Is the Semantic Web what we expected? Adoption Patterns and Content-driven Challenges (ISWC 2016 Keynote)

Bizer: Is the Semantic Web what we Expected? ISWC 2016, 10/20/2016 Slide 1
15th International Semantic Web Conference (ISWC 2016)
Kobe, Japan, 10/20/2016
Is the Semantic Web what we Expected?
Deployment Patterns and Data-driven Challenges
Prof. Dr. Christian Bizer

Ian‘s Keynote Last Year in Bethlehem
http://videolectures.net/
iswc2015_horrocks_semantic_technology/

This Year
Use cases
 on the public Web
 many data sources
 no central control

Outline
1. What did we expect the Semantic Web to be?
2. What does the Semantic Web actually look like?
1. Linked Data
2. HTML-embedded Data
3. Why is this the case?
4. What does this mean for Semantic Web applications?

1. What did we expect the Semantic Web to be?

2001 Article: The Semantic Web
Envisions three things to happen:
 people publish structured data
on the Web
 ontologies are used to enable
shared understanding
 people implement cool applications that
do smart things with the available data
Tim Berners-Lee, James Hendler, Ora Lassila:
The Semantic Web. Scientific American, May 2001.

Expectation: Hyperlinks are Set on Data Level
https://www.w3.org/History/1989/proposal.html
https://www.w3.org/DesignIssues/LinkedData.html

Expectation: High Quality Content / Provenance Metadata
 Publishers provide high quality content
 Publishers support applications in determining trustworthiness
• by providing provenance metadata
• using digital signatures
Layer Cake, 2001

Check List: Our Expectations about the Semantic Web
1. People publish structured data on the Web
2. Ontologies are used to enable shared understanding
3. Hyperlinks are set on data level
4. People publish high quality content / metadata
5. Cool applications do smart things with the data

2. What does the Semantic Web actually look like?
Linked Data HTML-embedded Data

2.1 Linked Data Deployment
Schmachtenberg, Bizer, Paulheim: Adoption of the Linked Data Best Practices. ISWC2014.
Ermilov, Lehmann, Martin, Auer: LODStats: The Data Web Census Dataset. ISWC 2016.
Topics # of Datasets Percentage
Media 24 2 %
Government 199 18 %
Publications/Library 138 13 %
Geographic 27 2 %
Life Sciences 85 8 %
Cross‐domain 47 4 %
Unser‐generated Content 51 5 %
Social Networking 520 48 %
LODStats 2016: 2740 datasets
LOD Cloud 2014: 1091 datasets

Ontological Agreement
 Strong agreement on some vocabularies
• base terminology, people, publications
 Proprietary vocabularies are used in
addition to common ones,
as data is often very specific
http://linkeddatacatalog.dws.informatik.uni-mannheim.de/state/
https://lov.okfn.org/dataset/lov/

 Some datasets put a lot of effort into linking
 Many datasets only link to a small number of
other datasets or do not set RDF links at all
RDF Links
Link to # of Datasets Percentage
more than 10 datasets 79 8 %
6 to 10 datasets 81 8 %
5 datasets 31 3 %
4 datasets 42 4 %
3 datasets 54 5 %
2 datasets 106 10 %
1 datasets 176 17 %
0 datasets 445 44 %
http://linkeddatacatalog.dws.informatik.uni-mannheim.de/state/
71 %
16 %

Cool Applications
Prototypes of Semantic Web browsers and search engines.

2.2 HTML-embedded Data
Microformats
Microdata
RDFa
JSON-LD

Overall Adoption 2015
Web Data Commons, 2015:
 2.72 million pay-level-domains (PLDs) out of the 14.41
million PLDs provide HTML-embedded data (19%)
 540 million HTML pages out of the 1.7 billion pages
provide HTML-embedded data (30%)
Guha/Brickley/Macbeth, 2015:
12 million websites provide schema.org data
Guha/Brickley/Macbeth: Schema.org. ACM queue, 2015
http://webdatacommons.org/structureddata/2015-11/

Number of PLDs providing HTML-embedded Data
http://webdatacommons.org/structureddata/

Widely-used Classes
Class # of Domains
WDC 2015
# of Domains
schema.org 2015
schema:PostalAddress 124,000 > 1,000,000
schema:Product 108,000 > 1,000,000
schema:Offer 82,000 > 1,000,000
schema:LocalBusiness 77,000 500,000 ‐ 1,000,000
schema:Person 74,000 > 1,000,000
schema:Review 28,000 250,000 ‐ 500,000
schema:GeoCoordinates 17,000 100,000 ‐ 250,000
schema:Event 12,000 100,000 ‐ 250,000
schema:Hotel 5,300 10,000 ‐ 50,000
schema:Restaurant 3,800 10,000 ‐ 50,000
schema:JobPosting 3,600 10,000 ‐ 50,000

