev:rating 10 .
}
```

Example Query: Henry Levin's Blood Pressure

PREFIX dc: <http://purl.org/dc/elements/1.1/>
PREFIX edns: <http://www.loa-cnr.it/ontologies/ExtendedDnS.owl#>
PREFIX foaf: <http://xmlns.com/foaf/0.1/>
PREFIX galen: <http://www.co-ode.org/ontologies/galen#>
PREFIX r: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>
PREFIX snomed: <http://termhost.example/SNOMED/>

SELECT ?date ?sys ?dias ?position {
?p    r:type galen:Patient ;
      foaf:family_name "Levin" ;
      foaf:firstName "Henry" .
?c    edns:patient ?p ;
      edns:screeningBP ?scr .
?scr  dc:date ?date ;
      edns:systolic [ r:value ?sys ] ;
      edns:diastolic [ r:value ?dias ] ;
      edns:posture ?position .
} ORDER by ?date

The sample query can be run against this sample data.

Or try sparql.org with this query against this data.

Using GRDDL to get RDF from XML, XHTML

GRDDL (Gleaning Resource Descriptions from Dialects of Languages) is a way to boostrap RDF out of XML and in particular XHTML data by explicitly indicating transformations from RDF to XML. GRDDL relies on:

Source Document: an XHTML or XML document which references at least one GRDDL transformation and hence licenses a GRDDL-aware agent to extract RDF.
GRDDL-aware agent: a software agent able to identify GRDDL transformations and run them to extract RDF.
GRDDL Transformation: an algorithm--usually expressed in XSLT--for getting RDF from a source document

GRDDLing HTML

GRDDL can extract RDF from both XML and (X)HTML.

Patient	Systolic BP	Diastolic BP
Henry Levin	132	86
...	...	...

<html>
  <head profile="http://www.w3.org/2003/g/data-view">
    <title>Clinical Study 8B1a: Patient BP</title>
    <link rel="transformation" href="bp-html-to-pomr.xslt" />
  </head>
  ...

GRDDL-aware Querying

Some SPARQL engines can directly query GRDDL source documents.

PREFIX dc: <http://purl.org/dc/elements/1.1/>
PREFIX edns: <http://www.loa-cnr.it/ontologies/ExtendedDnS.owl#>
PREFIX foaf: <http://xmlns.com/foaf/0.1/>
PREFIX galen: <http://www.co-ode.org/ontologies/galen#>
PREFIX r: <http://www.w3.org/1999/02/22-rdf-syntax-ns#>

SELECT ?date ?sys ?dias ?location ?position {
?p    r:type galen:Patient ;
      foaf:family_name "Levin" ;
      foaf:firstName "Henry" .
?c    edns:patient ?p ;
      edns:screeningBP ?scr .
?scr  dc:date ?date ;
      edns:systolic [ r:value ?sys ] ;
      edns:diastolic [ r:value ?dias ] ;
      edns:location ?location ;
      edns:posture ?position .
}

The actual query against the actual XML source is more complex. You can try it live against Virtuoso with this live query.

UltraLink

Novartis solution for cross-linking over 1,500,000 biologic and chemical terms, including synonyms, taxonomies, and pointers into data repositories
RDF (in particular the SKOS vocabulary) models the network of domain-specific terms
Text-mining annotators enable one-click access to enterprise-wide information on a particular term in context

UltraLink in Action

annotated screenshot of Novartis's UltraLink

Why RDFa?

UltraLink is only as good as its text-mining software. What if an acquisition brings with it a new Web-based corpus of pathway data that uses terms not recognized by the annotators?
What if content publishers could indicate the semantics of their content inline in a standard fashion? Then we'd only need to write a single recognizer...
RDFa (RDF in attributes) allows exactly this.

RDFa Example: Chemicals

InChI is a textual identifier for chemical substances. Consider inchi.html:

<table>
<tr>
  <th>Familiar name</th><th>InChI</th>
</tr><tr>
  <td>Methane</td>
  <td about="http://example.org/methane" property="chem:inchi"
      xmlns:chem="http://www.blueobelisk.org/chemistryblogs/">
    InChI=1S/CH4/h1H4
  </td>
...

This RDFa encodes the single RDF triple:

<http://example.org/methane> chem:inchi "InChI=1S/CH4/h1H4" .

Using RDFa: In context

The Firefox Operator extension finds structured data in Web pages and enables domain-specific actions.
On the current Web: content publishers decide what can be done with the data (via links, script)
On the Semantic Web: content publishers publish actionable data; content consumers decide how to act on it

See inchi.html.

Using RDFa: Query

There are various ways to query Web pages marked up with RDFa:

Use an RDFa-aware SPARQL engine (as we did with GRDDL).
Use a GRDDL-aware SPARQL engine, and publish RDFa data with the profile: http://ns.inria.fr/grddl/rdfa/
Wrap the Web page(s) in an online RDFa extraction service:

# Find propane's InChI string
PREFIX chem: <http://www.blueobelisk.org/chemistryblogs/>
PREFIX ex:   <http://example.org/>
SELECT ?inchi 
FROM <http://www.w3.org/2007/08/pyRdfa/extract?uri=http://www.w3.org/2008/Talks/0305-C-SHALS/inchi.html>
WHERE {
  ex:propane chem:inchi ?inchi .
}

You can try it against sparql.org.

RDF Schema (RDFS) modeling

type labels classify things in your data.
- :Fido rdf:type :dog .
subClassOf creates hierarchy of these types.
- :dog rdfs:subClassOf :mammal .
- ⇒ :Fido a :mammal .
domain/range identify the type of the subject/object of a predicate.
- :Bob foaf:givenName "Bob" .
- foaf:givenName rdfs:domain foaf:Person .
- ⇒ :Bob a foaf:Person .

Groups/Sets/Classes

[Named] collections of things with similar attributes.

Groups/Sets/Classes

[Named] collections of things with similar attributes.

Groups/Sets/Classes

[Named] collections of things with similar attributes.

OWL Expressivity

Allows you to define classes.

(a,b,c) set enumeration

union

disjunction

algebraics

intersection

complement

restriction

cardinality

equivalence

OWL: Identity

identity assertion
```
:Bob owl:sameAs :BobSmith .
```

inverse functional properties

foaf:mbox a owl:inverseFunctionalProperty .
:Bob foaf:mbox <mailto:bobS@foo.example> .
:BobSmith foaf:mbox <mailto:bobS@foo.example> .
⇒ :Bob owl:sameAs :BobSmith

cardinality constraints

:patient owl:maxCardinality 1 .
:patientEncounter5 :patient :Bob .
:patientEncounter5 :patient :BobSmith .
⇒ :Bob owl:sameAs :BobSmith

OWL: Types within context

RDFS: The object of a :Specialist's :specialty is a :Disease
OWL: If the :Specialist is an :Oncologist, then the object of :specialty is :Cancer

:DrJones :specialty :Cancer .
⇒ :DrJones rdf:type :Oncologist .

And there's much more...

Tools (and a potential W3C working group) that expose existing relational databases as RDF
Semantics for Web Services - SAWSDL
Semantic rules languages - e.g. SWRL, RIF, N3 rules
OWL 2 - improved datatype handling, new expressivity, better metamodeling, ...
SPARQL 2 - update, aggregates, ...
Description resources for asserting semantics of large swaths of Web pages (POWDER)
Many, many existing RDF vocabularies and ontologies - both general and domain-specific