Adoption by E-Commerce Websites
Alexa Top‐15 Website schema:Product
Amazon.com 
Ebay.com 
NetFlix.com 
Amazon.co.uk 
Walmart.com 
etsy.com 
Ikea.com 
Bestbuy.com 
Homedepot.com 
Target.com 
Groupon.com 
Newegg.com 
Lowes.com 
Macys.com 
Nordstrom.com 
Adoption:
60 %

Properties used to Describe Products
Properties PLDs
# %
schema:Product/name 78,292 87 %
schema:Product/image 59,445 66 %
schema:Product/description 58,228 65 %
schema:Product/offers 57,633 64 %
schema:Offer/price 54,290 61 %
schema:Offer/availability 36,789 41 %
schema:Offer/priceCurrency 30,610 34 %
schema:Product/url 23,723 26 %
schema:Product/aggregateRating 21,166 24 %
schema:Product/manufacturer 10,150 11 %
schema:Product/brand 9,739 11 %
schema:Product/productID 9,221 10 %
schema:Product/sku 7955 9 %
schema:Product/gtin13 935 1 %

Challenge: Small Amount of Identifiers and Data Links
1. Small amount of product identifiers
2. Hardly any schema:sameAs links
Definition: URL of a reference Web page that unambiguously
indicates the item's identity.
Properties PLDs
# %
schema:Product/sameAs 85 0.07 %
schema:LocalBusiness/sameAs 655 0.8 %
schema:Organization/sameAs 3,900 5 %
Properties PLDs
# %
schema:Product/productID 9,221 10 %
schema:Product/gtin13 935 1 %

Challenge: Flat Data Structures
Websites do not explicitly annotate product features
but mention them in product names and descriptions.
<div itemtype="http://schema.org/Product">
<span itemprop="name">Apple MacBook Air A1370 Intel Core i5
1.60GHz 64GB SSD 11.6 Laptop
</span>
<span itemprop=“description"> Catch up on work, school, or socializing
on the Apple MacBook Air A1370 11.6-inch laptop. This handy
computer features 2GB DDR3 RAM, an Intel Core i5 560UM
processor, 64GB hard drive, and the Mac OS …
</span>
</div>
Petrovski, Bryl, Bizer: Integrating Product Data from Websites Offering Microdata Markup.
DEOS 2014.

Challenge: Product Categorization
1. Small amount of websites publishing categorization information
2. Heterogeneity of the product taxonomies
Properties PLDs
# %
schema:Offer/category 2200 2 %
schema:WebPage/breadcrumb 460 0.4 %
Home > Shop > Outdoor & Garden > Barbecues & Outdoor
Living > Garden Furniture > Tables > Dining Tables
Philadelphia Eagles > Philadelphia Eagles Mens > Philadelphia
Eagles Mens Jerseys > over $60
Meusel, Primpeli, Meilicke, Paulheim, Bizer: Exploiting Microdata Annotations to
Consistently Categorize Product Offers at Web Scale. EC-Web 2015.

Adoption by Travel Websites
Top 15 Travel Websites schema:Hotel
Booking.com 
TripAdvisor 
Expedia 
Agoda 
Hotels.com (uses OGP) 
Kayak 
Priceline 
Travelocity 
Orbitz 
ChoiceHotels 
HolidayCheck 
ChoiceHotels 
InterContinental Hotels Group 
Marriott International 
Global Hyatt Corp. 
Adoption:
86 %
Kärle, Fensel, Toma, Fensel: Why Are There More Hotels in Tyrol than in Austria?
Analyzing Schema. org Usage in the Hotel Domain. ICTT 2016.

Properties used to Describe Hotels
Top 10 Properties PLDs
# %
schema:Hotel/name 4173 88,35 %
schema:Hotel/address 3311 70,10 %
schema:Hotel/telephone 2488 52,68 %
schema:PostalAddress/streetAddress 2362 50,01 %
schema:PostalAddress/addressLocality 2231 47,24 %
schema:Hotel/url 2102 44,51 %
schema:PostalAddress/postalCode 2096 44,38 %
schema:AggregateRating/ratingValue 1952 41,33 %
schema:Hotel/aggregateRating 1866 39,51 %
schema:AggregateRating/bestRating 1697 35,93 %
Might improve in the future as new schema.org accommodation
vocabulary was released August 2016.

Adoption by Job Portals
Adoption:
70 %
Top‐10 Employment Websites schema:JobPosting
Indeed.com 
Monster.com 
Careerbuilder.com 
Snagajob.com 
Jobsdb.com 
Jobsearch.about.com 
Jobs.net 
Internships.com 
Jobs.aol.com 
Quintcareers.com 

Properties used to Describe Job Postings
Properties PLDs
# %
JobPosting/title 2588 91 %
JobPosting/hiringOrganization 1412 49 %
JobPosting/description 1192 41 %
JobPosting/jobLocation 1062 37 %
Organization/name 862 30 %
JobPosting/datePosted 793 27 %
Place/address 471 16 %
JobPosting/baseSalary 227 8 %
JobPosting/industry 209 7 %
JobPosting/educationRequirements 145 5 %
JobPosting/occupationalCategory 105 0.3 %
JobPosting/skills 56 0.2 %

Cool Applications
1. Rich snippets within search results
2. Knowledge graph
panels
https://developers.google.com/
structured-data/

Cool Applications
 Open Graph Protocol allows site owners to
determine how entities are displayed in Facebook
 uses RDFa for marking up data in HTML pages
 used by over 200,000 websites (WDC 2015)

Our Expectations Revisited
Expectation Linked Data HTML‐embedded Data
1. People publish
structured data
> 1000 sources,
wide range of
specific topics
Millions of sources,
focused on search
engines and Facebook
2. Ontologies enable
understanding
Partial agreement,
complex data structures
Strong agreement,
flat data structures
3. Hyperlinks on data
level
Some data links Hardly any data links
4. High quality content
Web quality,
partly outdated
Web quality,
some SPAM
5. Cool applications
various application
prototypes
strong application pull
by search engines

3. Why is this the Case?
Expectation Linked Data HTML‐embedded Data
1. People publish
structured data
> 1000 sources,
wide range of
specific topics
Millions of sources,
focused on search
engines and Facebook
2. Ontologies enable
understanding
Partial agreement,
complex data structures
Strong agreement,
flat data structures
3. Hyperlinks on data
level
Some data links Hardly any data links
4. High quality content
Web quality,
partly outdated
Web quality,
some SPAM
5. Cool applications
various application
prototypes
strong application pull
by search engines

Benefits and Costs for Data Providers
Making the Web a better place isn’t enough
motivation for most data providers.

Benefits of Publishing HTML-Embedded Data
Get richer visibility in search results and
potentially more clicks.

Effort for Publishing HTML-Embedded Data
 Most providers just change their HTML templates
 If more effort is required, most providers just do not do it
1. Annotate specific product features
• given free text descriptions in the backend database
2. Map to common product taxonomy like GS1 GPC
• given local categorizations
3. Annotate skills and occupational categories
• given free text descriptions in the backend database
<div itemtype="http://schema.org/Hotel">
<span itemprop="name">Vienna Marriott Hotel</span>
<span itemprop="address" itemscope="" itemtype="http://schema.org/PostalAddress">
<span itemprop="streetAddress">Parkring 12a</span>
<span itemprop="addressLocality">Vienna</span>
</span></div>

Effort for Setting Data Links
 Effort:
1. Decide which data sources to link to
2. Compare schemata and develop a matching rule for each class
3. Run link generation algorithm
4. Publish resulting link set on the Web
 Benefits:
• You increase the value of your data as it becomes
easier to use it together with data from other sources
• You reduce the integration costs for the data consumer
<http://dbpedia.org/resource/Berlin> owl:sameAs
<http://sws.geonames.org/2950159> .
Data publisher
provides links
Effort
Distribution
Data consumer
generates links

For Whom does the Linking Effort pay off?
 Scientists
• Innovation becomes possible by
connecting datasets
• My impact / prestige grows if
my data is used for cool things
 Librarians
• Have the mission to catalog artefacts
• Traditionally use shared identifiers
 E-Commerce Vendors
• Benefits of setting data links are unclear
• Just want to look nice on Google
• Might not want to be comparable on
price portals

Effort of Maintaining Links
 We want to be nice!
• we want to link to everybody
 We set instance- and schema-level links!
• created and collected 37 link sets
• over 20 million RDF links
• http://wiki.dbpedia.org/services-resources/interlinking
https://github.com/dbpedia/links
 We would likely need a full-time volunteer to maintain all these links
 Result: Many dead links
1. because target data source has changed
2. because we used bad linkage rules due to insufficient domain knowledge

Hypothesis
 Missing links and shared identifiers
 Flat data structures
 Heterogeneity of taxonomies
 Mixed data quality
We will keep on seeing similar adoption patterns,
as we need to be realistic about the effort
spent by data publishers

4. What does this mean for Semantic Web Applications?

Be happy about all semantic clues
(integration hints) provided
But do not expect the clues to be perfect
4. What does this mean for Semantic Web Applications?

Applications should be happy about …
… all effort that data providers put into setting data links
• but treat links with caution as they might be wrong / outdated
… all effort that data providers put into using common vocabularies
• but still try to understand proprietary vocabularies / taxonomies
… all effort that data providers put into structuring their data
• but still try to understand flat free-text descriptions
Treat all statements on the Web as a claims
• whose trustworthiness needs to be verified
Data Publisher’s
Effort
Data Consumer’s
Effort
Effort
Distribution

Semantic Web Clients need to be FAT Clients
There are no shortcuts!
1. Crawl data
2. Normalize vocabularies
6. Resolve data conflicts
3. Parse flat descriptions
4. Verify existing data links
5. Create missing data links

Parsing Flat Descriptions
Ristoski, Mika: Enriching Product Ads with Metadata from HTML Annotations. ESWC 2016.
Foley, et al.: Learning to Extract Local Events from the Web. SIGIR 2015.
Dictionary
Parser
schema.org
schema.org data suitable as
distant supervision:
 schema:Product/brand
 schema:Product/manufacturer
 schema:JobPosting/industry
 Schema:JobPosting/skills
 schema:Event/name

Create and Verify Data Links
 Supervised learning of detailed
matching rules leads to F1>95%
(e.g. Silk and LIMES frameworks)
 Sources of supervision
1. Data links and shared identifiers
• owl:sameAs
• schema:Product/productID
• schema:Product/gtin13
2. Human guidance via active learning
 How to generalize matching rules to data
from multiple sources?
Isele, Bizer: Active Learning of Expressive Linkage Rules using Genetic Programming. JWS 2013.
Stonebraker, et al.: Data Curation at Scale: The Data Tamer System. CIDR 2013.

Resolve Data Conflicts
 Do you have some data that you already trust?
 Knowledge-based Trust
• determine trustworthiness of a data source by comparing its content
with trusted data (ground truth)
• outperforms PageRank and voting
Dong, et al.: Knowledge-based Trust: Estimating the Trustworthiness of Web Sources. VLDB 2015.
Web Data Source
Country City
Germany Berlin
France Paris
United Kingdom London
Canada Ottawa
USA Washington D.C.
Mexico Ecatepec




?
?
Trusted Data
Country Capital
Germany Berlin
France
United Kingdom London
Canada
USA Washington D.C.
Mexico Mexico City

Google Knowledge Vault
 Extends Freebase with data from
one billion web pages
1. Web text (TXT): Entity linking,
relationship extraction
2. HTML trees (DOM): Wrapper induction
3. HTML tables (TBL): Relational tables
4. Semantic Annotations (ANO): schema.org, OGP
 Employs knowledge-based trust for ranking
 Results:
• 271 million facts with confidence >90%
• 90 million facts not in Freebase before
Dong, et al.: Knowledge Vault: A Web-scale Approach to Probabilistic Knowledge Fusion.
SIGKDD 2014.

The Structural Continuum
Be open to different forms of
“structured” Web content.
HTML tables
DOM Trees
CSV Tables
HTML-embedded Data
Linked Data Upper
Ontologies

Exploit Schema.org and HTML Tables Together
Qui, et al.: DEXTER: Large-Scale Discovery and Extraction of Product Specifications on the Web. VLDB 2015.
Petrovski, et al: The WDC Gold Standards for Product Feature Extraction and Product Matching. ECWeb 2016.
s:name
HTML table
s:breadcrumb

Conclusions

 The Semantic Web contains more data than most people like
 exciting test-bed for research on data profiling, cleansing
and integration
 endless data pool for commercial applications (product
comparison, business listings, job search, …)
The Semantic Web is Huge
Billions of product offers
A description of every
hotel in the world
Lots of data about local businesses

We will keep on seeing Similar Adoption Patterns
 as we need to be realistic about the effort spent by data publishers
 be happy about any semantic clues (integration hints) provided
 design algorithms to work despite the scarcity and noisiness of clues
Flat Data Structures
Missing Links and Identifiers
Mixed Data Quality
Heterogeneity of Taxonomies

Semantic Web Clients need to be FAT Clients
There are no shortcuts!
1. Crawl data
2. Normalize vocabularies
6. Resolve data conflicts
3. Parse flat descriptions
4. Verify existing data links
5. Create missing data links

Thank you!

Is the Semantic Web what we expected? Adoption Patterns and Content-driven Challenges (ISWC 2016 Keynote)

Recommended

More Related Content

What's hot (20)

Similar to Is the Semantic Web what we expected? Adoption Patterns and Content-driven Challenges (ISWC 2016 Keynote) (20)

More from Chris Bizer (12)

Recently uploaded (20)

Is the Semantic Web what we expected? Adoption Patterns and Content-driven Challenges (ISWC 2016 Keynote)